---
schema_version: 7
rubric_version: "0.3"
canonical_url: "https://podium.do/compare/agents/products/swe-af/"
data_url: "https://podium.do/compare/agents/products/swe-af/data.json"
generated_at: 2026-11-22
oldest_checked_at: 2026-11-16
last_checked_at: 2026-11-16
publisher: "Podium"
---

# SWE-AF: ADE findings

> Podium publishes this catalog and sells an agent development environment. Podium chooses the products, criteria, sources and wording, so treat these pages as vendor-authored research. Check important claims against the linked primary sources. Sourced findings follow the same evidence rules for Podium and its competitors. Facts labeled Podium team research come from Podium's own team and have no linked source. None of this removes the conflict of interest.

> Based on data from adecomparison.com by Podium (https://adecomparison.com; source repository https://github.com/madeinorbit/ade-comparison, snapshot 4322f0c of 2026-11-22), used under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Source links and check dates are kept. Podium arranges the findings into pairs, applies its publication checks, and added its own research where marked "Podium team research". Data license terms: https://github.com/madeinorbit/ade-comparison/blob/4322f0c035f303791413639d1462a138443d8c1e/DATA-LICENSE.md

50 of 51 criteria have a sourced finding; 50 were also cross-checked by a second agent. Source checks: 2026-11-16. A missing sourced finding is not evidence that a capability is absent.

- Website: https://github.com/Agent-Field/SWE-AF
- Canonical HTML: https://podium.do/compare/agents/products/swe-af/
- Machine-readable JSON: https://podium.do/compare/agents/products/swe-af/data.json
- Rubric: https://podium.do/compare/agents/rubric/index.md
- Comparisons with SWE-AF on podium.do: 1 (listed in the JSON and HTML)
- Compare SWE-AF with any other product on adecomparison.com by Podium: https://adecomparison.com/compare/?products=swe-af

## Access and agent choice

### ACC-01: Client access

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

SWE-AF is headless: builds start through the AgentField HTTP API or af CLI. The AgentField control plane web UI shows runs, DAG graph and notes; AgentField Desktop (macOS dmg/zip, Windows nsis) installs and starts swe-planner. No SWE-AF mobile app or editor extension exists.

Availability: beta
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with, AgentField Desktop (Electron companion that installs and starts swe-planner); version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- desktop os: SWE-AF has no desktop client of its own; the AgentField Desktop companion (a separate Electron app that installs swe-planner) targets macOS (dmg, zip) and Windows (nsis) with no Linux target.
- mobile apps: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a mobile client; AgentField ships no iOS or Android app in desktop/ or control-plane/web.
- mobile browser: The control plane web client declares a viewport meta tag and uses responsive utility classes, but this pass does not establish phone-width behavior of the run views.
- editor extensions: Bounded search of swe_af/ and go/ at the pinned commit found no implementation for editor extension terms (vsix, vscode extension, jetbrains plugin).
- task actions: Start via API or af call, poll executions and notes, cancel an execution tree through the control plane, resume a crashed build with resume_build, approve a plan through the optional Hax gate; no in-product
- notifications: Execution webhooks can be passed to implement_issue and Hax approval requests reach a person; there are no product-owned desktop or mobile push notifications for SWE-AF builds.

- desktop_os — Desktop os [partial]: SWE-AF has no desktop client of its own; the AgentField Desktop companion (a separate Electron app that installs swe-planner) targets macOS (dmg, zip) and Windows (nsis) with no Linux target. Sources: swe-af:afdesk, swe-af:afdeskpkg.
- mobile_apps — Mobile apps [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a mobile client; AgentField ships no iOS or Android app in desktop/ or control-plane/web. Sources: swe-af:afdesk, swe-af:readme.
- browser — Browser [supported]: The AgentField control plane serves a web UI with runs, workflow DAG, step trace and execution notes for SWE-AF builds. Sources: swe-af:afui, swe-af:afreadme.
- mobile_browser — Mobile browser [unknown]: The control plane web client declares a viewport meta tag and uses responsive utility classes, but this pass does not establish phone-width behavior of the run views. Sources: swe-af:afui.
- cli — Cli [supported]: af CLI (af call swe-planner.build, af install, af run, af ls --entrypoints) triggers and discovers SWE-AF; the swe-af console script and python -m swe_af start the node. Sources: swe-af:readme, swe-af:pyproject, swe-af:surface.
- editor_extensions — Editor extensions [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation for editor extension terms (vsix, vscode extension, jetbrains plugin). Sources: swe-af:readme, swe-af:pyproject.
- task_actions — Task actions [partial]: Start via API or af call, poll executions and notes, cancel an execution tree through the control plane, resume a crashed build with resume_build, approve a plan through the optional Hax gate; no in-product steer or review UI. Sources: swe-af:readme, swe-af:app, swe-af:afcancel, swe-af:hitl.
- notifications — Notifications [partial]: Execution webhooks can be passed to implement_issue and Hax approval requests reach a person; there are no product-owned desktop or mobile push notifications for SWE-AF builds. Sources: swe-af:readme, swe-af:hitl.
- cross_device_continuity — Cross device continuity [supported]: Build state lives in the control plane and in .artifacts checkpoints on the node host, so any client that reaches the control plane URL can poll the same execution. Sources: swe-af:readme, swe-af:app.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:afdesk, swe-af:afdeskpkg, swe-af:readme, swe-af:afui, swe-af:afreadme, swe-af:pyproject, swe-af:surface, swe-af:app, swe-af:afcancel, swe-af:hitl

### ACC-02: Agent and model choice

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Five selectable runtimes: aforge (AgentField AForge, default when OPENROUTER_API_KEY is set), claude_code, open_code, codex and command_code (Python only); roles can use different models, but one build uses one runtime.

Availability: beta
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- custom adapters: Runtimes are a fixed enum validated in BuildConfig; OpenCode model ids pass through, and execute_fn_target lets a caller redirect execution to another reasoner, but no generic harness adapter registry ex
- mixed harness operation: A single build resolves one runtime for all roles; separate calls or nodes can use different runtimes, but one build does not mix two harnesses.

- harnesses — Harnesses [supported]: runtime values aforge, claude_code (Claude Code via claude-agent-sdk), open_code (OpenCode), codex (Codex CLI) and command_code (Command Code, Python node only). Sources: swe-af:readme, swe-af:go, swe-af:deploy.
- model_providers — Model providers [supported]: Anthropic, OpenRouter, OpenAI, Google, Infron and MiniMax providers are documented, with per-role model overrides in the flat models map. Sources: swe-af:readme, swe-af:deploy.
- custom_adapters — Custom adapters [partial]: Runtimes are a fixed enum validated in BuildConfig; OpenCode model ids pass through, and execute_fn_target lets a caller redirect execution to another reasoner, but no generic harness adapter registry exists. Sources: swe-af:schemas, swe-af:readme.
- session_integration — Session integration [supported]: SWE-AF controls harness sessions programmatically through AgentField SDK harness providers (claude-agent-sdk, aforge exec --json, codex exec, opencode, command-code -p), not through a terminal wrapper. Sources: swe-af:pyproject, swe-af:readme, swe-af:codexpatch.
- mixed_harness_operation — Mixed harness operation [partial]: A single build resolves one runtime for all roles; separate calls or nodes can use different runtimes, but one build does not mix two harnesses. Sources: swe-af:readme, swe-af:schemas.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:go, swe-af:deploy, swe-af:schemas, swe-af:pyproject, swe-af:codexpatch

### ACC-03: Subscriptions, API keys and vendor credits

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Bring your own: Anthropic API key, OpenRouter or other provider keys, a Claude Code subscription token (CLAUDE_CODE_OAUTH_TOKEN, Pro/Max credits) and Codex ChatGPT login (SWE_CODEX_AUTH_MODE). No vendor credits; no local-model route documented.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- vendor credits: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of vendor-sold credits or hosted billing; the manifest requires the user to supply a provider key.
- local models: Bounded search of swe_af/ and go/ at the pinned commit found no implementation for local-model runtimes (ollama, llama.cpp, lm studio); open_code model ids point at hosted gateways.
- account switching: Provider and runtime are switched per request through config.runtime and config.models, and the Codex auth mode is an environment setting; no multi-account manager exists.
- provider restrictions: README warns ANTHROPIC_BASE_URL is process-wide and that an Anthropic credential left set can be sent to a non-Anthropic endpoint; no provider terms on subscription reuse are recorded here.

- agent_subscriptions — Agent subscriptions [supported]: CLAUDE_CODE_OAUTH_TOKEN from claude setup-token uses Claude Pro/Max credits; Codex with ChatGPT subscription auth after codex login and SWE_CODEX_AUTH_MODE=chatgpt or auto. Sources: swe-af:readme, swe-af:deploy, swe-af:codexpatch.
- api_accounts — Api accounts [supported]: ANTHROPIC_API_KEY, OPENROUTER_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, INFRON_API_KEY and MINIMAX_API_KEY are documented. Sources: swe-af:readme, swe-af:manifest.
- vendor_credits — Vendor credits [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of vendor-sold credits or hosted billing; the manifest requires the user to supply a provider key. Sources: swe-af:manifest, swe-af:readme.
- local_models — Local models [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation for local-model runtimes (ollama, llama.cpp, lm studio); open_code model ids point at hosted gateways. Sources: swe-af:readme, swe-af:deploy.
- account_switching — Account switching [partial]: Provider and runtime are switched per request through config.runtime and config.models, and the Codex auth mode is an environment setting; no multi-account manager exists. Sources: swe-af:readme, swe-af:go.
- vendor_account_requirement — Vendor account requirement [supported]: At least one provider key is required (manifest require_one_of); GH_TOKEN is only needed for private repos, pushes and PRs; no SWE-AF account exists. Sources: swe-af:manifest, swe-af:readme.
- provider_restrictions — Provider restrictions [partial]: README warns ANTHROPIC_BASE_URL is process-wide and that an Anthropic credential left set can be sent to a non-Anthropic endpoint; no provider terms on subscription reuse are recorded here. Sources: swe-af:deploy, swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:deploy, swe-af:codexpatch, swe-af:manifest, swe-af:go

### ACC-04: Onboarding and project prerequisites

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Documented path: af install https://github.com/Agent-Field/SWE-AF, af run swe-planner, set one provider key, then af call swe-planner.build with a goal and repo_url or repo_path. A local directory needs at least one commit; GitHub token only for private repos, pushes and PRs.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- no git: implement_issue requires repo_path to be a checkout with at least one commit; build initializes git for new work in a repo_path, but a plain directory without Git history is not a supported implement_issue target
- managed cloud setup: No SWE-AF-operated cloud exists in the inspected repositories; Railway one-click deploys the user's own services (control plane, PostgreSQL and node) in the user's Railway account.
- byo vps onboarding: Docker Compose and Railway are documented first-class deploy paths and deployment.md covers a new server, but there is no VPS provisioning installer.

- first_task_setup — First task setup [supported]: Install (af install or Docker/Railway), run the control plane (af) and node, set a provider key, then POST swe-planner.build with goal plus repo_url or repo_path. Sources: swe-af:readme, swe-af:deploy, swe-af:manifest.
- no_git — No git [unsupported]: implement_issue requires repo_path to be a checkout with at least one commit; build initializes git for new work in a repo_path, but a plain directory without Git history is not a supported implement_issue target. Sources: swe-af:readme, swe-af:issuebuild.
- local_git_without_remote — Local git without remote [supported]: repo_path mode runs on a local checkout; implement_issue returns a branch and pushes nothing unless enable_github_pr is set (needs an origin remote). Sources: swe-af:readme, swe-af:issue.
- hosted_git — Hosted git [supported]: repo_url clones from GitHub, pushes the integration branch and opens a PR through gh using GH_TOKEN; multi-repo builds use repos with primary and dependency roles. Sources: swe-af:readme, swe-af:schemas.
- prerequisite_guidance — Prerequisite guidance [supported]: README and docs/deployment.md list Python 3.12+, Docker and Compose, the AgentField control plane, provider keys, GH_TOKEN scope and Codex bubblewrap requirements. Sources: swe-af:readme, swe-af:deploy.
- managed_cloud_setup — Managed cloud setup [unsupported]: No SWE-AF-operated cloud exists in the inspected repositories; Railway one-click deploys the user's own services (control plane, PostgreSQL and node) in the user's Railway account. Sources: swe-af:readme.
- byo_vps_onboarding — Byo vps onboarding [partial]: Docker Compose and Railway are documented first-class deploy paths and deployment.md covers a new server, but there is no VPS provisioning installer. Sources: swe-af:deploy, swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:deploy, swe-af:manifest, swe-af:issuebuild, swe-af:issue, swe-af:schemas

### ACC-05: Programmable access

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Programmable by design: AgentField HTTP execute APIs (sync and async), af CLI, webhook callbacks, discovery endpoint, and control-plane MCP at <server>/mcp. SWE-AF exposes no SDK of its own and registers no MCP server in node code.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- sdk: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a published SWE-AF client SDK; the AgentField SDKs are the platform's, and SWE-AF depends on the Python and Go SDKs.
- mcp server: The AgentField control plane serves MCP at <server>/mcp for harnesses; node code in swe_af/ and go/ contains no MCP server.
- operations: Create (build, plan, implement_issue, resolve), inspect (executions, notes), stop (cancel execution tree) and resume (resume_build) exist; there is no mid-run steer operation beyond Hax approval forms.
- stability: Public beta: V2 config migration is breaking and AgentField publishes rc builds daily; reasoner names are kept stable between the Python and Go nodes.

- cli — Cli [supported]: af call swe-planner.build --in {...}, af ls --entrypoints and af install/run drive SWE-AF. Sources: swe-af:readme, swe-af:afreadme.
- api — Api [supported]: POST /api/v1/execute/async/swe-planner.<reasoner> returns an execution_id; GET /api/v1/executions/<id> and /notes poll status; GET /api/v1/discovery/capabilities lists entrypoints. Sources: swe-af:readme, swe-af:surface.
- sdk — Sdk [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a published SWE-AF client SDK; the AgentField SDKs are the platform's, and SWE-AF depends on the Python and Go SDKs. Sources: swe-af:pyproject, swe-af:go.
- mcp_server — Mcp server [partial]: The AgentField control plane serves MCP at <server>/mcp for harnesses; node code in swe_af/ and go/ contains no MCP server. Sources: swe-af:afreadme.
- webhooks — Webhooks [supported]: A webhook url on async execute posts the result; approval responses arrive at /api/v1/webhooks/approval-response. Sources: swe-af:readme, swe-af:hitl.
- operations — Operations [partial]: Create (build, plan, implement_issue, resolve), inspect (executions, notes), stop (cancel execution tree) and resume (resume_build) exist; there is no mid-run steer operation beyond Hax approval forms. Sources: swe-af:readme, swe-af:app, swe-af:afcancel.
- authentication — Authentication [supported]: Optional X-API-Key or bearer token on the control plane via AGENTFIELD_API_KEY; GH_TOKEN and provider keys are node environment secrets. Sources: swe-af:afenv.
- stability — Stability [partial]: Public beta: V2 config migration is breaking and AgentField publishes rc builds daily; reasoner names are kept stable between the Python and Go nodes. Sources: swe-af:changelog, swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:afreadme, swe-af:surface, swe-af:pyproject, swe-af:go, swe-af:hitl, swe-af:app, swe-af:afcancel, swe-af:afenv, swe-af:changelog

### ACC-06: External tools and extensions

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

SWE-AF is not an MCP client in its own code. Tools come from the selected harness (Claude Code, OpenCode, Codex); optional Exa web search is wired for the OpenCode runtime. No SWE-AF plugin system.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- mcp client: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an MCP client; MCP use is inherited from the underlying harness (grep mcp in swe_af and go returns nothing).
- plugins: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a plugin or extension registry in the node; af install installs whole nodes, not SWE-AF plugins.
- custom tools: OPENCODE_ENABLE_EXA plus EXA_API_KEY enables OpenCode websearch and webfetch for every reasoner; other tools come from the harness.
- instructions import: Bounded search of swe_af/ and go/ at the pinned commit found no implementation that reads CLAUDE.md, AGENTS.md or skill files into prompts; the harness may read repository instructions itself.
- authentication: EXA_API_KEY and provider keys are environment variables; no per-integration auth UI.
- permission scope: Codex runs with --sandbox workspace-write and read-only for review roles; tool permissions otherwise follow the harness and permission_mode config.

- mcp_client — Mcp client [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an MCP client; MCP use is inherited from the underlying harness (grep mcp in swe_af and go returns nothing). Sources: swe-af:readme, swe-af:app.
- plugins — Plugins [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a plugin or extension registry in the node; af install installs whole nodes, not SWE-AF plugins. Sources: swe-af:manifest, swe-af:readme.
- custom_tools — Custom tools [partial]: OPENCODE_ENABLE_EXA plus EXA_API_KEY enables OpenCode websearch and webfetch for every reasoner; other tools come from the harness. Sources: swe-af:readme.
- instructions_import — Instructions import [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation that reads CLAUDE.md, AGENTS.md or skill files into prompts; the harness may read repository instructions itself. Sources: swe-af:readme.
- authentication — Authentication [partial]: EXA_API_KEY and provider keys are environment variables; no per-integration auth UI. Sources: swe-af:readme, swe-af:manifest.
- permission_scope — Permission scope [partial]: Codex runs with --sandbox workspace-write and read-only for review roles; tool permissions otherwise follow the harness and permission_mode config. Sources: swe-af:codexpatch, swe-af:schemas.
- integration_owner — Integration owner [supported]: Web search is wired by SWE-AF through OpenCode env vars; all other tool connections are owned by the harness runtime. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:app, swe-af:manifest, swe-af:codexpatch, swe-af:schemas

### ACC-07: Client implementation

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No SWE-AF desktop or mobile client. The nodes are Python 3.12 and Go services; the AgentField control plane UI is a React web client served by the Go control plane, and AgentField Desktop is Electron.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with, AgentField Desktop (Electron companion that installs and starts swe-planner); version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- mobile implementation: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a mobile app implementation in the inspected repositories.

- desktop_implementation — Desktop implementation [supported]: AgentField Desktop is Electron (electron-vite, electron-builder); SWE-AF has no desktop UI of its own. Sources: swe-af:afdeskpkg, swe-af:afdesk.
- mobile_implementation — Mobile implementation [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a mobile app implementation in the inspected repositories. Sources: swe-af:afdesk.
- hybrids — Hybrids [supported]: Hybrid platform: Go control plane serving a React/Vite web client, Electron wrapper, and SWE-AF Python and Go nodes. Sources: swe-af:afui, swe-af:go, swe-af:pyproject.
- implementation_evidence — Implementation evidence [supported]: Pinned package.json build config, electron-builder targets and the control-plane web client sources. Sources: swe-af:afdeskpkg, swe-af:afui.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:afdeskpkg, swe-af:afdesk, swe-af:afui, swe-af:go, swe-af:pyproject

## Planning and context

### PLAN-01: Intake and external tracker relationship

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Intake is a prompt or goal passed to build, plus a fully specified issue passed to implement_issue; resolve takes an existing PR with review comments and CI failures. No tracker import or sync in node code.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- local creation: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a UI or API to create saved tasks; builds are one-off executions that generate an internal issue DAG.
- external import: Bounded search of swe_af/ and go/ at the pinned commit found no implementation that imports Linear, Jira or GitHub issues (terms linear, jira, github issue in swe_af); the control plane has Linear and Se
- chat git triggers: resolve is invoked with a PR URL and review comments or a free-form goal from a caller such as a comment bot; the repository ships no bot of its own, and control-plane sources (github, slack, linear) c
- authoritative tracker: No tracker is authoritative; the internal issue DAG and checkpoint.json belong to one build.
- stable ids: Issues have names and a build_id; implement_issue branches use issue/<build_id>-<slug>; execution ids are stable in the control plane; no cross-build issue IDs.
- sync: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of two-way sync with an external tracker.
- supported edits: Planned issues cannot be edited by a person between plan and execute except by passing a modified plan_result to execute.
- external roundtrips: No tracker integration exists, so no tracker round trip applies.

- local_creation — Local creation [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a UI or API to create saved tasks; builds are one-off executions that generate an internal issue DAG. Sources: swe-af:readme, swe-af:app.
- prompt_intake — Prompt intake [supported]: goal string (build/plan) or issue object (implement_issue) is the intake; additional_context is appended. Sources: swe-af:readme, swe-af:issue.
- external_import — External import [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation that imports Linear, Jira or GitHub issues (terms linear, jira, github issue in swe_af); the control plane has Linear and Sentry source triggers but not a SWE-AF importer. Sources: swe-af:afcron, swe-af:app.
- chat_git_triggers — Chat git triggers [partial]: resolve is invoked with a PR URL and review comments or a free-form goal from a caller such as a comment bot; the repository ships no bot of its own, and control-plane sources (github, slack, linear) can route events to reasoners. Sources: swe-af:app, swe-af:afcron.
- authoritative_tracker — Authoritative tracker [unsupported]: No tracker is authoritative; the internal issue DAG and checkpoint.json belong to one build. Sources: swe-af:arch, swe-af:dag.
- stable_ids — Stable ids [partial]: Issues have names and a build_id; implement_issue branches use issue/<build_id>-<slug>; execution ids are stable in the control plane; no cross-build issue IDs. Sources: swe-af:issuebuild, swe-af:dag.
- sync — Sync [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of two-way sync with an external tracker. Sources: swe-af:app.
- supported_edits — Supported edits [unsupported]: Planned issues cannot be edited by a person between plan and execute except by passing a modified plan_result to execute. Sources: swe-af:app, swe-af:arch.
- external_roundtrips — External roundtrips [not-applicable]: No tracker integration exists, so no tracker round trip applies. Sources: swe-af:app.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:app, swe-af:issue, swe-af:afcron, swe-af:arch, swe-af:dag, swe-af:issuebuild

### PLAN-02: Durable task plans

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

build generates a PRD, architecture and issue specs with acceptance criteria, saved as artifacts under .artifacts/plan/; plans can be human-approved through the optional Hax gate (revised plan loop) and execute accepts an edited plan_result. No plan editor UI or revision history.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- editors: With HAX_API_KEY, a Hax plan-review form lets a person approve or request revision; otherwise plan files are plain artifacts a caller may edit before execute.
- revisions: Architecture revisions snapshot and restore on failure and the Hax loop iterates revisions; no versioned history beyond files.
- implementation links: Issue results record branches, commits and files; the PR body includes the PRD and architecture summary and accumulated debt.
- amendment: Failed issues trigger advisor adaptation and replanning of the remaining DAG during a run, but a person cannot amend a running plan.

- plans — Plans [supported]: Product manager, architect, tech lead and sprint planner produce prd.md, architecture.md and per-issue specs saved in .artifacts/plan/. Sources: swe-af:arch, swe-af:readme.
- acceptance_criteria — Acceptance criteria [supported]: PRD and every planned issue carry acceptance criteria; the verifier checks them against the repo and fix issues are generated for failures. Sources: swe-af:arch, swe-af:readme.
- editors — Editors [partial]: With HAX_API_KEY, a Hax plan-review form lets a person approve or request revision; otherwise plan files are plain artifacts a caller may edit before execute. Sources: swe-af:app, swe-af:hitl.
- revisions — Revisions [partial]: Architecture revisions snapshot and restore on failure and the Hax loop iterates revisions; no versioned history beyond files. Sources: swe-af:app.
- implementation_links — Implementation links [partial]: Issue results record branches, commits and files; the PR body includes the PRD and architecture summary and accumulated debt. Sources: swe-af:arch.
- amendment — Amendment [partial]: Failed issues trigger advisor adaptation and replanning of the remaining DAG during a run, but a person cannot amend a running plan. Sources: swe-af:readme, swe-af:dag.
- restart_persistence — Restart persistence [supported]: Plan artifacts and checkpoint.json persist on disk under .artifacts for resume_build. Sources: swe-af:app, swe-af:dag.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:arch, swe-af:readme, swe-af:app, swe-af:hitl, swe-af:dag

### PLAN-03: Hierarchy, dependencies and evolving scope

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Issues form a dependency-sorted DAG executed in levels with depends_on edges; blockers are enforced by scheduling, failures propagate to dependents, and replanning restructures the remaining DAG. No epics or user-defined hierarchy.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- epics: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of epic entities; the unit is one build with a flat issue DAG.
- hierarchy: One flat issue DAG per build; the issue advisor can split an issue into smaller issues, but there is no arbitrary parent/child depth.
- task to epic: The issue advisor can split an issue; no promotion of a simple task to an epic while keeping identity.
- identity retention: Split and replanned issues keep the build checkpoint but get new issue names.
- workflow constraints: The build pipeline (plan, execute, verify, finalize, PR) is fixed; implement_issue and fast mode skip planning.

- epics — Epics [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of epic entities; the unit is one build with a flat issue DAG. Sources: swe-af:arch, swe-af:schemas.
- hierarchy — Hierarchy [partial]: One flat issue DAG per build; the issue advisor can split an issue into smaller issues, but there is no arbitrary parent/child depth. Sources: swe-af:readme, swe-af:arch.
- blockers — Blockers [supported]: PlannedIssue.depends_on edges define blockers; levels are computed topologically. Sources: swe-af:arch, swe-af:dag.
- readiness — Readiness [supported]: An issue runs only when its dependency levels complete; a level aborts when failures reach level_failure_abort_threshold (0.8). Sources: swe-af:dag, swe-af:schemas.
- enforcement — Enforcement [supported]: The executor schedules levels sequentially and runs issues in a level in parallel, so a blocked issue is never dispatched. Sources: swe-af:dag.
- state_propagation — State propagation [supported]: Issue outcomes (completed, failed, skipped) propagate to dependents; unrecoverable failures escalate to the replanner. Sources: swe-af:dag, swe-af:arch.
- task_to_epic — Task to epic [partial]: The issue advisor can split an issue; no promotion of a simple task to an epic while keeping identity. Sources: swe-af:readme.
- identity_retention — Identity retention [partial]: Split and replanned issues keep the build checkpoint but get new issue names. Sources: swe-af:arch.
- workflow_constraints — Workflow constraints [partial]: The build pipeline (plan, execute, verify, finalize, PR) is fixed; implement_issue and fast mode skip planning. Sources: swe-af:readme, swe-af:arch.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:arch, swe-af:schemas, swe-af:readme, swe-af:dag

### PLAN-04: Shared context and memory

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Context is passed through artifacts and prompts. Optional enable_learning injects conventions and failure patterns from earlier issues into later ones within one build. No cross-build memory, no repo instruction import by SWE-AF itself.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- instructions: Bounded search of swe_af/ and go/ at the pinned commit found no implementation that loads CLAUDE.md or AGENTS.md; a harness such as Claude Code may read them itself.
- skills: SWE-AF ships a delegate-issue skill for callers; its own agents do not import skills.
- project docs: Planning agents read the repository and write PRD and architecture documents into .artifacts/plan/ for later roles.
- memory: enable_learning (default false) enables cross-issue shared memory of conventions, failure patterns and bug patterns within one build.
- scope: Learning memory is scoped to one build; artifacts are scoped to the repo .artifacts directory.
- provenance: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of source citations on learned conventions.
- correction: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an editor or delete command for learned memory.
- deletion: Memory ends with the build; artifacts can be removed with make clean-examples or by deleting .artifacts.
- freshness: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of stale-context checks on learned conventions.
- takeover continuity: A new caller can resume_build from checkpoint.json and artifacts; the runtime is re-resolved from the current environment, not frozen.

- instructions — Instructions [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation that loads CLAUDE.md or AGENTS.md; a harness such as Claude Code may read them itself. Sources: swe-af:readme.
- skills — Skills [partial]: SWE-AF ships a delegate-issue skill for callers; its own agents do not import skills. Sources: swe-af:skill.
- project_docs — Project docs [partial]: Planning agents read the repository and write PRD and architecture documents into .artifacts/plan/ for later roles. Sources: swe-af:arch.
- memory — Memory [partial]: enable_learning (default false) enables cross-issue shared memory of conventions, failure patterns and bug patterns within one build. Sources: swe-af:schemas, swe-af:readme.
- scope — Scope [partial]: Learning memory is scoped to one build; artifacts are scoped to the repo .artifacts directory. Sources: swe-af:schemas.
- provenance — Provenance [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of source citations on learned conventions. Sources: swe-af:schemas.
- correction — Correction [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an editor or delete command for learned memory. Sources: swe-af:schemas.
- deletion — Deletion [partial]: Memory ends with the build; artifacts can be removed with make clean-examples or by deleting .artifacts. Sources: swe-af:readme.
- freshness — Freshness [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of stale-context checks on learned conventions. Sources: swe-af:schemas.
- takeover_continuity — Takeover continuity [partial]: A new caller can resume_build from checkpoint.json and artifacts; the runtime is re-resolved from the current environment, not frozen. Sources: swe-af:deploy, swe-af:app.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:skill, swe-af:arch, swe-af:schemas, swe-af:deploy, swe-af:app

### PLAN-05: Architecture and API awareness

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Agents read the repository directly through the selected harness; SWE-AF builds no code index, retrieval API or citation layer, and the verifier checks acceptance criteria against the code.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- code search: Search is whatever the harness provides inside the worktree; SWE-AF adds no index.
- indexing: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a repository index or embedding store.
- external knowledge: Optional Exa websearch/webfetch for the OpenCode runtime when OPENCODE_ENABLE_EXA and EXA_API_KEY are set.
- multimodal inputs: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of image or screenshot inputs in the goal or issue schemas.
- freshness: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of source freshness tracking for external knowledge.
- access scope: Agents work in a worktree of the target repo; read scope is not restricted unless the Go SWE_FENCE option is enabled.
- api retrieval: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an API that retrieves repository context for other agents.
- citations: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of citation of retrieved sources in agent output.
- conflict detection: Sprint planning detects file conflicts between issues scheduled at the same parallel level and the merger resolves branch conflicts; no architecture-conflict detector.
- review enforcement: The code reviewer and verifier agents check quality and acceptance criteria; they are prompts, not a deterministic architecture rule engine.

- code_search — Code search [partial]: Search is whatever the harness provides inside the worktree; SWE-AF adds no index. Sources: swe-af:readme.
- indexing — Indexing [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a repository index or embedding store. Sources: swe-af:schemas, swe-af:arch.
- external_knowledge — External knowledge [partial]: Optional Exa websearch/webfetch for the OpenCode runtime when OPENCODE_ENABLE_EXA and EXA_API_KEY are set. Sources: swe-af:readme.
- multimodal_inputs — Multimodal inputs [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of image or screenshot inputs in the goal or issue schemas. Sources: swe-af:issue, swe-af:schemas.
- freshness — Freshness [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of source freshness tracking for external knowledge. Sources: swe-af:readme.
- access_scope — Access scope [partial]: Agents work in a worktree of the target repo; read scope is not restricted unless the Go SWE_FENCE option is enabled. Sources: swe-af:go.
- api_retrieval — Api retrieval [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an API that retrieves repository context for other agents. Sources: swe-af:readme.
- citations — Citations [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of citation of retrieved sources in agent output. Sources: swe-af:schemas.
- conflict_detection — Conflict detection [partial]: Sprint planning detects file conflicts between issues scheduled at the same parallel level and the merger resolves branch conflicts; no architecture-conflict detector. Sources: swe-af:arch, swe-af:app.
- review_enforcement — Review enforcement [partial]: The code reviewer and verifier agents check quality and acceptance criteria; they are prompts, not a deterministic architecture rule engine. Sources: swe-af:coding, swe-af:arch.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:schemas, swe-af:arch, swe-af:issue, swe-af:go, swe-af:app, swe-af:coding

### PLAN-06: Native issue management and idea backlog

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No native issue management: planned issues are internal build data (state in DAGState and checkpoint.json). No human create, edit, prioritize, label, assign or board views.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- create: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a user-facing issue creation UI; issues are generated by the sprint planner or supplied inline to implement_issue.
- edit: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of issue editing apart from passing an altered plan_result.
- discuss: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of comments or discussion on issues.
- prioritize: Sprint planner orders issues into levels by dependency; there is no user priority field.
- label: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of labels or tags on issues.
- assign: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of human assignment of issues.
- search: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of issue search.
- filter: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of issue filters; the control plane filters runs by status and time range.
- archive: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of archive of issues.
- reopen: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of reopening a closed issue; resume_build re-executes unfinished issues from a checkpoint.
- project epic views: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of project or epic views.
- idea parking: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of saving an idea without starting a run.
- explicit dispatch: plan can run alone and execute dispatches a prior plan later; there is no persisted backlog.
- task session separation: Issues, per-issue branches and execution ids are separate entities within a build; multiple agent calls run per issue.
- metadata: PlannedIssue carries acceptance criteria, files, dependencies, guidance and testing strategy.
- durable states: Issue outcomes persist in checkpoint.json and per-issue logs for the build.

- create — Create [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a user-facing issue creation UI; issues are generated by the sprint planner or supplied inline to implement_issue. Sources: swe-af:app, swe-af:issue.
- edit — Edit [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of issue editing apart from passing an altered plan_result. Sources: swe-af:app.
- discuss — Discuss [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of comments or discussion on issues. Sources: swe-af:app.
- prioritize — Prioritize [partial]: Sprint planner orders issues into levels by dependency; there is no user priority field. Sources: swe-af:arch.
- label — Label [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of labels or tags on issues. Sources: swe-af:schemas.
- assign — Assign [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of human assignment of issues. Sources: swe-af:schemas.
- search — Search [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of issue search. Sources: swe-af:app.
- filter — Filter [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of issue filters; the control plane filters runs by status and time range. Sources: swe-af:afui.
- archive — Archive [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of archive of issues. Sources: swe-af:app.
- reopen — Reopen [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of reopening a closed issue; resume_build re-executes unfinished issues from a checkpoint. Sources: swe-af:app, swe-af:dag.
- project_epic_views — Project epic views [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of project or epic views. Sources: swe-af:afui.
- idea_parking — Idea parking [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of saving an idea without starting a run. Sources: swe-af:app.
- explicit_dispatch — Explicit dispatch [partial]: plan can run alone and execute dispatches a prior plan later; there is no persisted backlog. Sources: swe-af:app.
- task_session_separation — Task session separation [partial]: Issues, per-issue branches and execution ids are separate entities within a build; multiple agent calls run per issue. Sources: swe-af:dag, swe-af:schemas.
- metadata — Metadata [partial]: PlannedIssue carries acceptance criteria, files, dependencies, guidance and testing strategy. Sources: swe-af:arch, swe-af:schemas.
- durable_states — Durable states [partial]: Issue outcomes persist in checkpoint.json and per-issue logs for the build. Sources: swe-af:dag.
- external_tracker_requirement — External tracker requirement [supported]: No external tracker is required for a build; no tracker integration exists. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:app, swe-af:issue, swe-af:arch, swe-af:schemas, swe-af:afui, swe-af:dag, swe-af:readme

### PLAN-07: Agent-native issue operations

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Agents operate on the internal issue DAG during a build (advisor, replanner, fix generator create and modify issues) and a caller agent can delegate issues through implement_issue. There are no durable shared issue IDs that humans and agents both edit.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- shared ids: Issue names are shared across planner, executor and PR text within one build; no IDs persist across builds.
- read: Agents read issue specs from the plan artifacts; a caller reads results from the execution record and notes.
- claim: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of agent claim of tracker issues; executor dispatch is internal.
- assign: The executor assigns issues to coder roles internally; there is no assignee field.
- children: The advisor can split an issue into smaller issues; there is no parent/child record.
- human visibility: The control plane run view and execution notes show progress; a person cannot browse or filter an issue list there.
- attribution: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-agent attribution fields on issues beyond role names in logs.
- permissions: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-issue permissions.
- conflicts: File conflicts between concurrent issues are detected at planning and branch conflicts are resolved by the merger agent.

- shared_ids — Shared ids [partial]: Issue names are shared across planner, executor and PR text within one build; no IDs persist across builds. Sources: swe-af:arch, swe-af:dag.
- read — Read [partial]: Agents read issue specs from the plan artifacts; a caller reads results from the execution record and notes. Sources: swe-af:readme.
- create — Create [supported]: The sprint planner, issue advisor (split) and fix generator create issues programmatically. Sources: swe-af:arch, swe-af:readme.
- update — Update [supported]: The issue advisor adapts acceptance criteria, approach and scope; the replanner restructures remaining issues. Sources: swe-af:readme, swe-af:arch.
- claim — Claim [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of agent claim of tracker issues; executor dispatch is internal. Sources: swe-af:dag.
- assign — Assign [partial]: The executor assigns issues to coder roles internally; there is no assignee field. Sources: swe-af:dag.
- state_changes — State changes [supported]: The executor records completed, failed or skipped outcomes and escalates failures. Sources: swe-af:dag.
- dependencies — Dependencies [supported]: Planner and replanner create depends_on edges. Sources: swe-af:arch.
- children — Children [partial]: The advisor can split an issue into smaller issues; there is no parent/child record. Sources: swe-af:readme.
- results — Results [supported]: Results (branch, commits, files, debt items, verification) return in the execution output and PR body. Sources: swe-af:readme, swe-af:issue.
- human_visibility — Human visibility [partial]: The control plane run view and execution notes show progress; a person cannot browse or filter an issue list there. Sources: swe-af:afui.
- attribution — Attribution [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-agent attribution fields on issues beyond role names in logs. Sources: swe-af:schemas.
- permissions — Permissions [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-issue permissions. Sources: swe-af:schemas.
- conflicts — Conflicts [partial]: File conflicts between concurrent issues are detected at planning and branch conflicts are resolved by the merger agent. Sources: swe-af:arch.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:arch, swe-af:dag, swe-af:readme, swe-af:issue, swe-af:afui, swe-af:schemas

### PLAN-08: Living product specifications

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No living product specification: the PRD is a per-build document, and the verifier checks acceptance criteria only for that build.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- persistent requirements: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of persistent requirements across builds; the PRD is regenerated per build.
- human edits: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a spec editor.
- agent edits: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of agent edits to a persistent spec.
- revisions: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of spec revision history.
- approvals: Optional Hax approval of the plan before coding starts; not a persisted spec approval.
- implementation links: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of requirement-to-code links beyond acceptance criteria text in issue results.
- later retrieval: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of retrieving a prior PRD in later builds.
- drift detection: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of drift checks between specification and code beyond one build's verifier pass.
- review integration: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of review integration with a persistent spec.

- persistent_requirements — Persistent requirements [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of persistent requirements across builds; the PRD is regenerated per build. Sources: swe-af:arch.
- human_edits — Human edits [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a spec editor. Sources: swe-af:app.
- agent_edits — Agent edits [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of agent edits to a persistent spec. Sources: swe-af:arch.
- revisions — Revisions [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of spec revision history. Sources: swe-af:arch.
- approvals — Approvals [partial]: Optional Hax approval of the plan before coding starts; not a persisted spec approval. Sources: swe-af:app, swe-af:hitl.
- implementation_links — Implementation links [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of requirement-to-code links beyond acceptance criteria text in issue results. Sources: swe-af:arch.
- later_retrieval — Later retrieval [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of retrieving a prior PRD in later builds. Sources: swe-af:arch.
- drift_detection — Drift detection [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of drift checks between specification and code beyond one build's verifier pass. Sources: swe-af:arch.
- review_integration — Review integration [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of review integration with a persistent spec. Sources: swe-af:arch.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:arch, swe-af:app, swe-af:hitl

## Coordination and automation

### COORD-01: Parallel work

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Parallelism is the core design: issues in a dependency level run concurrently (max_concurrent_issues default 3, 0 for unlimited) in isolated worktrees, and several builds can run across nodes. Status is visible in control plane runs and notes.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- quotas: max_concurrent_issues (default 3) caps parallel issues per level; the README advises a handful of concurrent implement_issue calls and says the node bounds its own concurrency; no account quotas.
- observed concurrency: README reports 79 agent invocations for one build and 400-500+ typical; those are vendor-reported figures, not measured here.

- concurrent_sessions — Concurrent sessions [supported]: Many builds and issue-level calls can execute concurrently; docker compose --scale swe-agent=3 adds worker nodes. Sources: swe-af:readme.
- concurrent_tasks — Concurrent tasks [supported]: Issues in one level run in parallel via asyncio gather bounded by a semaphore. Sources: swe-af:dag.
- quotas — Quotas [partial]: max_concurrent_issues (default 3) caps parallel issues per level; the README advises a handful of concurrent implement_issue calls and says the node bounds its own concurrency; no account quotas. Sources: swe-af:schemas, swe-af:readme.
- observed_concurrency — Observed concurrency [partial]: README reports 79 agent invocations for one build and 400-500+ typical; those are vendor-reported figures, not measured here. Sources: swe-af:readme.
- status_visibility — Status visibility [supported]: Control plane runs show per-step status and workflow DAG; execution notes stream progress tags per phase. Sources: swe-af:afui, swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:dag, swe-af:schemas, swe-af:afui

### COORD-02: Ownership and dispatch

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Dispatch is automatic by the DAG executor. No human assignment, claims or queues; duplicate work is prevented by per-issue worktrees and branches and file-conflict detection.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- human assignment: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of human assignment of issues.
- agent assignment: The executor assigns each issue to coder, reviewer and QA roles with per-role model settings.
- claims: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of claims or locks on issues by external agents.
- queues: Issues wait for their dependency level and the semaphore; there is no persisted queue.
- routing: The sprint planner sets guidance needs_deeper_qa and complexity that select the coder-reviewer or QA-synthesizer path.
- duplicate prevention: Per-issue worktrees and issue/<build_id>-<slug> branches isolate concurrent work and planning detects same-level file conflicts.
- ownership enforcement: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of enforced ownership beyond branch separation.

- human_assignment — Human assignment [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of human assignment of issues. Sources: swe-af:dag.
- agent_assignment — Agent assignment [partial]: The executor assigns each issue to coder, reviewer and QA roles with per-role model settings. Sources: swe-af:dag, swe-af:coding.
- claims — Claims [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of claims or locks on issues by external agents. Sources: swe-af:dag.
- queues — Queues [partial]: Issues wait for their dependency level and the semaphore; there is no persisted queue. Sources: swe-af:dag.
- routing — Routing [partial]: The sprint planner sets guidance needs_deeper_qa and complexity that select the coder-reviewer or QA-synthesizer path. Sources: swe-af:coding.
- duplicate_prevention — Duplicate prevention [partial]: Per-issue worktrees and issue/<build_id>-<slug> branches isolate concurrent work and planning detects same-level file conflicts. Sources: swe-af:issuebuild, swe-af:arch.
- ownership_enforcement — Ownership enforcement [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of enforced ownership beyond branch separation. Sources: swe-af:dag.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:dag, swe-af:coding, swe-af:issuebuild, swe-af:arch

### COORD-03: Inspectable delegation and communication

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Product-orchestrated multi-agent system: roles (product manager, architect, tech lead, sprint planner, coder, QA, reviewer, synthesizer, advisor, replanner, merger, integration tester, verifier, PR agent) are separate reasoner calls with per-role models; results return as structured outputs and artifacts. Auth follows control plane auth; no cross-harness teams in one build.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- model harness effort: models map sets model per role (pm, architect, coder, qa, verifier and others); one runtime per build; effort is set only through the engine variant or model ids.
- messages: Roles exchange structured outputs and artifact files; there is no free-form agent-to-agent messaging channel.
- cross harness: Different runtimes are selectable per build but not mixed inside a build.
- authentication: Calls pass through the AgentField control plane (optional API key); node roles use the build environment secrets.
- followup: resume_build and resolve act on prior state or an existing PR; a finished child agent cannot be reopened for a follow-up question.

- roles — Roles [supported]: 16+ named role reasoners with fixed responsibilities listed in the README agent table. Sources: swe-af:readme.
- model_harness_effort — Model harness effort [partial]: models map sets model per role (pm, architect, coder, qa, verifier and others); one runtime per build; effort is set only through the engine variant or model ids. Sources: swe-af:readme, swe-af:pro.
- messages — Messages [partial]: Roles exchange structured outputs and artifact files; there is no free-form agent-to-agent messaging channel. Sources: swe-af:arch.
- results_return — Results return [supported]: Each reasoner returns schema-validated results to the executor; the executor merges and reports in BuildResult. Sources: swe-af:dag, swe-af:schemas.
- cross_harness — Cross harness [partial]: Different runtimes are selectable per build but not mixed inside a build. Sources: swe-af:readme.
- orchestration_owner — Orchestration owner [supported]: SWE-AF code (build, execute, dag_executor) owns orchestration; delegating a scoped issue to implement_issue from a main harness is a supported sub-harness use. Sources: swe-af:app, swe-af:skill.
- activation — Activation [supported]: Active by default in build; enable_issue_advisor and enable_replanning default true; the Go pro engine is beta and opt-in outside af install. Sources: swe-af:schemas, swe-af:pro.
- authentication — Authentication [partial]: Calls pass through the AgentField control plane (optional API key); node roles use the build environment secrets. Sources: swe-af:afenv.
- live_progress — Live progress [supported]: app.note emits tagged progress events per phase and issue; the control plane shows them in run view and /executions/<id>/notes. Sources: swe-af:readme, swe-af:afui.
- retained_results — Retained results [supported]: Plan, execution logs per issue, checkpoints and verification results persist under .artifacts/. Sources: swe-af:readme.
- followup — Followup [partial]: resume_build and resolve act on prior state or an existing PR; a finished child agent cannot be reopened for a follow-up question. Sources: swe-af:app.
- invocation — Invocation [supported]: Specialists are callable directly as swe-planner.<agent>; discovery marks internal roles do-not-call-directly. Sources: swe-af:readme, swe-af:surface.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:pro, swe-af:arch, swe-af:dag, swe-af:schemas, swe-af:app, swe-af:skill, swe-af:afenv, swe-af:afui, swe-af:surface

### COORD-04: Unattended triggers

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

SWE-AF has no scheduler of its own. The AgentField control plane ships trigger sources (cron, GitHub, Linear, Slack, Stripe, generic webhook) that can invoke reasoners such as swe-planner.build; completion conditions and overlap handling are the control plane's, and not SWE-AF-specific.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- schedules: Control plane cron trigger source can dispatch to a reasoner; SWE-AF code adds no scheduling.
- events: Control plane sources for github, linear, slack, sentry, stripe and generic webhook/HMAC events dispatch to reasoners.
- recurring tasks: Recurring runs depend on cron triggers configured in the control plane, not SWE-AF configuration.
- completion conditions: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of completion conditions for recurring tasks.
- overlaps: Bounded search of swe_af/ and go/ at the pinned commit found no implementation in SWE-AF of overlap policy for repeated triggers.
- retries: Execution retry exists in the control plane run view (Retry now, Restart run, Replay) and resolve/CI-fix loops retry bounded times.

- schedules — Schedules [partial]: Control plane cron trigger source can dispatch to a reasoner; SWE-AF code adds no scheduling. Sources: swe-af:afcron, swe-af:pro.
- events — Events [partial]: Control plane sources for github, linear, slack, sentry, stripe and generic webhook/HMAC events dispatch to reasoners. Sources: swe-af:afcron.
- recurring_tasks — Recurring tasks [partial]: Recurring runs depend on cron triggers configured in the control plane, not SWE-AF configuration. Sources: swe-af:afcron.
- completion_conditions — Completion conditions [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of completion conditions for recurring tasks. Sources: swe-af:app.
- scheduler_location — Scheduler location [supported]: Triggers run in the AgentField control plane process on the operator host. Sources: swe-af:afcron, swe-af:afui.
- overlaps — Overlaps [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation in SWE-AF of overlap policy for repeated triggers. Sources: swe-af:app.
- retries — Retries [partial]: Execution retry exists in the control plane run view (Retry now, Restart run, Replay) and resolve/CI-fix loops retry bounded times. Sources: swe-af:afui, swe-af:ci.
- client_independence — Client independence [supported]: Triggers and node execution run on the server side, independent of any client. Sources: swe-af:afcron.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:afcron, swe-af:pro, swe-af:app, swe-af:afui, swe-af:ci

### COORD-05: Intervention and attention

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

A running build can be cancelled through the control plane, resumed with resume_build, and paused for human input when the Hax gate is enabled. No in-run steering or chat; plan approval and ask_user forms are the intervention points.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- pause: Workflow pause occurs for Hax plan approval and ask_user forms (app.pause) when HAX_API_KEY is set; no general pause control.
- steer: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of mid-run steering messages to coder agents.
- notifications: Approval requests go out through Hax, webhooks report completion; no SWE-AF desktop notifications.
- pending decisions: Hax forms surface pending plan or clarification decisions with a 72-hour default expiry.
- chat terminal parity: SWE-AF has no chat or terminal surface; all control goes through API calls.

- pause — Pause [partial]: Workflow pause occurs for Hax plan approval and ask_user forms (app.pause) when HAX_API_KEY is set; no general pause control. Sources: swe-af:hitl, swe-af:app.
- stop — Stop [supported]: The control plane can cancel an execution and its child tree. Sources: swe-af:afcancel.
- resume — Resume [supported]: resume_build resumes from the last checkpoint; the Go and Python docs note the runtime is re-resolved. Sources: swe-af:app, swe-af:deploy.
- steer — Steer [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of mid-run steering messages to coder agents. Sources: swe-af:app, swe-af:coding.
- notifications — Notifications [partial]: Approval requests go out through Hax, webhooks report completion; no SWE-AF desktop notifications. Sources: swe-af:hitl, swe-af:readme.
- pending_decisions — Pending decisions [partial]: Hax forms surface pending plan or clarification decisions with a 72-hour default expiry. Sources: swe-af:schemas, swe-af:hitl.
- escalation — Escalation [supported]: Failures escalate inner loop to issue advisor to replanner and finally to a human via the open PR with failing checks. Sources: swe-af:readme, swe-af:ci.
- visible_errors — Visible errors [supported]: Failures, compromise debt items and CI failures are recorded in results, notes and the PR body. Sources: swe-af:arch, swe-af:ci.
- chat_terminal_parity — Chat terminal parity [not-applicable]: SWE-AF has no chat or terminal surface; all control goes through API calls. Sources: swe-af:readme.
- retained_history — Retained history [supported]: Execution notes, run records and .artifacts keep history after cancellation or failure. Sources: swe-af:readme, swe-af:afui.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:hitl, swe-af:app, swe-af:afcancel, swe-af:deploy, swe-af:coding, swe-af:readme, swe-af:schemas, swe-af:ci, swe-af:arch, swe-af:afui

### COORD-06: Fleet and epic management UI

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

The AgentField control plane shows runs with status, step trace and workflow DAG for SWE-AF builds, plus execution history and filters. It is a run viewer, not an epic or fleet board.

Availability: generally-available
Delivery: native
Scope: surfaces=AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with, SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- fleet fixture: Control plane Executions and Workflows pages list many runs with status and duration; the pinned screenshot shows sample non-SWE-AF agents.
- dependencies: The DAG graph shows call relations, not issue depends_on edges.
- attention: Failed and degraded states are visible in run lists; no dedicated attention queue.
- workload: Dashboard summarises agents online and executions; no per-person workload.
- navigation errors: Failed steps are visible in the trace with error logs; no one-click failure navigation was verified.
- fixture mode: No fixture mode applies to a source-and-documentation pass; no hands-on fixture was run.

- fleet_fixture — Fleet fixture [partial]: Control plane Executions and Workflows pages list many runs with status and duration; the pinned screenshot shows sample non-SWE-AF agents. Sources: swe-af:afui, swe-af:afshot.
- parent_child — Parent child [supported]: Workflow DAG and step trace show parent and child agent calls inside a build. Sources: swe-af:afui.
- dependencies — Dependencies [partial]: The DAG graph shows call relations, not issue depends_on edges. Sources: swe-af:afui.
- attention — Attention [partial]: Failed and degraded states are visible in run lists; no dedicated attention queue. Sources: swe-af:afui.
- history — History [supported]: Executions and workflows are retained and filterable by time range and status. Sources: swe-af:afui, swe-af:afshot.
- workload — Workload [partial]: Dashboard summarises agents online and executions; no per-person workload. Sources: swe-af:afui.
- navigation_errors — Navigation errors [partial]: Failed steps are visible in the trace with error logs; no one-click failure navigation was verified. Sources: swe-af:afui.
- fixture_mode — Fixture mode [not-applicable]: No fixture mode applies to a source-and-documentation pass; no hands-on fixture was run. Sources: swe-af:afui.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:afui, swe-af:afshot

### COORD-07: Workflow flexibility and orchestration visibility

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Entry points are fixed (build, plan, execute, implement_issue, resolve, resume_build) with config for models per role and runtime. No natural-language workflow configuration or in-run role changes; the control plane DAG shows which roles ran.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- growing epic: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of promoting a quick task to a larger plan with retained identity.
- presets: Mode choice (swe-planner vs swe-fast, single vs multi-repo) is required per call; there are no saved presets.
- natural language: The goal is natural language, but workflow settings are JSON config keys.
- in run changes: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of changing config during a running build.
- auth friction: Needs the control plane, node, a provider key and optionally GH_TOKEN and a Hax key before a first build.

- quick_task — Quick task [supported]: implement_issue and swe-fast.build provide lighter paths than the full planner build. Sources: swe-af:readme, swe-af:issue.
- growing_epic — Growing epic [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of promoting a quick task to a larger plan with retained identity. Sources: swe-af:app.
- presets — Presets [partial]: Mode choice (swe-planner vs swe-fast, single vs multi-repo) is required per call; there are no saved presets. Sources: swe-af:readme.
- natural_language — Natural language [partial]: The goal is natural language, but workflow settings are JSON config keys. Sources: swe-af:readme.
- agent_configuration — Agent configuration [supported]: models map overrides each role; runtime, iterations, advisor and replanning budgets are configurable. Sources: swe-af:readme, swe-af:schemas.
- in_run_changes — In run changes [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of changing config during a running build. Sources: swe-af:app.
- live_roles — Live roles [supported]: The control plane trace and notes show which role reasoners are running. Sources: swe-af:afui, swe-af:readme.
- retained_relationships — Retained relationships [supported]: Run DAG and artifacts retain role call relationships after completion. Sources: swe-af:afui, swe-af:readme.
- discoverability — Discoverability [supported]: af ls --entrypoints and GET /api/v1/discovery/capabilities list entry points with routing descriptions. Sources: swe-af:surface, swe-af:readme.
- auth_friction — Auth friction [partial]: Needs the control plane, node, a provider key and optionally GH_TOKEN and a Hax key before a first build. Sources: swe-af:readme, swe-af:manifest.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:issue, swe-af:app, swe-af:schemas, swe-af:afui, swe-af:surface, swe-af:manifest

## Execution and reliability

### RUN-01: Laptop-independent execution

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Local-first service architecture: node and control plane run on the operator machine, a Docker host or Railway, and executions continue server-side after the caller disconnects. No vendor-operated cloud for SWE-AF; payer is the operator.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- managed cloud: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a SWE-AF-managed cloud; AgentField offers Enterprise private deployment by contact only, which is a business offering outsi
- hybrid: Host control plane plus Docker node via docker-compose.local.yml is supported; hybrid cloud variants are not described.
- provisioning: Compose and Railway provision services; no dynamic compute provisioning per task.
- resources: Resource sizing is the operator's; deploy docs list software prerequisites, not CPU/RAM.
- region: Region selection belongs to the operator host; SWE-AF defines none.
- expiry: Default per-agent timeout is 2700 s (1800 s for issue-level); the pause clock covers Hax waits up to 72 hours; there is no sandbox expiry.

- local — Local [supported]: python -m swe_af, af run swe-planner and make run-planner run on the local machine. Sources: swe-af:readme, swe-af:go.
- byo_remote — Byo remote [supported]: Docker Compose on any host and Railway deploy the node and control plane on user-controlled infrastructure. Sources: swe-af:readme, swe-af:deploy.
- self_hosted — Self hosted [supported]: The whole stack (control plane, PostgreSQL, node) is self-hosted from Apache-2.0 sources. Sources: swe-af:readme, swe-af:license.
- managed_cloud — Managed cloud [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a SWE-AF-managed cloud; AgentField offers Enterprise private deployment by contact only, which is a business offering outside SWE-AF code. Sources: swe-af:readme.
- hybrid — Hybrid [partial]: Host control plane plus Docker node via docker-compose.local.yml is supported; hybrid cloud variants are not described. Sources: swe-af:readme.
- provisioning — Provisioning [partial]: Compose and Railway provision services; no dynamic compute provisioning per task. Sources: swe-af:readme, swe-af:deploy.
- payer — Payer [supported]: The operator pays the host and provider tokens; GH_TOKEN and provider keys are the operator's. Sources: swe-af:readme.
- resources — Resources [partial]: Resource sizing is the operator's; deploy docs list software prerequisites, not CPU/RAM. Sources: swe-af:deploy.
- region — Region [not-applicable]: Region selection belongs to the operator host; SWE-AF defines none. Sources: swe-af:deploy.
- expiry — Expiry [partial]: Default per-agent timeout is 2700 s (1800 s for issue-level); the pause clock covers Hax waits up to 72 hours; there is no sandbox expiry. Sources: swe-af:schemas, swe-af:issue, swe-af:hitl.
- client_disconnect — Client disconnect [supported]: Async execute returns an execution_id and the control plane keeps running the build when the caller disconnects. Sources: swe-af:readme.
- other_device_continuation — Other device continuation [supported]: Any client reaching the control plane can poll or cancel the same execution. Sources: swe-af:readme, swe-af:afui.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:go, swe-af:deploy, swe-af:license, swe-af:schemas, swe-af:issue, swe-af:hitl, swe-af:afui

### RUN-02: Workspace and process isolation

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Each issue runs in its own git worktree on its own branch; an opt-in Go-only SWE_FENCE wraps the coding engine in sandbox-exec (macOS) or bwrap (Linux); Codex uses --sandbox workspace-write. The Docker deployment contains the whole node, not each task. Not a hostile-code security boundary.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- container: docker-compose runs the nodes in containers, but individual builds are not placed in separate containers.
- vm: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of VM isolation (terms firecracker, qemu, microvm, virtual machine).
- os sandbox: SWE_FENCE=1 (Go node, opt-in) wraps codeaf with sandbox-exec or bwrap; Codex sandbox flags for the codex runtime.
- filesystem: SWE_FENCE blocks reads and writes to sibling builds under SWE_WORKSPACE_ROOT with documented read-scope limits; otherwise agents can read the host per harness permissions.
- process: Processes are separated only when the fence or container is used.
- network: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of network isolation or egress policy for build agents.
- secrets: Scout-negotiated credentials are scoped and redacted from logs; provider keys and GH_TOKEN are inherited by agent processes.
- security boundary: README and go docs document the fence read-scope limitation; no separate hostile-code containment evidence is claimed.

- checkout — Checkout [supported]: build clones to a workspace and plans against it; repo_path mode uses an existing checkout. Sources: swe-af:readme, swe-af:app.
- branch — Branch [supported]: Integration branch plus issue branches per issue (issue/<build_id>-<slug>). Sources: swe-af:issuebuild.
- worktree — Worktree [supported]: git worktree add per issue with retries for concurrent adds; deterministic_git default true. Sources: swe-af:dag, swe-af:issuebuild.
- container — Container [partial]: docker-compose runs the nodes in containers, but individual builds are not placed in separate containers. Sources: swe-af:readme.
- vm — Vm [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of VM isolation (terms firecracker, qemu, microvm, virtual machine). Sources: swe-af:go.
- os_sandbox — Os sandbox [partial]: SWE_FENCE=1 (Go node, opt-in) wraps codeaf with sandbox-exec or bwrap; Codex sandbox flags for the codex runtime. Sources: swe-af:go, swe-af:codexpatch.
- filesystem — Filesystem [partial]: SWE_FENCE blocks reads and writes to sibling builds under SWE_WORKSPACE_ROOT with documented read-scope limits; otherwise agents can read the host per harness permissions. Sources: swe-af:go.
- process — Process [partial]: Processes are separated only when the fence or container is used. Sources: swe-af:go.
- network — Network [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of network isolation or egress policy for build agents. Sources: swe-af:go, swe-af:readme.
- secrets — Secrets [partial]: Scout-negotiated credentials are scoped and redacted from logs; provider keys and GH_TOKEN are inherited by agent processes. Sources: swe-af:creds, swe-af:readme.
- code_separation — Code separation [supported]: Worktrees separate concurrent edits per issue; a failed build with no commits deletes its branch. Sources: swe-af:issuebuild, swe-af:readme.
- security_boundary — Security boundary [unsupported]: README and go docs document the fence read-scope limitation; no separate hostile-code containment evidence is claimed. Sources: swe-af:go.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:app, swe-af:issuebuild, swe-af:dag, swe-af:go, swe-af:codexpatch, swe-af:creds

### RUN-03: Recovery, expiry and continuity

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Checkpointed execution supports resume_build after a crash; the control plane can retry or restart runs. Resume re-resolves the runtime from current config; worktrees and artifacts persist on the node host.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- expiry: No sandbox expiry exists to recover from; per-agent timeouts apply.
- transcript reload: Execution notes and artifact files can be re-read; harness conversation transcripts are not stored by SWE-AF.
- process reconnect: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of reconnecting to a running harness process; resume restarts the unfinished issues.
- context loss: The rebuilt plan has an empty PRD and architecture dict (state held in summaries), so resumed roles lose original planning detail.
- duplicate actions: Completed issues are skipped on resume; the runtime note warns a resumed build can use a different harness than the original run.
- manual repair: resume_build needs repo_path and artifacts_dir from the caller; failed builds with commits return the branch for triage.
- multi day continuity: Hax requests stay open 72 hours by default and the pause clock handles waits; long builds rely on checkpoints.

- client_disconnect — Client disconnect [supported]: Executions continue server-side when the caller disconnects. Sources: swe-af:readme.
- worker_crash — Worker crash [supported]: DAGState checkpoints after levels let resume_build continue; AgentField also dispatches interrupted work to a replacement agent instance after AGENTFIELD_RESUME_INTERRUPTED_DELAY. Sources: swe-af:dag, swe-af:app, swe-af:afenv.
- service_restart — Service restart [supported]: Checkpoint.json and .artifacts persist across restarts; resume_build is invoked by a caller. Sources: swe-af:app.
- expiry — Expiry [not-applicable]: No sandbox expiry exists to recover from; per-agent timeouts apply. Sources: swe-af:schemas.
- transcript_reload — Transcript reload [partial]: Execution notes and artifact files can be re-read; harness conversation transcripts are not stored by SWE-AF. Sources: swe-af:readme.
- process_reconnect — Process reconnect [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of reconnecting to a running harness process; resume restarts the unfinished issues. Sources: swe-af:app.
- task_resume — Task resume [supported]: resume_build calls execute with resume=True over the saved plan and DAGState. Sources: swe-af:app.
- state_reconstruction — State reconstruction [supported]: Plan_result is rebuilt from the checkpoint's all_issues and levels. Sources: swe-af:app.
- context_loss — Context loss [partial]: The rebuilt plan has an empty PRD and architecture dict (state held in summaries), so resumed roles lose original planning detail. Sources: swe-af:app.
- duplicate_actions — Duplicate actions [partial]: Completed issues are skipped on resume; the runtime note warns a resumed build can use a different harness than the original run. Sources: swe-af:deploy, swe-af:dag.
- manual_repair — Manual repair [partial]: resume_build needs repo_path and artifacts_dir from the caller; failed builds with commits return the branch for triage. Sources: swe-af:app, swe-af:readme.
- multi_day_continuity — Multi day continuity [partial]: Hax requests stay open 72 hours by default and the pause clock handles waits; long builds rely on checkpoints. Sources: swe-af:schemas, swe-af:hitl.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:dag, swe-af:app, swe-af:afenv, swe-af:schemas, swe-af:deploy, swe-af:hitl

### RUN-04: Development environment reliability

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No dependency-installation, service or port management for the target project; environment setup is left to the repo and agents. The environment scout agent can request missing credentials from a human; compose manages SWE-AF's own services.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- dependencies: An environment_scout planning role inspects what the build needs and can ask for env vars through a form; installs happen in the agents' shell.
- caches: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of dependency cache handling for target repos.
- snapshots: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of environment snapshots.
- services: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of starting databases or services for target projects.
- ports: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of port allocation for target projects; node ports are fixed defaults (8003, 8005, 8006).
- secrets: Scout-negotiated credentials are stored per build with redaction; node secrets come from the environment or af install encrypted storage.
- duplicate installs: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of install deduplication across worktrees.
- collisions: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of port collision handling.
- remote setup: deployment.md documents server setup for SWE-AF itself, not per-task remote environments.
- automatic repair: The CI gate fix loop and verify-fix loop repair failing checks and criteria within bounded cycles.

- dependencies — Dependencies [partial]: An environment_scout planning role inspects what the build needs and can ask for env vars through a form; installs happen in the agents' shell. Sources: swe-af:creds, swe-af:hitl.
- caches — Caches [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of dependency cache handling for target repos. Sources: swe-af:go.
- snapshots — Snapshots [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of environment snapshots. Sources: swe-af:go.
- services — Services [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of starting databases or services for target projects. Sources: swe-af:go.
- ports — Ports [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of port allocation for target projects; node ports are fixed defaults (8003, 8005, 8006). Sources: swe-af:go, swe-af:readme.
- secrets — Secrets [partial]: Scout-negotiated credentials are stored per build with redaction; node secrets come from the environment or af install encrypted storage. Sources: swe-af:creds, swe-af:readme.
- duplicate_installs — Duplicate installs [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of install deduplication across worktrees. Sources: swe-af:dag.
- collisions — Collisions [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of port collision handling. Sources: swe-af:dag.
- remote_setup — Remote setup [partial]: deployment.md documents server setup for SWE-AF itself, not per-task remote environments. Sources: swe-af:deploy.
- automatic_repair — Automatic repair [partial]: The CI gate fix loop and verify-fix loop repair failing checks and criteria within bounded cycles. Sources: swe-af:ci, swe-af:arch.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:creds, swe-af:hitl, swe-af:go, swe-af:readme, swe-af:dag, swe-af:deploy, swe-af:ci, swe-af:arch

### RUN-05: Limits and failure handling

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Per-agent timeout (2700 s) and turn budget (150), bounded iteration, advisor, replan and CI-fix counts are enforced; there is no spend or token limit in the Python node, and an optional per-run cost ceiling (SWE_PRO_MAX_COST) applies to the Go pro engine only.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- token limits: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of token ceilings in BuildConfig.
- spend limits: SWE_PRO_MAX_COST forwards a USD ceiling to the beta Go pro engine per dispatch; the classic loop has no spend limit.
- enforcement: Time, turn and iteration limits are enforced stops; cost control is provider-side billing apart from the Go engine ceiling.

- time_limits — Time limits [supported]: agent_timeout_seconds 2700 (1800 for implement_issue) wraps each agent call in asyncio.wait_for. Sources: swe-af:schemas, swe-af:coding, swe-af:issue.
- turn_limits — Turn limits [supported]: agent_max_turns 150 (50 for implement_issue). Sources: swe-af:readme, swe-af:issue.
- token_limits — Token limits [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of token ceilings in BuildConfig. Sources: swe-af:schemas.
- spend_limits — Spend limits [partial]: SWE_PRO_MAX_COST forwards a USD ceiling to the beta Go pro engine per dispatch; the classic loop has no spend limit. Sources: swe-af:go.
- timeouts — Timeouts [supported]: Agent calls time out; CI watch caps at 1500 s; Hax requests expire after 72 h. Sources: swe-af:schemas, swe-af:ci, swe-af:hitl.
- retries — Retries [supported]: max_coding_iterations, max_advisor_invocations, max_replans and max_ci_fix_cycles bound retries. Sources: swe-af:schemas, swe-af:readme.
- provider_failures — Provider failures [supported]: fatal_error classifies fatal harness errors (auth, rate limits) and aborts instead of retrying blindly. Sources: swe-af:dag.
- enforcement — Enforcement [partial]: Time, turn and iteration limits are enforced stops; cost control is provider-side billing apart from the Go engine ceiling. Sources: swe-af:schemas, swe-af:go.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:schemas, swe-af:coding, swe-af:issue, swe-af:readme, swe-af:go, swe-af:ci, swe-af:hitl, swe-af:dag

### RUN-06: Truthful status and operational statistics

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Per-phase notes, per-issue logs, planning raw-response logs and run results persist; the control plane adds run, trace and status views. No per-task statistics dashboard or usage export in SWE-AF; the control plane ingests token and cost_usd entries.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- usage export: Control plane ingests per-entry cost_usd and tokens; no export command for SWE-AF was found in the node.
- task stats: BuildResult and execution records carry counts and durations; README cost tables are vendor-reported.
- agent stats: Control plane usage ingest attributes entries to executions and agents; no SWE-AF per-role dashboard exists.
- project stats: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-project statistics.
- time stats: Control plane shows duration and start times per execution; dashboard velocity charts exist.
- approval status: Hax request state (submitted, timeout, cancelled, error) returns in the workflow, not as a board state.
- reconnect consistency: State is reconstructed from checkpoint.json and the control plane record after restart.
- performance method: README benchmark figures are vendor-reported; this catalog measures nothing.
- responsiveness: No responsiveness claims are sourced beyond vendor benchmarks; none measured here.

- logs — Logs [supported]: .artifacts holds plan logs, per-issue execution logs and agent outputs; AgentField exposes execution notes and node logs. Sources: swe-af:readme, swe-af:afui.
- traces — Traces [supported]: Control plane run trace and DAG graph show step timing and parent/child calls. Sources: swe-af:afui.
- failure_reasons — Failure reasons [supported]: Results carry issue outcomes, error messages and typed compromise debt; CI results record failing checks. Sources: swe-af:arch, swe-af:ci.
- usage_export — Usage export [partial]: Control plane ingests per-entry cost_usd and tokens; no export command for SWE-AF was found in the node. Sources: swe-af:afusage.
- task_stats — Task stats [partial]: BuildResult and execution records carry counts and durations; README cost tables are vendor-reported. Sources: swe-af:schemas, swe-af:readme.
- agent_stats — Agent stats [partial]: Control plane usage ingest attributes entries to executions and agents; no SWE-AF per-role dashboard exists. Sources: swe-af:afusage.
- project_stats — Project stats [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-project statistics. Sources: swe-af:schemas.
- time_stats — Time stats [partial]: Control plane shows duration and start times per execution; dashboard velocity charts exist. Sources: swe-af:afui.
- failure_status — Failure status [supported]: Execution status values distinguish succeeded, failed and degraded; empty builds report failed through ReasonerFailed. Sources: swe-af:pyproject, swe-af:afui.
- approval_status — Approval status [partial]: Hax request state (submitted, timeout, cancelled, error) returns in the workflow, not as a board state. Sources: swe-af:hitl.
- reconnect_consistency — Reconnect consistency [partial]: State is reconstructed from checkpoint.json and the control plane record after restart. Sources: swe-af:app, swe-af:dag.
- performance_method — Performance method [unsupported]: README benchmark figures are vendor-reported; this catalog measures nothing. Sources: swe-af:readme.
- responsiveness — Responsiveness [unsupported]: No responsiveness claims are sourced beyond vendor benchmarks; none measured here. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:afui, swe-af:arch, swe-af:ci, swe-af:afusage, swe-af:schemas, swe-af:pyproject, swe-af:hitl, swe-af:app, swe-af:dag

### RUN-07: Per-task API and compute cost

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

AgentField ingests usage entries with cost_usd and token counts per execution; SWE-AF documents one measured build cost ($19.23, 79 invocations) by role. The Python node does not aggregate cost per issue or epic and billing reconciliation is manual.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- task cost: Control plane records usage entries per execution; per-issue rollups are not provided by SWE-AF.
- session cost: Usage is attached to executions in the control plane; the SWE-AF README shows a single build total as example.
- child cost: Child execution usage is retained per execution, with parent totals not double-added per the telemetry doc.
- epic cost: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of epic-level cost aggregation.
- harness aggregation: Harness entries are kept in local accounting; the Codex patch estimates CLI cost with estimate_cli_cost and the Command Code patch reports null cost.
- billed vs estimated: SDK-native llm usage is recorded; Codex costs are estimates; Command Code cost is unknown.
- subscription allowance: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of subscription allowance reporting; subscription-token runs have no marginal USD cost shown.
- provider reconciliation: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of provider invoice reconciliation.
- compute cost: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of compute-cost tracking; infra is operator-paid.
- refresh delay: Usage arrives in the execution result envelope, so a killed agent can lose entries.
- alerts: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of budget alerts.
- export: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of CSV or API export of usage in SWE-AF.
- drilldown: UI usage handlers exist in the control plane; drill-down by SWE-AF role was not verified.

- task_cost — Task cost [partial]: Control plane records usage entries per execution; per-issue rollups are not provided by SWE-AF. Sources: swe-af:afusage.
- session_cost — Session cost [partial]: Usage is attached to executions in the control plane; the SWE-AF README shows a single build total as example. Sources: swe-af:afusage, swe-af:readme.
- child_cost — Child cost [partial]: Child execution usage is retained per execution, with parent totals not double-added per the telemetry doc. Sources: swe-af:afusage.
- epic_cost — Epic cost [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of epic-level cost aggregation. Sources: swe-af:schemas.
- harness_aggregation — Harness aggregation [partial]: Harness entries are kept in local accounting; the Codex patch estimates CLI cost with estimate_cli_cost and the Command Code patch reports null cost. Sources: swe-af:afusage, swe-af:codexpatch.
- billed_vs_estimated — Billed vs estimated [partial]: SDK-native llm usage is recorded; Codex costs are estimates; Command Code cost is unknown. Sources: swe-af:codexpatch, swe-af:afusage.
- tokens — Tokens [supported]: Token counts ride with usage entries into control plane accounting. Sources: swe-af:afusage.
- subscription_allowance — Subscription allowance [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of subscription allowance reporting; subscription-token runs have no marginal USD cost shown. Sources: swe-af:codexpatch.
- provider_reconciliation — Provider reconciliation [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of provider invoice reconciliation. Sources: swe-af:afusage.
- compute_cost — Compute cost [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of compute-cost tracking; infra is operator-paid. Sources: swe-af:readme.
- refresh_delay — Refresh delay [partial]: Usage arrives in the execution result envelope, so a killed agent can lose entries. Sources: swe-af:afusage.
- alerts — Alerts [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of budget alerts. Sources: swe-af:schemas.
- export — Export [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of CSV or API export of usage in SWE-AF. Sources: swe-af:afusage.
- drilldown — Drilldown [partial]: UI usage handlers exist in the control plane; drill-down by SWE-AF role was not verified. Sources: swe-af:afusage.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:afusage, swe-af:readme, swe-af:schemas, swe-af:codexpatch

### RUN-08: Compute-aware scheduling and verification

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Concurrency is bounded per level (max_concurrent_issues), with a semaphore and retry on concurrent worktree adds. No CPU/RAM awareness, check reuse or fair scheduling across builds.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- queues: Excess issues wait on the semaphore in-process; not a persisted queue.
- cpu ram: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of CPU or memory limits.
- backpressure: The semaphore throttles issue starts; README advises limiting delegated fan-out.
- host awareness: The Furrow store manager checks free space and evicts mirrors by budget in the Go node; no CPU awareness.
- check reuse: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of reuse of test or CI results across issues.
- check deduplication: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of check deduplication.
- bounded tests: CI gate waits are capped (ci_wait_seconds 1500) and fix cycles are capped; no test-resource bounding.
- fairness: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of fairness policy across builds.
- shared limits: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of shared limits across concurrent builds.

- concurrency — Concurrency [supported]: max_concurrent_issues default 3, 0 for unlimited; semaphore when more issues are active. Sources: swe-af:dag, swe-af:schemas.
- queues — Queues [partial]: Excess issues wait on the semaphore in-process; not a persisted queue. Sources: swe-af:dag.
- cpu_ram — Cpu ram [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of CPU or memory limits. Sources: swe-af:dag.
- backpressure — Backpressure [partial]: The semaphore throttles issue starts; README advises limiting delegated fan-out. Sources: swe-af:dag, swe-af:readme.
- host_awareness — Host awareness [partial]: The Furrow store manager checks free space and evicts mirrors by budget in the Go node; no CPU awareness. Sources: swe-af:go.
- check_reuse — Check reuse [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of reuse of test or CI results across issues. Sources: swe-af:ci.
- check_deduplication — Check deduplication [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of check deduplication. Sources: swe-af:ci.
- bounded_tests — Bounded tests [partial]: CI gate waits are capped (ci_wait_seconds 1500) and fix cycles are capped; no test-resource bounding. Sources: swe-af:ci, swe-af:schemas.
- cancellation — Cancellation [supported]: The control plane cancels an execution tree. Sources: swe-af:afcancel.
- fairness — Fairness [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of fairness policy across builds. Sources: swe-af:dag.
- recovery — Recovery [supported]: resume_build recovers from checkpoints. Sources: swe-af:app.
- shared_limits — Shared limits [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of shared limits across concurrent builds. Sources: swe-af:dag.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:dag, swe-af:schemas, swe-af:readme, swe-af:go, swe-af:ci, swe-af:afcancel, swe-af:app

## Human collaboration and control

### TEAM-01: Live human collaboration

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No multi-user collaboration inside SWE-AF. Hax forms let an approver respond to a plan or clarification; the control plane UI shows runs to anyone with access; no shared prompting or presence.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- visibility: Anyone with control-plane access can view runs; no sharing controls beyond deployment access.
- discussion: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of discussion threads.
- comments: The resolve flow replies to and resolves GitHub review threads; there are no in-product comments.
- presence: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of presence indicators.
- coprompting: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of simultaneous prompting of one build.
- ordering: No co-prompting exists, so input ordering does not apply.
- conflicts: No co-prompting exists, so edit conflicts do not apply.
- attribution: Hax requests carry sender names and approver user id env settings; no general attribution.
- permissions: Control plane API-key auth is all-or-nothing; AgentField has separate VC authorization docs not evaluated here.

- visibility — Visibility [partial]: Anyone with control-plane access can view runs; no sharing controls beyond deployment access. Sources: swe-af:afui, swe-af:afenv.
- discussion — Discussion [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of discussion threads. Sources: swe-af:app.
- comments — Comments [partial]: The resolve flow replies to and resolves GitHub review threads; there are no in-product comments. Sources: swe-af:app.
- presence — Presence [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of presence indicators. Sources: swe-af:afui.
- coprompting — Coprompting [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of simultaneous prompting of one build. Sources: swe-af:app.
- ordering — Ordering [not-applicable]: No co-prompting exists, so input ordering does not apply. Sources: swe-af:app.
- conflicts — Conflicts [not-applicable]: No co-prompting exists, so edit conflicts do not apply. Sources: swe-af:app.
- attribution — Attribution [partial]: Hax requests carry sender names and approver user id env settings; no general attribution. Sources: swe-af:hitl, swe-af:go.
- permissions — Permissions [partial]: Control plane API-key auth is all-or-nothing; AgentField has separate VC authorization docs not evaluated here. Sources: swe-af:afenv.
- plan_restrictions — Plan restrictions [supported]: Open source with no plans; no plan restrictions apply. Sources: swe-af:license.
- owner_online — Owner online [supported]: Executions run server-side without the originating caller online. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:afui, swe-af:afenv, swe-af:app, swe-af:hitl, swe-af:go, swe-af:license, swe-af:readme

### TEAM-02: Roles and access boundaries

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No organization, project or task roles in SWE-AF. The control plane offers API-key access; AgentField SSO or RBAC for the platform was not established in the pinned docs (a grep of docs/README for sso, oidc, rbac found none).

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- organization roles: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of organization roles in SWE-AF or the AgentField README and docs.
- project roles: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of project roles.
- task roles: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of task roles.
- membership: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of membership management.
- sso: Pinned AgentField docs and README mention no SSO or OIDC login for the control plane UI.
- provisioning: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of SCIM or user provisioning.
- unauthorized access: With AGENTFIELD_API_KEY set the control plane requires it on most endpoints; without a key, package-install, secret-store, agent env and config endpoints accept only local-host callers and refuse oth
- revocation: Revoking access means rotating the single API key.

- organization_roles — Organization roles [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of organization roles in SWE-AF or the AgentField README and docs. Sources: swe-af:afenv.
- project_roles — Project roles [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of project roles. Sources: swe-af:afenv.
- task_roles — Task roles [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of task roles. Sources: swe-af:afenv.
- membership — Membership [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of membership management. Sources: swe-af:afenv.
- sso — Sso [unsupported]: Pinned AgentField docs and README mention no SSO or OIDC login for the control plane UI. Sources: swe-af:afenv, swe-af:afreadme.
- provisioning — Provisioning [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of SCIM or user provisioning. Sources: swe-af:afenv.
- unauthorized_access — Unauthorized access [partial]: With AGENTFIELD_API_KEY set the control plane requires it on most endpoints; without a key, package-install, secret-store, agent env and config endpoints accept only local-host callers and refuse others with 401. Sources: swe-af:afenv.
- revocation — Revocation [partial]: Revoking access means rotating the single API key. Sources: swe-af:afenv.
- plan_scope — Plan scope [supported]: No paid plan tiers exist for the open-source stack; enterprise features are sales-led for AgentField only. Sources: swe-af:license.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:afenv, swe-af:afreadme, swe-af:license

### TEAM-03: Approvals and temporary access

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Approval is limited: a plan-approval gate and ask_user forms through Hax; agent command approvals follow the harness (Codex sandbox flags, permission_mode config). No product-enforced network or secret approvals; default posture lets agents act without per-command approval.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- command: permission_mode config and Codex --sandbox modes select harness command policy; there is no SWE-AF approval prompt per command.
- tool: Tool permissions are inherited from the harness.
- network: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of network approvals.
- secret: Scout credential requests go to a human through a Hax form; the store redacts them from logs.
- delivery: PRs open ready for review and are never merged by SWE-AF; CI-gate fixes push to the same branch without approval.
- approver identity: Hax approval user id is configurable (AGENTFIELD_APPROVAL_USER_ID); no identity model in SWE-AF.
- lifetime: Hax requests expire after approval_expires_in_hours (72).
- allow: Approve a plan or answer a form to proceed.
- deny: Rejecting a Hax form returns status cancelled to the reasoner.
- revoke: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of revoking an approval after the fact.
- tunnels: SWE-AF provides no tunnels or temporary shared access.
- defaults: Harness defaults apply; the README benchmark runs Claude with --dangerously-skip-permissions for comparison and Codex has an explicit bypass option.
- safer modes: Codex workspace-write and read-only sandboxes, opt-in SWE_FENCE and per-role read-only review.
- enforcement: Sandbox flags and the fence are enforced by the harness or OS when enabled; plan approval is enforced by pausing the workflow.

- command — Command [partial]: permission_mode config and Codex --sandbox modes select harness command policy; there is no SWE-AF approval prompt per command. Sources: swe-af:schemas, swe-af:codexpatch.
- tool — Tool [partial]: Tool permissions are inherited from the harness. Sources: swe-af:codexpatch.
- network — Network [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of network approvals. Sources: swe-af:codexpatch, swe-af:go.
- secret — Secret [partial]: Scout credential requests go to a human through a Hax form; the store redacts them from logs. Sources: swe-af:creds, swe-af:hitl.
- delivery — Delivery [partial]: PRs open ready for review and are never merged by SWE-AF; CI-gate fixes push to the same branch without approval. Sources: swe-af:ci, swe-af:app.
- approver_identity — Approver identity [partial]: Hax approval user id is configurable (AGENTFIELD_APPROVAL_USER_ID); no identity model in SWE-AF. Sources: swe-af:go, swe-af:hitl.
- lifetime — Lifetime [partial]: Hax requests expire after approval_expires_in_hours (72). Sources: swe-af:schemas.
- allow — Allow [partial]: Approve a plan or answer a form to proceed. Sources: swe-af:hitl.
- deny — Deny [partial]: Rejecting a Hax form returns status cancelled to the reasoner. Sources: swe-af:hitl.
- revoke — Revoke [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of revoking an approval after the fact. Sources: swe-af:hitl.
- tunnels — Tunnels [not-applicable]: SWE-AF provides no tunnels or temporary shared access. Sources: swe-af:readme.
- defaults — Defaults [partial]: Harness defaults apply; the README benchmark runs Claude with --dangerously-skip-permissions for comparison and Codex has an explicit bypass option. Sources: swe-af:readme, swe-af:codexpatch.
- safer_modes — Safer modes [partial]: Codex workspace-write and read-only sandboxes, opt-in SWE_FENCE and per-role read-only review. Sources: swe-af:codexpatch, swe-af:go.
- enforcement — Enforcement [partial]: Sandbox flags and the fence are enforced by the harness or OS when enabled; plan approval is enforced by pausing the workflow. Sources: swe-af:go, swe-af:hitl.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:schemas, swe-af:codexpatch, swe-af:go, swe-af:creds, swe-af:hitl, swe-af:ci, swe-af:app, swe-af:readme

### TEAM-04: Decision and activity history

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

History lives in execution notes, run records, artifacts and the resulting git history and PR text. No user-attributed audit log or export; integrity depends on files and the control plane database.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- actions: Notes record phase and role events with tags; no human actor attribution.
- approvals: Hax approval decisions flow back into the workflow; not stored as a separate log by SWE-AF.
- edits: Git commits and the PR carry agent changes; plan revisions are logged in artifacts.
- discussion: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of in-product discussion history.
- decisions: Compromise debt and replanning decisions are typed records in results and the PR body.
- cross user persistence: Control plane runs persist for anyone with access.
- export: Artifacts are plain files and the API returns execution JSON; no packaged export.
- mutability: Artifacts and git history are mutable by anyone with host access.
- retention: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of retention policy for notes or artifacts.
- integrity: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of tamper-evidence for artifacts; AgentField DID/VC records for executions are a platform feature not evaluated here.

- actions — Actions [partial]: Notes record phase and role events with tags; no human actor attribution. Sources: swe-af:readme, swe-af:afui.
- approvals — Approvals [partial]: Hax approval decisions flow back into the workflow; not stored as a separate log by SWE-AF. Sources: swe-af:hitl.
- edits — Edits [partial]: Git commits and the PR carry agent changes; plan revisions are logged in artifacts. Sources: swe-af:arch.
- discussion — Discussion [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of in-product discussion history. Sources: swe-af:app.
- decisions — Decisions [partial]: Compromise debt and replanning decisions are typed records in results and the PR body. Sources: swe-af:arch.
- cross_user_persistence — Cross user persistence [partial]: Control plane runs persist for anyone with access. Sources: swe-af:afui.
- cross_session_persistence — Cross session persistence [supported]: Artifacts, checkpoints and run records persist across sessions. Sources: swe-af:readme.
- export — Export [partial]: Artifacts are plain files and the API returns execution JSON; no packaged export. Sources: swe-af:readme.
- mutability — Mutability [partial]: Artifacts and git history are mutable by anyone with host access. Sources: swe-af:readme.
- retention — Retention [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of retention policy for notes or artifacts. Sources: swe-af:readme.
- integrity — Integrity [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of tamper-evidence for artifacts; AgentField DID/VC records for executions are a platform feature not evaluated here. Sources: swe-af:arch.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:afui, swe-af:hitl, swe-af:arch, swe-af:app

### TEAM-05: Review experience and guidance

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Review happens in GitHub: SWE-AF opens a PR with summary, changes, test plan, PRD and debt; the resolve flow addresses review comments and replies to threads. No in-product review UI, keyboard model or review queue.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- pending reviews: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a pending-review inbox; PRs appear on GitHub.
- change understanding: The PR body summarises changes, test plan, PRD, architecture and debt; the verifier reports acceptance-criteria results.
- evidence navigation: Verification results and artifacts are files; navigation is manual.
- shipping guidance: PR is left open for human merge; the CI gate leaves failing checks visible.
- effort: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of review-effort estimates.
- missed evidence: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of missed-evidence detection.
- keyboard: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of review keyboard shortcuts in SWE-AF.
- narrow screen: GitHub provides the review UI; SWE-AF documents no narrow-screen review support.
- context switching: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of context-switch helpers.
- roles: The reviewer, QA and synthesizer roles are named agents; the PR does not attribute per-role findings.
- retained context: Review replies keep the PR thread; .artifacts keep issue logs.

- pending_reviews — Pending reviews [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a pending-review inbox; PRs appear on GitHub. Sources: swe-af:app.
- change_understanding — Change understanding [partial]: The PR body summarises changes, test plan, PRD, architecture and debt; the verifier reports acceptance-criteria results. Sources: swe-af:app.
- evidence_navigation — Evidence navigation [partial]: Verification results and artifacts are files; navigation is manual. Sources: swe-af:readme.
- comments — Comments [supported]: resolve fetches review comments, fixes, replies and resolves threads. Sources: swe-af:app.
- request_changes — Request changes [supported]: A reviewer's GitHub comments are fed to resolve to request changes. Sources: swe-af:app.
- shipping_guidance — Shipping guidance [partial]: PR is left open for human merge; the CI gate leaves failing checks visible. Sources: swe-af:ci.
- effort — Effort [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of review-effort estimates. Sources: swe-af:app.
- missed_evidence — Missed evidence [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of missed-evidence detection. Sources: swe-af:app.
- keyboard — Keyboard [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of review keyboard shortcuts in SWE-AF. Sources: swe-af:afui.
- narrow_screen — Narrow screen [unsupported]: GitHub provides the review UI; SWE-AF documents no narrow-screen review support. Sources: swe-af:app.
- context_switching — Context switching [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of context-switch helpers. Sources: swe-af:app.
- roles — Roles [partial]: The reviewer, QA and synthesizer roles are named agents; the PR does not attribute per-role findings. Sources: swe-af:readme.
- retained_context — Retained context [partial]: Review replies keep the PR thread; .artifacts keep issue logs. Sources: swe-af:app.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:app, swe-af:readme, swe-af:ci, swe-af:afui

### TEAM-06: Colleague takeover and continuity

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No colleague takeover feature. Another operator can poll the same execution, resume from artifacts, or resolve an existing PR, but SWE-AF has no ownership or transfer model.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- ongoing takeover: Another caller can resume_build from the same repo_path and artifacts_dir or call resolve on the PR.
- same task: Execution id and branch stay the same for observers; there is no transfer of ownership.
- session history: Notes and artifacts are accessible to anyone with access.
- decisions: Debt and replanning decisions are in artifacts and the PR.
- permissions: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-person permissions.
- runtime context: Worktrees may remain with keep_worktree; the runtime is re-resolved on resume.
- ownership: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of ownership transfer.
- attribution: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of takeover attribution.
- context loss: Resume loses the original PRD detail kept only in summaries.
- harness transfer: resume_build can run under a different runtime than the original, per deployment docs.

- ongoing_takeover — Ongoing takeover [partial]: Another caller can resume_build from the same repo_path and artifacts_dir or call resolve on the PR. Sources: swe-af:app.
- same_task — Same task [partial]: Execution id and branch stay the same for observers; there is no transfer of ownership. Sources: swe-af:app.
- session_history — Session history [partial]: Notes and artifacts are accessible to anyone with access. Sources: swe-af:readme.
- decisions — Decisions [partial]: Debt and replanning decisions are in artifacts and the PR. Sources: swe-af:arch.
- artifacts — Artifacts [supported]: Artifacts persist under .artifacts on the node host. Sources: swe-af:readme.
- permissions — Permissions [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of per-person permissions. Sources: swe-af:afenv.
- runtime_context — Runtime context [partial]: Worktrees may remain with keep_worktree; the runtime is re-resolved on resume. Sources: swe-af:issue, swe-af:deploy.
- owner_disconnect — Owner disconnect [supported]: Execution continues without the originator. Sources: swe-af:readme.
- ownership — Ownership [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of ownership transfer. Sources: swe-af:app.
- attribution — Attribution [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of takeover attribution. Sources: swe-af:app.
- context_loss — Context loss [partial]: Resume loses the original PRD detail kept only in summaries. Sources: swe-af:app.
- harness_transfer — Harness transfer [partial]: resume_build can run under a different runtime than the original, per deployment docs. Sources: swe-af:deploy.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:app, swe-af:readme, swe-af:arch, swe-af:afenv, swe-af:issue, swe-af:deploy

## Verification and delivery

### SHIP-01: Retained artifacts and outputs

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Retained outputs are files: PRD, architecture, issue specs, execution logs, verification results under .artifacts, a PR body, git branches and commit lists. No screenshot, recording or preview capture, no artifact gallery.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- screenshots: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of screenshot capture.
- recordings: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of recordings.
- previews: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of previews.
- gallery: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an artifact gallery; artifacts are plain directories.
- task links: Artifacts are in per-build directories keyed by build_id.
- revision links: Results list commit SHAs per issue branch.
- prompt leakage: Scout-negotiated credentials are redacted from logs, but README warns it is not a general-purpose scrubber.

- diffs — Diffs [supported]: Branches, commits and files_changed are returned and the PR shows the diff on GitHub. Sources: swe-af:readme, swe-af:issue.
- documents — Documents [supported]: PRD, architecture and issue specs are saved under .artifacts/plan/. Sources: swe-af:readme.
- screenshots — Screenshots [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of screenshot capture. Sources: swe-af:readme.
- recordings — Recordings [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of recordings. Sources: swe-af:readme.
- previews — Previews [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of previews. Sources: swe-af:readme.
- test_outputs — Test outputs [supported]: QA, integration test and CI results are recorded in results and artifacts. Sources: swe-af:arch, swe-af:ci.
- reviewer_outputs — Reviewer outputs [supported]: Code review and QA synthesizer results are structured outputs retained in execution logs. Sources: swe-af:coding, swe-af:arch.
- gallery — Gallery [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of an artifact gallery; artifacts are plain directories. Sources: swe-af:readme.
- task_links — Task links [partial]: Artifacts are in per-build directories keyed by build_id. Sources: swe-af:readme, swe-af:issue.
- revision_links — Revision links [partial]: Results list commit SHAs per issue branch. Sources: swe-af:issue.
- persistence — Persistence [supported]: Artifacts persist on disk after the build. Sources: swe-af:readme.
- child_findings — Child findings [supported]: Per-issue review and QA findings persist in execution logs. Sources: swe-af:readme.
- prompt_leakage — Prompt leakage [partial]: Scout-negotiated credentials are redacted from logs, but README warns it is not a general-purpose scrubber. Sources: swe-af:readme, swe-af:creds.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:issue, swe-af:arch, swe-af:ci, swe-af:coding, swe-af:creds

### SHIP-02: Verification and spec gates

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Verification is built into the pipeline: QA and tests per issue, reviewer, integration tester, verifier against PRD criteria with a fix loop, and a post-PR CI gate with a bounded fix loop. Checks are agent-run; lint and type checking are not distinct gates.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- lint: Lint is whatever the repository CI runs; SWE-AF has no separate lint step.
- types: Type checks run only if the repo CI or tests include them.
- spec checks: The verifier checks PRD acceptance criteria for the build only.
- required checks: Default pipeline stages are mandatory in build; implement_issue verify can be disabled.
- failure blocking: A BLOCK or failed criteria triggers fix issues; failed CI leaves the PR open with failing checks rather than blocking the PR.
- requirement conflicts: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of persistent-requirement conflict checks.
- enforced policy: Stages are enforced by code, but pass/fail judgments are model outputs and the PR is not merge-gated by SWE-AF.

- tests — Tests [supported]: Coder writes tests; QA, integration tester and verifier run them. Sources: swe-af:arch, swe-af:coding.
- lint — Lint [partial]: Lint is whatever the repository CI runs; SWE-AF has no separate lint step. Sources: swe-af:ci.
- types — Types [partial]: Type checks run only if the repo CI or tests include them. Sources: swe-af:ci.
- ci — Ci [supported]: check_ci default true: wait for GitHub Actions on the PR with polling and capped fix cycles. Sources: swe-af:ci, swe-af:schemas.
- independent_review — Independent review [supported]: A code reviewer agent reviews each issue; the synthesizer returns FIX, APPROVE or BLOCK. Sources: swe-af:coding.
- spec_checks — Spec checks [partial]: The verifier checks PRD acceptance criteria for the build only. Sources: swe-af:arch.
- required_checks — Required checks [partial]: Default pipeline stages are mandatory in build; implement_issue verify can be disabled. Sources: swe-af:schemas, swe-af:issue.
- failure_blocking — Failure blocking [partial]: A BLOCK or failed criteria triggers fix issues; failed CI leaves the PR open with failing checks rather than blocking the PR. Sources: swe-af:ci, swe-af:arch.
- requirement_conflicts — Requirement conflicts [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of persistent-requirement conflict checks. Sources: swe-af:arch.
- configured_policy — Configured policy [supported]: check_ci, verify, max cycles and review path are configurable. Sources: swe-af:schemas, swe-af:issue.
- enforced_policy — Enforced policy [partial]: Stages are enforced by code, but pass/fail judgments are model outputs and the PR is not merge-gated by SWE-AF. Sources: swe-af:ci, swe-af:coding.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:arch, swe-af:coding, swe-af:ci, swe-af:schemas, swe-af:issue

### SHIP-03: Git and pull request workflow

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Integrated Git workflow: deterministic worktree and merge helpers, per-issue branches, merger agent for conflicts, integration branch, gh-based push and PR creation (ready for review), resolve for CI failures and review comments with merge-not-rebase of base.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- rebase: resolve merges the base branch into the PR head and never rebases.

- branches — Branches [supported]: Integration branch and per-issue branches. Sources: swe-af:arch, swe-af:issuebuild.
- worktrees — Worktrees [supported]: Worktree per issue with plain git helpers when deterministic_git is true. Sources: swe-af:dag.
- commits — Commits [supported]: Coders commit; results list commits. Sources: swe-af:issue.
- pull_requests — Pull requests [supported]: A GitHub PR agent pushes and runs gh pr create without --draft. Sources: swe-af:app, swe-af:schemas.
- comments — Comments [supported]: resolve replies to and resolves review threads. Sources: swe-af:app.
- ci_response — Ci response [supported]: CI gate watches checks and fixes with a CI-fixer agent for max_ci_fix_cycles. Sources: swe-af:ci.
- rebase — Rebase [unsupported]: resolve merges the base branch into the PR head and never rebases. Sources: swe-af:app.
- conflicts — Conflicts [supported]: The merger agent resolves branch conflicts and resolve reports merge_state clean, merged, conflict or skipped. Sources: swe-af:app, swe-af:arch.
- product_state — Product state [supported]: Results record PR URL, number, merge state and branches in structured outputs. Sources: swe-af:app, swe-af:schemas.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:arch, swe-af:issuebuild, swe-af:dag, swe-af:issue, swe-af:app, swe-af:schemas, swe-af:ci

### SHIP-04: Merge and deployment authority

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

SWE-AF opens PRs and never merges them; approval, merge, release and deploy stay with humans and repo CI. No exact-revision binding or authority enforcement in SWE-AF.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- approve: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of approval controls; GitHub reviews apply.
- merge: github_pr prompt says do not merge the PR; no merge operation exists.
- release: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of release operations.
- deploy: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of deploy operations.
- exact revision: The CI gate checks the PR head just pushed, with a grace period to avoid stale check states.
- authority enforcement: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of authority enforcement beyond GitHub token scope.

- approve — Approve [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of approval controls; GitHub reviews apply. Sources: swe-af:ci, swe-af:app.
- merge — Merge [unsupported]: github_pr prompt says do not merge the PR; no merge operation exists. Sources: swe-af:app.
- release — Release [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of release operations. Sources: swe-af:app.
- deploy — Deploy [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of deploy operations. Sources: swe-af:app.
- exact_revision — Exact revision [partial]: The CI gate checks the PR head just pushed, with a grace period to avoid stale check states. Sources: swe-af:ci, swe-af:schemas.
- authority_enforcement — Authority enforcement [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of authority enforcement beyond GitHub token scope. Sources: swe-af:app.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:ci, swe-af:app, swe-af:schemas

### SHIP-05: Completion evidence and rollback

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Results link issues to branches, commits and the PR; no build, deployment or rollback records. Revisions after review rely on GitHub.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- build links: CI gate returns check results from gh pr checks.
- deployment links: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of deployment records.
- delivery status: Build success flags and PR URL indicate delivery to PR; merge status is not tracked.
- rollback: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of rollback operations.
- revert: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of revert operations.
- stale approval invalidation: SWE-AF has no approvals bound to revisions.

- commit_links — Commit links [supported]: Issue results list commit SHAs and files changed. Sources: swe-af:issue.
- pr_links — Pr links [supported]: BuildResult and implement_issue return the PR URL and number when created. Sources: swe-af:app, swe-af:issue.
- build_links — Build links [partial]: CI gate returns check results from gh pr checks. Sources: swe-af:ci.
- deployment_links — Deployment links [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of deployment records. Sources: swe-af:app.
- delivery_status — Delivery status [partial]: Build success flags and PR URL indicate delivery to PR; merge status is not tracked. Sources: swe-af:schemas.
- retained_evidence — Retained evidence [supported]: Artifacts hold verification results and debt records. Sources: swe-af:readme.
- rollback — Rollback [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of rollback operations. Sources: swe-af:app.
- revert — Revert [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of revert operations. Sources: swe-af:app.
- stale_approval_invalidation — Stale approval invalidation [not-applicable]: SWE-AF has no approvals bound to revisions. Sources: swe-af:app.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:issue, swe-af:app, swe-af:ci, swe-af:schemas, swe-af:readme

### SHIP-06: Batch review and high-volume shipping

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No batch review queue. Each build produces one PR with summary and typed debt; many builds produce many PRs reviewed in GitHub. No grouping, risk filters or dependency-aware merge queue.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- review queue: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a review queue.
- grouping: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of grouping builds.
- risk size filters: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of risk or size filters.
- findings: Findings are retained in artifacts and the PR body.
- artifacts: Artifacts are files in .artifacts per build.
- exact revision approval: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of exact-revision approval.
- dependency queues: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of dependency-aware merge queues.
- next actions: The result lists the PR URL and failing checks.
- stale approval invalidation: No approvals are stored, so stale-approval invalidation does not apply.
- check deduplication: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of CI check deduplication.

- review_queue — Review queue [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a review queue. Sources: swe-af:app.
- summaries — Summaries [supported]: The PR body summarises work, changes, tests and debt. Sources: swe-af:app.
- grouping — Grouping [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of grouping builds. Sources: swe-af:app.
- risk_size_filters — Risk size filters [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of risk or size filters. Sources: swe-af:app.
- findings — Findings [partial]: Findings are retained in artifacts and the PR body. Sources: swe-af:arch.
- artifacts — Artifacts [partial]: Artifacts are files in .artifacts per build. Sources: swe-af:readme.
- request_changes — Request changes [supported]: resolve addresses review comments. Sources: swe-af:app.
- exact_revision_approval — Exact revision approval [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of exact-revision approval. Sources: swe-af:app.
- dependency_queues — Dependency queues [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of dependency-aware merge queues. Sources: swe-af:app.
- next_actions — Next actions [partial]: The result lists the PR URL and failing checks. Sources: swe-af:ci.
- stale_approval_invalidation — Stale approval invalidation [not-applicable]: No approvals are stored, so stale-approval invalidation does not apply. Sources: swe-af:app.
- check_deduplication — Check deduplication [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of CI check deduplication. Sources: swe-af:ci.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:app, swe-af:arch, swe-af:readme, swe-af:ci

## Ownership, privacy and economics

### OWN-01: Source and license

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Apache-2.0 across the SWE-AF repository (Python node, Go node, tests, docs). The vendored swe-pro, furrow and furrowd binaries under go/bin are distributed in the repository; their source is not in this repository and is not inspected, and the AgentField control plane is a separate Apache-2.0 repository.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with, AgentField Desktop (Electron companion that installs and starts swe-planner); version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- hosted extensions: No SWE-AF hosted service is documented; AgentField Enterprise is a contact-sales offering outside the code.
- source visibility: Public source for both nodes; prebuilt swe-pro, furrow and furrowd binaries are vendored without source in this repository.
- hosted parity: No hosted SWE-AF service exists to compare.

- client_license — Client license [supported]: No SWE-AF client; AgentField desktop and web UI sit in the Apache-2.0 agentfield repository. Sources: swe-af:license, swe-af:repo.
- server_license — Server license [supported]: Apache-2.0 for the SWE-AF nodes and for AgentField per its repository license. Sources: swe-af:license, swe-af:repo.
- hosted_extensions — Hosted extensions [unsupported]: No SWE-AF hosted service is documented; AgentField Enterprise is a contact-sales offering outside the code. Sources: swe-af:repo.
- source_visibility — Source visibility [partial]: Public source for both nodes; prebuilt swe-pro, furrow and furrowd binaries are vendored without source in this repository. Sources: swe-af:pro, swe-af:license.
- source_revision — Source revision [supported]: Pinned to commit b281a4c (2026-11-12); no tags or releases. Sources: swe-af:repo.
- hosted_parity — Hosted parity [not-applicable]: No hosted SWE-AF service exists to compare. Sources: swe-af:repo.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:license, swe-af:repo, swe-af:pro

### OWN-02: Hosting and exit

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Fully self-hostable: control plane, PostgreSQL and nodes run on user infrastructure via Docker Compose, Railway or af install. Data are files and execution records; exports are artifacts and git branches.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- vendor dependencies: Needs a model provider; the AForge binary downloads from agentfield.ai at image build; optional Hax service for approvals; GitHub for PRs.
- export: Artifacts are plain files and branches are git; the control plane API returns execution JSON.
- import: resume_build and execute accept saved artifacts and plan_result.
- cancellation: No SWE-AF account exists to cancel.
- history portability: Run history lives in the control plane database; artifacts are portable files.

- self_hosting — Self hosting [supported]: docker compose up -d runs control plane, build-db and node; af install and Railway are alternatives. Sources: swe-af:readme, swe-af:deploy.
- vendor_dependencies — Vendor dependencies [partial]: Needs a model provider; the AForge binary downloads from agentfield.ai at image build; optional Hax service for approvals; GitHub for PRs. Sources: swe-af:readme, swe-af:go.
- export — Export [partial]: Artifacts are plain files and branches are git; the control plane API returns execution JSON. Sources: swe-af:readme.
- import — Import [partial]: resume_build and execute accept saved artifacts and plan_result. Sources: swe-af:app.
- formats — Formats [supported]: Markdown, JSON checkpoints and git branches. Sources: swe-af:readme, swe-af:dag.
- cancellation — Cancellation [not-applicable]: No SWE-AF account exists to cancel. Sources: swe-af:readme.
- account_independence — Account independence [supported]: No SWE-AF account is required; af install secrets are node-scoped. Sources: swe-af:manifest.
- history_portability — History portability [partial]: Run history lives in the control plane database; artifacts are portable files. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:deploy, swe-af:go, swe-af:app, swe-af:dag, swe-af:manifest

### OWN-03: Data handling and telemetry

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

SWE-AF code sends no telemetry (no matches for telemetry, analytics, posthog or sentry in node code). Code, prompts and diffs go to the chosen model provider and GitHub. The AgentField control plane sends anonymous usage telemetry by default with documented opt-out. Retention and residency are the operator's.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- logs: Logs stay on the node host under .artifacts and in the control plane database; optional LiteLLM callbacks can export traces to LangFuse in the Python SDK.
- retention: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of retention settings; operators control disks.
- deletion: Delete artifacts or database rows manually; make clean targets exist for examples.
- training: Training use depends on the chosen model provider, outside SWE-AF.
- residency: Residency depends on the operator host and model provider; SWE-AF defines none.
- subprocessors: No SWE-AF service exists; model providers are the operator's choice.
- consent prompt: No first-run consent dialog is documented for the control plane.

- code_data — Code data [supported]: Repository content is read by the selected harness and sent to its model provider; PRs push to GitHub. Sources: swe-af:readme.
- prompts — Prompts [supported]: Prompts go to the selected provider; AgentField telemetry excludes prompts and outputs. Sources: swe-af:afenv.
- credentials — Credentials [supported]: Provider keys and GH_TOKEN stay in the node environment; scout credentials are redacted in logs. Sources: swe-af:creds, swe-af:manifest.
- logs — Logs [partial]: Logs stay on the node host under .artifacts and in the control plane database; optional LiteLLM callbacks can export traces to LangFuse in the Python SDK. Sources: swe-af:readme, swe-af:afenv.
- artifacts — Artifacts [supported]: Artifacts stay on the node host. Sources: swe-af:readme.
- retention — Retention [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of retention settings; operators control disks. Sources: swe-af:readme.
- deletion — Deletion [partial]: Delete artifacts or database rows manually; make clean targets exist for examples. Sources: swe-af:readme.
- training — Training [unsupported]: Training use depends on the chosen model provider, outside SWE-AF. Sources: swe-af:readme.
- residency — Residency [unsupported]: Residency depends on the operator host and model provider; SWE-AF defines none. Sources: swe-af:deploy.
- subprocessors — Subprocessors [unsupported]: No SWE-AF service exists; model providers are the operator's choice. Sources: swe-af:readme.
- telemetry_disclosure — Telemetry disclosure [supported]: AgentField README and environment docs disclose default-on privacy-minimized telemetry. Sources: swe-af:afreadme, swe-af:afenv.
- telemetry_defaults — Telemetry defaults [supported]: Anonymous telemetry is enabled by default in the control plane and records startup, registration, SDK language, runtime and execution status buckets. Sources: swe-af:afenv.
- telemetry_controls — Telemetry controls [supported]: AGENTFIELD_TELEMETRY_ENABLED=false disables it; the AgentField CodeAF engine honors DO_NOT_TRACK and CI. Sources: swe-af:afenv.
- consent_prompt — Consent prompt [unsupported]: No first-run consent dialog is documented for the control plane. Sources: swe-af:afenv, swe-af:afreadme.
- provider_separation — Provider separation [supported]: Provider policy is separate from orchestration; telemetry and usage entries exclude raw model names and prompts. Sources: swe-af:afenv.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:afenv, swe-af:creds, swe-af:manifest, swe-af:deploy, swe-af:afreadme

### OWN-04: Price and included usage

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Free and open source (Apache-2.0); model tokens, hosting and GitHub are the operator's costs. README cites one build at $19.23 on claude-haiku-4-5. No paid SWE-AF plan appears in the repository; AgentField Enterprise is contact-sales.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- currency: No price is charged for SWE-AF.
- interval: No recurring charge exists.
- seats: No seat pricing.
- seat minimum: No seat minimum.
- workspace fees: No workspace fee.
- included usage: No vendor usage is included; provider tokens are the operator's.
- credits: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of vendor credits.
- trial: No trial applies to free software.
- overages: No overages.
- taxes: No vendor taxes apply.

- currency — Currency [not-applicable]: No price is charged for SWE-AF. Sources: swe-af:license.
- interval — Interval [not-applicable]: No recurring charge exists. Sources: swe-af:license.
- seats — Seats [not-applicable]: No seat pricing. Sources: swe-af:license.
- seat_minimum — Seat minimum [not-applicable]: No seat minimum. Sources: swe-af:license.
- workspace_fees — Workspace fees [not-applicable]: No workspace fee. Sources: swe-af:license.
- included_usage — Included usage [not-applicable]: No vendor usage is included; provider tokens are the operator's. Sources: swe-af:readme.
- credits — Credits [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of vendor credits. Sources: swe-af:readme.
- trial — Trial [not-applicable]: No trial applies to free software. Sources: swe-af:license.
- overages — Overages [not-applicable]: No overages. Sources: swe-af:license.
- taxes — Taxes [not-applicable]: No vendor taxes apply. Sources: swe-af:license.
- model_cost — Model cost [supported]: Model cost is billed by the operator's provider; README shows a $19.23 example and a roughly $6 MiniMax run, vendor-reported. Sources: swe-af:readme.
- compute_cost — Compute cost [supported]: Compute is operator-paid; no vendor compute charge. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:license, swe-af:readme

### OWN-05: Scenario cost and controls

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No priced scenarios: the software is free and costs depend on model tokens and hosting the catalog does not model. README gives vendor-reported examples (one build $19.23 and a benchmark build about $20 or $6).

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- one person: No catalog scenario was priced; costs are model tokens and hosting only.
- two people: No catalog scenario was priced.
- five people: No catalog scenario was priced.
- task assumptions: No standard task assumptions are published; README examples use specific builds.
- concurrency: Default concurrency is 3 issues per level.
- tokens: Token counts per build are not published beyond invocation counts.
- compute hours: Compute hours are not published.
- variable cost: Variable cost is model tokens, vendor-reported at $6 to $20 per example build.
- included cost: Nothing is included by a vendor.
- unknown cost: Hosting, GitHub Actions minutes and Hax usage are unknown.

- one_person — One person [unsupported]: No catalog scenario was priced; costs are model tokens and hosting only. Sources: swe-af:readme.
- two_people — Two people [unsupported]: No catalog scenario was priced. Sources: swe-af:readme.
- five_people — Five people [unsupported]: No catalog scenario was priced. Sources: swe-af:readme.
- task_assumptions — Task assumptions [unsupported]: No standard task assumptions are published; README examples use specific builds. Sources: swe-af:readme.
- concurrency — Concurrency [partial]: Default concurrency is 3 issues per level. Sources: swe-af:schemas.
- tokens — Tokens [unsupported]: Token counts per build are not published beyond invocation counts. Sources: swe-af:readme.
- compute_hours — Compute hours [unsupported]: Compute hours are not published. Sources: swe-af:readme.
- fixed_cost — Fixed cost [supported]: Fixed software cost is zero under Apache-2.0. Sources: swe-af:license.
- variable_cost — Variable cost [partial]: Variable cost is model tokens, vendor-reported at $6 to $20 per example build. Sources: swe-af:readme.
- included_cost — Included cost [not-applicable]: Nothing is included by a vendor. Sources: swe-af:readme.
- unknown_cost — Unknown cost [partial]: Hosting, GitHub Actions minutes and Hax usage are unknown. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:schemas, swe-af:license

### OWN-06: Availability and support

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Public beta, open source, single repository maintained by Agent-Field with commits as recent as 2026-11-12; no tagged releases, no stated response commitments beyond a security-policy acknowledgment window.

Availability: beta
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- support: GitHub issues and Discord (AgentField) are channels; SECURITY.md covers vulnerabilities.
- response commitments: Security policy: acknowledgment within 3 business days and triage within 7; no general SLA.
- release activity: Last push 2026-11-12; 0.1.0 released 2026-02-16 with no later tags; AgentField publishes daily rc builds.
- status history: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a status page.
- deprecation: V2 config migration removed legacy keys; the Python node is superseded by the Go node under af install.
- export options: Artifacts and branches are portable.

- release_stage — Release stage [supported]: README badge says public beta; Go pro engine beta. Sources: swe-af:readme, swe-af:pro.
- access_restrictions — Access restrictions [supported]: Public repository and Apache-2.0; no gating. Sources: swe-af:repo.
- support — Support [partial]: GitHub issues and Discord (AgentField) are channels; SECURITY.md covers vulnerabilities. Sources: swe-af:security.
- response_commitments — Response commitments [partial]: Security policy: acknowledgment within 3 business days and triage within 7; no general SLA. Sources: swe-af:security.
- release_activity — Release activity [partial]: Last push 2026-11-12; 0.1.0 released 2026-02-16 with no later tags; AgentField publishes daily rc builds. Sources: swe-af:repo, swe-af:changelog.
- status_history — Status history [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of a status page. Sources: swe-af:readme.
- deprecation — Deprecation [partial]: V2 config migration removed legacy keys; the Python node is superseded by the Go node under af install. Sources: swe-af:changelog, swe-af:manifest.
- export_options — Export options [partial]: Artifacts and branches are portable. Sources: swe-af:readme.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:readme, swe-af:pro, swe-af:repo, swe-af:security, swe-af:changelog, swe-af:manifest

### OWN-07: Enterprise assurance

Verification status: No sourced public finding (not-published)
Support: not-published

No publishable sourced finding is available for this criterion. This does not mean the product lacks the capability.

### OWN-08: Useful free tier and evaluation access

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

Open-source self-hosting is free; costs are provider tokens and hosting. No trial, vendor credits or account required. No card is needed because nothing is sold.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- trial: No trial applies to free software.
- credits trial: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of trial credits.
- paid only: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of paid-only access.
- free scenarios: Single-user remote execution and BYO subscription (Claude token, Codex login) work free of platform charges; no collaboration, mobile or issue-management scenario applies.
- card required: No payment is requested by the software.
- usage caps: Only operator-set limits (concurrency, timeouts) exist; provider caps apply.
- paid evaluation: Evaluation needs a paid provider key; no sales approval is needed.

- ongoing_free — Ongoing free [supported]: Apache-2.0 free software with no usage cap from the vendor. Sources: swe-af:license, swe-af:repo.
- trial — Trial [not-applicable]: No trial applies to free software. Sources: swe-af:license.
- credits_trial — Credits trial [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of trial credits. Sources: swe-af:readme.
- self_host_cost — Self host cost [supported]: Self-hosting costs host, provider tokens and optional Railway and Hax charges. Sources: swe-af:readme, swe-af:deploy.
- paid_only — Paid only [unsupported]: Bounded search of swe_af/ and go/ at the pinned commit found no implementation of paid-only access. Sources: swe-af:readme.
- early_access — Early access [supported]: Public beta with no gating. Sources: swe-af:readme.
- free_scenarios — Free scenarios [partial]: Single-user remote execution and BYO subscription (Claude token, Codex login) work free of platform charges; no collaboration, mobile or issue-management scenario applies. Sources: swe-af:readme, swe-af:deploy.
- card_required — Card required [not-applicable]: No payment is requested by the software. Sources: swe-af:license.
- usage_caps — Usage caps [partial]: Only operator-set limits (concurrency, timeouts) exist; provider caps apply. Sources: swe-af:schemas.
- paid_evaluation — Paid evaluation [partial]: Evaluation needs a paid provider key; no sales approval is needed. Sources: swe-af:manifest.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:license, swe-af:repo, swe-af:readme, swe-af:deploy, swe-af:schemas, swe-af:manifest

### OWN-09: Pricing transparency

Verification status: Agent cross-checked (agent-cross-checked)
Support: partial

No ADE price to disclose for SWE-AF: the software is free and model costs pass to the provider directly. AgentField Enterprise is sales-led with no public price.

Availability: generally-available
Delivery: native
Scope: surfaces=SWE-AF node (swe-planner and swe-fast, Python and Go) driven through the AgentField HTTP API and af CLI, AgentField control plane web UI (runs, workflow DAG, notes) that SWE-AF registers with; version=SWE-AF default branch at commit b281a4c (2026-11-12; no tagged release); node ids swe-planner and swe-fast, Python and Go implementations; AgentField control plane 161cab6 for platform behavior; plan=not specified; os=not specified; execution_modes=not specified; region=not specified; harness=not specified.

Limitations:
- units: No units are charged for SWE-AF.
- included limits: No included limits.
- overages: No overage rates.
- minimum commitment: No commitment.
- cancellation: No account to cancel.
- sales only unknowns: AgentField Enterprise is a "Talk to us" offering with no public price; it is a platform offering outside SWE-AF.
- transparency classification: Open-source free software with a separate sales-only enterprise offer from the platform vendor.

- units — Units [not-applicable]: No units are charged for SWE-AF. Sources: swe-af:license.
- included_limits — Included limits [not-applicable]: No included limits. Sources: swe-af:license.
- overages — Overages [not-applicable]: No overage rates. Sources: swe-af:license.
- compute_markup — Compute markup [supported]: No compute markup because compute is operator-paid. Sources: swe-af:readme.
- model_markup — Model markup [supported]: No model markup; provider keys are used directly. Sources: swe-af:manifest.
- minimum_commitment — Minimum commitment [not-applicable]: No commitment. Sources: swe-af:license.
- cancellation — Cancellation [not-applicable]: No account to cancel. Sources: swe-af:license.
- sales_only_unknowns — Sales only unknowns [partial]: AgentField Enterprise is a "Talk to us" offering with no public price; it is a platform offering outside SWE-AF. Sources: swe-af:repo.
- transparency_classification — Transparency classification [partial]: Open-source free software with a separate sales-only enterprise offer from the platform vendor. Sources: swe-af:license, swe-af:repo.

Checked 2026-11-16 (6 days old on 2026-11-22); agent cross-checked 2026-11-17.

Sources: swe-af:license, swe-af:readme, swe-af:manifest, swe-af:repo

## Sources

### swe-af:readme

SWE-AF README. Headless runtime: builds are triggered with af call or HTTP POST to swe-planner.build; runtimes aforge, claude_code, open_code, codex and command_code; Docker/Railway/af install deployment; PR creation and CI gate; resume_build; issue-level implement_issue sub-harness.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/README.md
- Locator: One-Call DX, Operating Modes, Quick Start, API Reference, Artifacts, Go implementation
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:app

swe-planner reasoners. Entry points build, plan, resolve (update existing PR: merge base, fix CI, answer review comments), resume_build (from checkpoint.json) and execute; optional HAX plan approval gate pauses the workflow.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/app.py
- Locator: build, plan, execute, resolve, resume_build reasoners and plan-approval gate
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/app.py

### swe-af:surface

Discovery surface metadata. Entrypoint tag marks build, implement_issue, plan, resolve and resume_build for discovery via af ls --entrypoints; internal role reasoners are tagged and described as do-not-call-directly.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/surface.py
- Locator: TAG_ENTRYPOINT, TAG_INTERNAL, internal_role
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/surface.py

### swe-af:hitl

Human-in-the-loop ask_user form and pause. When HAX_API_KEY is set, planning and scout roles can request a Hax form; the workflow pauses via app.pause until a person responds or the request expires; without the key the gate is disabled.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/hitl/ask_user.py
- Locator: build_hax_client_from_env, request_user_input_and_pause
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/hitl/ask_user.py

### swe-af:pyproject

Python project metadata. Version 0.1.0; depends on agentfield>=0.1.145, claude-agent-sdk 0.1.20 and hax-sdk for the HITL gate; Python 3.12+.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/pyproject.toml
- Locator: dependencies
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: pyproject.toml

### swe-af:afreadme

AgentField README. Control plane exposes POST /api/v1/execute and async execution endpoints, serves MCP at <server>/mcp, installs nodes with af install, and states default-on privacy-minimized telemetry with an opt-out variable.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/README.md
- Locator: Install nodes, execute API, MCP, observability, telemetry notice
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:afui

AgentField run detail page. Web UI for runs: status, step trace, workflow DAG graph, execution notes, webhooks, retry, restart and replay controls.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/control-plane/web/client/src/pages/RunDetailPage.tsx
- Locator: Run detail with trace, graph, notes, triggers
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: 161cab6a948b3cc7677a86d20c1a119ef56f6e4c
- Path: control-plane/web/client/src/pages/RunDetailPage.tsx

### swe-af:afcancel

Cancel execution tree. Control plane can cancel an execution and its child executions.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/control-plane/internal/handlers/execute_cancel_tree_test.go
- Locator: cancel tree tests
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: 161cab6a948b3cc7677a86d20c1a119ef56f6e4c
- Path: control-plane/internal/handlers/execute_cancel_tree_test.go

### swe-af:afdesk

AgentField Desktop README. Electron desktop shows dashboard, agents and live activity, installs bundled swe-planner on first launch and starts agents with af run; Mac-first with DMG and NSIS builds.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/desktop/README.md
- Locator: Views; Bundled agent nodes
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:afdeskpkg

AgentField Desktop build config. electron-builder targets macOS (dmg, zip) and Windows (nsis); no Linux target.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/desktop/package.json
- Locator: build.mac.target dmg zip; build.win.target nsis
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: 161cab6a948b3cc7677a86d20c1a119ef56f6e4c
- Path: desktop/package.json

### swe-af:schemas

BuildConfig and result schemas. Config fields: max_concurrent_issues=3, max_coding_iterations=5, agent_timeout_seconds=2700, max_advisor_invocations=2, enable_learning=false, check_ci, max_ci_fix_cycles, enable_github_pr, repos with primary/dependency roles, approval_expires_in_hours=72, deterministic_git.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/execution/schemas.py
- Locator: class BuildConfig / ExecutionConfig fields
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/execution/schemas.py

### swe-af:codexpatch

Codex harness patch. Codex runtime runs the Codex CLI with --sandbox workspace-write by default, read-only for review roles, with an explicit bypass option; auth mode chatgpt/api_key/auto.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/runtime/codex_harness_patch.py
- Locator: cmd construction with --sandbox workspace-write and bypass flag
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/runtime/codex_harness_patch.py

### swe-af:go

SWE-AF Go node. Go port registering swe-planner/swe-fast with identical API; opt-in OS fence via sandbox-exec or bwrap; SWE_PRO_MAX_COST ceiling; vendored swe-pro engine (beta) on macOS and Linux only.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/go/README.md
- Locator: Two binaries, Environment variables, SWE_FENCE, SWE_PRO_ENGINE
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:deploy

Deployment guide. Docker 20.10+ with Compose v2; provider keys incl. CLAUDE_CODE_OAUTH_TOKEN (Pro/Max credits); resume_build resolves the runtime from supplied config or environment; Railway uses the same Dockerfile.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/docs/deployment.md
- Locator: Prerequisites, environment variables, runtime selection, resume_build runtime note
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:manifest

af install manifest. Requires one LLM provider key; GH_TOKEN optional; the root manifest redirects af install to the Go node.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/agentfield-package.yaml
- Locator: user_environment require_one_of; superseded_by
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: agentfield-package.yaml

### swe-af:issue

Issue-level build schemas. implement_issue accepts an issue with title, description, acceptance criteria, files and testing strategy; config includes verify, enable_github_pr, agent_timeout_seconds=1800, max_coding_iterations=3, keep_worktree.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/issue/schemas.py
- Locator: ImplementIssueConfig and issue fields
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/issue/schemas.py

### swe-af:issuebuild

implement_issue entry point. Creates its own git worktree and issue/<build_id>-<slug> branch from base_branch, runs the coder-reviewer loop, optional verifier, removes the worktree and returns branch, commits and files changed without touching the caller checkout.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/issue/build.py
- Locator: build_issue orchestration
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/issue/build.py

### swe-af:changelog

SWE-AF changelog. Unreleased V2 config migration to runtime plus flat models; 0.1.0 released 2026-02-16; only dated release.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/CHANGELOG.md
- Locator: [Unreleased] and [0.1.0]
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:afenv

AgentField environment variables. X-API-Key or bearer auth when AGENTFIELD_API_KEY is set (local-host restriction on install and secret endpoints otherwise); AGENTFIELD_RESUME_INTERRUPTED_DELAY for replacement instances; anonymous telemetry on by default with AGENTFIELD_TELEMETRY_ENABLED=false opt-out, no prompts, outputs or secrets in payloads.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/docs/ENVIRONMENT_VARIABLES.md
- Locator: API key auth; Anonymous Telemetry; AGENTFIELD_TELEMETRY_ENABLED
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:arch

SWE-AF architecture. Six-phase build pipeline with checkpoints, issue DAG with dependency levels, three nested control loops, merger, integration tester, verifier and draft PR.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/docs/ARCHITECTURE.md
- Locator: Build Pipeline, Planning to Issue DAG, Execution Engine
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:dag

DAG executor. Executes issues by dependency level, caps parallel issues with a semaphore, persists DAGState to .artifacts/execution/checkpoint.json after levels, uses git worktrees per issue.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/execution/dag_executor.py
- Locator: _save_checkpoint, bounded concurrency with asyncio.Semaphore, worktree helpers, level execution
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/execution/dag_executor.py

### swe-af:afcron

Cron trigger source. Control plane ships a cron trigger source alongside github, linear, slack, stripe and generic webhook sources that dispatch to reasoners.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/control-plane/internal/sources/cron/cron_test.go
- Locator: TestCron_MetadataAndSchema
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: 161cab6a948b3cc7677a86d20c1a119ef56f6e4c
- Path: control-plane/internal/sources/cron/cron_test.go

### swe-af:skill

delegate-issue skill. Ships a Claude Code skill that lets a main coding harness delegate scoped issues to implement_issue, poll execution notes and merge the returned branches.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/.claude/skills/delegate-issue/SKILL.md
- Locator: When to delegate; preflight; fan-out and merge
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:coding

Per-issue coding loop. Inner loop: coder then reviewer (two calls) or coder then QA plus reviewer plus synthesizer when needs_deeper_qa; 2700 s default call timeout; middle advisor and outer replanner loops.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/execution/coding_loop.py
- Locator: module docstring; _call_with_timeout; default and flagged paths
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/execution/coding_loop.py

### swe-af:pro

Pro engine (beta). Prebuilt swe-pro binaries vendored under go/bin for darwin and linux amd64/arm64; on by default only via af install or AgentField Desktop; classic loop elsewhere; cron triggers and reasoner calls behave identically.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/go/docs/pro-engine.md
- Locator: Opting out, supervised sidecar
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:ci

Post-PR CI gate. After the PR is pushed, watches gh pr checks with poll and wall-clock caps and runs a bounded fix-and-repush agent that must not silence tests.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/execution/ci_gate.py
- Locator: watch and fix loop
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/execution/ci_gate.py

### swe-af:afshot

AgentField workflows screenshot. First-party control plane Workflows screen with sample sec-af, voice-agent and healthcare entries; shows no SWE-AF build.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/docs/screenshots/screenshot-workflows.png
- Locator: repository screenshot asset
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:license

Apache License 2.0. Full Apache-2.0 text covering the repository (Python node, Go node, vendored binaries are in the same repository).

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/LICENSE
- Locator: LICENSE file
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: LICENSE

### swe-af:creds

Scout-negotiated credential store. Environment-scout credentials supplied by a human are scoped to the build and redacted from planning logs; not a general secret scrubber.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/swe_af/hitl/credentials_store.py
- Locator: credential handling and redaction
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: b281a4c50b38b32dd85636120de816e441adaa82
- Path: swe_af/hitl/credentials_store.py

### swe-af:afusage

Usage ingest handler. Control plane ingests SDK usage entries with cost_usd and token counts per execution for local accounting and UI usage views.

- URL: https://github.com/Agent-Field/agentfield/blob/161cab6a948b3cc7677a86d20c1a119ef56f6e4c/control-plane/internal/handlers/usage_ingest.go
- Locator: cost_usd and token fields
- Evidence kind: source-inspected
- Source date: Not stated
- Retrieved: 2026-11-16
- Commit: 161cab6a948b3cc7677a86d20c1a119ef56f6e4c
- Path: control-plane/internal/handlers/usage_ingest.go

### swe-af:repo

Agent-Field/SWE-AF repository. Apache-2.0 repository with 1,130 stars on 2026-11-16, last push 2026-11-12, not archived, no GitHub releases; description: autonomous software engineering fleet of AI agents.

- URL: https://github.com/Agent-Field/SWE-AF
- Locator: Repository header and GitHub API (stars, license, pushed_at)
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16

### swe-af:security

Security policy. Early public beta; fixes land on main first; private GitHub vulnerability reporting; acknowledgment within 3 business days, triage within 7.

- URL: https://github.com/Agent-Field/SWE-AF/blob/b281a4c50b38b32dd85636120de816e441adaa82/SECURITY.md
- Locator: Supported Versions; Response Targets
- Evidence kind: documented
- Source date: Not stated
- Retrieved: 2026-11-16
