podium

OpenAI Codex
vs Podium

A source-linked comparison across the complete Podium ADE rubric. 42 criteria sourced for both products; 29 criteria also agent cross-checked for both. 49 rows have at least one sourced finding, including 48 with an agent-cross-checked finding. No score or winner is inferred.

Each row shows a finding's verification status, support, check date and answer. Its scope, limitations, atomic answers and sources are on the product's page. Humans approve the rubric, templates and prototype; they do not review every finding. An agent source cross-check is not a hands-on product test. A missing sourced finding is not evidence that a capability is absent.

Rubric
0.3 · 51 criteria
Last source check
2026-11-13
Oldest source check
2026-11-06 (16 days old)
Agent cross-checked for both
29 criteria
OpenAI Codex source checks
2026-11-07 to 2026-11-13
Podium source checks
2026-11-06

Access and agent choice

ACC-01

Client access

OpenAI Codex

Agent cross-checked partial

Codex is available inside the ChatGPT desktop app on macOS and Windows; Linux is a preview on the distributions documented separately. The shared application is named ChatGPT desktop app; its exact instruction is "Choose ChatGPT or Codex." Codex remains a separately selectable developer experience, so this is evidence of a shared desktop surface rather than an explicit product-wide rename or merger announcement. Codex remote access supports ChatGPT iOS and Android apps, subject to rollout; connected execution hosts must stay awake and online. Codex Cloud tasks are available through ChatGPT on the web. Codex CLI has macOS, Linux and Windows installers. Codex has an IDE extension with in-place code context and edits; compatibility details are bounded by each editor's installation guide. iOS Codex can start tasks, send instructions, approve requested actions and review diffs or test results. Remote clients receive notifications on completion or when input is needed. iOS can continue a cloud task or one on a connected Mac or Windows PC under the same account and workspace.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-07. Scope and sources

Podium

Sourced partial

Podium ships desktop, CLI and browser clients, including phone-browser control. The native iPhone app is private beta; the complete download offering contains no native Android app.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Sourced partial

Codex supplies its own coding harness. In the inspected CLI/core, agent delegation creates Codex threads through AgentControl; model-provider configuration and MCP tools do not replace the native harness with another coding CLI.

Checked 2026-11-07 (15 days old). Scope and sources

Podium

Agent cross-checked partial

Podium launches the native Claude Code, Codex, Grok, OpenCode, Cursor and Pi CLIs in persistent terminals on machines you control. The 0.2.0 release calls Pi fully supported. Models come from each harness catalog, and capabilities differ per harness: Pi declares no quota, handoff or hooks. Mixed-harness operation was not exercised.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Supported local Codex clients can sign in with ChatGPT. API-key sign-in uses API billing rather than included ChatGPT usage. API-key authentication works for CLI, SDK and IDE but excludes cloud features such as GitHub review and Slack integration. Ollama and LM Studio are built-in local model providers that do not require OpenAI sign-in. An experimental app-server API switches between ChatGPT workspaces of one managed login; switching among separate accounts is not documented. The local CLI can run with an API key, Bedrock or Vertex credentials, or a local Ollama/LM Studio model; codex cloud needs a ChatGPT sign-in.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium supports native coding-agent logins and managed provider credentials. Claude subscriptions (via native login or a managed setup-token) and local ChatGPT/Codex login are supported paths, alongside Anthropic and OpenAI API keys. Docs list billing and hosted metering as not implemented product claims.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Install the CLI, open a project directory, run codex and authenticate; elapsed setup time was not tested. Local non-Git projects are supported for scheduled work; worktree mode specifically requires Git. Worktree instructions require a Git repository and do not require a hosted remote. Choose GitHub repositories, let Codex prepare dependencies and tools, review setup and test results, then publish the environment. Configure an SSH alias, verify login, install/authenticate Codex remotely, then add the host and remote project in Settings Connections. Cloud tasks use repositories from a connected GitHub or GitLab account, and codex cloud exec targets an environment and branch; other hosts are not documented. codex cloud tells signed-out users to run codex login, and codex doctor checks installation, config, auth and network.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

Docs describe a first-run flow (install, podium, register a Git repository, confirm a detected logged-in harness) and a headless VPS installer with machine joining. Setup time and failure rate were not measured, and no self-serve managed cloud setup is documented. The current README and installation guide use podium.do/download and podium.do/install.sh as the canonical stable entrypoints.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

codex exec runs non-interactive tasks in scripts and CI. The app-server serves JSON-RPC over stdio, Unix sockets or WebSocket, with generated TypeScript and JSON Schema; the commands are labelled experimental. TypeScript @openai/codex-sdk (Node.js 18+) and Python openai-codex. The SDKs start, run, continue and stream threads with structured output; codex cloud submits, checks status, lists, diffs and applies tasks. ChatGPT sign-in by browser OAuth or device code, API keys, and file or keyring credential storage; remote exec-server connections use bearer tokens. app-server, remote-control, cloud and exec-server are labelled experimental; exec and the SDKs carry no experimental label.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

The podium CLI covers issue create, show, update, claim, dependencies, artifacts, session send/status/read/stop and review actions, and the server exposes the same issue commands as a bearer-token MCP tool endpoint. No documented public HTTP API, SDK or webhooks were found.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

The CLI supports packaged skills and plugins. codex mcp manages external MCP servers, and /mcp login signs in from a session. SKILL.md skills (including project .agents/skills) and lifecycle hooks extend the agent alongside MCP tools. AGENTS.md and AGENTS.override.md are loaded, and a TUI migration flow imports configuration from other coding agents; importing their memories is still an off-by-default flag under development. MCP servers can use OAuth, including pre-registered client secrets; enterprise MCP can use Codex account grants. MCP config sets per-server and per-tool approval modes plus tool allow and deny lists; approval prompts include tool names and annotation hints. Users configure integrations in config.toml; admins can restrict hooks to managed ones via requirements.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium gives agents its own issue tools over CLI and MCP, and agent profiles import and export Claude Code agent files. No Podium documentation for user-configured MCP clients, plugins or custom tools was found.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-07. Scope and sources

OpenAI Codex

Agent cross-checked partial

Client implementation: The desktop and mobile apps act as remote clients of the app-server daemon; their own UI technology is not in the public source. Pinned source ccf9be8 holds the Rust CLI/TUI and only a desktop launcher; desktop and IDE client code is not public. Not established: desktop implementation, mobile implementation.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Desktop releases use a Tauri 2 shell that serves the Podium web UI; mobile source uses Expo and React Native. A native iPhone release is described as private beta.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

Planning and context

OpenAI Codex

Agent cross-checked partial

Intake and external tracker relationship: Tasks are threads started from the CLI, SDK or codex cloud exec; there is no separate issue object. Prompts come as arguments or stdin, with optional images. Sessions have IDs or names, and cloud tasks have task IDs. Not established: external import, chat git triggers, authoritative tracker, sync, supported edits, external roundtrips.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium issue records are native and authoritative with stable references, CLI creation and editing, and agent-driven filing from conversation. The in-product assistant has Linear search, create and move tools when a key is configured, but no import or two-way tracker sync was found.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Durable task plans: The update_plan tool publishes and updates a step plan in the thread. Thread goals persist in the local state database and saved sessions resume; step plans live in the session transcript. Not established: acceptance criteria, editors, revisions, implementation links, amendment.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Issue brief, acceptance criteria, state paragraph and todos persist on the durable issue record. Release 0.3.0 removed the separate design and notes fields and appended their text to the brief. On the human issue page the title, description and acceptance are editable and the brief is read-only; agents edit the brief and other fields through issue commands.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-11. Scope and sources

OpenAI Codex

Agent cross-checked partial

Hierarchy, dependencies and evolving scope: Subagent threads persist under parents and come back when a parent is unarchived; this is agent nesting, not a task hierarchy. Not established: epics, blockers, readiness, enforcement, state propagation, task to epic, identity retention, workflow constraints.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium supports persistent parent and child issues, phases and typed dependency links, including reparenting an existing issue without replacing its identity. Since 0.3.0 issue types are task, phase and decision: epic and other kinds are retired, and any issue with children acts as a parent. Dependency types are fixed to blocked-by, related, discovered-from, supersedes and duplicates. Cycles are refused and docs list an open blocker as a reason an issue cannot start. No mandatory workflow preset applies in the baseline.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Shared context and memory: Codex reads AGENTS.md before beginning work. SKILL.md skills, including project .agents/skills, plus bundled system skills. AGENTS.md files in the repository act as project docs; no other document store was found. A two-phase memory pipeline is stable but off by default in the feature table; when memories are enabled, fresh profiles default to memory v2. Global guidance is followed by repository-root to working-directory instructions; nearer files override earlier guidance. The memory read path parses memory citations; what users see was not established. Edit or remove an AGENTS.override.md file before the next startup to change guidance. Instruction discovery runs once per session startup; automatic rereading during a run is not established. Not established: deletion, takeover continuity.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

v0.3.0 routes Podium issue instructions and relay environment through terminal and engine session launch. Claude stream sessions use appended instructions and boundary callbacks; Codex server sessions use developer instructions, Grok ACP uses rules, and OpenCode uses the prompt system field. Claude and Codex boundary hooks can re-prime after compaction. Transcript mirroring is on by default; full-text indexing needs the command-palette flag, which defaults off. Agent profiles can carry markdown instructions. Curated memory, skills and memory correction or deletion controls remain unestablished. No instruction-delivery or compaction scenario was run.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

codex --search enables live web search with visible search activity. codex --image accepts images with a prompt. A file-search crate and shell tools search the local workspace.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

Not established. Podium mirrors daemon transcripts by default, keeps a full-text index only when the command-palette flag is on, and stores session attachments. No repository code search, architecture awareness or API retrieval feature was found.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Sourced partial

The public CLI supplies a per-turn update_plan checklist inside a conversation event stream. It does not establish separate persistent user-managed issues/tasks across sessions. The hosted Cloud and closed client implementation remains unestablished.

Checked 2026-11-13 (9 days old). Scope and sources

Podium

Agent cross-checked partial

Podium exposes native issue creation with an optional start flag, separate dispatch, editing, comments, labels, assignment, search, board filters, archive, sibling ordering and an issue tree. Issue records have their own identity separate from sessions and stages are durable. The idea-to-mission scenario was not run. v0.3.0 refuses an agent close or done update while issue sessions are computing unless --confirm-interrupt is supplied; guidance says to request review and close only on user instruction.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

No sourced public finding not-published

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

Podium

Agent cross-checked partial

CLI and MCP address the same canonical issue records for reads, creation, edits, claims, state, dependencies and artifacts. Relayed agents have worker capability scoped to their issue subtree, with --outside-scope confirmation for other existing issues. attachSession is not subtree-gated, but direct rehoming from a real issue is refused and an agent attach naming a Proposed target is refused. That does not make every Proposed subtree inert: operator attachment or spawn-time cwd binding can leave a session attached to a Proposed issue; newSubissue without targetId then bypasses the proposed-parent guard, creates a human-audience backlog child and attaches the existing session. This path remains at v0.3.0 source and was not exercised live. Concurrent-edit handling remains unknown.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

No sourced public finding not-published

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

Podium

Sourced partial

The spec CLI and service (living specs under pspec with stable SP IDs) are in v0.3.0. Specs is an experimental, default-off feature listed on the edge channel: the flag controls the Specs UI, the spec line in CLI help and whether agents are told about the spec CLI. The podium spec command itself is dispatched, and the server specs service has no flag check, so direct CLI use does not need the flag. This agrees with the docs, which call specs not baseline rather than blocked. Not run end to end.

Checked 2026-11-06 (16 days old). Scope and sources

Coordination and automation

COORD-01

Parallel work

OpenAI Codex

Agent cross-checked partial

Each cloud task has its own workspace; the page gives no concurrency quota. The agent command center browses and switches saved sessions, and up to 6 subagent threads run by default; concurrent execution of several top-level sessions was not established. Subagents default to 6 threads at depth 1, both configurable. Plus and Standard Business usage is estimated per five-hour period; Pro currently has no five-hour limit, weekly limits may apply, and Enterprise/Edu with flexible pricing have no fixed rate limits. No fixed concurrency quota is published. The command center marks unread results, list_agents reports live agents, and codex cloud status shows task state.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Docs and release notes describe many sessions running as independent terminals, one worktree per issue, and several paired machines under one server, with session and issue state visible on the flight deck, board and CLI. Agent-initiated spawns are limited to 10 per target issue per UTC day. Achieved concurrency is unmeasured.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Ownership and dispatch: spawn_agent takes an optional agent role; there is no human assignment or task claim object. User-defined roles carry descriptions that guide which role the model spawns, plus their own config layers; routing is model-driven. Not established: human assignment, claims, queues, duplicate prevention, ownership enforcement.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Issues carry an assignee, claim sets assignee and in_progress, a coordinator session can be set per issue, sessions have seats, and work can be placed on a chosen machine or agent profile. claim has no exclusivity check of its own, locks are advisory leases and Podium does not serialize agents editing one worktree, so exclusive ownership is not enforced.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Local custom agents can set model and reasoning effort; unspecified children inherit their parent. The main agent collects subagent results into a consolidated response. Further scoped facts are recorded in the atomic answers. User-defined roles with descriptions, config layers and nicknames. send_input, send_message and followup_task deliver messages to existing agents. The parent thread orchestrates through tools; spawn depth defaults to 1. Subagent relationships persist, and unarchiving a parent restores descendant threads. followup_task and resume_agent continue existing agents. The model calls spawn_agent, usually when asked; the public source shows no direct user control to spawn a child.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium orchestrates delegation itself: podium agent spawn starts a full child session on an issue with a chosen harness, model, effort or agent profile, mail carries messages, and bounded await reports the child result. Children are real sessions shown as seats on the flight deck. Agent-initiated spawns take a daily budget per issue. A complete delegated task with retained child findings was not run.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Desktop Codex supports recurring tasks with an RRULE custom cadence. Standalone scheduled tasks start a new chat per run; an in-chat schedule retains the conversation. Local project schedules require the desktop app running on the execution computer. Local scheduled tasks require the computer on and app running; cloud web scheduling is a separate workflow. Thread goals hold an objective, an optional token budget and a status: active, paused, blocked (shown as stalled in the TUI), usage-limited (set when a usage limit stops the active goal), budget-limited (token budget reached) or complete.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

The server owns a scheduler for cron and one-time automations that spawn agent sessions, skipping overlaps and not backfilling occurrences more than an hour late. The Automations UI is an edge-visibility feature that defaults off and docs list automations as non-baseline. Agents can subscribe to issue and session events.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

The interactive CLI accepts steering during the active turn. Scheduled acts as an inbox with unread indicators for findings requiring attention. Interrupting a turn also stops Code Mode cells and nested tool calls; agents can be interrupted or closed. codex resume, fork and queue --resume reopen sessions, and resume_agent reopens subagents. The TUI sends terminal desktop notifications (on by default), a notify command can be configured, and the command center marks unread results; push or email notifications are not in the public source. On-request approval is the default. The stable guardian_approval feature flag is on by default, which makes automatic review available, but approval requests go to the user unless approvals_reviewer is set to auto_review; managed checkpoints force a human decision. Safety blocks explain themselves and offer safer follow-ups; subagents receive their preparation failures. Sessions are saved under CODEX_HOME/sessions and can be archived, unarchived or deleted.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Chat and native CLI views share a session. v0.3.0 defaults human controller chat to steer and immediately types it into running Claude Code, Codex, Grok and OpenCode TUIs that declare native busy-input queuing; undeclared harnesses wait for a boundary, and compaction waits. Mail and automation keep when-ready delivery. Stop, resume, questions, decision issues and offers remain available. This is inspected implementation, not a cross-harness runtime parity result.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Fleet and epic management UI: Subagent threads are tied to parents and restored with them; the default TUI workspace shows subagent transcripts beside the main conversation. A nested tree view is not documented. The command center and session picker mark unread results. The command center pages through and searches older tasks, and tasks can be pinned. Directory grouping and a repository filter span sessions across linked worktrees. Not established: fleet fixture, dependencies, navigation errors, fixture mode.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Release 0.3.0 ships a flight deck showing a whole mission in one view: phases, every task and its state, who works on what, what is blocked and by what, and proposals. Done work stays visible. The CLI prints the same snapshot. The waterfall source is dormant and the Timeline view was deleted, so per-task timeline history is not shown. The shared 12-task UX scenario has not been run.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-11. Scope and sources

OpenAI Codex

Agent cross-checked partial

A scheduled task can choose default or explicit model and reasoning effort. Desktop tasks select projects and local or worktree execution. A single codex exec prompt or one interactive turn. The model spawns subagents through tools when asked in chat. Reasoning effort can change mid-conversation by default, and follow-ups steer a running turn. Subagent relationships are persisted. /config searches settings and new sessions show occasional tips.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

A quick task can be an interactive session or an issue; an issue can grow into a mission by gaining children while keeping its identity; no mandatory workflow preset applies in the baseline. Per-agent harness, model and effort are chosen at spawn or through agent profiles, and the deck shows seats. The coordinator-with-two-reviewers scenario was not run.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

Execution and reliability

OpenAI Codex

Agent cross-checked partial

The CLI inspects, edits and runs tools in the local project. Codex can execute against an SSH host with Codex installed and authenticated on its PATH. Cloud tasks can keep working while the originating computer is asleep. The desktop UI can control Codex tools and repository files on an SSH host. A paired iOS or Android app can steer and approve work on an awake Mac/Windows host. The Apache-2.0 agent runs on any machine the user controls; no self-hosted cloud service was found. Local use bills the signed-in ChatGPT plan or the API or cloud-provider account; cloud tasks need ChatGPT sign-in and share the plan allowance. A background server starts automatically for eligible interactive sessions and serves remote desktop and mobile clients; queued messages resume after reconnection. Continued execution after the client disconnects was not established.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

The published Podium execution choices are local user machines and a user-owned VPS or paired machine. No vendor-operated agent host is in the complete download offering.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

Codex supports local OS sandboxing, user-managed devcontainers and Git worktrees. The current Cloud environment guide explicitly gives a VM per cloud task with plan-dependent resources; local and cloud isolation scopes are separate. On Linux, bubblewrap adds a PID namespace and seccomp filters syscalls; macOS uses Seatbelt. Isolation between tasks was not established. Credentials are masked in supported remote shells and can live in the keyring. In Codex Cloud (Legacy) environments, which serve Code Review and the Linear and GitHub integrations, secrets are available only to setup scripts and are removed before the agent phase; secret handling in the current Cloud environments guide was not checked.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

Podium separates issue files with Git branches and worktrees, but its released daemon launches native host processes without product-managed container or VM provisioning.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

Phone worktree control still requires the execution computer awake; this is distinct from cloud tasks that run while it sleeps. codex resume reopens saved local chats; this does not establish worker-crash recovery. Remote control stops if the host sleeps, disconnects or closes the app and becomes available when the host returns; successful recovery was not tested. Handoff moves a chat between its Local checkout and a worktree while Codex performs the Git operations. An existing cloud task resumes with its own saved files, uncommitted changes and installed tools. /daemon restarts the background server, warns about interruptible work and reconnects. Context is compacted automatically, and a failed compaction leaves no partial summaries. Uncertain submissions are resolved before queued messages resend. codex doctor diagnoses problems, and damaged SQLite databases are kept as backups.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Closing a browser tab does not stop the agent process on a running Podium host. Idle sessions may hibernate under configurable policies; a stopped session keeps its transcript, branch and resume reference. Release 0.2.0 says a failed update rolls back and running sessions keep running through the upgrade; database rollback remains a manual restore. Terminal reattachment goes through the driver adopt contract with process-identity and observation-lease checks. Worker-crash and multi-day behavior were not tested.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Cloud setup identifies and installs repository runtimes, dependencies and tools. Background repository refresh preserves dependency caches without rerunning installation. Publish captures prepared filesystem state for new tasks. The start skill documents how to start services and check readiness. New tasks reuse the published prepared filesystem; background refresh preserves caches. Transient environment startup failures retry on the next turn.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Only machine setup is documented: the installer installs missing runtime tools and pairs a host, and the state directory holds server-managed secrets. Dependency caching, services, ports and automatic repair for project environments were not found.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Provider SSE idle timeout defaults to 300,000 ms and is configurable. Provider HTTP request retry count defaults to four; interrupted SSE streams default to five retries. Retries for HTTP and SSE failures are configurable; successful recovery was not tested. Automatic compaction, a context-remaining tool and optional goal token budgets; no hard per-task cap was found.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium exposes retryable errors, optional auto-continue with backoff, bounded waits and provider quota. Claude stream and hosted headless drivers enforce a supplied turn deadline, with no deadline when timeoutMs is absent. The one-shot headless server runner supplies its budget, default ten minutes. Agent spawns have a daily per-issue budget. Operator-configurable task-wide turn, token and spend caps remain unestablished; quota display is not enforcement.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Opt-in OTel exports structured API, tool, approval and stream events. Telemetry records request and tool durations and status; full distributed-trace parity is not implied. API events include status, duration and error details. Configured OTel stream events include token counts at response completion. Tool-decision events identify approval or denial and whether configuration or user made the decision.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-07. Scope and sources

Podium

Agent cross-checked partial

Podium offers local logs and crash export, structured session status, issue stats, a usage and analytics sheet with hourly list-price cost, per-task costs and host load panels. The 0.3.0 notes report a vendor benchmark of about 330 MB server memory against about 1.2 GB for an earlier Podium build, and 8 to 52 percent faster start; the method is published in the repository and was not reproduced here. Usage export and cross-reconnect consistency were not established.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-11. Scope and sources

OpenAI Codex

Agent cross-checked partial

Per-task API and compute cost: turn.completed (TypeScript) and TurnResult (Python) report token usage. Not established: task cost, session cost, child cost, epic cost, harness aggregation, billed vs estimated, subscription allowance, provider reconciliation, compute cost, refresh delay, alerts, export, drilldown.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Since 0.2.0 Podium shows what each task costs, on the task header and the usage sheet, with per-session rows and descendant rollups. The 0.3.0 deck adds a mission cost chip. Figures are token estimates at public model list prices, not subscription invoices or provider bills.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Compute-aware scheduling and verification: agent_max_threads defaults to 6 and is configurable. Interrupt stops Code Mode cells and nested tools, and dropped remote MCP requests are cancelled. Not established: queues, cpu ram, backpressure, host awareness, check reuse, check deduplication, bounded tests, fairness, recovery, shared limits.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium reports host memory, load and disk, auto-hibernates idle sessions under pressure, and provides named advisory resource leases. The merge-queue panel was removed; a shipping queue remains behind the Shipping flag. Concurrency limits and bounded test queuing were not established.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

Human collaboration and control

OpenAI Codex

No sourced public finding not-published

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

Podium

Agent cross-checked partial

The issue UI displays owner and visibility read-only and deliberately omits sharing controls. Backend share handlers and a Members settings section with invites and admin or member roles exist; they do not establish a finished human collaboration workflow. Docs say small teams can share an instance but the interface prioritizes one operator.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Business includes SAML SSO; Enterprise adds SCIM and RBAC. Workspace switching checks membership before switching (experimental API).

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

Podium defines member/admin instance roles and admin-only membership management. The released client authentication uses password-backed sessions and paired-device tokens without enterprise SSO.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

Local approval policies control escalation beyond the sandbox; /permissions exposes read-only and other configurations. Network permissions and approval policy are separate controls. Remote projects use SSH; device control uses an authorized secure relay rather than exposing the host publicly. A read-only sandbox mode is documented for local Codex use. MCP tools have per-tool approval modes, and approval prompts include tool names and annotation hints. Managed checkpoints require a human decision that Guardian automatic review cannot satisfy; approver identity is not recorded in the sources read. Session and persistent approvals exist and cannot satisfy managed checkpoints. Exec-policy rules allow commands, and codex execpolicy check tests them. Granular policy auto-rejects categories, and approved commands keep explicit filesystem denials. The default approval policy is on-request. The OS sandbox and network proxy enforce boundaries locally; no bypass test was run.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

Agent-session management requests go through an operator approval broker with approve and deny; the privacy docs say lifecycle commands and automation requests may require operator approval. Harness permission prompts appear as a Permission needed state; credential copying during pairing is a per-grant opt-in. Command-level policy enforcement beyond the harness own modes was not established.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

Opt-in OTel tool-decision events retain approved or denied decisions and decision source. OTel exports selected events to user-controlled collectors. Saved local session rollouts record agent actions. Saved sessions persist across runs. Threads can be archived, deleted or reverted.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Sourced partial

Issue actions, approval steps, comments and mail are recorded server-side. Issue activity events, including approval steps, are pruned by the server's janitor once they are older than 14 days or fall outside the newest 50,000 rows. That policy does not define all content retention, export or tamper-evident audit guarantees.

Checked 2026-11-06 (16 days old). Scope and sources

OpenAI Codex

Agent cross-checked partial

Review experience and guidance: Patch details expand in the transcript during approval. codex review reports on uncommitted changes, a base branch or a commit. Keyboard transcript navigation, a configurable leader key and keymap. Not established: pending reviews, evidence navigation, comments, request changes, effort, missed evidence, narrow screen, context switching, roles, retained context.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Review is a documented loop: issue stage review, diff and checks, artifacts, offers with merge or send-back buttons, and a Tray of review-ready work. Release 0.3.0 adds quoting text from a reply or review finding into the next message and a full-height reading view with an outline. The board has a key map with a help sheet. Guided-review quality and narrow-screen use were not measured.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

No sourced public finding not-published

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

Podium

Agent cross-checked partial

The durable issue, session transcript, artifacts and mail persist for any later agent, and a live session can be handed to another machine in an edge feature. Colleague takeover by a second person was not established.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

Verification and delivery

OpenAI Codex

Agent cross-checked partial

Cloud review shows changed files and check results. Viewed and generated images preview in the transcript, and stored thread attachments are managed through app-server. Local rollout files keep thread items; cloud retention was not established. wait_agent can return a subagent's final message to its parent.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Artifact commands store immutable server-side copies of issue-linked files, including images, video, HTML and documents, which render in the issue sidebar and offers and stay readable after the worktree is gone. Each task has an artifacts page grouped into screens, documents and reports, with filters and search. Revision linking remains untested.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-07. Scope and sources

OpenAI Codex

Agent cross-checked partial

A dedicated local review can inspect uncommitted changes, a commit or a base branch and return prioritized findings. The stable, default-on shell tool runs test commands inside the sandbox; no dedicated test gate was found.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

In the experimental Shipping path a configured validation profile runs before landing (default argv bun run test with a 10 minute timeout) and a failed validation holds the order. Lint, type, CI and spec conformance gates and independent review requirements were not found as product-enforced policy.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Cloud results can be committed or opened as pull requests after review. Cloud tasks take a branch, and local reviews compare against a base branch. Cloud users can request follow-up changes in the same chat; pull-request review comments were not established.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Issues own branches and worktrees, and issue actions rebase, open a pull request with gh (pr create --fill) and merge, subject to a merge-style setting (ff-only by default, with auto-rebase before merge). Podium does not resolve branch conflicts, and PR review comments and CI response were not established.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

No sourced public finding not-published

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

Podium

Agent cross-checked partial

The operator accepts work through offers and runs the merge action under a branch merge-lock; in the experimental Shipping path approval is frozen to an exact head and delivery lands on a local target branch by default. Release and deployment are not product features found in the sources.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Completion evidence and rollback: thread/revert rewinds a thread in the app-server; code rollback was not established. Not established: commit links, pr links, build links, deployment links, delivery status, retained evidence, rollback, stale approval invalidation.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Shipping orders record approved base and head SHAs, tested integration SHA, landed ref and validation result in a durable delivery receipt, and stale approvals hold the order. Shipping is experimental. The pr action records a PR URL on the issue. Build and deployment links and rollback were not found.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

No sourced public finding not-published

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

Podium

Agent cross-checked partial

The shipping service records revision-specific approval, configured validation and delivery receipts. The Shipping panel is an experimental edge toggle, off by default; the separate Queues panel was removed.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-07. Scope and sources

Ownership, privacy and economics

OpenAI Codex

Agent cross-checked partial

The repository LICENSE at fc01c073b9ea09c06347a234a1ca231a9ef65bf5 is Apache-2.0; hosted components and separate clients require their own licensing checks. The IDE extension and Codex Cloud are explicitly not open source. CLI, SDK and App Server sources are public; IDE extension and cloud service are not open source. Public CLI/App Server source does not establish hosted Cloud or IDE implementation parity. Latest stable release rust-v0.171.0 (2026-11-20) maps to commit ccf9be8.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

The public repository is licensed Apache 2.0, and the website presents Podium as an open-source agent development environment. Hosted extensions and hosted-service terms are not established.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

The Codex CLI and App Server core can run on user-controlled infrastructure under the public source distribution. This hosts the coding loop and local session/control process, while inference uses its configured provider. Codex Cloud and the closed IDE/desktop components are not claimed self-hostable. Sessions are local JSONL rollout files under CODEX_HOME; a hosted export was not found. A TUI flow migrates configuration from other coding agents; importing their memories is behind an off-by-default flag under development. JSONL session rollouts and JSONL exec events. The Ollama and LM Studio providers need no OpenAI sign-in; this was not tested.

Checked 2026-11-13 (9 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Local desktop and self-hosted installs run on user-controlled machines with a documented state directory. The docs describe stopped full-state backup and automatic pre-migration database-only snapshots retaining three copies. Those snapshots omit artifacts, transcripts, uploads, configuration and secrets; no portable structured export/import contract was established.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-11. Scope and sources

OpenAI Codex

Agent cross-checked partial

The Business plan states no training on business data by default; consumer routes are not covered by this statement. Built-in Statsig analytics and independently configured OTel exporters have distinct controls. The default Statsig metrics exporter follows analytics.enabled; the reference says unset analytics uses the client default. OTel log export is separately opt-in. Set analytics.enabled=false for built-in Statsig metrics, and set OTel log, metrics and trace exporters to none to disable explicitly configured telemetry exports. OTel prompt and response logging default to off and need opt-in. Credentials can be stored in a file, the OS keyring or keyring with fallback, and are masked in supported remote shells. ChatGPT workspace routing reports us, us_cr or NO_CONSTRAINT.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Docs state repositories and native harness credentials stay on machines you control and server state lives in a documented directory. Since 0.2.0 a daily usage report is on by default and can be turned off; crash reports stay opt-in. Usage reports carry an install id, schema version, Podium version, OS, architecture, bucketed install age and machine count, session counts per harness and whether issues were used. Crash reports carry the same install id, versions, OS and architecture, a fixed error type and up to 20 stack frames, each with a Podium source file path, line and function when there is one; error messages and frames outside Podium are dropped. Neither tier sends prompts, transcripts, your code or paths outside the Podium source tree. Model-provider handling is the harness own. Hosted analytics and website analytics are separate paths.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-21. Scope and sources

OpenAI Codex

Sourced partial

Codex offers Free and recurring paid ChatGPT plans, with separate API-key billing. Plus/Pro users can purchase additional credits after included limits, and eligible workspace plans can buy shared credits. Existing USD plan prices retain their recorded scope.

Checked 2026-11-10 (12 days old). Scope and sources

Podium

Agent cross-checked partial

No Podium price is published: podium.do/pricing returns 404 and the docs list billing as not implemented. The software is open source and free to download; model and compute costs are the user's own.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

No sourced public finding not-published

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

Podium

No sourced public finding not-published

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

OpenAI Codex

Agent cross-checked partial

The Linux desktop client is in preview, with named x64/ARM64 distribution packages. Linux preview supports Ubuntu 24.04/26.04, Debian 13, Fedora 43/44 and current Arch. Public GitHub issues with per-surface templates, and a Bugcrowd security channel; no response commitment. Stable releases from 0.157.0 (2026-09-25) to 0.171.0 (2026-11-20), with 0.173.0 alphas through 2026-11-21. Release code rejects the removed ollama-chat provider and names ollama as the replacement. The docs label Codex Cloud (Legacy) and point to a newer guide, without an end date.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium is pre-1.0. Stable v0.3.1 (2026-11-06), an update fix, follows v0.3.0 (2026-11-05), v0.2.0 (2026-10-11) and v0.1.0 (2026-08-21), with edge and dev prerelease channels. Windows and Linux desktop are Experimental, the native iPhone app is in private beta and hosted Podium is unconfirmed. Support runs through GitHub issues, Discord and private vulnerability reporting.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

No sourced public finding not-published

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

Podium

No sourced public finding not-published

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

OpenAI Codex

Agent cross-checked partial

A $0 monthly plan includes limited desktop Codex usage, subject to rollout. Free access is limited to listed desktop model/speed eligibility; detailed task allowances depend on the plan and model. The Apache-2.0 CLI has no license fee; with Ollama or LM Studio inference runs locally. Not tested.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-22. Scope and sources

Podium

Agent cross-checked partial

Podium is open-source Apache-2.0 software, and the download page calls it free and open source. Download links for macOS, Linux and Windows redirect to GitHub release assets without a login (observed by HTTP requests; not installed). No Podium charge is documented for self-hosting; agent provider usage is separate. The native iPhone app is private beta and hosted access is unconfirmed.

Checked 2026-11-06 (16 days old); agent cross-checked 2026-11-06. Scope and sources

OpenAI Codex

Agent cross-checked partial

Prices distinguish individual subscriptions, Business seats and separate API usage. Enterprise and Edu require contacting sales.

Checked 2026-11-07 (15 days old); agent cross-checked 2026-11-07. Scope and sources

Podium

No sourced public finding not-published

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