Feed initial prompt to Codex on autolaunch (CROW-493)

[dgershman]

Summary

OpenAICodexAgent.autoLaunchCommand returned "codex\n" for .work and nil for every other session kind. Result: skill-created Codex coding sessions opened a bare codex TUI sitting at >_ waiting for the user to type, and app-created Codex .job sessions never auto-launched at all (SessionService logged a skip and pasted a ⚠️ echo).

Switches to a kind-based dispatch that mirrors CursorAgent.autoLaunchCommand:

• .work — bare codex\n (skill seeds the prompt at first launch; in-app resume just reopens the TUI)
• .job — first launch passes .crow-job-prompt.md as a positional argv (codex "$(cat …)"); subsequent launches fall back to bare codex (no --continue in MVP)
• .review — nil (review-on-Codex stays gated per the ticket's acceptance criterion 3)
• .manager — nil

The implementation now also honors defaults.binaries.codex via findBinary() ?? "codex", matching the Claude/Cursor convention — minor adjacent bug fixed along the way.

The skill side (crow-workspace/setup.sh::launch_codex + the Resources/.template mirror) is updated to pass the prompt the same way launch_claude_code / launch_cursor already do: cd \$WORKTREE_PATH && \$bin "\$(cat \$prompt_path)". Codex v0.139.0's --help confirms codex [OPTIONS] [PROMPT] is supported.

CROW-424 (Cursor .work)

Bundled-check requested in plan: after tracing both paths, Cursor's .work doesn't actually have the same Codex bug. The skill (launch_cursor, setup.sh:822) already feeds the prompt at first launch, and the in-app resume returning bare agent\n is the intentional MVP behavior — same shape Codex .work now adopts. The deeper "auto-feed prompt on .work resume" feature would need a new .crow-work-prompt.md convention + SessionService writer, applicable to all three agents simultaneously; out of scope here. Verify the CROW-424 repro against this PR before closing.

Test plan

• swift build --arch arm64 --target CrowCodex — passes
• arch -arm64 swift test --arch arm64 --filter OpenAICodexAgent (from Packages/CrowCodex) — 11/11 pass, including new .job first-launch / .job resume / .manager / binary-override-on-.work cases
• bash -n skills/crow-workspace/setup.sh — clean
• codex --help confirms codex [OPTIONS] [PROMPT] positional form is supported on v0.139.0
• Manual end-to-end: open a Codex coding session for a crow:auto ticket and confirm the agent receives the prompt as its first message and starts working without manual paste (post-merge acceptance per ticket criterion 1)

Closes #493

[dhilgaertner]

Code & Security Review

Critical Issues

None.

Security Review

Strengths:

• The .job first-launch command codex "$(cat <promptPath>)" is the established cross-agent convention (shared with Claude/Cursor). It is injection-safe: the file content is the output of a double-quoted command substitution, so backticks/$(…)/$VAR inside the generated prompt are not re-evaluated by the shell. No new attack surface.
• No secrets, no telemetry/OTEL prefix, no remote-control flag leaked into the Codex command (asserted by autoLaunchCommandIgnoresTelemetryAndRemoteControl).
• SessionService.launchAgent preflights .crow-job-prompt.md existence before dispatch (OpenAICodexAgent.swift:74 ↔ SessionService.swift:604-617), so a missing file surfaces a ⚠️ echo instead of launching Codex against an empty string.

Concerns:

• None.

Code Quality

• autoLaunchCommand now mirrors CursorAgent.autoLaunchCommand precisely — kind-based dispatch with .work → bare codex, .job → positional prompt on first launch / bare on resume (gated by reviewPromptDispatched), .review/.manager → nil. Good parity, clear comments.
• Adjacent bug correctly fixed: .work now resolves through findBinary() ?? "codex" instead of hardcoding the token, restoring defaults.binaries.codex honoring (CROW-484), with a dedicated regression test (autoLaunchCommandHonorsBinaryOverride).
• The skill-side launch_codex in both skills/crow-workspace/setup.sh and the Resources/…template mirror are byte-identical and match launch_cursor's shape.
• End-to-end wiring verified: job-prompt file is written agent-agnostically at job creation (SessionService.swift:1892-1902), so app-created Codex .job sessions get the file.

Verification performed:

• swift test --filter OpenAICodexAgent → 11/11 pass (new .job first/subsequent, .manager, binary-override cases included).
• bash -n clean on both setup scripts.
• Confirmed launch_codex blocks identical across skill + template.

Summary Table

Color	Meaning	Verdict effect
Red	Must fix	Request changes
Yellow	Should fix	Request changes
Green	Consider	Approve allowed

Recommendation: Approve — driven by [0 Red, 0 Yellow, 0 Green] findings. Clean, well-tested change with strong parity to the existing Cursor/Claude launch paths. (Note the PR's own unchecked manual end-to-end item is a post-merge acceptance step, not a blocker.)

---

🐦‍⬛ Reviewed by Crow via Claude Code