Ship codex/agent-fleet-deep-worker-respec - #137
Conversation
- add managed agent deep-worker (opus/high -> gpt-5.6-terra xhigh, workspace-write) - move fast-worker to opus/medium with explicit gpt-5.6-luna max target override - opus family default projection gpt-5.6-sol -> gpt-5.6-terra; sol stays for fable - document the two per-agent overrides as the only effort remaps; drop pinned model versions
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b90f9beefe
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| "explorer", | ||
| "deep-reasoner", | ||
| "fast-worker", | ||
| "deep-worker", |
There was a problem hiding this comment.
When project initialization or workflow repair writes a downstream policy, scripts/lib/project-init-lib.sh:1974 and scripts/ensure-task-workflow.sh:1279 still emit the previous six-agent roster. Because resolveManagedAgents() in scripts/check-agent-tooling.sh:841-847 trusts that configured list, a newly generated repository can pass strict readiness without deep-worker installed even though the package contract now declares it managed; update both generators, their asset mirror, and the corresponding runtime expectations.
AGENTS.md reference: AGENTS.md:L54-L56
Useful? React with 👍 / 👎.
| const HOME = os.homedir(); | ||
| const MANAGED_AGENTS = ["explorer", "deep-reasoner", "fast-worker", "gatekeeper", "root-cause-prover", "harness-evaluator"]; | ||
| const WRITABLE_AGENTS = new Set(["fast-worker", "root-cause-prover", "harness-evaluator"]); | ||
| const MANAGED_AGENTS = ["explorer", "deep-reasoner", "fast-worker", "deep-worker", "gatekeeper", "root-cause-prover", "harness-evaluator"]; |
There was a problem hiding this comment.
When repo-harness install invokes this expanded fleet installer, src/cli/installer/install-profile.ts still omits deep-worker from installProfileHostMutationPaths() (line 510), transaction component classification (line 617), and completeAgentFleetEvidence() (line 937). Consequently a failed install cannot roll back newly created deep-worker files or record them in the ownership manifest, while profile status can still call the old six-role fleet complete; update all three install-profile projections alongside this canonical roster.
AGENTS.md reference: AGENTS.md:L54-L56
Useful? React with 👍 / 👎.
Release scope v0.11.3..main (PR #137-#143, squash bd2155d, 4a79587, 8b506da, 3c99146, 83792a8, a37c16e, eb6c001): the breaking adopt-to-init cutover, the deep-worker fleet addition with Codex model projection respec, the docs/reference-configs generated-projection convergence, and four production defect fixes (MCP allowed-root canonicalization, ensure-task-workflow write-order race, acceptance-receipt fingerprint key-order normalization, contract-worktree squash-absorption cleanup). Version sources, localized READMEs, changelog, and the 0.12.0 release filing are updated; publish remains pending until the filing's release sequence completes.
* chore(workflow): capture release-0-12-0 plan and contract * chore(release): prepare 0.12.0 Release scope v0.11.3..main (PR #137-#143, squash bd2155d, 4a79587, 8b506da, 3c99146, 83792a8, a37c16e, eb6c001): the breaking adopt-to-init cutover, the deep-worker fleet addition with Codex model projection respec, the docs/reference-configs generated-projection convergence, and four production defect fixes (MCP allowed-root canonicalization, ensure-task-workflow write-order race, acceptance-receipt fingerprint key-order normalization, contract-worktree squash-absorption cleanup). Version sources, localized READMEs, changelog, and the 0.12.0 release filing are updated; publish remains pending until the filing's release sequence completes. * docs(review): record release-0-12-0 acceptance
1 participant
Agent fleet respec
deep-worker(heavy execution: cross-module refactors, one-pass-correct changes; workspace-write; sameRESULT:protocol as fast-worker).AGENT_TARGET_OVERRIDESin the generator holds the only two effort remaps; every other combination stays on the family default with effort carried through, fail-closed validation unchanged.docs/reference-configs/external-tooling.mdmapping table gains the missing fable→sol row; "Current assignments" resynced from actual fleet frontmatter; pinned model versions removed everywhere.check-agent-tooling.sh,.ai/harness/policy.json, helper mirrors, and the four hardcoded test tables.Verification:
bun test2094 pass / 1 skip / 0 fail (163 files); all required checks green (check-task-workflow --strictclean, architecture-sync advisory WARN for the pendingworkflow-engine-contract-assetsrequest is expected).The captured plan's contract-closure ceremony is intentionally still open; plan + slice notes ship with this PR.