Codex-Co-Engineer 3.4.0

Codex-Co-Engineer 3.4.0

Repository-side GitHub Release body for Codex-Co-Engineer 3.4.0.

A clearer Co-Engineer experience

3.4.0 gives Codex a visible team of external co-engineers while preserving a
simple contract: up to eight isolated assignments, one bounded submission,
one coordinated wait, and one verified decision. The plugin page, README,
skills, inline cards, accessibility states, wordmark, and final static artwork
now use the same language:

• Delegating to Co-Engineer
• Chatting with Co-Engineer
• Using Grok Co-Engineer
• Using Cursor Co-Engineer
• Using Muse Co-Engineer

Codex remains chief engineer. External workers operate on their own branches
and managed worktrees. Worker commits and draft PRs are evidence; they do not
silently become accepted product changes.

Luna-first, interrupt-driven coordination

Luna Max can serve as the low-cost relay and control layer. It receives typed
task-to-task events, maintains delivery and acknowledgement cursors, and wakes
Sol only for meaningful boundaries such as completed, blocked, failed,
question, timeout, user update, or merge-ready. Attempt and generation fences
prevent a late worker or replaced Luna task from settling the current run.

Sol High/XHigh remains the exception layer and regular-merge authority. A
merge requires the expected and current branch head/tree, fresh green checks,
accepted verifier evidence, and safe topology. Provider prose and thread
messages cannot override those facts.

Less context, better evidence

• Deterministic reducers aggregate typed receipts without replaying every
  worker transcript through another model.
• UsageLedgerV1 records provider-reported tokens, host-visible model bytes,
  evidence bytes, invocations, waits, attention rounds, retries, and unknowns
  separately.
• Progressive lane health exposes provider, branch, head, and state first;
  large diffs and logs remain retrievable by reference.
• The PR-ready final card reports exact blockers or a verified
  ready_for_sol_merge result.
• Cleanup planning revalidates task/worktree/lock/ref identity and uses
  expected-SHA compare-and-swap semantics before any deletion.

Truthful attention and compatibility

Grok questions that cannot enter the structured question channel no longer
project as successful work. They remain unresolved attention with bounded
question evidence. The five MCP tools remain status, delegate, task,
tasks, and cancel; omitting additive run fields preserves the 3.2.1
single-task path.

Validation

The release candidate is accepted only after the repository-owned exact-tree
gate, current GitHub CI, package inventory, and identity checks pass. The full
integrated feature candidate passed 2,079/2,079 tests before the version-only
release commit was formed.

Maintainer publication

After the exact candidate is reviewed, replace EXACT_REVIEWED_MAIN_SHA with
that immutable commit and verify it before publishing:

test "$(git rev-parse HEAD)" = "EXACT_REVIEWED_MAIN_SHA"
gh release create v3.4.0 \
  --target EXACT_REVIEWED_MAIN_SHA \
  --title "Codex-Co-Engineer 3.4.0" \
  --notes-file docs/releases/v3.4.0.md

This document does not independently authorize tagging, pushing, or release
publication.

Codex-Co-Engineer 3.3.0

Codex-Co-Engineer 3.3.0

Repository-side GitHub Release body for Codex-Co-Engineer 3.3.0.

Bounded runs on the five-tool catalog

3.3.0 adds a first-class bounded run: 1–8 independent assignments against one
immutable repository/base identity. The MCP catalog remains status,
delegate, task, tasks, and cancel. A run is additive parameters on
those tools, not a sixth tool.

• Submit with delegate run (1–8 lanes). Direct mode, replay, fallback,
  merge, push, and create-PR are rejected on that path.
• Inspect with status or task plus run_id.
• Wait with wait_until: "decision_or_attention"; routine progress never
  wakes.
• Latch attention and deliver one same-session run reply with task
attention / run_reply.
• Cancel or proof-bound cleanup with cancel run_id.

Provider and model are explicit on each assignment or filled from one named
data-only profile. VerificationPolicyV1 is the only executable command
catalog. Codex remains the only final acceptance and merge authority. Frozen
verified child deltas may be composed into one run-owned, single-parent,
non-authoritative candidate; that candidate is never the integration
authority.

Every 3.3.0 run Cloud lane must pin one exact already-pushed
provider-visible SHA in starting_ref. Individual 3.2.1 Cloud tasks still
treat starting_ref as optional.

The repository argument remains the literal MCP property repo:

{
  "task_id": "review-auth-refactor",
  "provider": "grok",
  "repo": "/absolute/path/to/git-worktree",
  "role": "review",
  "workspace_mode": "managed",
  "prompt": "Review the current branch and return concrete evidence.",
  "expected_duration_ms": 600000
}

Omit additive run fields to keep that exact 3.2.1 single-task path.

Wait on a run without waking on routine text:

{
  "run_id": "auth-split",
  "wait_until": "decision_or_attention"
}

Compatibility

• The MCP tool count remains five.
• Omitting run fields preserves exact 3.2.1 single-task, compact, wait-any,
  structured transport, DSH Muse default, and Ox Alpha selector behavior.
• Direct mode remains available only for 3.2.1 single-task delegate.
• Run submissions never use direct mode.
• Gate B (context-efficiency) and Gate C (credit economics) stay advisory.

Out of scope

3.3.0 does not add semantic memory, cross-run search, learned routing,
protected-branch integration, or automatic garbage collection.

Validation

npm --prefix plugins/codex-co-engineer test
node scripts/validate-release.mjs
npm --prefix plugins/codex-co-engineer run setup:check

This document does not publish live provider transcripts or host
measurements.

Maintainer publication

After the exact candidate is reviewed, replace EXACT_REVIEWED_MAIN_SHA with
that immutable commit and verify it before publishing:

test "$(git rev-parse HEAD)" = "EXACT_REVIEWED_MAIN_SHA"
gh release create v3.3.0 \
  --target EXACT_REVIEWED_MAIN_SHA \
  --title "Codex-Co-Engineer 3.3.0" \
  --notes-file docs/releases/v3.3.0.md

This document does not authorize tagging, pushing, or release publication.

Codex-Co-Engineer 3.2.1

Codex-Co-Engineer 3.2.1

Repository-side GitHub Release body for Codex-Co-Engineer 3.2.1.

Optional Ox Alpha for DSH

3.2.1 keeps Muse Spark 1.2 Contributor as the default DSH model and adds an
explicit per-task Ox Alpha choice through OpenRouter:

{
  "task_id": "ox-review",
  "provider": "dsh",
  "dsh_model": "stealth/ox-alpha",
  "repo": "/absolute/path/to/git-worktree",
  "role": "review",
  "prompt": "Review the current branch and return concrete evidence.",
  "expected_duration_ms": 600000
}

The setup command creates a separate owner-only Ox Alpha ACP configuration.
Its credential is read from OPENROUTER_API_KEY,
CODEX_CO_ENGINEER_OPENROUTER_API_KEY_FILE, or the default owner-only file at
~/.config/codex-co-engineer/openrouter-api-key. Muse continues to use its
existing configuration and credential.

The generated Ox profile follows OpenRouter's published model metadata:

• context window: 1,048,576 tokens;
• maximum completion: 131,072 tokens;
• mandatory reasoning with max as the default effort;
• supported efforts: low, high, and max;
• native sampling defaults: temperature 1, top-p 0.95.

DSH continues to use the pinned rc.7 ACP composition through ACPX. Once ACPX
spawns, dispatch remains uncertain and Co-Engineer never replays the prompt
through the CLI or another model.

If ACPX cannot start for an explicit Ox Alpha task, the task fails before
prompt dispatch. Co-Engineer does not use the model-blind DSH CLI fallback for
that task.

Compatibility

• The MCP tool count remains five.
• Omitting dsh_model keeps the existing Muse route, credentials, and
  pre-spawn CLI fallback behavior.
• dsh_model is rejected for non-DSH providers.
• Unknown DSH model identifiers fail before workspace creation and dispatch.
• Task and diagnostic receipts record the selected DSH model.

Validation

npm --prefix plugins/codex-co-engineer test
node scripts/validate-release.mjs
npm --prefix plugins/codex-co-engineer run setup:check

Live Ox Alpha acceptance requires an explicit OpenRouter credential and is
recorded separately from credential-free CI.

Maintainer publication

After the exact candidate is reviewed, replace 99498a5a6e417fd81c72d8072103993ddb4d86c3 with
that immutable commit and verify it before publishing:

test "$(git rev-parse HEAD)" = "99498a5a6e417fd81c72d8072103993ddb4d86c3"
gh release create v3.2.1 \
  --target 99498a5a6e417fd81c72d8072103993ddb4d86c3 \
  --title "Codex-Co-Engineer 3.2.1" \
  --notes-file docs/releases/v3.2.1.md

This document does not authorize tagging, pushing, or release publication.

Codex-Co-Engineer 3.2.0

Codex-Co-Engineer 3.2.0

Repository-side GitHub Release body for Codex-Co-Engineer 3.2.0.
This file is the publication source. It does not create a GitHub Release,
tag, or remote by itself.

Date: 2026-08-20
Package / plugin / MCP id: codex-co-engineer
Skill: control-codex-co-engineer-agents
Tools: status, delegate, task, tasks, cancel

What changed

3.2.0 keeps the thin five-tool catalog and makes parallel coordination cheaper
without changing the 3.1.1 default omitted-mode response shapes.

Measured efficiency

Compact readiness and list paths avoid constructing full receipts for omitted
rows. With full JSON-RPC text duplication, measured sizes stay within:

• readiness-only status ≤ 8,192 bytes
• compact status with 20 cards ≤ 24,576 bytes
• compact tasks page with 20 cards ≤ 32,768 bytes

The single-task view: "compact" structured payload is capped at 8,192 UTF-8
bytes by the MCP server. That is a server guarantee, not a measured Codex
Desktop renderer hard limit.

Compact task and list pagination

• status: detail: "compact", task_limit 0–20, or include_tasks: false
  for readiness-only checks.
• task: view: "compact" for routine progress; keep
  view: "diagnostics" for attention and failure evidence.
• tasks: detail: "compact", limit 1–20, and opaque keyset next_cursor
  with provider/state filters bound into the cursor.

Compact status/task cards preserve the complete valid task ID. Pass that ID
unchanged to task, cancel, or wait-any; the card's normalized state and
updated_at carry the routine lifecycle signal without duplicating raw status
or terminal timestamp fields.

Opt-in structured transport

Clients that consume structuredContent may set
response_mode: "structured" on any of the five tools. Text becomes a bounded
fallback; structuredContent remains authoritative. Omitting the property
preserves the exact 3.1.1-compatible full JSON text duplication. The
efficiency harness records at least 30% aggregate JSON-RPC reduction versus
the immutable 3.1.1 duplication baseline when structured mode is opted in.

Bounded, redacted terminal evidence

Terminal provider results are redacted and bounded, including nested objects.
When evidence is clipped, receipts report result_truncated and, when the
source size is known, result_original_chars (Unicode code points).

Managed workspace and Cursor Cloud preflight

Managed local tasks verify worktree identity before provider launch. Cursor
Cloud preflight hardens provider-visible origin checks and fails closed on
credential-bearing or unsupported origins. Cloud still requires the local
checkout in repo and a pushed immutable commit SHA in starting_ref.

Wait-any semantics

Coordinate 1–8 exact tasks through one tasks call:

{
  "task_ids": [
    "change-api-validation",
    "update-operator-guide"
  ],
  "wait_until": "terminal",
  "wait_ms": 3600000
}

Supply per-task event cursors when continuing a wait. Wait-any options cannot
be mixed with list filters or pagination. Each target snapshot and live event
preview is bounded; when a preview is present, progress.detail_hint directs
the caller to task for that target's full live event detail. Aggregate
structured wait-any responses stay within a 72 KiB cap. Disconnecting the
waiter does not cancel provider work.

Install

Requires Node.js 24+, Git, Codex CLI, and worktree-bootstrap for managed
local tasks.

git clone https://github.com/ajhcs/Codex-Co-Engineer.git
cd Codex-Co-Engineer
codex plugin marketplace add "$PWD"
codex plugin add codex-co-engineer@codex-co-engineer
npm --prefix plugins/codex-co-engineer run setup
npm --prefix plugins/codex-co-engineer run setup:check

Start a new Codex session after installing or upgrading the plugin.

The repository argument remains the literal MCP property repo:

{
  "task_id": "review-auth-refactor",
  "provider": "grok",
  "repo": "/absolute/path/to/git-worktree",
  "role": "review",
  "workspace_mode": "managed",
  "prompt": "Review the current branch and report concrete correctness risks.",
  "expected_duration_ms": 600000
}

See CHANGELOG.md for the complete change history, the
efficient dogfood workflow for coordination
examples, and the README for provider setup and safety
boundaries.

This file is the publication source for the GitHub Release. Publishing it is a
separate explicit action against an exact reviewed main SHA. Capture the
exact-tree gate receipt first (not run by documentation-only work):

release-gate run --repo "$PWD" \
  --receipt /tmp/codex-co-engineer-v3.2.0-release-gate.json

Suggested later publication command, with the reviewed SHA substituted for the
placeholder (not run by this change):

gh release create v3.2.0 \
  --title "Codex-Co-Engineer 3.2.0" \
  --notes-file docs/releases/v3.2.0.md \
  --target EXACT_REVIEWED_MAIN_SHA

Codex-Co-Engineer 3.1.1

Codex-Co-Engineer 3.1.1

Repository-side GitHub Release body for Codex-Co-Engineer 3.1.1.

What changed

3.1.1 fixes an ambiguity in the delegation instructions. The MCP tool has
always required the absolute Git worktree path in the property named repo,
but the 3.1.0 control skill described that value only as an "absolute Git
root." A caller could reasonably translate that prose into git_root, which
the strict schema rejects before a task receipt exists.

The public contract now shows the literal argument everywhere it matters:

{
  "task_id": "review-auth-refactor",
  "provider": "grok",
  "repo": "/absolute/path/to/git-worktree",
  "role": "review",
  "workspace_mode": "managed",
  "prompt": "Review the current branch and report concrete correctness risks.",
  "expected_duration_ms": 600000
}

Use the property name repo exactly. git_root, repository, and other
aliases are not accepted.

Cursor Cloud also requires repo, identifying the clean local checkout whose
provider-accessible origin is used. Its already-pushed immutable commit SHA is
a separate, Cursor Cloud-only starting_ref property:

{
  "task_id": "cloud-auth-refactor",
  "provider": "cursor-cloud",
  "repo": "/absolute/path/to/clean-checkout",
  "starting_ref": "0123456789abcdef0123456789abcdef01234567",
  "role": "implement",
  "prompt": "Implement the requested change, run tests, and commit the result.",
  "expected_duration_ms": 3600000,
  "create_pr": true
}

Regression coverage now pins this language in the control skill and MCP tool
schema so the call-shape mismatch cannot recur.

Install

Requires Node.js 24+, Git, Codex CLI, and worktree-bootstrap for managed
local tasks.

git clone https://github.com/ajhcs/Codex-Co-Engineer.git
cd Codex-Co-Engineer
codex plugin marketplace add "$PWD"
codex plugin add codex-co-engineer@codex-co-engineer
npm --prefix plugins/codex-co-engineer run setup
npm --prefix plugins/codex-co-engineer run setup:check

Start a new Codex session after installing or upgrading the plugin.

See CHANGELOG.md for the complete change history and the
README for provider setup, safety boundaries, and usage.

This file is the publication source for the GitHub Release. Publishing it is a
separate explicit action:

gh release create v3.1.1 --title "Codex-Co-Engineer 3.1.1" --notes-file docs/releases/v3.1.1.md