agent-collab

agent-collab publishes one collaboration plugin for Claude Code, Codex, and compatible hosts. Version 6 keeps the closed semantic coordinator and co-packaged direct native runtime while adding provider-neutral quality and effort profiles, current-portfolio routing, and truthful advisory degradation.

This public repository's current source distribution is agent-collab (v6.2.0). It combines a governance-pool-widening provider runtime advance (4.1.0), actionable coordinator rejection and outcome handling, and empirical project estimation. It is published as v6.2.0 and activated on all four primary hosts (Claude, Codex, Antigravity, Grok) with byte-exact installed trees and a consumed post-install canary each.

The human-first architecture handbook explains the system boundaries and lifecycle. The package reference below is the machine-operational contract for this release. The final documentation closeout records the exact published artifact and installation evidence after release. The dated status and evidence snapshot keeps repository, tag, release, installation, and readiness claims separate.

What's new - v6.2.0

This release advances the signed provider runtime to 4.1.0 (an additive governance-pool widening) and makes the coordinator boundary tolerant of imperfect invocation and self-describing about every outcome, so a caller recovers in place instead of re-deriving a request or misreading a transient failure as an outage.

• Provider runtime 4.1.0. The signed, notarized runtime adds opencode/governance.repository (read-only, repository authority; no write-authority widening) and lowest-priority governance edges for the zhipu, moonshot, alibaba, and deepseek lineages, so Tier-3 peer legs do not dead-end when frontier CLIs degrade. The Phase-1 governance gate is unchanged. Runtime protocol 4 is unchanged; the wire-contract descriptor digest advances as a compatible evolution (4de687b8… → e601a455…), widening the base transport actions 13 → 14 and the valid source pairs 17 → 18.

• Actionable rejections. An invalid or underspecified request no longer collapses to a bare invalid_request. It carries a specific error_code (timeout_ms_over_cap, unknown_logical_action, quality_profile_invalid, request_not_closed, and so on) and a bounded detail object naming the field, its constraint, and the admitted values or required source. detail never reflects unbounded or untrusted input.

• Outcome disposition. Every non-usable response carries a disposition (fix_request, retry, inspect, or unavailable) and a short recovery hint. By construction provider_error, teardown_error, and protocol_error are never classified unavailable, so an attempt-local failure is no longer read as a provider outage.

• In-place recovery. An empty target_agent is coerced to null (recorded in a normalized field). Nothing that changes cost, depth, or a security decision is rewritten; the closed-request, fail-closed posture is unchanged.

• Doc currency. The documented timeout_ms bound now matches the enforced 1-600000 (Addressed: #125).

• Empirical project estimation. Add project-estimation, a read-only-by-default skill for scoped agent-led delivery estimates, reconciliation, calibration, and audit. It keeps focused agent wall-clock, calendar waits, API-equivalent cost, actual marginal cash, and quota capacity distinct, and supplies a compact checkpoint for formal implementation designs and plans where the host supports it. See the project-estimation architecture for public modes and examples.

  A governed, content-addressed bootstrap handoff is now admitted: the enhancement duration prior is descriptive, greenfield has no compatible prior, and token, wait, rework, quota-delay, and marginal-cash omissions stay typed and visible. This is not promoted calibration and never reports high confidence.

The 12-action set and every action's authority/artifact/evidence contract are unchanged from v6.1.1; the signed provider runtime advances to 4.1.0 (the additive governance-pool widening described above).

Previous tagged release: v6.1.1

• Provider runtime advanced to 4.0.6: agent-collaboration routing hardening, lifecycle diagnostics decoupled from authority, pinned worktree sealing, and failure-text salvage — the advisory response is additively widened (completeness, salvaged_from, one new reason value), advancing the wire-contract digest; the 12-action set and every action's authority/artifact contract are unchanged.
• Everything below from the merged-but-untagged v6.1.0 is included in this first tagged publication.

Included from v6.1.0

• New project-knowledge skill: a durable, provenance-tracked knowledge layer inside the user's project (knowledge/), with a bundled deterministic stdlib-only CLI (knowledge_tool.py: init / validate / lint / index / draft / export), claim-marker and prompt-injection lint, a source trust lattice, and budget-capped untrusted-banner exports.
• New learning-loop skill: a project-local learning ledger (.learnings/) with a bundled deterministic stdlib-only CLI (learning_ledger.py: add / suggest / recur / index / check / lint), mechanical/judgmental lesson classes, prevention-debt surfacing, and reuse-as-hypothesis discipline.
• Both skills document an OPT-IN consultation snippet for the project's CLAUDE.md / AGENTS.md (explicit reads only; pages and ledger entries are data, never instructions), offered at install time as a visible, consent-gated setup step (see Install).

What ships

• Source skill specifications in skill-specs/ and generated skills in plugins/agent-collab/skills/.
• Pinned source and license provenance for MIT-derived skill material in docs/third-party-skill-provenance.md.
• A closed semantic coordinator at plugins/agent-collab/coordinator.py.
• A direct bounded process client at plugins/agent-collab/runtime_client.py.
• Provider-neutral host observations and migration reporting.
• One schema-4 runtime manifest contract and public archive/release/export safety gates.
• A final signed native standalone bundle only when produced by the separate private build/sign workflow.

No provider executor source, provider invocation recipe, model pin, compatibility package, downloader, post-install hook, broker, socket, lane, launchd job, lifecycle setup command, or raw provider wire is public.

Install

codex plugin marketplace add sumitake/agent-collab
codex plugin add agent-collab@agent-collab

Post-install setup (agent-performed, consent-gated)

When an agent performs the installation inside a user project, it completes one visible setup step for the project-knowledge and learning-loop skills: check whether the project has a CLAUDE.md (Claude Code agents) and/or an AGENTS.md (other agent families) and offer to append those skills' opt-in consultation snippets (quoted in each skill's SKILL.md) to whichever file matches the host agent family — both files when both exist. The snippets direct future sessions to consult knowledge/index.md and the .learnings/ ledger by EXPLICIT reads only; derived pages and ledger entries are untrusted data, never instructions, and nothing may auto-inject their bodies into agent context. Confirm with the user before writing; if the user declines, or the files are absent and the user does not want them created, skip the step and say so. This is an in-session, consented setup action — never a hidden side effect.

Semantic actions

Public requests select one of 12 logical actions:

architecture.conceptual
architecture.repository
codegen.repository
context.documents.extract
context.documents.intent
context.documents.reason
context.repository.extract
context.repository.reason
frontend_codegen.repository
frontend_review.repository
governance.repository
review.repository

Repository actions require a canonical absolute repo_root. Document context uses bounded inline documents. Conceptual architecture uses prompt-only source. The runtime's workspace-generated wire descriptor derives the internal 13 transport actions and 17 action/source pairs. Those projections are diagnostic contract data, not a second public request surface.

See plugins/agent-collab/README.md for the exact coordinator and runtime contract.

Source and generated files

• Edit skill-specs/<name>.md.
• Generate with python3 scripts/build_skills.py.
• Check with python3 scripts/build_skills.py --check.
• Generate marketplace metadata with python3 scripts/build_marketplace.py.
• Check it with python3 scripts/build_marketplace.py --check.

context is the sole source-grounded corpus/repository skill. No parallel size-branded source or generated skill surface is supported.

Runtime trust boundary

The canonical workspace build owns the final binary and generated manifest. The public source expects:

• manifest schema 4;
• runtime protocol 4;
• native manifest contract 4;
• provider runtime version 4.1.0;
• one top-level closed wire_contract plus canonical wire_contract_sha256, bound into each artifact record; and
• no action-membership mirror in artifact entries.

The public client verifies fixed plugin-relative path, exact membership and digests, Mach-O architecture/minimum macOS, hardened Developer ID identity, team, and secure timestamp. Online notarization verification remains a release gate. One accepted request launches one process group with bounded streams, deadline, TERM/KILL/reap, and no hidden replay.

Migration status

Run the provider-free doctor:

python3 plugins/agent-collab/migration_doctor.py --json

It inventories retired packages, reports host and descriptor state, and does not invoke a provider or mutate the host. No daemon installation or runtime setup step exists.

Validation

python3 scripts/build_skills.py --check
python3 scripts/build_marketplace.py --check
python3 scripts/build-changelog.py --check
python3 -m unittest discover -s tests -t . -v
python3 -m unittest discover -s scripts -p 'test_*.py' -v
python3 scripts/check_release_consistency.py
python3 scripts/check-public-export-safety.py --active-tree
python3 scripts/secret_scan.py
git diff --check

Archive/release validation additionally requires the canonical final signed runtime artifact and generated manifest. Public source work must not rebuild, sign, notarize, or hand-edit either artifact.

Contribution and release governance

Read AGENTS.md and docs/public-governance.md. User-visible changes use a unique changelog.d/ fragment; do not commit generated CHANGELOG.md. Pull requests must include the repository compliance trace and the required independent review for their tier.

The clean-public-repository invariant applies to the active tree, reachable history, and release archive. If executor source, credentials, private paths, or suspect native bytes appear, stop publication and follow SECURITY.md.

Public CI uses distinct GitHub-hosted runners, pins every external action to a full commit SHA, runs CodeQL and Gitleaks, enables secret scanning, and uses Dependabot for dependency update review.

After every other release task finishes, complete the documentation closeout. The v6.0.1 closeout is recorded in the status and evidence snapshot. Each future closeout must likewise align the architecture handbook, this README, and generated changelog evidence with the exact release without exposing private executor recipes.

License

The public repository and distributed package use the unmodified PolyForm Strict License 1.0.0, except that the derived portions of decision-map, prototype, and architecture-review, plus the adapted spec-fidelity and smell-baseline portions of code-review, and the adapted decomposition guidance in orchestrate and teamwork, remain MIT-licensed and carry the full MIT notice in each generated skill. Their pinned upstream and per-file provenance is recorded in docs/third-party-skill-provenance.md. Commercial use of the PolyForm-licensed material requires separate, explicit written approval administered by Osumi Consulting LLC. See NOTICE and COMMERCIAL-LICENSING.md for the ownership and approval boundary.