Agent Collab Skills
繁體中文 · Public harness contract · Native lanes · 0.6 migration · 0.5 goal-slice migration · 0.4 migration
Provider-neutral collaboration skills plus an optional, standard-library-first
Python harness for bounded, resumable, human-authorized agent work.
This project is a governance layer, not a general agent runtime. A host supplies
models, tools, sessions, sandboxes, and tracing. Agent Collab supplies portable
roles, policy evaluation, checkpoints, proposal-only memory, reconciliation,
and acceptance evidence.
Architecture
flowchart LR
H[Human decision owner] --> P[Primary agent]
P --> S[Task splitter]
S --> E[Delegated executor]
S --> R[Reviewer]
E --> C[Task checkpoint]
R --> C
C --> Y[Canonical policy evaluation]
Y -->|continue| S
Y -->|checkpoint| P
Y -->|stop| Q{Stop scope}
Q -->|action| X[Diagnose or wait]
Q -->|goal| B[Explicit blocker]
E --> O[Acceptance gate: reconcile]
R --> O
O --> A[Acceptance evidence]
A --> H
H -->|approve / decline / revise| D[Append-only human decision]
The deterministic layer controls state and limits. Agents produce prose,
implementation, review, and synthesis. A human remains responsible for
breaking contracts, memory promotion, semantic acceptance, and shipping.
Skills
Each skill fires on a situation, not only on a request:
| Skill | Use when | Responsibility |
|---|
agent-task-splitter | about to run two or more subagents, workflow lanes, or delegates on one goal | Fan-out planner: one lane table with role, model tier, disjoint write scope, bounded brief, and capped return per lane, plus the merge and verification step fixed before launch. |
agent-context-budget | launching parallel or background lanes, or a result or handoff is larger than a short summary | Cap returns (summary inline, detail by path), choose cheap tiers for mechanical lanes, write primers, and record durable decisions as proposals. |
agent-debate | a consequential decision has two plausible options | Two opposed reviewers on the same evidence with a round cap stated before launch; the synthesis goes to the human. Never decides truth by vote. |
agent-acceptance-gate | about to commit or push merged lane output, locale variants, catalog entries, or claims about named frontier models (benchmarks, release dates, licenses) | Reconcile results (failures preserved), run the triggered preset, and return immutable PASS, CONDITIONAL PASS, or FAIL evidence. |
Redirect stubs, removed in 0.7.0 (migration 0.6):
agent-output-reconciler → agent-acceptance-gate;
agent-shared-memory → agent-context-budget;
agent-plan-act-reflect → deprecated, its harness loop now lives in
agent-context-budget.
Native mode and harness mode
Native mode is the default. When every lane runs inside one host session with
the user present, the lane table and briefs live in the conversation, the
host's own limits and permissions apply, and no plan, packet, or checkpoint is
written to .coord/ (only a memory proposal may be).
Native lanes maps the roles to Claude Code's Agent,
SendMessage, Workflow, and worktree mechanisms.
Harness mode adds the policy and checkpoint contract below when a round spans
sessions or tools, runs unattended, must resume after an interruption, needs a
signed human gate, or AGENT_COLLAB_POLICY is configured. A configured policy
always applies and fails closed when unreadable.
Public harness package
Distribution: agent-collab-harness
Import: agent_collab_harness
Command: agent-collab
python -m pip install agent-collab-harness
agent-collab policy validate --policy "$AGENT_COLLAB_POLICY" --json
agent-collab checkpoint validate --checkpoint .coord/checkpoint.json --json
agent-collab policy evaluate \
--policy "$AGENT_COLLAB_POLICY" \
--checkpoint .coord/checkpoint.json \
--json
agent-collab checkpoint migrate \
--checkpoint .coord/checkpoint-v1.json \
--policy .coord/policy-v2.json \
--metadata .coord/migration-request.json \
--output .coord/checkpoint-v2.json
agent-collab checkpoint advance \
--checkpoint .coord/checkpoint-v2.json \
--policy .coord/policy-v2.json \
--request-id stable-transition-id \
--expected-sha256 RAW_FILE_SHA256
agent-collab doctor --json
Version 0.5 adds opt-in v2 goal slices; installing it does not reinterpret v1
state. checkpoint migrate writes a separate v2 checkpoint and leaves the v1
source untouched. checkpoint advance records an eligible slice transition;
stable request IDs are idempotent, while stale file hashes and lock contention
fail closed. See the 0.5 migration guide.
Runtime dependencies are standard-library only. Policy files use strict JSON;
a .yaml suffix is accepted only when the file content is valid JSON. An
unreadable configured policy fails closed.
AGENT_COLLAB_POLICY selects the one machine-readable budget source for the
run. The Codex portable adapter points it to
${CODEX_HOME}/portable-harness/policies/agent-budget.yaml; that v1 shape and
the public v1 JSON shape are both accepted. Human decisions and overrides are
action-hash bound and require an HMAC key supplied by the trusted host through
AGENT_COLLAB_HUMAN_KEYS_JSON. The matching key digest must already be pinned
in the canonical policy's human_authorization.key_hashes; caller-created keys
or checkpoint text alone cannot grant authority. The legacy portable v1 shape
normalizes to an empty trust root, so it can evaluate budgets but human records
fail closed until the canonical policy is explicitly upgraded.
Role contract
Plans use four public roles:
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
