Execution Log: the contribution loop — joining existing work items

Append-only log of progress. Checked in at docs/specs/issue-185/execution-log.md.

Phase transitions

Phase	Entered	Reviewed/approved by	Notes
phase-selection	2026-08-09	— (see Blockers)	run in-session; the full chain was walked, no phase declared away — this change adds a control keyword (trust boundary) and touches the cli-config schema
requirements-definition	2026-08-09	pending (PR)	five requirements: the third graph, no-goal-no-start, trigger parity, intervention-sized phases, done-means-criteria-met. Risk tier 4
design	2026-08-09	pending (PR)	one graph + the smallest selection seams; auto-detection explicitly rejected — joining is a declared decision (decision-070)
test-planning	2026-08-09	pending (PR)	reviewed with the design, one gate for both (decision-060 D2)
tasks-breakdown	2026-08-09	pending (PR)	9 tasks; T1 → {T2,T3,T4} → T5 → T6 → {T7,T8} → T9
implementation	2026-08-09	—	T1–T8
verification	2026-08-09	—	T9; every planned activity ran, red recorded before green
needs-review	2026-08-09	pending	3 self-review rounds below; critic rounds unavailable (no critic configured in reviews.critics)

Pull requests

PR	Scope / tasks	Status
(opened from branch claude/github-issue-185-9utd4r)	the whole work item — T1–T9, in MadaraUchiha-314/the-loop (origin and only contributing repository)	open

Progress entries

2026-08-09 — spec chain, implementation and verification

• Phase: requirements-definition → verification
• Did: wrote the four-artifact chain; shipped pdlc-contribution-loop.yaml (14 nodes, 2 required gates, 2 skip sets), the goal gate (hooks/goal.py: post-goal-request + classify-goal), the contribute control keyword (arming + spawn-arming at both dispatcher seams, schema property), durable loop choice (GraphState.loop, build_runtime(loop=…), GraphLink state-first resolution, core/graphs.py verbs), the contribute-to plugin command, the contribution.md template, and the docs sweep (README, guides, SKILL, workflow reference, two capability docs, config reference, decision-070).
• Red → green: three reds during TDD, all fixed in code, never by weakening a test: the goal parser rejected **Goal:** (bold wrapping the colon); the thread path dropped authorized goals because the GitHub API returns bare logins while allowlists conventionally carry @ (normalised on that path only); and the walk test exposed both at once. See evidence/tests.md.
• Next: reviewer briefing on the PR; human review.

Review cycles

Round	Reviewer	Findings	Outcome
self-1	the-loop (session)	stale "four commands" prose in control.py and a dispatcher comment left wrong by the widened vocabulary — fixed; confirmed the poller's parse_command call needs no change	fixed in-place
self-2	the-loop (session)	a session spawned for a contribution item was still steered to /the-loop:work-on by the spawn template; gave the two human gates command: contribute-to so the rendered graph context names the right command — and confirmed graph complete routes through advance, so goal freezing works on the claim path too	fixed in-place
self-3	the-loop (session)	history-table row initially landed above the separator row in process-graph.md (broken table) — fixed; pyright signature drift in two test fakes injecting _build_runtime — fixed; checked checkmarks: complete cannot trip on the template's comments	fixed in-place
critic-1..3	—	not run: reviews.critics is empty in this repository's config (no critic harness configured); recorded per the escalation rule rather than silently skipped	n/a
human-1	@MadaraUchiha-314 (PR #187)	the target repo may not have the-loop initialized at all — no harness-config.yaml, no specDir to depend on; writing contribution.md into it would make the PR unclean, so "the contribution can just be commented on the PR"	fixed: R6 added; reads already degraded to defaults (decision-044) and now the distinction is recorded (repoInitialized); the spec tree is excluded from git at Runtime.start (structural — info/exclude, worktree-aware), and a new best-effort publish-artifact hook posts the plan / verification results to the thread at plan-approval / human-approval, skipping in adopted repos; 7 tests added (40 total), decision-070 amended

Security review (gate)

Per reference/security.md and the requirements' threat model:

• New attack surface: one word in an existing fixed vocabulary (contribute), one new comment-reading gate (classify-goal). Both sit strictly behind the existing guards: self-authored-marker drop, then named-actor authorizedUsers check; ambiguity (two different keywords) still refuses outright. Verified by tests (test_an_unauthorized_goal_is_not_read, test_a_self_authored_goal_is_not_read, test_contribute_plus_another_command_is_refused).
• Parsed text is data, never a destination: the goal reaches graph state, a confirmation comment and an artifact section; no argv, path, label or routing target derives from it. Routing stays with declared edges.
• Agent-writable state cannot choose arbitrary graphs: GraphState.loop is honoured only for shipped names, at all three read sites (bootstrap, GraphLink, core verbs) — each fail-closed to the default and covered by a test.
• Fail-closed defaults hold: no goal → wait forever; empty authorizedUsers → no goal accepted; integration outage → wait, never a guess.
• Schema change (keywords.contribute) matches sensitivePaths → risk tier 4, human PR approval required before completion. Gate: pass, with the human sign-off pending on the PR (tier 4 ≥ humanSignOffMinTier).

Final validation evidence

All planned verification activities ran; results in testing-plan.md § Verification results and raw output in evidence/tests.md: new suite 40/40 (33 + 7 from review round human-1), full suite 1565 passed / 1 skipped (pre-existing), ruff clean, pyright 0 errors, markdownlint 0 issues across the changed/added markdown files.

Capability docs

• process-graph — the third loop's behaviour block (goal gate, one artifact, durable loop choice) + history row.
• webhook-triggers — the widened control vocabulary + history row.

Documentation

User-facing surfaces updated in this PR, per the issue-174 gate:

• README.md and docs/guide/what-is-the-loop.md — "Two loops" became three, with the contribution loop described.
• docs/guide/how-it-works.md — the shipped-graph list and repo layout.
• skills/the-loop/SKILL.md — the graph paragraph and the commands list (/the-loop:contribute-to).
• skills/the-loop/reference/workflow.md — new § The contribution loop.
• docs/config/cli/routing-options.md — control.keywords.contribute reference.
• New: commands/contribute-to.md, skills/the-loop/templates/contribution.md, docs/decisions/decision-070.md (+ index row).