Execution Log: the graph runs the PDLC

Append-only log of progress. The-loop keeps the work item's loop:<phase> label in sync with the phase front-matter above.

Phase transitions

Phase	Entered	Reviewed/approved by	Notes
requirements-definition	2026-08-04	MadaraUchiha-314 ("go ahead", PR #149 comment)	Tier 5 (human-approves-spec-and-pr): each phase artifact needs explicit human approval. Requirements drafted from the issue's audit of the ingress↔graph seam.
design	2026-08-04	MadaraUchiha-314 ("go ahead with implementation", PR #149 comment)	Seven decisions (D1–D7); both open questions from requirements resolved: completion signal is a CLI verb (D1), no consume-only routes (D4).
tasks-breakdown	2026-08-04	MadaraUchiha-314 (same comment — implementation explicitly green-lit)	9 tasks. Owner directives recorded: no backwards-compatibility shims; remove unused baggage in touched code.
implementation	2026-08-04	—	Executing the DAG.
needs-review	2026-08-04	pending	All 9 tasks complete; suite 1168 passed / 2 skipped; ruff, pyright, markdownlint clean. Tier 5: named human security sign-off requested with the PR review.
complete

Pull requests

PR	Scope / tasks	Status
#149	Phase-1 requirements artifact	open, in review

Progress entries

2026-08-04 — requirements drafted

• Phase: requirements-definition
• Did: Audited the ingress↔graph seam (graphlink.py, dispatcher.py:1089-1143, poller.py, runtime.py) and confirmed the issue's three claims: the agent walks the PDLC from prose while the graph records node one; events are delivered before the graph is consulted; Runtime.resolve_session() has zero callers. Drafted requirements.md around the missing primitive — a node-completion signal (R1) — plus pointer-as-authority (R2), consult-before-deliver (R3), gate-first event routing (R4), session: inherit honoured (R5), one source of truth for the process (R6), and the safety invariants that must survive the inversion (R7). Risk tier 5.
• Checkpoint/tests: none yet (no code). Next: human review of requirements on the PR; do not derive design.md until locked.

2026-08-04 — artifact finalized for review; gate green

• Phase: requirements-definition
• Did: The CI gate (the-loop check --recompute --fail-on block) blocked on status: draft — by design: the requirements node's exit chain requires the lock, and the human gate is the next node, exercised as the PR review (precedent: issue-142). Finalized the artifact (status: approved, approvedBy deferred to the PR review) and re-ran the gate locally: the item now parks at requirements-approval waiting on a human, which --fail-on block treats as the normal state of an open PR. Tier 5 still means a human must approve this phase on PR #149 before design begins.
• Checkpoint/tests: uv run the-loop check issue-148 --recompute --fail-on block green locally; markdownlint clean.

2026-08-04 — requirements approved; design derived

• Phase: design
• Did: The owner approved the requirements on PR #149 ("go ahead") — approvedBy recorded, phase advanced, label moved to loop:design. Derived design.md from the locked requirements: D1 completion signal as the-loop graph complete (JSON envelope, node-named idempotent claims), D2 read-only GraphContext, D3 one $graph_context template variable (empty ⇒ byte-identical prompts for out-of-graph repos), D4 consult-first only at human gates (open question 2 answered: no consume-only routes), D5 read-before-render / enter-after-success spawn, D6 session binding in graph state + resolve_session's first caller, D7 phase parity test. Two-writer concurrency handled with a stdlib flock; error handling enumerated per call site.
• Checkpoint/tests: the-loop check issue-148 --recompute --fail-on block and markdownlint to be re-run before push. Next: human review of design on the PR; do not derive tasks.md until locked.

2026-08-04 — design + tasks approved; implementation complete

• Phase: needs-review
• Did: Owner approved design and green-lit implementation on PR #149 ("go ahead with implementation. No need to think of backwards compatibility… remove any extra baggage from the past that's not used"). Derived tasks.md (9 tasks) and executed the DAG:
  • T2/D1 the-loop graph complete — claims ledger (completions), Runtime.complete (already-past no-op, wrong-node refusal, not-started refusal, busy), one JSON envelope, exit 0 for results.
  • T1/D2 GraphContext + GraphLink.context() — read-only, behind the full _guarded gate order, None on every skip/fault.
  • T3/D3 $graph_context in both prompt templates and both built-in constants; the spawn template's hard-coded phase-flow line removed (R6.3).
  • T4/D4 consult-first at human gates in the dispatcher; on_event now returns the NodeReport; advance_after threaded through the respawn/occupant paths so a gate-consumed event is never advanced twice.
  • T5/D5 spawn resolves context after workspace prep, before render; on_spawn (the write) still last, on success only.
  • T6/D6 session binding recorded on spawn/respawn, flipped dead on close (on_close), resolve_session called at human-gate entry and recorded as graph.gate_session.
  • T7/D7 parity P4: pdlc.yaml phase order enforced against both harness configs; SKILL.md + reference/workflow.md now defer to the graph.
  • T8 capability docs updated (process-graph, webhook-triggers, CLI graph page + history row). Baggage found and removed in passing: tmux spawns never called on_spawn — a whole runner whose items never entered the graph; fixed with the binding change.
• Deviations from design, with reasons:
  • The two-writer lock is taken by _guarded for the daemon's write actions (start/advance/close) rather than inside each runtime method — flock is not reentrant across file handles, so the lock lives at the outermost writer on each path (Runtime.complete for claims). context reads unlocked: a stale read costs a stale prompt, never a wrong pointer.
  • graph-state.lock added to .gitignore — coordination, not record.
• Self-review: 3 rounds (correctness / security / minimalism-docs). Round 1 found the design-vs-implementation lock gap above — fixed. Round 2 re-checked the trust boundaries: no payload text reaches GraphContext or the rendered block; the verdict is an outcome token, not comment text; the ownership-proof-before-checkout-read order is intact on the new context and on_close paths. Round 3: stdlib only, no new dependencies, docs and CLI reference updated. reviews.critics is empty in config, so no critic harness was run (noted per config).
• Checkpoint/tests: full suite 1168 passed, 2 skipped; ruff + format clean; pyright 0 errors; markdownlint 0 errors; the-loop check issue-148 --recompute --fail-on block exit 0 (parked at requirements-approval — the shipped graph has no pointer for this item on main yet; the phase front-matter above is the record). Evidence: the new suites test_graph_drive.py (19) and test_graph_drive_integration.py (8) map one-to-one onto R1–R7; P4 covers R6.
• Next: reviewer briefing on the PR, then human review + tier-5 security sign-off.