Tasks: the graph runs the PDLC

Phase 3 of 3. DAG of small, verifiable tasks; each references the design decision and requirements it delivers. Owner directives recorded with the approval: no backwards-compatibility shims (no shadow reads, no dual formats), and remove unused baggage encountered in the touched code.

DAG

graph TD
    T1[T1 GraphContext + context] --> T3[T3 prompt integration]
    T1 --> T4[T4 consult-first dispatch]
    T1 --> T5[T5 spawn read-before-render]
    T2[T2 graph complete verb + claims] --> T4
    T1 --> T6[T6 session binding + resolve_session]
    T7[T7 phase parity + prose]
    T3 --> T8[T8 capability docs + cleanup]
    T4 --> T8
    T5 --> T8
    T6 --> T8
    T2 --> T8
    T7 --> T8
    T8 --> T9[T9 full suite + evidence + reviews]

Tasks

• [x] T1 — GraphContext and GraphLink.context() (D2; R3.1, R3.3, R3.4) Frozen dataclass in graphlink.py; read-only resolve behind _guarded's gate order; every skip → None; never raises. Unit tests: each skip path, field extraction (node/phase/status/reason/messages/next_command/actor).
• [x] T2 — the-loop graph complete + completion claims (D1; R1.1–R1.6, abuse cases 2–3)completions ledger + advisory flock in graph/state.py; Runtime.complete(item, node) wrapping advance with claim semantics (already-past no-op, wrong-node refusal); CLI verb with one JSON envelope, exit 0 for refusals/blocks. Tests: idempotent replay, wrong-node, block leaves pointer, lock busy path.
• [x] T3 — $graph_context in both prompts (D3; R3.1, R3.2, R3.4, R6.3)_render_prompt gains the variable; template files updated; the spawn template's hard-coded phase-flow line removed (owner: no compat baggage). Test: absent context ⇒ byte-identical prompt modulo empty substitution.
• [x] T4 — consult-first at human gates (D4; R4.1–R4.4)on_event returns Optional[NodeReport]; dispatcher: resolve → (gate? advance first, re-resolve) → deliver → (else advance after). Tests: both orderings; gate fault still delivers; unauthorized comment neither resolves nor is lost.
• [x] T5 — spawn reads context before render (D5; R3.2, R7.2) Fresh item ⇒ start prompt; mid-graph respawn ⇒ resume-at-node directive from next_command. Orderings asserted: start recorded before spawn; on_spawn after success only.
• [x] T6 — session binding + resolve_session's caller (D6; R5.1–R5.3)on_spawn writes state.session; new best-effort on_close flips alive; gate-node entry calls resolve_session and event-logs the resolution; registry wins on disagreement.
• [x] T7 — phase parity test + prose deferral (D7; R6.1, R6.2)test_graph_parity.py: pdlc phase order ⊂ workflow.phases (template + this repo), complement pinned. One authoritative sentence in SKILL.md + reference/workflow.md.
• [x] T8 — capability docs + baggage removal (SKILL rule: same-PR docs)docs/capabilities/process-graph.md behaviour + history row; webhook-triggers.md prompt-variable note. Remove dead code encountered in touched files. (Found and fixed in passing: tmux spawns never entered the graph — on_spawn was only on the process-runner tail.)
• [x] T9 — full suite, evidence, self/critic reviews (ready-to-ship gate)uv run pytest, ruff, pyright, markdownlint; evidence in execution log; 3 self-reviews (critic list is empty in config — noted); PR briefing before human review.