Skip to content

docs(dispatcher-skill): Stage 0 R-phase — loop design + primitive-reality + BFR=BUILD - #402

Merged
artyhoo merged 2 commits into
stagingfrom
research/dispatcher-skill-rphase
Jun 3, 2026
Merged

docs(dispatcher-skill): Stage 0 R-phase — loop design + primitive-reality + BFR=BUILD#402
artyhoo merged 2 commits into
stagingfrom
research/dispatcher-skill-rphase

Conversation

@artyhoo

@artyhoo artyhoo commented Jun 3, 2026

Copy link
Copy Markdown
Owner

Stage 0 — R-phase (dispatcher-skill umbrella)

Designs the aif-control loop /dispatcher will run (dispatch → monitor → Q&A → harvest → Phase-1 review → stage-gate → advance) and verifies the 4 CLI primitives' runtime reality. No skill code — that is Stage 1.

Deliverable: docs/meta-factory/dispatcher-skill-rphase.md (355 lines) — primitive-reality table, loop design, BFR verdict, DN-A/DN-B recommendations, T15 self-application.

Headline verdict: BFR = BUILD (loop-wiring layer). DeepWiki ≥3 phrasings + WebSearch ≥3 phrasings + SSOT layer-5 consult (rows #27/#28/#64/#65/#67/#83/#88/#109) — none covers the cross-boundary dispatch→harvest→advance loop. New SSOT #111 deferred to the Stage 1 capability commit (per CLAUDE.md: new SSOT entry in the same commit as the capability artifact). Because verdict ≠ ADOPT/REFERENCE, the meta-launch §8 "upstream exists → halt Stage 1" stop condition does not fire.

Review trail: Phase -1 cold-review of the dispatch prompt caught 3 BLOCKER + 4 MAJOR (board-contamination risk) before dispatch; an independent Phase-1 cold-review of the deliverable caught 1 BLOCKER (missing SSOT consult + BUILD/ADAPT contradiction) → fixed in f31b2fa.

§1.7 Forward-check applied

Checked the existing disciplines that govern a build-vs-reuse design decision before settling the loop verdict and the delivery-channel call: BFR (verdict + SSOT layer-5 consult), dual-implementation (CC-skill vs portable for DN-A), and the no-paid-LLM policy (all primitives are session-bound).

  • build-first-reuse-default.md §3 — BFR verdict = BUILD; all 6 mechanism layers run incl. SSOT consult; no upstream covers loop-wiring. file:line: docs/meta-factory/dispatcher-skill-rphase.md:223
  • dual-implementation-discipline.md §3/§6 — DN-A = CC-skill, @cc-only-rationale required (brainstorm not meaningful outside CC). file:line: docs/meta-factory/dispatcher-skill-rphase.md:311
  • no-paid-llm-in-ci.md §1 — every primitive is session-bound; harvest.ts does zero LLM. file:line: packages/runtime-bridge/src/cli/harvest.ts:18

§1.7 Backward-check applied

Swept the artefacts under this umbrella's scope to confirm the R-phase supersedes nothing and only wires existing primitives: the complementary plan-skill, the 4 CLI primitives it wires, and the SSOT register it consulted.

  • .claude/skills/pipeline/SKILL.md/dispatcher executes, /pipeline plans; complementary, no supersession. file:line: .claude/skills/pipeline/SKILL.md:313
  • packages/runtime-bridge/src/cli/dispatch.ts — wired as the dispatch primitive; state machine verified in section (a). file:line: packages/runtime-bridge/src/cli/dispatch.ts:50
  • docs/meta-factory/prior-art-evaluations.md — SSOT layer-5 consult; conclusion = no row covers loop-wiring. file:line: docs/meta-factory/dispatcher-skill-rphase.md:249

🟡 Maintainer decisions to confirm (surfaced, not decided — reviewer-discipline §2)

  • DN-A (delivery channel): R-phase recommends CC-skill only (brainstorm integration is the core value-add; portable fallback would be degraded). Wrong if non-CC consumers are an imminent use-case.
  • DN-B (technical-fork autonomy boundary): R-phase recommends technical-fork → auto-brainstorm-and-answer; strategic-fork → park-and-surface. 3 options laid out; you pin the boundary.

Scope-out

  • No skill code, no new CLI primitives, no npm deps, no ~/.claude/skills/** edits (anti-scope §9, verified).

artyhoo added 2 commits June 3, 2026 18:14
…lity + BFR verdict

Prior-art: prior-art-evaluations.md#27 (HANDOFF_MODE, verdict DEFER — aif-internal mode
suppression, different problem class from operator-side control loop).
Prior-art: prior-art-evaluations.md#28 (paused:true/false, verdict DEFER — consumed as
dependency by answer.ts --decision resume, not rebuilt).
Prior-art: prior-art-evaluations.md#64 (SDD skill, verdict ADOPT — inner loop vocabulary
only; cross-boundary egress not covered, different problem class).
Prior-art: prior-art-evaluations.md#67 (aif Kanban runtime, verdict REJECT — infra mismatch
on 3 axes; autoQueueMode vocabulary sub-component separately SSOT #88).
Prior-art: prior-art-evaluations.md#88 (autoQueueMode LOGIC, verdict ADOPT VOCABULARY —
covers umbrella sequencing vocabulary, not operator execution loop).
Prior-art: prior-art-evaluations.md#109 (park primitive, verdict BUILD — already built;
consumed by /dispatcher loop, not rebuilt).
New SSOT entry #111 required in Stage 1 capability commit: /dispatcher loop-wiring BUILD
verdict — no upstream covers cross-boundary dispatch→monitor→Q&A→harvest→stage-gate→advance.
Corrects the R-phase's Option-A (CC-only) lean. Skills ship to consumers
via install.sh:207, so /dispatcher is consumer-facing → dual-implementation
§3 default dual + BFR §1.1 shipped-axis agnostic invariant. CC-native primary
+ portable fallback that degrades (technical forks surface when no companion).
Maintainer decision 2026-06-03.

Prior-art: skipped — doc edit recording a maintainer decision, no new capability.
@artyhoo
artyhoo merged commit a8e8050 into staging Jun 3, 2026
22 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant