research(aif-handoff-bridge): Sub-wave A — Variant A (MCP-consumer) deep evaluation

[artyhoo]

Summary

R-phase Sub-wave A for umbrella aif-handoff-as-runtime-bridge. Evaluates Variant A: aif-handoff as MCP-consumer of meta-kickoffs (bridge script calls handoff_create_task → aif-handoff Planner/Implementer/Reviewer cycle → WebSocket tail → status write-back to state.md).

Verdict: REFERENCE (not ADOPT, not BUILD). Match score: 28%.

Body-format falsifier (kickoff line 76): CONDITIONAL PASS — handoff_create_task description field = z.string().optional() with no .max() (Zod schema) — our kickoff body fits. BUT accept_existing_plan bypass requires physical PLAN.md on disk in .ai-factory/ (not pure MCP), and aif-handoff Reviewer in autoMode conflicts with reviewer-discipline.md §2 strategy-fork-surface gate.

Key findings:

• MCP schema compatible: minimum call = {projectId, title}, description accepts arbitrary markdown
• accept_existing_plan provides Planner-bypass but requires filesystem coupling (disk-resident PLAN.md)
• WebSocket = broadcast channel (no topic filter — bridge must filter by taskId client-side)
• Docker-free deployment feasible (Node.js + SQLite minimum)
• Bridge LOC estimate: ~175 LOC, 3 MCP calls + 1 WebSocket (within targets)
• Admission gate §2.4 CLEAR: PR Wave 10.6 — port hook-stub-completeness audit to principle 16 #127/docs(automerge): codify branch-from-main staging flow + resync discipline #128 touch packages/runtime/ only — no packages/mcp/ or apps/coordinator/ changes
• SKILL.md:404 vs :429 inconsistency resolved: both consistent (BFR verdict = REFERENCE for SP skill; operational action = ADOPT template)

Evidence base: 7 DeepWiki probes + 2 WebSearch queries + SSOT rows #27/#28/#30/#43/#44/#46/#67/#80 consult + SKILL.md:404/429 + prior-art-evaluations.md:95-148 direct read.

Output

Single research-patch: docs/meta-factory/research-patches/2026-05-29-aif-handoff-bridge-variant-a-mcp.md (393 LOC). No code changes, no skill edits, no SSOT row landings — additive-note proposals only for Sub-wave D to land.

§1.7 Forward-check applied

• build-first-reuse-default.md §1 (verdict ladder): REFERENCE chosen per §1 — ADOPT blocked by PLAN.md disk coupling + Reviewer autoMode conflict. BFR §3 6-layer search: (a) SSOT consult prior-art-evaluations.md:95-148; (b) DeepWiki ≥5 probes (7 executed with URLs); (c) WebSearch ≥2 phrasings (both executed); (d) own-stack sweep SKILL.md:404 + SKILL.md:429.
• no-paid-llm-in-ci.md §1: all evidence via DeepWiki MCP + WebSearch + gh CLI + bash grep. Zero API-billed calls.
• reviewer-discipline.md §2: patch surfaces aif-handoff Reviewer conflict as a risk to maintainer; does NOT pick between our Phase -1 and aif-handoff Reviewer. That is maintainer's decision.
• phase-research-coverage.md §1.7: §7+§8 self-reflexive walk in the research-patch.
• ai-laziness-traps.md §3: T1 (7 probes ≥5), T3 (all findings have file:line or DeepWiki URL), T7 (adversarial counter-probes 6-7), T11 (SSOT consult + WebSearch), T12 (DeepWiki at R-phase time), T13 (SP requesting-code-review re-verified at research-patches/2026-05-23-meta-orchestrator-prior-art.md:73), T15 (§7-§8 self-application), T16 (§3 T16 table), T17 (markdown only), T19 (cold-QA noted), T20 (§5 verdict cites SSOT row + DeepWiki URL + falsifier outcome).
• Admission gate §2.4: gh api repos/lee-to/aif-handoff/commits?since=2026-04-29 run — commits = PR docs(automerge): codify branch-from-main staging flow + resync discipline #128 (51ce96a) touching packages/runtime/ only. Gate CLEAR.

§1.7 Backward-check applied

• SSOT rows docs(research-patches): Wave 6 AI-doc cold audit (§13.26) #27/docs(wave-6): close §13.26 — AI-doc effectiveness cold audit + review verdicts #28/docs(wave-5): readiness REVISE fixes — SSOT ID remap + §13.25 status #30/Wave 8.4: pre-push §7 Prior-art escape-hatch substance arm #43/Arm Phase 10 (§13.32) + ai-laziness-traps project rule #44/docs(refactor): split open-questions.md into open + closed-questions archive #46/docs: post-1A follow-up (Class C reviewer-discipline, Q3/Q5, README absolutism) #67/feat(agents): resolve C-1 agent-name collision with AI Factory (KEEP-AIF / skill-context / rename) #80 (prior-art-evaluations.md:95-148): none silently superseded. §6 additive notes are additive-only — DEFER/REJECT verdicts unchanged.
• No .claude/rules/* modified. No .claude/skills/* modified. No agents/* modified. No packages/* modified. No install.sh modified. No kickoff.md modified.
• SKILL.md:404 and SKILL.md:429 read but not modified — inconsistency resolved analytically in the patch, not via edit.
• Scope: Variant A only. Variants B/C/D have separate Sub-waves.
• T15 self-application confirmed: BFR §3 applied to this patch itself.

Pre-flight grep

grep -nE '^### §1\.7 (Forward|Backward)-check applied$' — PR body (this body):
  line: "### §1.7 Forward-check applied" ✓
  line: "### §1.7 Backward-check applied" ✓

file:line citations in §1.7 sections:
  prior-art-evaluations.md:95-148 ✓
  SKILL.md:404 ✓
  SKILL.md:429 ✓
  research-patches/2026-05-23-meta-orchestrator-prior-art.md:73 ✓

Test plan

• Research-patch first line = <!-- scope:aif-handoff-bridge-variant-a-mcp --> (principle 10 regex)
• §0-§9 sections present (12 elements per kickoff §11)
• §4 5-criteria scoring with file:line evidence per cell (15 cells)
• §5 verdict (REFERENCE) per BFR-default ladder
• Falsifier explicitly applied and stated as CONDITIONAL PASS
• §1.7 Forward + Backward in research-patch (§7+§8) AND PR body
• ≥5 DeepWiki probes cited with quoted excerpts (7 probes)
• ≥2 WebSearch queries cited (2 executed)
• Admission gate §2.4 re-verified (CLEAR)
• markdownlint-cli2: 0 errors (pre-commit hook passed)
• All principle tests: 153 passed / 4 skipped
• PR opened against staging
• Phase -1 cold-review by orchestrator/maintainer before merge (T19 — worker does not self-approve)