oleg.m/waggle-os
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
master
waggle-os/docs/briefs/2026-04-22-cc-sprint-11-kickoff.md
Oleg Maslov 0c3e2ead3b
Some checks failed
Installer Smoke / installer-smoke (push) Has been cancelled
Details
moving
2026-09-02 10:10:29 +02:00

16 KiB
Raw Permalink Blame History

CC-1 Sprint 11 Kickoff Brief — Pre-flight Readiness Sprint

Sprint ID: 11 Sprint type: Pre-flight Readiness Datum brief-a: 2026-04-22 Author: PM (Claude Opus 4.7, Cowork mode) For execution: CC-1 (fresh context, post Sprint 10 full-close + PR #2 merged) Scope authority: decisions/2026-04-22-sprint-11-scope-locked.md (LOCKED) Stage 2 config authority: decisions/2026-04-22-stage-2-primary-config-locked.md (LOCKED 2026-04-22, on/64K) Supersedes: strategy/2026-04-22-sprint-11-kickoff-memo.md (DRAFT, judge-methodology axis) Update 2026-04-22: Task A1 reasoning_content handling add-on + Task B1 config LOCKED (on/64K).

1. Sprint primary axis

Sprint 11 je Pre-flight Readiness Sprint. Exit kriterijum = green-light autorizacija za H-42a/b benchmark ($1500-2600 Stage 2 full-run). Sprint 11 ne izvršava H-42a/b.

Sve što CC-1 radi kroz Sprint 11 mora da se mapira na 1 od 10 hard gate kriterijuma iz scope lock dokumenta §4. Ako task ne mapuje — CC-1 ne pokreće bez PM ratifikacije.

2. Sprint 11 entry state (Sprint 10 full-close verifikacija)

Pre nego što Sprint 11 kick-uje, CC-1 verifikuje:

  1. Sprint 10 Task 1.1 CLOSED sa PASS verdiktom (5 safe Qwen configs) — sessions/2026-04-22-sprint-10-task-1-1-exit.md landed.
  2. Sprint 10 Task 1.5 Phase 3 CLOSED i push-ovan na origin (ili eksplicitno deferred ka Sprint 11 kao carry-over sa PM ratifikacijom).
  3. PR #2 merged u main (per sessions/2026-04-22-task-18-closed.md).
  4. Sprint 10 close-out report finalizovan i pushed.

Ako bilo šta od gornjeg nije landed — CC-1 NE kick-uje Sprint 11. Ping PM sa blokerom.

3. Task lista (10 gate kriterijuma iz scope lock §4)

Track A — Audit-grade traceability

Task A1 — H-AUDIT-1 design doc (1 page). Predaj docs/plans/H-AUDIT-1-DESIGN-DOC-2026-04-XX.md sa sledećim sadržajem (i ničim više):

  • turnId generation point: gde se generiše (orchestrator turn entry) i kako (crypto.randomUUID() v4, ne v7, ne custom).
  • Propagation surface: lista 5+ fajlova u kojima turnId mora biti vidljiv i funkcija/metoda signature-a gde ulazi kao parametar. Očekivani fajlovi iz project_h_audit_1_not_implemented.md: orchestrator.ts, cognify.ts, tools.ts, combined-retrieval.ts, prompt-assembler.ts, agent-loop (entry point). Ako se CC-1-ov kodni pregled pokaže da neki od ovih ne postoji sa tim imenom ili da postoje dodatni fajlovi — ažuriraj listu u design doc-u i pomeni razliku.
  • Persistence format: kako turnId se upisuje u existing trace store u chat.ts. Struktura polja, tip, retention pravila ako postoje.
  • Reasoning_content handling (dodato per decisions/2026-04-22-stage-2-primary-config-locked.md §5): Stage 2 config je LOCKED na thinking=on, max_tokens=64000, što znači da Qwen response sadrži reasoning_content polje izvan finalnog answer payload-a. Design doc mora eksplicitno adresirati tri pravila:
    • Persistence rule: gde se reasoning_content persista u trace store-u (pod istim turnId u poseban slot turn.reasoningContent ili potpuno separatan store). Predloži izbor i razlog.
    • Retention policy: koliko dugo se čuva, da li ulazi u audit export-ove, da li se brise posle N dana.
    • Exclusion rule: gde NE ulazi (npr. user-facing output log-ovi, public trace viewer-i, MCP response payload-i). Trace reconstruction iz turnId-a mora moći da opciono includuje ili excluduje reasoning_content per caller permission.
  • Test scenario: unit test koji reconstruct-uje full turn graph iz single turnId-a, uključujući reasoning_content slot kao opcioni payload. Sample code-om opiši test, ne implementiraj još.
  • Acceptance kriterijumi: grep target (≥6 hits za turnId u ≥5 fajlova), zero regresija na postojećim suite-ovima, tsc clean, reasoning_content handling dokumentovan u design doc-u.

CRITICAL: A1 NE kreće u A2 implementation dok PM ne ratifikuje design doc. PM ratifikacija gate je tvrd. Ne preempty.

Budget: $0 (markdown only). Wall-clock estimate: 1-2h za design doc draft + PM iteration turnaround.

Task A2 — H-AUDIT-1 implementation. Start SAMO posle A1 PM ratifikacije.

Implementiraj turnId propagaciju per ratifikovani design doc. Acceptance:

  • grep -n "turnId" packages/**/*.ts | wc -l vraća ≥6.
  • Unique fajlovi sa match-om ≥5.
  • Novi unit test u existing test suite lokaciji pokazuje reconstruct full turn graph iz single turnId-a.
  • pnpm test vraća zero regresija.
  • pnpm tsc --noEmit clean.
  • Commit message: feat(audit): implement H-AUDIT-1 turnId propagation per design doc 2026-04-XX.

Nakon merge-a, posti exit ping u sessions/2026-04-XX-sprint-11-h-audit-1-exit.md sa grep output-om i test log-om.

Budget cap: $0.10 (unit test API calls). Wall-clock estimate: 0.5-1 dan.

Task A3 — H-AUDIT-2 bench-spec resolution.

Ova task zahteva 30-min Marko + PM call (bench-spec odluka "wall-clock vs recall correctness only", per project_audit_findings.md). CC-1 ne izvršava task samostalno — CC-1 STANDS BY dok PM ne dostavi decisions/2026-04-XX-bench-spec-wall-clock-resolution.md.

Po landing-u decision dokumenta, CC-1 (ako treba) ažurira Stage 2 kickoff memo (B4) da reflektuje bench-spec izbor.

Budget: $0. Wall-clock estimate: 0.5 dan (posle Marko+PM call).

Track B — Methodology locks

Task B1 — Stage 2 Qwen config apply. (LOCKED 2026-04-22)

Authority: decisions/2026-04-22-stage-2-primary-config-locked.md — LOCKED na thinking=on, max_tokens=64000 na route qwen3.6-35b-a3b-via-openrouter. Marko je override-ovao PM preporuku (off/16K) sa rationale-om iz v05 eval rezultata.

Apply akcija:

  • Ažuriraj LiteLLM default config za Stage 2 batch runs: 
    thinking: true (ili ekvivalent parameter name u harness-u)
max_tokens: 64000
model: qwen3.6-35b-a3b-via-openrouter
  • Verifikuj da C2 i C3 harness koristi taj config eksplicitno, ne nasleđeno iz drugog lokala.
  • Smoke test 1 poziv da se config aktivirao. Log cost, latency, i reasoning_content size (polje koje dolazi sa thinking=on).
  • Exit ping: sessions/2026-04-XX-sprint-11-b1-config-applied.md sa cost/latency/reasoning_content size zapisom.

Anti-pattern lock (iz §8 decision doc-a): Ako C3 faila sa 3 attempts na on/64K, HARD STOP i PM review — NE automatska reformulacija na off/16K sredinom sprinta. Thinking=off je legitimana contingency samo posle Sprint 11/12 retrospektive.

Budget cap: $0.05 (smoke test). Budget impact downstream: Stage 2 4-cell mini expected ~$8-14 (vs $4-6 za PM preporuku). Cap $134 nepromenjen. Wall-clock estimate: 30 min.

Task B2 — Tie-break policy implementacija.

Implementiraj per PM preporuku u scope lock §2.B2 (Opcija 3: Sonnet 4.6 kao fourth vendor) sa Opcija 2 fallback (PM escalation na 1-1-2 quadri-vendor split).

  • Novi fajl: packages/server/src/benchmarks/judge/ensemble-tiebreak.ts.
  • Interface: funkcija resolveTieBreak(votes: Vote[]): TieBreakResult gde TieBreakResult = { verdict: string; path: 'majority' | 'quadri-vendor' | 'pm-escalation'; votes: Vote[]; }.
  • 4 unit testa pokrivajući: 1-1-1 (trigger quadri-vendor Sonnet call), 1-1-2 (već majority, no tiebreak), 2-1-1 (majority wins), 3-0 (consensus, no tiebreak).
  • Integracija u postojeći ensemble orchestration path.
  • Observability: logovi koji path je uzet (majority/quadri/pm-esc) per case.

PM će LOCK-ovati policy kroz decisions/2026-04-XX-tie-break-policy-locked.md pre merge-a.

Budget cap: $0.20 (Sonnet 4.6 calls u unit testu). Wall-clock estimate: 2-3h.

Task B3 — Opus 4.6 route audit.

Per superseded memo bolt-on C:

  1. grep -rn "claude-opus\|claude-sonnet-4" packages/server/ packages/cli/ → klasifikuj svaku referenca kao:
    • (a) dated snapshot (npr. claude-opus-4-6-20251014, claude-opus-4-7-20260201),
    • (b) floating alias (npr. claude-opus-4-6, claude-opus-4-7),
    • (c) hardcoded mismatch (npr. reference na deprecated model).
  2. Za (b) i (c), dokumentuj Why floating je prihvatljiv (npr. dev convenience path) ili migriraj na dated snapshot.
  3. LOCK naming convention u docs/BENCHMARK-INFRASTRUCTURE.md ili ekvivalent canonical fajl.
  4. Trigger-on-first-caller-trip observability verify: postoji li log alert kada floating alias failuje na provider side?

Deliverable: docs/reports/opus-4-6-route-audit-2026-04-XX.md + PM izdaje decisions/2026-04-XX-model-route-naming-locked.md posle audit review-a.

Budget cap: $0.10 (verification test calls). Wall-clock estimate: 1-2h.

Task B4 — Stage 2 kickoff memo draft (saradnja sa PM).

PM piše prvi draft u strategy/2026-04-XX-stage-2-kickoff-memo.md sa 4-nedeljnim planom:

  • Week 1: Qwen3.6-35B-A3B × full LoCoMo corpus (500 instances ili full set po local zip-u).
  • Week 2: Gemma Week 3 probe comparison (per decisions/2026-04-20-gemma-week3-probe-locked.md).
  • Week 3: Three-model comparison analysis.
  • Week 4: Failure mode distribution + publishable results draft.

CC-1 dodaje harness readiness assessment na kraj memo-a (jednokratno): da li sav tooling za 4-week plan postoji u repo-u, gde su gap-ovi, procenjeni wall-clock za gap closure.

Budget: $0. Wall-clock estimate: 1h CC-1 assessment + PM iteracije posebno.

Track C — Pre-flight gate execution

Task C1 — Stage 0 verifikacija.

Stage 0 je CLOSED per memory (project_preflight_gate.md). CC-1:

  • Link postojeće artifacts (stage 0 results).
  • Verifikuj da nije regredirao (run poslednji known-good smoke test ako postoji, ili uporedi current build vs Stage 0 artifact hash).
  • Ako regredirao → re-run Stage 0 po originalnom scope-u (~$5).

Deliverable: kratak check-in ping u sessions/2026-04-XX-sprint-11-stage-0-verify.md sa "still green" ili "regression detected, re-run triggered".

Budget: $0 ako green, ~$5 ako re-run. Wall-clock estimate: 15-30 min verify, +30 min ako re-run.

Task C2 — Stage 1 mikro-eval.

Izvršava SAMO posle:

  • A2 CLOSED (turnId landed),
  • B1 CLOSED (Qwen config LOCKED),
  • B2 CLOSED (tie-break implementovan),
  • C1 CLOSED (Stage 0 verified).

Scope: kratka eval baseline per project_preflight_gate.md Stage 1 spec. Verifikacija da je engine + methodology stack operativan pre Stage 2 main run-a.

Deliverable: preflight-results/stage-1-microeval-2026-04-XX.md sa results tabelom i PASS/HARD STOP verdiktom.

Budget: $5-10. Wall-clock estimate: 1-2h.

Task C3 — Stage 2 4-cell mini.

Izvršava SAMO posle C2 PASS.

Scope per decisions/2026-04-20-preflight-stage2-4cell-amendment.md:

  • 50 pitanja (fiksni seed=42, pre-registrovan sample).
  • 4 cells: raw / memory-only / evolve-only / full-stack.
  • Sonnet 4.6 judge.
  • Pass kriterijum: full-stack ≥85% recall + ordinal consistency (full ≥ memory, full ≥ evolve, full > raw sa delta ≥10pp, bar jedan layer > raw).
  • Re-run policy: max 3 attempts, sample curation samo za legitimate scope gap ≤5 pitanja (dokumentovano).

Deliverable: preflight-results/stage-2-4cell-mini-2026-04-XX.md sa 4-cell tabelom, ordinal consistency check-om, i verdiktom.

Ako PASS → ovo je poslednji gate kriterijum; Sprint 11 blizu full-close-a. Ako HARD STOP (3 attempts faili) → ne zatvaraj Sprint 11. PM review.

Budget cap: $134 (per ratifikovani preflight gate amendment). Wall-clock estimate: 3-5h (worst case sa retries).

4. Sekvenciranje (dependencies)

C1 ──────────────────────────┐
                              │
A1 design doc → PM ratify → A2 ──┐
                                  │
B1 ratified ──────────────────────┼→ C2 → C3 → Sprint 11 full-close
                                  │
B2 impl + LOCK ───────────────────┘

A3 (after PM+Marko call) ──────── informational, ne blocks C3
B3 route audit ──────────────── informational, ne blocks C3
B4 kickoff memo draft ────── informational, ne blocks C3

CC-1 redosled preporučen:

  1. Day 1 AM: Entry state verify + A1 design doc draft (uključuje reasoning_content handling per LOCK) + B1 apply (config je već LOCKED, nema čekanja).
  2. Day 1 PM: C1 verify paralelno sa čekanjem na PM A1 ratifikaciju. B1 smoke test push-ovan.
  3. Day 2 AM: A2 implementation (posle A1 ratifikacije) + B2 impl start.
  4. Day 2 PM: A2 finish + B2 finish + B3 route audit paralelno.
  5. Day 3 AM: B2 merge pending LOCK + C2 Stage 1 mikro-eval.
  6. Day 3 PM: C3 Stage 2 4-cell mini start.
  7. Day 4: C3 finish + Sprint 11 close-out report + push.

A3 (bench-spec) ume da ispadne iz calendar-a ako Marko+PM call nije zakazan do Day 2; to je sledeće za eskalaciju PM-u ne CC-1-u. B4 (Stage 2 kickoff memo) PM vodi; CC-1 asistira kad memo bude spreman za harness assessment add-on.

5. Budget summary

Track Low High
A (A1+A2+A3) $0 $0.10
B (B1+B2+B3+B4) $0.05 $0.35
C (C1+C2+C3) $67 $149

Sprint 11 total ceiling: ~$150. Sprint 10+11 cumulative ceiling: ~$152 vs $15k Stage 2 full-run envelope. Radikalno unutar granica.

Per-task hard alarm: ako bilo koji pojedinačni task premaši 130% svoje cap vrednosti, CC-1 HARD STOP i PM ping.

6. Reporting cadence

  • Day 1 EOD: CC-1 posti sessions/2026-04-XX-sprint-11-day-1-status.md sa (a) entry verify summary, (b) A1 draft link, (c) C1 verify outcome.
  • Day 2 EOD: status ping sa A2 + B2 progress, B3 audit outcome (ako CLOSED).
  • Day 3 EOD: status ping sa B1 applied, C2 verdikt.
  • Day 4 EOD ili sprint close: sessions/2026-04-XX-sprint-11-close.md sa svih 10 gate kriterijuma check-in tabelom i green-light preporukom.

Svaki exit ping po tasku: sessions/2026-04-XX-sprint-11-<task-id>-exit.md.

7. Anti-patterns (hard ograničenja)

  • Anti-pattern #1: No H-42a/b execution u Sprint 11. Ako neki signal u Sprint 11 iskače u smeru "hajde odmah da pokrenemo full run", CC-1 stand by i PM ping.
  • Anti-pattern #2: No scope creep. Task lista iz §3 je exhaustive. Bilo koji dodatak zahteva novi decision dokument potpisan od PM-a.
  • Anti-pattern #3: No self-merge na PR-ovima vezanim za Sprint 11 deliverables. CC-1 predaje za PM review; PM izdaje merge verdikt.
  • Anti-pattern #4: No post-hoc reformulation rezultata C3 4-cell mini. Pre-registered thresholds (full ≥85%, ordinal consistency) stoje. Re-run pravilo: max 3, curation ≤5 pitanja na legitimate scope gap, eksplicitno dokumentovano. Ne "hajmo samo da slušamo ovaj cell".
  • Anti-pattern #5: No turnId implementacija bez A1 PM ratifikacije. Implementation-first → design-doc-second je refused workflow.

8. Exit criteria check-list (za Sprint 11 close-out report)

Svih 10 mora biti ✅:

  1. A1 CLOSED — H-AUDIT-1 design doc ratifikovan (link)
  2. A2 CLOSED — turnId landed, grep ≥6, tests green (commit SHA + grep output)
  3. A3 CLOSED — bench-spec decision LOCKED (decision dokument link)
  4. B1 CLOSED — Stage 2 config LOCKED (decision dokument link)
  5. B2 CLOSED — tie-break policy LOCKED + implementovan + tests green (commit + decision)
  6. B3 CLOSED — Opus 4.6 route audit + naming decision LOCKED (report + decision)
  7. B4 CLOSED — Stage 2 kickoff memo draft ratifikovan (memo link)
  8. C1 CLOSED — Stage 0 verify outcome (ping link)
  9. C2 CLOSED — Stage 1 mikro-eval PASS (results link)
  10. C3 CLOSED — Stage 2 4-cell mini PASS sa ordinal consistency (results link)

Green-light verdikt za H-42a/b izdaje PM čim svih 10 CLOSED + push.

9. Related

  • decisions/2026-04-22-sprint-11-scope-locked.md — scope authority (čitaj prvi)
  • decisions/2026-04-22-stage-2-primary-config-locked.md — Stage 2 config LOCK (on/64K), B1 authority
  • strategy/2026-04-22-stage-2-qwen-config-ratification-memo.md — B1 input memo (PM preporuka odbijena, Marko izbor ratifikovan)
  • strategy/2026-04-22-sprint-11-kickoff-memo.md — SUPERSEDED draft (audit trail only, ne izvršavati)
  • sessions/2026-04-22-sprint-10-task-1-1-exit.md — Qwen safe config pool
  • sessions/2026-04-22-sprint-10-task-1-5-phase-3-ratification.md — potential carry-over state
  • .auto-memory/project_h_audit_1_not_implemented.md — turnId implementacija kao ne-verifikacija
  • .auto-memory/project_preflight_gate.md — 3-stage gate struktura
  • .auto-memory/project_audit_findings.md — H-AUDIT-1 + H-AUDIT-2 Must-Fix
  • .auto-memory/project_target_model_qwen_35b.md — kanonski engine

End of Sprint 11 kickoff brief. Awaiting CC-1 entry state verify + A1 design doc draft.

Reference in New Issue View Git Blame Copy Permalink