waggle-os/docs/briefs/2026-04-21-cc-sprint-10-tasks.md

CC Sprint 10 — Task Brief

Datum: 2026-04-21 Preceded by: Sprint 9 final briefing (sessions/2026-04-21-sprint-9-final-briefing.md) · Stage 0 final close-out (sessions/2026-04-21-stage-0-final-close-out.md) Scope LOCKED memory: .auto-memory/project_sprint_10_scope_locked.md Wall-clock ceiling: 12–16 dana Cost ceiling: $140 hard stop (Sprint 10 total, across svih tasks)

---

0. Šta se menja u odnosu na Sprint 9

Sprint 9 se završio sa jednim klinički čistim rezultatom (10/10 Opus calibration) i jednim operativnim dugom od tri stavke. Sprint 10 ne širi scope narativno — namerno NE ulazi u landing copy, brand narrative, ili launch assets pre LoCoMo gate-a. Umesto toga, zatvara ceo operativni queue da Stage 2 LoCoMo full-run može krenuti bez ijedne produkcione nepoznate.

Četiri paralelne vektorske linije — svaka sa jasnim acceptance kriterijumom, svaka nezavisno od ostalih commit-able.

---

1. Vector 1 — Operativni queue (Sprint 9 §5.2 + §3.2, full-resolve)

Task 1.1 — Qwen3.6 thinking-mode stability matrix

Spec referenca: waggle-os/docs/plans/STAGE-2-PREP-BACKLOG.md (commit 813a4eb) Budget: $5 hard ceiling Effort estimate: 4–6h

Šta radi: Izvršava 2 × 4 × 5 test matrix (thinking on/off × max_tokens {8K, 16K, 32K, 64K} × prompt shape {direct, multi-anchor, chain-of-anchor, temporal-scope, null-result-tolerant}). Svaku ćeliju klasifikuje kao: converged / loop / truncated / empty-reasoning-only.

Deliverable:

• CSV heat-map: waggle-os/benchmarks/harness/data/qwen-stability-matrix-2026-04-XX.csv
• Markdown summary: waggle-os/docs/reports/qwen-thinking-stability-2026-04-XX.md
• Explicit flag lista "Stage-2-unsafe cells" sa obrazloženjem za svaku ne-converged ćeliju.

Acceptance:

• Sve 40 ćelija izvršene, zero empty cells.
• Svaka ćelija klasifikovana u jedan od četiri outcome bucket-a.
• Safety matrica: koje (thinking, max_tokens) kombinacije su pouzdane za Stage 2 LoCoMo scale.
• Spend ≤ $5.

Exit criterion: ako bilo koja od (thinking-off, max_tokens ≥ 16K) konfiguracija konvergira 100% across svih 5 prompt shapes — Stage 2 primary config LOCKED na taj config; ako ne, eskalira u PM za re-scoping.

---

Task 1.2 — Sonnet route repair

Budget: $0 (config edit + regression) Effort estimate: 30min

Šta radi: litellm-config.yaml alias claude-sonnet-4-6 trenutno pokazuje na decommissioned claude-sonnet-4-6-20250514. Repair: mapirati na trenutno živi Sonnet slug (verifikuj na https://docs.anthropic.com/en/docs/about-claude/models/overview). Ako nema direktnog zamenika, PR sa privremenim promote-om na claude-opus-4-7 i dokumentovana migraciona nota u waggle-os/ops/litellm/README.md.

Acceptance:

• litellm-config.yaml izmenjen, regression test na judge-client invokeJudge({ model: "claude-sonnet-4-6", ... }) prolazi bez model_not_found greške.
• PR sa commit message referencirajući ovaj brief.
• Marko ratifikuje mapping pre merge-a (PM review gate).

---

Task 1.3 — Sonnet calibration re-run posle Task 1.2

Depends on: Task 1.2 CLOSED. Budget: $0.50 (10 calibration pairs × Opus baseline + Sonnet diff) Effort estimate: 1h

Šta radi: Izvršava Task 4 calibration re-run (istih 10 ground-truth triples iz f9b98aa calibration artifact) na Sonnet ruti. Rezultat upisuje u isti calibration artifact format.

Acceptance:

• Calibration match rate zabeležen (očekivano 8–10/10 na osnovu Haiku 5/10 + Opus 10/10 tier pozicioniranja).
• Ako Sonnet ≥ 9/10 → Sonnet postaje Stage 2 primary default (brief konfiguraciju).
• Ako Sonnet 7–8/10 borderline → triggeruj Task 5 Fleiss' kappa ensemble (Sprint 9 brief conditional) na ensemble MULTI-VENDOR iz Vector 2, ne Claude-only.
• Ako Sonnet < 7/10 → stick with Opus, flag za PM review.

---

Task 1.4 — DashScope provisioning (paralelni track)

Budget: $0 (account provisioning) Effort estimate: Marko-side account work + CC-side config (30min CC effort kad key stigne)

Šta radi: DashScope direct key provisioning za qwen3.6-35b-a3b canonical slug (LiteLLM upstream). OpenRouter route ostaje bridge/failover.

Acceptance:

• litellm-config.yaml sadrži oba route-a: qwen3.6-35b-a3b (DashScope direct) + qwen3.6-35b-a3b-via-openrouter (bridge).
• Failover policy dokumentovan: DashScope primary, OpenRouter retry-on-rate-limit fallback.
• Regression test pokriva oba route-a sa istim probe prompt-om i pokazuje byte-equivalent inference output.

Napomena: Ovaj task ne blokira Stage 2 kickoff. OpenRouter route je dovoljan za 200+ Qwen calls per 4-cell batch. DashScope je on-prem parity hedge, ne critical path.

---

Task 1.5 — Harvest Claude artifacts adapter (hive-mind)

Spec referenca: hive-mind/BACKLOG.md (commit b3348fb) — tri source-path opcije u priority order Budget: $0 (adapter development + regression, no paid inference) Effort estimate: 6–10h

Šta radi: Implementira Opcija 1 (current export bundles artifacts dir) kao primary. Opcija 2 (Claude.ai API listing) i Opcija 3 (Computer Use scraping) ostaju kao backlog hedge ako Opcija 1 pokaže gap-ove.

Deliverable:

• hive-mind/packages/cli/src/commands/harvest-claude-artifacts.ts (nova datoteka)
• UniversalImportItem sa type: "artifact", parent_conversation_id, inherited timestamp
• Regression tests: 2 scenario-level cases (artifact sa valid parent, artifact bez parent fallback)

Acceptance:

• Re-harvest Marko personal corpus na fresh export bundle (Marko daje): artifacts se pojavljuju kao frames sa type-annotation.
• Dogfood probe: frame 421 (januar 2026) i njeni artifacts (MASTER_PLAN_REVIZIJE.md + drugi) svi accessibil u chat-text substrate.
• Zero regresija na postojeće 305 tests.
• tsc clean.

Non-goal: Ovaj task NE pokriva LoCoMo (LoCoMo je chat-text-only by construction). Adapter služi za buduće dogfood cikluse na real korpus, i za launch narrative o completeness-u substrate layer-a.

---

2. Vector 2 — Multi-vendor ensemble setup LOCKED

Zašto NE Claude-only trio (rejection rationale u memoriji)

Predlog "Sonnet + Opus + Haiku" kao tri-judge ensemble je eksplicitno odbijen. Razlog:

Tri modela iz iste organizacije, iste training distribucije, deljenih bias-a ne čine legitimni inter-rater agreement. Fleiss' kappa na toj konfiguraciji meri unutar-Claude konzistenciju, ne independent judgment agreement. Defensibility claim koji se oslanja na tu statistiku rizikuje trivijalnu kritiku "sva tri su Claude" od bilo kog spoljnog reviewer-a — akademskog ili komercijalnog.

Referenca: .auto-memory/project_sprint_10_scope_locked.md

Task 2.1 — Tri-vendor API integracije

Budget: $5 (integration probe + smoke tests across sva tri) Effort estimate: 6–8h

Vendori (latest frontier per vendor — namerno biran):

1. Anthropic Opus 4.7 — već integrisan (Sprint 9 Task 4 production)
2. OpenAI GPT-5.4 — nova integracija (latest, ne GPT-5 baseline)
3. Google Gemini 3.1 — nova integracija (latest, ne Gemini 3 Pro baseline)

Napomena: Ako bilo koji od tri latest model-a nije provisionable kroz OpenAI/Google API-je u trenutku Task 2.1 kickoff-a (npr. waitlist, regionalne restrikcije, pricing nepotvrđen), CC dokumentuje blocker + predloži najbliži available tier u fallback listi. PM ratifikuje fallback pre nastavka — ne unilateralno padati na stariji tier.

Šta radi:

• LiteLLM config: tri route-a sa provider-specific parameter normalizacijom (temperature, max_tokens, reasoning_effort, structured output schema per vendor).
• Judge-client (packages/harness ili gde je trenutno) proširen sa vendor-agnostic invokeJudge() koji normalizuje prompt, parsuje response, i nosi vendor-specific retry policy.
• Failure handling: vendor A timeout ne blokira vendor B/C; aggregator sakuplja N-of-3 where N ≥ 2 je operable ensemble.

Deliverable:

• Updated litellm-config.yaml sa svim tri route-a
• invokeJudgeEnsemble() metoda sa policy "minimum 2-of-3 required, fail-open on 3rd vendor"
• Smoke test suite: 5 ground-truth triples × 3 vendora = 15 invocations, verdict logging, cost logging
• docs/reports/multi-vendor-ensemble-baseline-2026-04-XX.md — per-vendor match rate vs PM ground truth

Acceptance:

• Sva tri vendora vraćaju parsable verdict sa istog prompt shape-a.
• Per-vendor match rate zabeležen. Očekivano: Opus 10/10 (iz Sprint 9), GPT-5.4 ~8-10/10, Gemini 3.1 ~7-10/10.
• Ensemble Fleiss' kappa izračunat na 10 triples minimum pre pravog Stage 2 run-a.

---

Task 2.2 — Fleiss' kappa ensemble baseline

Depends on: Task 2.1 CLOSED. Budget: $3 Effort estimate: 2h

Šta radi: Na 15 ground-truth triples (10 iz Sprint 9 calibration + 5 novih koje PM sastavlja za pokrivenost multi-category), izvršava full ensemble i računa Fleiss' kappa.

Pre-registered interpretation bands:

• κ ≥ 0.80 → strong agreement — ensemble ready for Stage 2 full-run, ensemble verdict primary
• 0.60 ≤ κ < 0.80 → substantial agreement — ensemble ready, ali dodaje tie-breaker policy dokumentovana
• 0.40 ≤ κ < 0.60 → moderate — flagged za PM review pre Stage 2 kickoff-a
• κ < 0.40 → fair ili worse → go/no-go review; scope pivot ka single-judge Opus + rubric refinement

Deliverable:

• docs/reports/multi-vendor-kappa-baseline-2026-04-XX.md sa per-pair kappa (A-B, A-C, B-C), aggregate Fleiss' kappa, interpretation band, i recommended Stage 2 policy.

Acceptance:

• Kappa izračunat, interpretation band pristiman, preporuka za Stage 2 jasna.
• PM review gate pre Stage 2 kickoff-a ako band je "moderate" ili niže.

---

3. Vector 3 — Launch prep SPLIT (narrative-agnostic only)

Ovaj vector NE ide kroz ovog CC. Ide paralelno kroz posebnu CC sesiju koju PM (Claude Opus 4.7 u Cowork mode) koordiniše.

Reason za split: ovaj brief je za CC koji drži tehnički execution i nema kontekst za brand voice, persona, i visual identity rad. Launch prep koji NE zavisi od LoCoMo rezultata pokriva:

• Persona research deep dives (nastavak project_persona_research_scope.md, 10 persona)
• Brand asset inventory (audit apps/www/public/brand + konsolidacija)
• Visual identity konsolidacija (extend Hive DS sa honeycomb motif, review globals.css tokens)
• Stripe/billing polish (ne-funkcionalni copy polish, legal review hooks)
• Legal/licensing prep (Apache 2.0 compliance audit hive-mind)
• i18n policy finalizacija (feedback_i18n_landing_policy.md — engleski first, locale-ready infra)

Eksplicitno DEFEROVANO u Sprint 11:

• Landing copy (zavisi od LoCoMo brojke)
• Brand narrative sync (zavisi od LoCoMo brojke)

---

4. Vector 4 — Harvest artifacts adapter

Pokriveno u Task 1.5 (Vector 1). Reprized ovde kao paralelna linija jer je CC može započeti bez čekanja Vector 1.1–1.4.

---

5. Pre-registered LoCoMo acceptance thresholds (informativna sekcija, CC ne mora da ih implementira, ali mora da ih poštuje u agregator output-u)

Pre Stage 2 full-run, Stage 2 aggregator mora emitovati eksplicitni banner na osnovu finalnog LoCoMo score-a:

Score	Banner	Consequence
≥ 91.6%	NEW_SOTA	Full launch narrative (Opus-class multiplier claim)
85.0–91.5%	SOTA_IN_LOCAL_FIRST	Narrower launch framing (sovereignty vs cloud-revenue positioning)
< 85.0%	GO_NOGO_REVIEW	Auto-halt, kompletna scope reklasifikacija sa PM pre bilo kakve javne komunikacije

Banner ulazi u agregator markdown report ## header. NEMA post-hoc narrative shifting. Ako rezultat padne ispod 91.6% — threshold bands se NE pomeraju. Workflow Reality Check anti-pattern #4 je na snazi.

---

6. Sequencing i zavisnosti

Task 1.1 (Qwen stability)  ─────┐
Task 1.2 (Sonnet route) → Task 1.3 (Sonnet calibration) ─┐
Task 1.4 (DashScope)  ───────────┤                        ├→ Gate: Stage 2 kickoff ready
Task 1.5 (Artifacts adapter) ────┤                        │
Task 2.1 (Tri-vendor setup) → Task 2.2 (Kappa baseline) ──┘

Task 1.1, 1.4, 1.5, 2.1 mogu krenuti paralelno dan-1. 1.2 → 1.3 je sekvencijalno. 2.2 zavisi od 2.1.

Stage 2 kickoff gate: sve od 1.1, 1.3, 2.2 mora biti CLOSED sa acceptance band na ili iznad minimuma (1.1 → bar jedna safe config; 1.3 → bar 7/10; 2.2 → bar moderate kappa).

---

7. Cost ceiling i governance

Task	Budget	Running total
1.1	$5	$5
1.2	$0	$5
1.3	$0.50	$5.50
1.4	$0 (Marko provisioning)	$5.50
1.5	$0	$5.50
2.1	$5	$10.50
2.2	$3	$13.50
Stage 2 projected (not Sprint 10 scope, informativno)	$90–130	—

Sprint 10 hard stop: $15 across svih tasks. Ako task prelazi budget, HARD STOP, PM review pre nastavka.

Stage 2 projekcija ($90–130) je van Sprint 10 ceiling-a. Kickoff Stage 2 full-run ide kao zaseban PM go-ahead nakon Sprint 10 close-a.

---

8. Reporting cadence

• Dan-1 EOD: kratka update nota u sessions/2026-04-22-sprint-10-day-1-status.md — koje task-ove si pokrenuo, koji commit-i su prošli.
• Mid-sprint: kada je 3 task-a CLOSED — update nota + preliminary findings.
• Sprint 10 close: sessions/2026-04-XX-sprint-10-final-briefing.md po istom template-u kao Sprint 9 final briefing.
• Anti-pattern check: ako u toku task-a identifikuješ novi substrate failure mode, feedback_workflow_reality_check.md anti-pattern #4 remains hard rule — document the mode, open backlog ticket, ne reformulisati gate ili acceptance threshold post-hoc.

---

9. Shta NE raditi u Sprint 10

• Ne pisati landing copy.
• Ne dirati brand narrative dokumente.
• Ne pokretati Stage 1 ili Stage 2 full-run dok gate uslovi iz §6 nisu ispunjeni.
• Ne širiti judge ensemble van tri-vendor konfiguracije LOCKED u Vector 2 (ne dodavati 4., 5. vendor "just in case").
• Ne lock-ovati Stage 2 primary judge na osnovu Task 1.3 single-vendor Sonnet result-a — tek posle Task 2.2 ensemble kappa.

---

10. Sprint 10 close criteria

Sprint 10 se zatvara kada:

1. Sve task-ove iz Vector 1 i Vector 2 CLOSED sa acceptance band ili iznad.
2. Sprint 10 final briefing napisan + PM ratifikovan.
3. Stage 2 kickoff memo (poseban brief) napisan i PM-ratified za narednu CC sesiju.
4. Zero test regressions, tsc clean, sve commit-e pushed.

Launch copy / brand narrative / Stage 2 full-run izvršenje → Sprint 11 scope, ne ovde.

---

End of Sprint 10 brief. Awaiting CC execution.