ci: Bump actions/upload-artifact from 4 to 7 - #4

…on — 17 madde fix

Sprint 168 v1→v4 zinciri:
- v1 (commit fc91fcd): brainstorming output
- v2 systematic-debugging eval (Agent A): 79/100, Phase 4.5 trigger, 5 critical/high
- v3 devil's advocate eval (Agent B): 22/100 — hedef <30 KARŞILANDI ✅
- v4 (bu commit): 17 madde fix integration
- v5 Alperen final approval bekliyor

17 madde fix entegre:
1. (Agent A #1 CRITICAL) C0e cross-sprint orphan handling — Option C selective filter + startup invocation
2. (Agent A #2 CRITICAL) C0a bundle split → C0a-1/C0a-2/C0a-3/C0a-4 (4 sub-anchor)
3. (Agent A #3 CRITICAL) C0c subscriber owner: decision-engine.ts designate + function signature + BRAIN→SPAWN:BLOCKED event mandatory test
4. (Agent A #4 + B V7 HIGH) ADR-047 Manuel Subagent Dispatch Protocol Sprint 168'de (önceki 168.5 ertelenmiş)
5. (Agent A #5 + B Saldırı #2 HIGH) Smoke test complex scenario: 3+ task (collision + crash + parallel)
6. (B V5 HIGH) Sprint 168 NO_GO → Sprint 168.5 fallback explicit (recursion paradox kabul)
7. (B V7 HIGH) TDD skip enforcement gate (skip artış 0 + Alperen review)
8. (Agent A #8 MED) checkSpawnLock singular helper eklendi
9. (Agent A #9 MED) auto_archive_directives Alperen decision NOW (spec yazımı sırasında)
10. (Agent B #5 MED) Subagent git branch isolation (worktree per cluster)
11. (Agent B #7 MED) ADR-046 invariant test C0a-2 + C0a-3 integration test
12. (Agent B #6 MED) ADR-048 multi-provider parity (Docker + Subprocess + Tmux)
13. (Agent B #4 MED) ADR-048 Wave 1.5 serial gate + Alperen CHECKPOINT
14. (P4.5 MED) Cross-cluster dependency graph spec içinde explicit (Section 2)
15. (Agent B #3 MED) Baseline tolerance "0 yeni skip" GO/NO_GO row
16. (Agent B V6 LOW) Effort yeniden tahmin 22-30h → 35h gerçekçi
17. (Agent B #1 LOW) Brain "kırık" iddiası modül-fonksiyon-satır kanıtı (Section 1)

v1 → v4 büyük yapı değişiklikleri:
- Scope 5 task → 8 task (C0a split + ADR-047 yeni)
- Effort tahmin 22-30h → 35h
- ADR sayısı 1 (ADR-048) → 2 (ADR-047 + ADR-048)
- Smoke test 2-task echo → 3+ task complex (collision + crash + parallel)
- Subagent dispatch "manuel" → "hardened (worktree + file authority + lock + TDD gate)"
- Sprint 168.5 dependency açıklama (NO_GO → manuel dispatch replay)
- Section 2 cross-cluster dependency graph yeni
- Section 3.2 (dispatch mechanism + lock pattern + TDD gate + fallback) yeni
- Section 5.3 complex smoke test suite yeni
- Pre-Flight checklist 11 → 14 madde (git worktree, dispatch locks, smoke test, Alperen auto_archive decision)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ical docs

Sub-project #1/4 (Embedded Web Terminal) shipped 2026-05-19→2026-05-20,
operationally smoke-confirmed by Alperen. Update reader-facing docs to
reflect the delivery honestly + the four-part path ahead.

- README + README-TR: new Highlights bullet (top), link to terminal guide
- VISION + VISION-TR: extend 'Where we are now' / 'Where we are going'
  with the embedded terminal as the first concrete step toward agentic-
  OS workflows + the 3 deferred sub-projects (#2 self-security,
  #3 multi-tenant/k8s, #4 enterprise integrations) with the seams
  (AuthProvider / SessionBackend / tenantId) called out
- docs/release/roadmap.md: new 'Phase 3.6: Embedded Web Terminal' before
  Phase 3.5, with delivered checklist + sub-project #2-4 backlog
- docs/release/beta-tracker.md + -tr.md: Last-updated header refreshed
  (2026-05-14→2026-05-20); new 'Sprint 175 — Embedded Web Terminal'
  section with metrics (46/46 tests, build/pack clean), honest debt
  (node-pty linux-x64 prebuild, DECKENT_API_AUTH_DISABLED=1 dashboard
  precondition), process learnings (two durable feedback memories)
- docs/ROADMAP-GOD-LEVEL.md: new ⚡ 2026-05-20 section at the top, full
  Sprint 175 record + sub-project backlog table + Sprint 176+ priorities
- docs/adr/README.md: regenerated index (already had ADR-062/063 after
  collision rename earlier)

No exaggeration: smoke confirmed, but DECKENT_API_AUTH_DISABLED=1 is
still required for dashboard data calls (frontend general API auth
plumbing is sub-project #2/#3 scope) and node-pty linux-x64 prebuild
needs the optionalDep fix in Sprint 176. Recorded as honest debt.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ical docs

Sub-project #1/4 (Embedded Web Terminal) shipped 2026-05-19→2026-05-20,
operationally smoke-confirmed by Alperen. Update reader-facing docs to
reflect the delivery honestly + the four-part path ahead.

- README + README-TR: new Highlights bullet (top), link to terminal guide
- VISION + VISION-TR: extend 'Where we are now' / 'Where we are going'
  with the embedded terminal as the first concrete step toward agentic-
  OS workflows + the 3 deferred sub-projects (#2 self-security,
  #3 multi-tenant/k8s, #4 enterprise integrations) with the seams
  (AuthProvider / SessionBackend / tenantId) called out
- docs/release/roadmap.md: new 'Phase 3.6: Embedded Web Terminal' before
  Phase 3.5, with delivered checklist + sub-project #2-4 backlog
- docs/release/beta-tracker.md + -tr.md: Last-updated header refreshed
  (2026-05-14→2026-05-20); new 'Sprint 175 — Embedded Web Terminal'
  section with metrics (46/46 tests, build/pack clean), honest debt
  (node-pty linux-x64 prebuild, DECKENT_API_AUTH_DISABLED=1 dashboard
  precondition), process learnings (two durable feedback memories)
- docs/ROADMAP-GOD-LEVEL.md: new ⚡ 2026-05-20 section at the top, full
  Sprint 175 record + sub-project backlog table + Sprint 176+ priorities
- docs/adr/README.md: regenerated index (already had ADR-062/063 after
  collision rename earlier)

No exaggeration: smoke confirmed, but DECKENT_API_AUTH_DISABLED=1 is
still required for dashboard data calls (frontend general API auth
plumbing is sub-project #2/#3 scope) and node-pty linux-x64 prebuild
needs the optionalDep fix in Sprint 176. Recorded as honest debt.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… NO_GO recovery + beta launch (10 task)

Sprint 183 = son beta-blocker sprint. Sub-project #3 + #4 Sprint 184+'ya iter.

Layer 1 — 3 P0 fix (Sprint 182 dogfood'dan keşfedildi):
- W1-1: Nervous PLAN-phase pasif (FSWatcher debounce + phase guard)
  Root: nervous=true PLAN'da 14+dk donuyor, 17 task JSON yazımı sırasında FS event amplification
- W1-2: DEPENDENCY_BLOCKED event spam debounce (state-change only emit)
  Root: 550+ event Sprint 182, 95% spam (her 5sn'de aynı state)
- W1-3: Worker timeout RC investigation + fix
  Root: 5 task "exit 0 without result" — prompt size? Docker overhead? heartbeat write?

Layer 2 — Sprint 182 5 NO_GO recovery:
- W2-1: mock hygiene orphan-cleaner-ipc + archive-debt (Sprint 181'den beri pre-existing)
- W2-2: vitest CI=true parity verify (W2-1 sonrası smoke)
- W2-3: title-prefix Dependencies resolver tamamla (Sprint 182 dosya yazıldı, impl eksik)
- W2-4: integration smoke regression tamamla (Sprint 182 dosya yazıldı, timeout)

Layer 3 — Beta Launch v1.0.0-beta.1:
- W3-1: validate:publish 6/6 GREEN recheck + Brain re-eval RC investigation
  (Sprint 182 W4-1 worker raporu GREEN ama Brain NO_GO; ironik durum)
- W3-2: npm pack + lint:adr + lint:link final hijyen
- W3-3: v1.0.0-beta.1 final smoke (build:all + vitest + dashboard + serve)

DIRECTIVES'te retro/stub task YOK (feedback_no_retro_task_in_directives memory rule).

Sub-project Roadmap (post-Sprint 183, post-beta):
- Sprint 184: Repo housekeeping + docs cleanup (Alperen analiz)
- Sprint 185: Sub-project #3 başlangıç (multi-tenant audit + mTLS scaffold)
- Sprint 186-188: Sub-project #3 tamamlama (k8s + cross-tenant + HW key)
- Sprint 189-192: Sub-project #4 (Enterprise SSO + SIEM + compliance)
- Sprint 193+: Nervous Faz 2/3 + AEGIS + daily-assistant + local LLM CUDA

Config Sprint 183:
- nervous_system.enabled: false (P0-1 fix LAND etmeden risk)
- dependency_pipeline_enabled: true (Sprint 182 dogfood'da kanıtlandı, korunur)
- directives_protection.auto_restore: true (Sprint 177-005 hook canlı)

June 1 OSS beta launch yolu net — Sprint 183 GO sonrası npm publish Alperen manuel.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- ROADMAP-GOD-LEVEL.md: ⚡ 2026-05-21 Crisis Stab Initiative (Sprint 177-183) 7-sprint summary, F1-F8 + ADR-064 + I1-I5 wins, Sub-project #3/#4 4-sprint detail each, Nervous Faz 2/3 + AEGIS Phase 1-3, Sprint 200 God-Level GA, Local LLM CUDA sub-project #5
- vision/roadmap.md: Trinity maturity update (Developer 95% beta READY, System Worker 55%, Assistant 30%), Crisis Stab closure section, Sprint 134-145 table replaced by Sprint 184-200 post-beta roadmap, MIT licensing callout for enterprise features

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

DEBT.md dosya tabanlı borç yönetimi → Memory V2 SQLite DB-first geçişi.
Kök neden: 7-vs-9-column bozulma + eksik satır sonu bug'ları + .md parse
fragility. Tüm okuma/yazma noktaları getDebtItems/recordRollbackDebt
üzerinden tek kaynağa geçti.

Task #4a — debt-manager.ts
  - recordRollbackDebt(): DB'ye idempotent rollback debt kaydı
  - getDebtItems(): DB-first DebtItem listesi (activeOnly filtresi)
  - debtEntryToDebtItem(): MemoryEntryV2 → legacy DebtItem shape dönüşümü

Task #4b — rollback.ts
  - recordRollbackInDebt: appendFileSync(DEBT.md) → recordRollbackDebt()
  - appendFileSync + DEBT_FILE import'ları kaldırıldı

Task #4c — sprint-docs-updater.ts
  - autoResolveDebt: readFileSync(DEBT.md) string-mangle → resolveDebt()

Task #4d — sprint-finalizer, sprint-phases, sprint-state-tracker, doctor, api/server
  - Her parseDebtTable(readFile(DEBT.md)) çağrısı → getDebtItems()
  - /api/debt endpoint: DEBT.md → exports/debt.md (generated view)
  - doctor checkDebt + checkBrainDir: DEBT_FILE referansları kaldırıldı

Tests
  - rollback.test.ts: DEBT.md mock testleri kaldırıldı
  - sprint-reporter.test.ts: dosya tabanlı autoResolveDebt testi kaldırıldı
  - sprint-044-modules.test.ts: dosya tabanlı DEBT auto-resolve testi kaldırıldı
  - debt-db-accessor.test.ts: DB-first davranışları için yeni 179-satır test

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

`getDebtItems()` + `recordRollbackDebt()` moved from orchestra/debt-manager
to core/debt-store.ts so any layer (CLI, MCP, API) can read debt without
importing orchestra modules. `deckent init` no longer writes a stub DEBT.md;
`deckent archive-debt` rewritten as a DB-first read-only status reporter.
ADR-009 content updated to reflect saf DB-first (Task #4 AMENDED).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

5 dosya düzeltildi, 1 doğrulandı (sprint-lifecycle.md sorunsuz):

architecture.md:
- Node.js ≥18 → ≥24.0.0 (IDENTITY.md ile eşleştirildi)
- coverage<80 → <90 (result-evaluator.ts:87 threshold=90 doğrulandı)
- CLI 46 → 55+ (2 yer)
- orchestra 76 → 78 modül (ls src/orchestra/*.ts | wc -l)
- monitor/ 1 → 5 modül (alert-emitter, dashboard-manager, sprint-state, index eklendi)

agents.md:
- routing_min_agent_score → agent_min_score (config.ts:740 doğrulandı)
- saveTempAgent() → saveTempAgentToPool() (agent-pool.ts:214 doğrulandı)

agent-skill-architecture.md:
- AgentRole types.ts → monitoring-types.ts:8 (grep doğrulandı)
- Last verified Sprint 172 → 186

memory-system.md:
- DEBT.md Sprint 186'da kaldırıldı (Task #4); tree + bölüm + budget tablosu güncellendi
- Last updated Sprint 172 → 186

authority-matrix.md:
- §10: ADR-038 naming karışıklığı çözüldü (kabul edilen ADR-038 farklı içerikle)
- Planned Evolution → Evolution History (durum sütunu eklendi, Sprint 186 gerçeklerle)
- DECKENT→USER:NOTIFY "tanımlandı" → "uygulandı" (notification-dispatcher.ts kanıtı)
- Rule 4 + src/** tablosu: "ADR-038 pending" referansı temizlendi

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Final Memory V2 cleanup — removes the legacy single-file knowledge
stores now that every writer/reader is DB-first (Task #4 + B6/B7/B8):

- .brain/PATTERNS.md, .brain/DEBT.md (git-tracked) — removed
- .brain/MEMORY.md, .brain/RETRO.md, .brain/PROJECT-IDENTITY.md
  (untracked working-tree files) — removed

memory.db is the single source of truth; .brain/exports/ holds the
generated views. .brain/ERRORS.md is kept — the error log is still
file-based (not migrated to the DB).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…status anchor + queue preview

Closes the orientation findings (Alperen live test):
- #1 visible cursor: InputBar renders an inverse-video caret → you always SEE
  where you are. Reuses the tested editInput reducer + InputHistory.
- #4 arrow keys: ←/→/Home/End move the cursor, ↑/↓ walk history, Backspace/
  Delete/Ctrl-A/E/U edit — full line editing (was append-only).
- #2 'working vs done': a PERSISTENT status anchor above the input shows the
  phase every frame — ⠋ düşünüyor… / üretiliyor… (animated, so a mid-reply
  network pause still reads as 'working') / ✓ hazır · sıra sende when idle.
- #3 queue visibility: queued continuation prompts render as '⋯ kuyrukta N: …'
  previews while busy (was invisible).

InputBar owns input only while active (isActive=false during the confirm modal);
batched/pasted Enter is split so completed lines submit. PTY-harness verified
(scripts/ink-pty-test.mjs now supports <LEFT>/<CR>/<BS> tokens): edit + reply +
clean exit code 0. i18n: tui.generating/ready keys. Opt-in DECKENT_INK=1.
tsc+build+tests/cli green.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…n (not just approve/reject)

The user wanted real back-and-forth, not just approve/reject. Now any authorized
non-command message drives deckent's native agentic chat engine and the reply comes
back over Telegram.

Safety posture (advisor-led — god-level keeps the guardrails):
- Chat inherits the SAME chat_id authorization chokepoint as commands. The router's
  auth check now gates BOTH paths: an unauthorized sender reaches neither the
  resolver nor the chat engine (silent — no oracle, and with chat, no RCE).
- Slice 1 provider is the SUBSCRIPTION claude adapter (API key stripped → session
  auth): the model can converse but NEVER emits tool_use, so it cannot itself
  decide to run a tool. Recognized read-only intents (status/recall/history) still
  ground answers via agenticDispatch; risky actions are DENIED by the default
  confirm (classifyActionRisk safe→allow, risky→deny). Zero destructive surface.
- Slice 2 (next) swaps in a tool_use provider and routes risky actions through the
  BOT-002 approve gate (the agenticConfirm seam is the hook).

Pieces:
- incoming-command-router: optional onChat(channelId, text), called only for
  authorized non-command messages; throwing onChat never escapes the handler.
- chat-bridge (new): makeChatResponder wraps runChatNativeLoop into
  (sessionId, text) -> reply. Per-session SERIALIZATION (a stateful chat corrupts
  under concurrent/overlapping turns; different sessions stay concurrent),
  gracefulErrors (the bot always replies, never silently dies), and chunkMessage
  (Telegram ~4096-char limit; tool output/logs are split on newline boundaries).
- connector-bootstrap: wires onChat -> thinking-ack + chat + chunked reply, fail-safe.
- bot listen: builds the responder so the long-lived host is a full conversation head.
- i18n: bot.chat_thinking/empty/error (en+tr).

Memory decision (advisor blind-spot #4, verified not assumed): chat turns persist as
type:'chat' entries with decay_exempt:false — i.e. subject to decay and living in the
same entries table sprint-232 just hardened. They are type-isolated from learnings/ADR
(no recall pollution), but to avoid adding decayable churn there, slice 1 ships
STATELESS (each message an independent turn). Multi-turn memory is slice 2 with a
proper bounded chat-retention policy.

TDD RED->GREEN: router onChat 6 + chat-bridge 7 (serialize, cross-session concurrency,
graceful error, chunking) + bootstrap chat-wire 2 (authorized->chat+reply,
unauthorized->not called) = new; 103 connector+cli tests green; tsc clean.

LIVE PROOF (real, multi-turn): chat-enabled 'deckent bot listen' against the real
Telegram token; user sent two natural-language messages from the phone and got two
real agentic replies; the listener survived both turns and exited cleanly on SIGTERM.
User-confirmed.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ch (gap #4)

Wires policy-engine.ts evaluatePolicy (dormant seam, 0 callers) into the
live autonomous decision path — RBAC goes from advisory (ADR-037 V1.0)
to ENFORCED for machine-initiated work, behind a default-off flag:

- config: autonomous.rbac_policy { enabled: false, role: 'viewer' }
  (+ validation). Role default 'viewer' = deny-by-default: viewer lacks
  the 'execute' permission, so enabling the flag without raising the
  role hard-blocks autonomous execution.
- autonomous-runtime: PolicyGateDecision gains 'deny' — a HARD refusal
  that finishes the cycle 'denied' with no approval detour (park keeps
  its defer-to-human semantics).
- runtime-loop: buildEngineRuntime's policyGate consults evaluatePolicy's
  RBAC layer (role x Permission.EXECUTE x entry.tenant) for every
  entry-carrying trigger (backlog / work-generator / reactive) BEFORE the
  G2/G3 policy decision. This is the path the trusted-internal authority
  wrap intentionally bypasses — the policy gate is now real governance.

The sprint worker-spawn path stays advisory (Task carries no
requirements/capabilities yet — per the known caveat); enforcement lands
where capabilities exist.

TDD: 8 new tests (cycle deny 1, engine-wiring rbac 3, config 4);
tsc clean; 391 targeted tests green. MASTER-PLAN ENT-1 + F10-001 updated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ody-read verified

- 075 [BOTH]: all 6 evolution wires LIVE — but API is CLASS-BASED (AgentGenealogy/
  AgentRetirement/SpecializationDriftDetector/PromptRollback in src/agents/,
  consumed by promotion-pipeline/sprint-reporter/prompt-evolution); ADR function
  names are conceptual — correct proof-pattern is class-name level (recorded so
  future shallow greps don't false-flag); affinity + countModules verified
- 076 [BOTH]: all 4 parts verified; Part-B LIVE-PROVEN (UX-audit token-mint);
  KEY product finding — Path-A chat-backend EXISTS+ready → dashboard-Chat-HOLLOW
  fix narrows to FRONTEND-wire only (ChatPage→Path-A); F1-CB extension noted
- 077 [BOTH]: all 4 parts verified; live evolution — S249 real mixed-fleet
  (15 tasks/4 providers), ADR-078 8-provider runtime, F1-CB 3rd-party billing
- 078 [BOTH]: Part-A battle-tested (hermeticity trio; governance heir ADR-087);
  Part-B HALF — bootstrap+auth ✓ but resolveWithOverflow 0-CALLER DORMANT (W-K #4,
  folds into deferred dormant-sweep); Part-C live w/ API-name nuance (F5-008
  identity-mutation in promotion-pipeline:273, not 'applyAdaptation'); Part-D ✓
  + EnterprisePage read-first was BY-DESIGN (UX#6 = V2-CRUD gap); lint warning cosmetic

memory.db synced, exports regenerated.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…rame throw (SP-1 M2p2a review)

Final-review findings #2 (collision-free synthesized ids for id-omitting OpenAI-compat backends) + #4 (Anthropic mid-stream error frame must throw, not silently complete as success — CLAUDE.md no-silent-failure).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

… router)

"enable autonomous" / "show pending approvals" / "nervous status" (+ TR: "otonom
durum", "bekleyen onay") artık 'task' sınıflanır → deckent tool yoluna gider, host-AI
chitchat'e düşmez. Otomasyon özellikleri non-coder/asistan yüzeyinden erişilebilir.
TDD 1 yeni test (6 ifade, TR+EN), 14/14 yeşil.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…steps + MASTER-PLAN batch notu

#1: `deckent init` next-steps'i artık otomasyon yeteneklerini keşfedilebilir kılar
("Otomasyon (opsiyonel · varsayılan KAPALI · insan-onaylı): deckent autonomous enable /
deckent nervous enable"). Düşük-risk (informative-append, interaktif-prompt yok, davranış
değişmez). TR+EN, TDD 2 test, 185 init test yeşil, tsc temiz.

MASTER-PLAN §4G'ye make-usable batch (4/4 TAMAM) ilerleme notu eklendi (lossless, additive)
— kalan M/L item'lar (AutonomousPage, process first-class, remote-approval, 21 nervous handler)
işaretlendi.

make-usable batch (#1-#4) TAMAMLANDI.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>