Skip to content

execution-notes.md

Latest commit

 

History

History
93 lines (76 loc) · 5.77 KB

execution-notes.md

File metadata and controls

93 lines (76 loc) · 5.77 KB
Raw

execution-notes.md — sprint-4 M1+M2

Active /goal

GOAL-sprint-4-M1-M2.md (mailed to Dais via Resend id 0493d1f1-...).

Sub-feature status

# feature phase notes
a2 earnings-to-settle-mirror ✅ COMPLETE (Phase 6) Full pipeline PROVEN LIVE on production gig: p-1782887987 roi 0→25000
a1 LAYER C STARTUP prompt update ✅ DEPLOYED 2026-07-01 3 injections in gig-cli.sh: PRE-B1 CURRENT_PASS_ID binding from oldest tasks/*.json; B2 per-apply task-request-map.jsonl append; EARNED CHECK jq exact-match historical pass_id lookup + task-request-map.errors.jsonl fallback. Restarted per REQ-L3, --status=ALIVE < 30s. M2 auto-closes on first real Coconala 検収.
b earn-roi-reconciler ✅ COMPLETE Feature (b) sprint-4 done
c dispatcher-live-dormant ✅ COMPLETE Phase 6 converged + live E2E on prod gig; .slot_created marker deployed
d recipe-6-actions ✅ COMPLETE 6 real wires (kill_server/send_keys/login/npm_install/git_checkout/escalate_via_bot2bot); 503 tests GREEN; INV-P1/INV-4 preserved

Milestone gates

  • ✅ M1 (settle pipeline ready) — reconciler + mirror COMPLETE. Full flow PROVEN LIVE.
  • ⏳ M2 (first real ¥) — pipeline FULLY WIRED including (a1) STARTUP update. Waiting on first real Coconala 検収 in production. Loop is LIVE via launchctl ai.anicca.gig-proactive (5-min tick) + hourly reconciler menu item.

Regression baseline

503/503 tests GREEN.

Block conditions

  1. No settle event in 30 days across ANY slot
  2. INV regression uncloseable in 3 iters
  3. crypto primitive fails

sprint-4 (a1) post-deploy notes (2026-07-01)

  • (a1) STARTUP deployed successfully; gig-cli.sh --status=ALIVE
  • task #6 hook fix: ~/.claude/settings.json PreToolUse:Bash hook was rtk hook claude (PATH-dependent). Headless tmux sessions couldn't resolve rtk → repeated "PreToolUse:Bash hook error". Fixed by pinning to /opt/homebrew/bin/rtk hook claude. Post-fix: session runs clean, no more hook errors.
  • First post-fix pass: correctly detected concurrency (multiple restarts + healthcheck cron) and skipped browser-driving to avoid collision. Registered cron 52b154a2 for future ticks. This is the CORRECT anti-collision behavior spec'd in HARD 0.36's INV-4.
  • task-request-map.jsonl materialization: waiting for first uncontested B2 apply pass (next hourly cron fire).
  • M2 auto-close path: unchanged. First real Coconala 検収 → gig-cli.sh a1 lookup → earnings row w/ pass_id → (a2) mirror → (b) reconciler → roi_jpy_realized > 0 → M2 satisfied.

task #2 diagnosis (2026-07-01)

  • ~/clips/queue/ is empty because it's fed by producer.sh (creates mp4+caption) while the proactive-loop dispatcher enqueues produce-clip task descriptors to ~/loops/clip/tasks/ (34 descriptors queued, 5-min tick, none consumed).
  • The clip LAYER C session's STARTUP (clip-cli.sh:21) only invokes run.sh (POSTS from queue), never producer.sh (FILLS queue). No consumer bridges ~/loops/clip/tasks/ → producer.sh.
  • FIX (out of scope for M2 sprint-4, deferred to sprint-5): mirror the gig-cli.sh a1 pattern in clip-cli.sh STARTUP:
    1. Read oldest ~/loops/clip/tasks/*.json
    2. If picked.name == "produce-clip", invoke producer.sh with the task's platform/params
    3. Move the consumed task descriptor to ~/loops/clip/tasks/done/
  • Impact: without this fix, produce-clip tasks pile up indefinitely and no clips are ever posted. NOT a M1/M2 blocker (gig is the M1/M2 primary slot).
  • Sprint-5 candidate feature: clip-cli.sh a1-equivalent for task-descriptor consumption + producer.sh wire.

task #5 status (山本さん #5123100 あい庵 SNS ¥40k/月) — CLOSED/lost

Full timeline from gig data (2026-06-30 → 07-01):

  1. Applied w/ site-specific 3-improvement proposal — buyer replied "契約手続きを進めたい"
  2. Formal 見積り sent via direct_offer/4857277 (¥40,000/月, 定期購入, 期限 7/7)
  3. Two follow-ups sent (2026-06-30 23:38, 07-01 01:17)
  4. Result: 公開募集終了 + direct_offer/4857277 → 404 (offer link dead) → outcome=ignored_closed, lesson: "高額見積りは決断を促す締切設定が必要"

M2 candidate: NO (deal closed). The gig-cli.sh a1 pipeline still stands ready for the next buyer that reaches 検収 stage. Cron 52b154a2 next :27 tick continues the discovery.

task #6 follow-up: hook wrapper (2026-07-01, deeper root cause)

  • Absolute-path fix (/opt/homebrew/bin/rtk hook claude) alone was NOT enough.
  • Continued observing "Failed with non-blocking status code: node:internal/modules/cjs/loader:1458" in gig session — Claude Code's own internal Node.js error trying to interpret rtk's empty stdout on non-rewrite pass-through cases.
  • FIX: /Users/anicca/.claude/hooks/rtk-hook-wrapper.sh (installed + settings.json pointed at it) — always emits valid JSON:
    • rtk rewrite? → forward rtk's JSON
    • rtk silent? → emit {"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"allow"}}
  • Will take effect on next gig-cli restart (currently: 28min-subagent completed with 0 applies but tmux still ALIVE; no active hook errors right now).

task #6 further follow-up: wrapper newline (2026-07-01, +30min)

  • After installing rtk-hook-wrapper.sh, hook errors REDUCED but not to zero (dropped from 8+/500lines to 3/500lines).
  • Root cause: wrapper output was missing trailing newline (printf '%s' → printf '%s\n'). Some Claude Code hook parser paths need line-terminated output.
  • After fix + gig-cli --restart: session progressing through Coconala gig pass 45 with proper phase structure (B1 nurture → B2 apply → B3 learn → B4 improve → B5 share → finalize + .last-pass).
  • Expected: task-request-map.jsonl materialized during B2 apply; task #1 verified on B5+finalize.