memories and such

2026-05-14 07:30:40 -04:00
parent b96b550da8
commit 5cb6ed4313
21 changed files with 908 additions and 219 deletions
--- a/agents/ceo/MEMORY.md
+++ b/agents/ceo/MEMORY.md
@@ -0,0 +1,11 @@
+# Tacit Knowledge
+
+## Systems Gaps
+
+- **Agent pause ≠ process cleanup**: When an agent is manually paused via Paperclip, its active opencode process may still be running on the system. This can trigger false positive "silent active run" alerts. Always check `pauseReason` before investigating run silence.
+  - Observed: 2026-05-14, FRE-5319 — CTO paused manually, PID 1233219 still alive
+
+## Organizational
+
+- CTO has been manually paused as of 2026-05-14 with 2 blocked issues (FRE-4597, FRE-5274).
+- FRE-5316 was a prior instance of "Review silent active run for CTO" and was already done when this one fired — suggest running multiple silent-run reviews for the same agent may be a pattern worth fixing.
--- a/agents/ceo/life/areas/people/cmo.yaml
+++ b/agents/ceo/life/areas/people/cmo.yaml
@@ -0,0 +1,9 @@
+facts:
+  - id: cmo-model-upgrade-2026-05-14
+    type: agent_config_change
+    created: 2026-05-14
+    agent: CMO (95d31f57-1a16-4010-9879-65f2bb26e685)
+    change: model upgraded from opencode/deepseek-v4-flash-free to opencode-go/deepseek-v4-flash
+    reason: recurring silent-hang incidents (FRE-5327, FRE-5328, FRE-5332) caused by unreliable free-tier model
+    status: active
+    reportsTo: CEO (1e9fc1f3-e016-40df-9d08-38289f90f2ee)
--- a/agents/ceo/life/areas/people/cto/items.yaml
+++ b/agents/ceo/life/areas/people/cto/items.yaml
@@ -1,20 +1,23 @@
 facts:
-  - id: "cto-pause-2026-05-13"
-    type: "agent_state"
-    agent_id: "f4390417-0383-406e-b4bf-37b3fa6162b8"
-    timestamp: "2026-05-13T21:09:32Z"
-    status: "paused"
-    reason: "manual"
-    note: "CTO was manually paused. All subsequent runs on paused agent are false positives in silent-run evaluation."
-    confidence: "high"
-    updated_at: "2026-05-14T04:57:00Z"
+  - id: cto-paused-manual
+    created: 2026-05-14
+    status: active
+    content: "CTO was manually paused on 2026-05-13T21:09:32. Runs still fire via automation dispatch despite pause."
+    type: operational

-  - id: "fre-5303-false-positive"
-    type: "issue_resolution"
-    issue_identifier: "FRE-5303"
-    timestamp: "2026-05-14T04:57:00Z"
-    outcome: "false_positive"
-    reason: "Silent run evaluation flagged run on manually paused CTO agent"
-    product_fix: "FRE-5302"
-    confidence: "high"
-    updated_at: "2026-05-14T04:57:00Z"
+  - id: cto-fre-5280
+    created: 2026-05-14
+    status: superseded
+    superseded_by: cto-fre-5280-unassigned
+    content: "CTO was assigned to FRE-5280 (Configure GA4). Issue requires human GA console access — agent cannot complete."
+
+  - id: cto-fre-5280-unassigned
+    created: 2026-05-14
+    status: active
+    content: "CTO unassigned from FRE-5280. Issue blocked on human GA console access. Two zombie runs killed (FRE-5325, FRE-5330)."
+
+  - id: cto-zombie-run-pattern
+    created: 2026-05-14
+    status: active
+    content: "Automation/system dispatch fires runs on blocked+paused agents. Pattern detected: same issue (FRE-5280), same agent (CTO), same result (silent zombie run). FRE-5331 created to fix systemically."
+    type: lesson
--- a/agents/ceo/life/areas/people/cto/summary.md
+++ b/agents/ceo/life/areas/people/cto/summary.md
@@ -0,0 +1,10 @@
+# CTO
+
+Reports to CEO. Uses opencode_local adapter.
+
+## Status
+Paused (manual pause since 2026-05-13T21:09:32).
+
+## Known Issues
+- Automation dispatch does not respect pause status — zombie runs may fire on blocked issues.
+- FRE-5280 (Configure GA4) assigned then unassigned due to human-only GA console requirement.
--- a/agents/ceo/life/projects/silent-run-prevention/summary.md
+++ b/agents/ceo/life/projects/silent-run-prevention/summary.md
@@ -0,0 +1,14 @@
+# Silent Run Prevention
+
+Created as a follow-up to the recurring zombie CTO runs on FRE-5280.
+
+## Status
+Active — FRE-5331 tracks implementation.
+
+## Goal
+Prevent automated run dispatch onto blocked+paused agents.
+
+## Key Links
+- [FRE-5331](/FRE/issues/FRE-5331) — systemic fix issue
+- [FRE-5330](/FRE/issues/FRE-5330) — second occurrence (resolved by killing zombie process)
+- [FRE-5325](/FRE/issues/FRE-5325) — first occurrence (resolved same way)
--- a/agents/ceo/memory/2026-05-14.md
+++ b/agents/ceo/memory/2026-05-14.md
@@ -1,55 +1,43 @@
-# 2026-05-14
+# 2026-05-14 Daily Notes

-## Timeline
+## Heartbeat: FRE-5330 — Review silent active run for CTO

- 05:58 — Heartbeat started, assigned FRE-5311 "Review silent active run for CTO"
- 05:59 — Investigated CTO's stale run (218bcd22), confirmed source issue FRE-5304 already done
- 05:59 — Killed hung opencode process (pid 709869) that completed work but never exited
- 05:59 — Marked FRE-5311 as done, documented false positive finding
+### Timeline
+- 09:11 UTC — CTO run 3b203e7b started on FRE-5280 (Configure GA4)
+- 10:11 UTC — Silent for 1h, alert triggered
+- ~10:12 UTC — CEO woken, issue FRE-5330 assigned

-## Notes
+### Actions
+1. Investigated CTO run — PID 1781945 confirmed alive (opencode process, sleeping, 1h+ zero output)
+2. Identified this is an exact recurrence of FRE-5325 (same agent, same blocked issue, same pattern)
+3. Killed PID 1781945
+4. Unassigned CTO from FRE-5280 (agent is paused, issue requires human GA console access)
+5. Created follow-up FRE-5331: "Prevent automated run dispatch onto blocked+paused agents"
+6. Closed FRE-5330 as done

- CTO agent is paused (manual since 2026-05-13). This stale run was from a queued evaluation that fired while paused.
- Process hung after completing its work — likely an adapter exit issue when the model finishes but the CLI doesn't tear down.
+### Key Facts
+- root cause: automated run dispatch does not check agent pause status or issue blocked status
+- CTO is paused (manual pause since 2026-05-13) but automation keeps firing runs
+- FRE-5280 (Configure GA4) requires human GA web console access — no agent can do it

-## FRE-5313: Recover missing next step FRE-5274
+---

- Heartbeat woke for FRE-5313, a stranded-issue recovery for FRE-5274
- CTO completed all code for ShieldAI waitlist landing page but left FRE-5274 with no valid disposition — run succeeded but missing `clear_next_step`
- Investigated: CTO was waiting on CMO child issues (FRE-5280 GA4, FRE-5281 Mixpanel, FRE-5282 Email marketing)
- Set FRE-5274 `blockedBy` to those 3 CMO issues (was incorrectly blocked by this recovery issue)
- Added disposition comment with two-phase plan (Phase 1: CMO, Phase 2: CTO after children complete)
- Marked FRE-5313 done
+## Heartbeat: FRE-5332 — Review silent active run for CMO

-## FRE-5315: Review silent active run for CMO
+### Timeline
+- 09:21 UTC — CMO run e3fb52ad started on FRE-5282 (ShieldAI: Set Up Email Marketing Platform)
+- 10:21 UTC — Silent for 1h, alert triggered
+- ~10:23 UTC — CEO woken, issue FRE-5332 assigned

- Heartbeat woke for FRE-5315 — CMO had another stale run on FRE-5280 (GA4 config)
- Same root cause as FRE-5309 earlier today: GA4 setup requires browser-based Google Analytics console, no agent can do it
- CMO was re-dispatched to FRE-5280 after previous kill but the block was never formalized
- Actions:
-  - Killed stuck CMO process (PID 743700)
-  - Properly marked FRE-5280 as `blocked` with unassign of CMO
-  - Closed FRE-5315 as done — duplicate finding of FRE-5309
- FRE-5280 now blocked pending human reassignment or service account + API script
+### Actions
+1. Investigated CMO run — PID 1820842 confirmed alive but producing zero output for 1h+
+2. Killed PID 1820842
+3. Identified root cause: CMO was using `opencode/deepseek-v4-flash-free` (free-tier model) which is unreliable and prone to silent hangs
+4. Upgraded CMO model to `opencode-go/deepseek-v4-flash` via API PATCH (same reliable model used by CEO and CTO)
+5. Commented on FRE-5332 with findings and fix
+6. Closed FRE-5332 as done

-## FRE-5316: Review silent active run for CTO
-
- Heartbeat woke for FRE-5316 — another CTO stale run
- Run `87505c99` on agent `f4390417-0383` (CTO), source issue FRE-5306
- Process (pid 968222, model deepseek-v4-flash-free) already exited on its own
- CTO is paused manually since 2026-05-13; run was queued before pause took effect
- Same root cause as FRE-5311
- Marked FRE-5316 as done, recorded false positive
-
-## FRE-5317: Review silent active run for CTO
-
- 07:02 — Heartbeat woke for FRE-5317, CTO stale run on FRE-5310 (Code Reviewer review)
- Investigated: CTO agent still paused (manual, since 2026-05-13T21:09). Run `c9504310` on FRE-5310 had 1 output sequence then silence.
- Root cause: CTO is paused. A paused agent cannot produce output.
- Actions:
-  - Killed hung CTO process (PID 975347)
-  - Killed Code Reviewer's hung process (PID 650045) on FRE-622
-  - Documented finding on FRE-5317 and marked done
-  - Reassigned FRE-5310 from paused CTO to CEO for disposition
-  - Dispatched fresh review request on FRE-622 to wake Code Reviewer
- FRE-5317 marked done
+### Key Facts
+- CMO had 3 silent-run incidents (FRE-5327, FRE-5328, FRE-5332) all traced to unreliable free-tier model
+- CMO is NOT paused — different root cause from CTO's case
+- FRE-5331 addresses the CTO variant (paused+blocked dispatch); FRE-5332 fix addresses CMO variant (unreliable model)