flower
/
1

review · segments

TarkovAI dormant project review + prod recovery + Codex worker dispatch

claude 1833 events 3 segments frontend-fixes

segment 1 of 3

Assess dormant tarkovai project state and production server status

Done

Read untracked docs (channel_user_analysis_report, twitch-user-view-performance-optimization, fix-cleanup-empty-userids-damage, tmi-namereply-race-condition-investigation, patch-details) via Codex subagent. Probed production server with read-only SSH: determined it runs Envoyer-style deploys, scp'd hotfix files match local commit fd5f0cc byte-for-better, second commit e6e108e not deployed, AppleDouble litter present, July 2025 vendor tree without May 2026 security patches, config cache stale. Provided comprehensive review summary covering local git state, doc statuses, prod config differences, and open issues.

outcome

Full inventory: local repo state (2 unpushed commits, untracked docs), prod git-less release tree, hotfix file diff (8 files, 3 differ by exact delta of not-yet-deployed commit), operational services (Horizon, TMI cluster, Reverb, scheduler all up), site HTTP 200, and list of open issues (ghost users, N+1 perf, TMI NameReply race, CleanupEmptyUserIds damage, stale May-2026 releases).

next steps

—

key decisions

  • scp'd hotfix files should not be reverted; they are fully captured in local commits
  • Do NOT mutate prod during this session (read-only SSH)

open questions

—

2 weeks ago → 2 weeks ago

segment 2 of 3

Verify prod is fully recovered and reconciled with local repo

Done

After user said they had fixed prod, re-checked local git state (synced with upstream, 0 unpushed commits, HEAD at 29f7689) and prod state via SSH: current symlink points to fresh release 20260620144147, full git-based deploy pipeline restored, TwitchApiLog.php is 134 lines (new), rate-limit migration present and ran (batch [47]), composer.lock matches content-hash, AppleDouble litter gone, site HTTP 200, Horizon/TMI/Reverb/scheduler all restarted ~17 minutes ago with new code.

outcome

Prod == local HEAD fully: all checked files byte-identical, security patches live, rate-limit migration applied, deployments working, no remaining divergence.

next steps

—

key decisions

  • Confirmed prod recovery complete; no further action needed on reconciliation

open questions

—

2 weeks ago → 2 weeks ago

segment 3 of 3

Dispatch four Codex workers via Solo MCP to review untracked docs against codebase

In progress

Read _solo-playbook.md, confirmed readiness (Solo MCP tools, project tarkovai id=20, Codex agent tool id=4, 95 tools enabled). Created 4 spec scratchpads (scratchpads 892-895) with detailed investigation instructions, saved to files under docs/_review/. Created 4 tracking todos (526-529) with 'in_progress' status and PID comments. Spawned 4 Codex workers (PIDs 832-835), each directed to its own spec. Sent kickoff prompts with read-only guardrails and output protocol. Set idle-wake timer (timer 894) watching all 4 workers to handle approval prompts and detect completion. All four workers reported 'Working' status.

outcome

Four Codex workers launched and actively reviewing: 832 (namereply-race, todo 529), 833 (channel-user-inflation, todo 526), 834 (twitchuserview-perf, todo 527), 835 (cleanup-empty-userids, todo 528). Workers accept read-only prod SSH commands (auto-approved), reject mutating commands. Idle-wake timer armed for 15 min max wait.

next steps

  • Wait for idle-wake timer to fire when any worker completes or hits a prompt
  • Handle any approval prompts (approve read-only, reject mutating)
  • Once all four report READY, read each findings scratchpad and cross-check verdicts against codebase
  • Present consolidated results to user for action decisions

key decisions

  • Each worker gets a self-contained spec file (scratchpad + local file) for fallback if Solo MCP not available
  • Workers write findings to scratchpads (primary) or fallback files under docs/_review/<slug>-findings.md
  • Orchestrator (Claude) handles ALL approval prompts via idle-wake timer, never polls
  • Workers left up after reporting for potential follow-up actions

open questions

  • Will the Codex workers successfully access Solo MCP scratchpads/todos, or will they use file fallbacks?
  • Does the cleanup-empty-userids worker (835) find any remnant damaged user records in prod DB?
  • Will any worker attempt a mutating command that needs to be blocked?

2 weeks ago → 2 weeks ago