[evolve] 14-day retrospective (Apr 4–17) — findings from CLI session 2026-04-18

[verkyyi]

Context

One-off CLI retrospective pass across state/usage_log.md (133 entries), state/research_log.md (152 entries), state/agent_log.md, and closed-issue / merged-PR history for the 14-day window 2026-04-04 → 2026-04-17. Analysis done interactively in a human-directed session rather than by evolve.yml.

Purpose: capture findings before they scroll out of context. Individual actionable findings are filed as separate issues (linked below).

Top-line numbers

• Cost: $173.62 over 8 days of data ($21.70/day avg, annualized ~$152/wk — slightly above $150/wk target)
• Commits: 889 total, 862 state: (97%), 27 non-state
• Issues closed: 13 (bot-authored fix PRs + human direct pushes)
• PRs merged: 13 (11 bot, 2 human) — 85% self-heal ratio
• Median bot-PR merge time: 1.6 minutes

New findings not previously surfaced

1. Opus is 100% unreachable since 2026-04-13T20:51Z

→ filed as issue #175 [pipeline] Opus unreachable since... 100% Haiku fallback for 4+ days.

project_state.md reports "26.8% Haiku dominance"; actual daily breakdown shows 100% Haiku for 4 consecutive days. Needs account/billing investigation.

2. Haiku fallback is NOT cost-effective on this workload

Per-turn cost on analyze/evolve/watcher: Haiku is 5-22% more expensive than Opus and takes similar or more turns. The fallback's "saves money" premise is empirically wrong here. (Part of #175.)

3. growth.yml saturates max-turns 9/9 runs, produces 0 actions

→ filed as issue #176 [pipeline] growth.yml saturates max-turns on 9/9 runs....

Either cap is too low OR scope is too broad. Hypothesis B (scope too broad) is the working theory: growth is doing surveillance when it should be acting.

4. watcher and reviewer are radically over-capped

→ filed as issue #177 [pipeline] Reduce watcher (50→35) and reviewer (45→25)....

Watcher hits cap 0/67 runs (utilization 56%). Reviewer hits cap 0/6 runs (utilization 34%). Caps should be tightened so they actually bind.

5. PATTERN_HUNT posture has 23 consecutive 0-yield runs

→ filed as issue #178 [evolve] Suspend/demote PATTERN_HUNT posture....

All 4 postures show 0-yield across 21 evolve runs in the window. PATTERN_HUNT is the most expensive and most repeatedly empty.

6. Docs-staleness is a repeating cycle (#156 after #155, #168 after #165)

→ filed as issue #179 [evolve] Add README regen as coder post-step for cron-cadence changes....

Each cron change produces a follow-up "README stale" finding 1-4 days later. ~$3-5 churn per cycle. Coder prompt should sync README frequency claims in the same PR.

7. Research yield is 4.6% (7 issues from 152 log entries)

Not actionable on its own, but grounds the "plateau" self-reports in data. Highest-yield source: anthropics/claude-code (5/8 adoptable). Most named sources yielded <1 action per week.

8. #173 was the outlier in resolution time (30h vs 4h median)

The system's three systemic failure modes converged on this issue: (a) scope too big for single-pass coder, (b) Haiku at max-turns, (c) workflow YAML edits need WORKFLOW_PAT + PR review. Worth considering a coder chunking strategy for >5-file changes.

Already-tracked items (not refiled)

• [pipeline] Stagger evolve and analyze cron schedules to prevent concurrent state file conflicts #172 — cron stagger for evolve/analyze race condition (still open, untriaged 40h+)
• build(deps): bump actions/setup-node from 4 to 6 #133 / build(deps): bump actions/checkout from 4 to 6 #135 / build(deps): bump actions/deploy-pages from 4 to 5 #136 — Dependabot PRs awaiting human merge 16+ days
• [growth] Submit to hesreallyhim/awesome-claude-code (36.9K stars) #22 / [growth] Submit to e2b-dev/awesome-ai-agents (26.7k stars) #48 / [growth] Submit to EvoMap/awesome-agent-evolution (agent evolution curated list) #149 — growth distribution channels blocked on human action 26d+

Filed as new issues

• [pipeline] Opus unreachable since 2026-04-13T20:51Z — 100% Haiku fallback for 4+ days #175 Opus unreachability (needs-human)
• [pipeline] growth.yml saturates max-turns on 9/9 runs, produces 0 actions — cap/scope mismatch #176 growth.yml saturation + scope
• [pipeline] Reduce watcher (50→35) and reviewer (45→25) max-turns — over-capped, never saturate #177 watcher/reviewer max-turn caps
• [evolve] Suspend/demote PATTERN_HUNT posture — 23 consecutive 0-yield runs, plateau confirmed empirically #178 PATTERN_HUNT suspend
• [evolve] Add README regen as coder post-step for cron-cadence changes — eliminates docs-staleness loop #179 README regen on cron changes

Prioritization suggestion

If re-enabling workflows and picking targets:

1. [pipeline] Opus unreachable since 2026-04-13T20:51Z — 100% Haiku fallback for 4+ days #175 — distorts every cost assumption, affects all other findings
2. [pipeline] growth.yml saturates max-turns on 9/9 runs, produces 0 actions — cap/scope mismatch #176 — pure waste stream, no action produced
3. [evolve] Suspend/demote PATTERN_HUNT posture — 23 consecutive 0-yield runs, plateau confirmed empirically #178 — free savings, zero risk
4. [pipeline] Reduce watcher (50→35) and reviewer (45→25) max-turns — over-capped, never saturate #177 — tail variance reduction, zero risk
5. [evolve] Add README regen as coder post-step for cron-cadence changes — eliminates docs-staleness loop #179 — eliminates a recurring cost loop

Not fileable as issues (observations only)

• State commit churn is high (97% of commits are state:) but not broken — just noise for humans doing git log
• Hourly distribution clusters correctly at cron-trigger times — no anomaly
• Failed-run rate is low (~5/130 runs) and all categorized

---

Generated from CLI retrospective session 2026-04-18, not evolve.yml. See individual child issues for actionable specifics.

[verkyyi]

Closing — repo pivoted away from the self-evolving scaffold. Current direction: docs/spec.md (harness for safe unattended claude -p). Legacy scaffold remains on branch feat/self-evolving-scaffold.