The Claude + Codex Local Stack Playbook That Cut One Loop by 78.5%

Save this runbook if you use Claude + Codex locally.

On February 19, 2026, one measured pass cut zsh -i -c exit from 1.794s to 0.386s(-78.5%). [E1]

Everyone says "fix prompts"; in practice, this is a systems issue, and teams without baselines + retrieval gates ship false confidence.

This is the same repeatable 5-step sequence I now reuse: baselines, gates, memory tiers, retrieval benchmarks, and drift checks.

Evidence Snapshot (Audited February 20, 2026)

• launch agents loaded: 8/8 [E2]
• daily memory coverage: 108 files (2025-09-27 through 2026-02-20) [E3]
• extracted totals: 7596 prompts, 22858 tool calls [E3]
• benchmark A (2026-02-20T04:16:45Z): hit_rate=1.0000, mrr=0.7500,ndcg=0.8143, precision=0.1250, recall=1.0000, triage=0 [E4]
• benchmark B (2026-02-20T04:56:59Z): hit_rate=0.9500, mrr=0.7125,ndcg=0.7727, precision=0.1188, recall=0.9500, triage=2 [E5]
• latest benchmark (2026-02-20T06:58:19Z): dense model nomic-embed-text:latest, vector_dim=768, dense enabled, gates passing [E6]

5-Step Saveable Implementation

Step 1: Baseline First

Measure before any config edits.

• zsh -i -c exit: 1.794s -> 0.386s (-78.5%) [E1]
• codex --help: 0.210s -> 0.116s (-44.8%) [E1]
• claude --help: 0.464s -> 0.312s (-32.8%) [E1]

Step 2: Add Reliability Gates

Replace reminder-based quality with execution checks:

• preflight checks
• apply + verify loops
• browser validation for UI work
• explicit done criteria

Step 3: Split Memory into 3 Tiers

• Tier 1: raw session history
• Tier 2: curated working memory
• Tier 3: hybrid retrieval + graph layer

This prevents noisy logs from polluting high-signal memory.

Step 4: Benchmark Retrieval Like Infra

Track:

• Hit@K
• MRR@K
• NDCG@K
• Precision@K
• Recall@K
• triage issue count

Back-to-back snapshots caught quality drift (triage: 0 -> 2) while gates remained green. [E4][E5]

Step 5: Fix Drift in Ingestion Paths

Incident found:

• active drafts in ~/docs/content/drafts
• sync defaults still pointed to old root

Fix:

• set canonical drafts root
• keep legacy remap compatibility
• re-sync and rebuild index

Outcome: graph alignment became explicit and auditable. [E7]

Copy-Paste Run Sequence

~/control/automation/ai-dev-optimization/scripts/manage_knowledge_launchagents.sh status python3 ~/control/automation/ai-dev-optimization/scripts/run_snapshot.sh 24 python3 ~/control/automation/ai-dev-optimization/scripts/sync_article_memory.py python3 ~/control/automation/ai-dev-optimization/scripts/build_memory_rag_index.py --backend hybrid --allow-sparse-fallback --output ~/control/knowledge/knowledge-memory/data/rag/hybrid-index.json python3 ~/control/automation/ai-dev-optimization/scripts/benchmark_memory_retrieval.py

Weekly Cadence (Use This As a Checklist)

1. Verify launch agents are loaded.
2. Refresh extraction + sync.
3. Rebuild retrieval index.
4. Run benchmark and compare triage delta.
5. Promote only high-signal memory updates.

Failure Patterns to Watch

• Health checks green, but capability degraded.
• “One memory file” turning retrieval into noise.
• Path drift between active drafts and sync roots.

Conclusion

The point of this setup is simple:

• less waiting
• less guesswork
• fewer silent regressions

Prompt quality still matters, but the durable gains came from instrumentation and operating discipline.

Use this as a working runbook before your next sprint, then iterate the thresholds and checks to match your stack.

Evidence key

Every [E] marker above is a first-party measurement from my own machine, not a published benchmark. They are reproducible in method, not in absolute numbers: your hardware, shell config, and corpus will move every figure. Treat the deltas as the claim and the absolutes as context.

• [E1] — Shell and CLI startup timing, measured 2026-02-19. Before/after pairs from zsh -i -c exit, codex --help, and claude --help, taken on the same machine in the same session, around one config pass.
• [E2] — macOS launch-agent load check, 2026-02-20: 8 of 8 agents loaded.
• [E3] — Memory-tier extraction totals, 2026-02-20: 108 daily files spanning 2025-09-27 to 2026-02-20, yielding 7,596 prompts and 22,858 tool calls.
• [E4] — Retrieval benchmark A, run 2026-02-20T04:16:45Z.
• [E5] — Retrieval benchmark B, run 2026-02-20T04:56:59Z. Compared against [E4] to surface drift; the triage count moved 0 → 2 while gates stayed green.
• [E6] — Latest retrieval benchmark, run 2026-02-20T06:58:19Z, on nomic-embed-text:latest at vector_dim=768 with dense retrieval enabled.
• [E7] — Graph-alignment pass following the same run, checked by inspection rather than by a scored metric.

Single-machine numbers. If you run this sequence and get different absolutes, that is the expected outcome — the point is that you now have baselines to compare against at all.