Skip to content

context-optimization-plan.md

Latest commit

 

History

History
70 lines (59 loc) · 3.18 KB

context-optimization-plan.md

File metadata and controls

70 lines (59 loc) · 3.18 KB
Raw

Potential Context Optimization Plan

Status: proposed for later review, not approved for implementation. Captured: 2026-07-03 Asia/Shanghai

Summary

Optimize Stillcurrent first by adding a small tracked resume index that becomes the default startup state for $resume-from-handoff and $studio-next. Keep src/README.md as full slice history, but stop reading it in full during ordinary resume and next-action routing.

Key Changes

  • Add tracked production/resume-index.md, target <=10 KB, maintained by $handoff.
  • Index fields: current slice version, real/stubbed one-liner, last reported or verified boot/playtest, owed verification, best next action, top 2-3 lanes, blockers/gates, evidence pointers, and source freshness.
  • Change $resume-from-handoff default to "light resume": read resume-index.md, session-handoff.md, sprint/stage/active state, and only a bounded top/current section of src/README.md for freshness checks.
  • Add explicit deep mode for $resume-from-handoff deep, historical questions, missing/stale index, or user-requested full catch-up.
  • Change $studio-next to read production/resume-index.md instead of full src/README.md.
  • Update $handoff to refresh production/resume-index.md after live handoff updates, verify its byte size, and commit it with the handoff when commits are authorized.
  • Update AGENTS/context docs to remove 1M/750k assumptions and use the observed practical budget: compact earlier, hard-handoff earlier, and keep default orchestration reads bounded.

Behavior Rules

  • production/session-handoff.md remains the canonical live narrative.
  • src/README.md remains canonical full slice history, but is no longer default resume context.
  • production/session-archive.md remains historical-only and is not read by default.
  • If the index is stale or mismatches the bounded current README header, the skill reports that plainly and either uses bounded state or asks for explicit deep resume.
  • If dirty local work has a newer slice header than the index, resume labels it as in-flight local state rather than silently treating the index as complete.

Test Plan

  • Confirm Stillcurrent branch is codex-game-studios-deployment.
  • Verify production/resume-index.md and production/session-handoff.md stay under size targets with wc -c.
  • Run Stillcurrent audits: ./.codex/audit.sh all --root "$PWD" and ./.codex/audit.sh skills --root "$PWD".
  • Search skill/docs text to confirm no default full-read instruction remains for src/README.md.
  • Prompt smoke: $resume-from-handoff should report state from the index and bounded freshness checks, then offer structured lanes.
  • Prompt smoke: "what should I do next?" should route through $studio-next and avoid full history loading.
  • Stale-index scenario: simulate missing/stale index and verify the skill labels it and falls back only with explicit deep intent.

Assumptions

  • First rollout is Stillcurrent only.
  • The index is tracked under production/, not ignored production/session-state/.
  • Splitting src/README.md is deferred unless the lightweight resume path is still too expensive.
  • No .claude/ or CLAUDE.md dependency is introduced.