Optimize agent context

Optimize Agent Context

Use ContextLean as the deterministic measurement, evaluation, and rollback layer. Optimize relevance, freshness, trigger precision, and tool behavior; never equate shorter context with better results.

Workflow

1. Establish the target Agent, project root, current canonical state, quality priority, and representative tasks. Keep account UI instructions, global/project rules, Skills, runtime configuration, Context Packs, and provider-fixed context separate.

2. Run the bundled CLI read-only:

   node <skill-directory>/scripts/contextlean.mjs doctor --root <project> --json
   

   Use audit --scope repo when personal Agent configuration is out of scope. Use snapshot --write <file> before an optimization experiment; snapshots contain metadata and hashes, not file contents.

3. Report Fact / Inference / Proposal / Blocked. Verify findings against real files and runtime state. Treat byte-to-token conversion and thresholds as heuristics.

4. Select one reviewed surface at a time. Preserve privacy, permission, destructive-action, business, canonical-source, network, model, dirty-worktree, and recovery boundaries. Move only task-specific detail into on-demand Skills or references.

5. For candidate generation, frozen A/B tasks, adapters, blind review, or Context Packs, read protocols-and-evaluation.md. Confirm the selected target is non-sensitive and the adapter/provider data path is authorized. Keep held-out tasks hidden from the optimizer and require a qualified runner sandbox.

6. Run smoke tasks before the full frozen suite. Track deterministic success, blind non-inferiority, input/cache tokens, latency samples, tool calls, repeated reads, errors, retries, and side effects. Do not promote a candidate without a material quality or efficiency gain.

7. Use experiment select to create a compatible v1 plan only when one candidate passes the quality-first gate. Review the full replacement and current SHA-256; selection never applies automatically.

8. If the user authorizes the reviewed change, run:

   node <skill-directory>/scripts/contextlean.mjs apply --plan <plan.json> --yes
   node <skill-directory>/scripts/contextlean.mjs verify --receipt <receipt.json>
   

9. Roll back when validation fails:

   node <skill-directory>/scripts/contextlean.mjs rollback --receipt <receipt.json> --yes
   

Read safety-and-method.md before any apply or rollback operation.

Boundaries

• Default to read-only audit, metadata-only snapshots, and dry plans.
• Never inspect auth files, keychains, .env files, private session transcripts, or secret values.
• Never publish local reports or plans without a separate redaction review.
• Do not disable plugins, Skills, hooks, MCP servers, memory, or reasoning globally from one benchmark.
• Treat Context Packs as derived, on-demand caches; canonical sources always win and candidate packs require explicit evaluation mode.
• Do not promise model-side inference acceleration or “100% capability.” Optimize the controllable Agent harness and verify the quality/latency tradeoff.