Cache-Efficient Mode

Non-negotiable for this repo: maximize cached knowledge; minimize tokens.

Load (do not skip)

Read manifest token_policy first. Spine (free): README index, .codebase-scan.txt, latest TODO, INDEX.md.

Capped loads (max_cache_files_default, default 2):

1. ≤2 files from docs/codebase/* — grep headings first, then section Read only.
2. ≤1 INDEX-selected memory (counts toward cap if 2 codebase docs already loaded).

Optional: IMPLEMENTATION_SUMMARY.md / BRANCH_ANALYSIS.md — one section only if task requires; prefer grep.

Hard rules:

• Grep before Read on registry, skills, or any file >80 lines.
• No application source until user confirms direction (no_source_until_confirmed) or names one exact path.
• At ≥100k session context (session_refresh_context_tokens), recommend .github/skills/chain/SKILL.md session-start in a fresh thread.

Respond

• Default cap: ≤120 words unless user asks for detail, code, or a plan.
• Format: bullets; one line per finding; cite path or path:line — no large code fences.
• Reference cache by name (“per CONCERNS §…”, “per ndestates-io-cache”) — do not quote paragraphs.
• End with Next (1–3 numbered actions) when helpful.
• Orchestrator JSON / handoff schemas: still valid when that skill requires them — keep JSON compact.

After load, confirm in one line

Example: Cache: INDEX + codebase-cache + esign memory + CONCERNS(§signing). Fresh. Ready.

Token meter (always-on)

When token_policy.report_tokens_estimate: true (manifest default), end every substantive response with the one-line footer from /token-usage-meter:

1. Run python3 .github/skills/token-usage-meter/scripts/token_monitor.py --sync --project . once per turn.
2. Append the printed line: Token meter: ctx … · +… this turn · … turns · STATUS.
3. Surface any WARNINGS above the footer.

Skip only for trivial acks or when user opts out (no token meter).

Chaining

• Auto-compose: .github/skills/chain/SKILL.md <intent> — discovers skills from chains/registry.yaml, one shared cache load, minimal handoffs (see CHAIN.md).
• Standup: .github/skills/chain/SKILL.md session-start or .github/prompts/daily-standup-with-cache.prompt.md then stay in cache-efficient tone.
• Deep work: user says “go deep” or .github/skills/chain/SKILL.md complex-task / .github/prompts/orchestrator-v2.prompt.md — expand stepwise, still cache-first per step.
• For full .github/prompts/read-codebase.prompt.md or major changes: start with cache-efficient load, then expand only as needed.