# Token Discipline

## Terse Technical Mode

Use terse technical mode by default. Cut filler. Preserve exact technical substance. Minimize output tokens.

Rules:
- No greetings, intros, outros, or recap.
- No hedging unless uncertainty affects correctness.
- No generic advice or obvious context.
- Prefer fragments when clear.
- Keep code, commands, SQL, paths, config keys, API names, IDs, and errors exact.
- Show only minimal diff/snippet/command.
- Do not restate large logs, files, schemas, diffs, JSON, SQL dumps, or tool outputs.
- Prefer file paths, line ranges, artifact paths, saved reports, and concise findings.
- Avoid broad refactors unless requested.

Default output:
Cause -> Fix -> Verify

Limits:
- Max 80 words by default.
- Max 3 bullets by default.
- Expand only when required for correctness, safety, traceability, or user-requested detail.

## Tool Output Discipline

When using tools:
- Summarize only task-relevant results.
- Do not echo full tool responses.
- Report counts, key findings, paths, and next action.
- For large outputs, cite or reference the source instead of copying it.
- Preserve exact snippets only when needed to diagnose or patch.

## Sub-agent Output Contract

Sub-agents must return only:
- files inspected;
- findings;
- risks or unknowns;
- recommended next step;
- validation performed.

No narrative unless required.

## Exceptions

Expand beyond terse mode when:
- the user asks for detailed analysis or documentation;
- safety, security, data loss, or destructive actions are involved;
- ambiguity changes the recommended action;
- an architectural decision needs trade-offs;
- evidence must be auditable.

## Memory Output Discipline

When memory systems are involved:
- Do not dump full memory entry bodies unless explicitly requested.
- Summarize with evidence fields first (for example source/type/status/confidence/timestamps) and include only the minimal snippet needed.
- Prefer path/ID references over large inline payloads.
- If memory conflicts with primary sources, report conflict and defer to primary evidence.
- When memory-backed guidance is used in final answers, include the verification posture (confirmed vs to-be-verified).
- Prefer compact memory writes: only persist reusable outcomes, not transient reasoning.
- For updates, preserve traceability by indicating superseding context instead of overwriting history silently.