map-token-context

Capsule

Use one capsule; put action fields in payload.

• With a virtual router, call skills {operation:"route",query:"English intent"}; its default is one domain-anchored match. Read only that skill_file; request a larger limit only when the task truly needs another specialist.
• Use cognition for exact cover, assignment, knapsack, shortest path, DAG scheduling, weighted decisions, information-gain checks, reasoning-governor status, or reusable decision kernels; trust its checksum certificate and reason only about unencoded judgment.
• Use run|batch for commands, file for files, index|search|remember for context, fetch for web text, and execute for derivation.
• Use progressive expand|diff for exact evidence and stats|gain|insight|doctor for measurement.
• Treat [Capsule replay], [Capsule delta], [Capsule poll: exactly unchanged], and [Capsule session transcript projection] as complete control tags: reuse them directly and call expand only when exact bytes are required. The current exact= capsule always contains the complete latest result.
• Treat [Capsule repeated failure] as proof that the same tool input produced the byte-identical error again. Use its short diagnostic, change input or external state, and call expand only if exact bytes are required.
• Treat [Capsule round-trip tax] and the sequence fuse as measured prompts to batch independent work or execute the next concrete change. A costly uncached suffix also tightens output/history budgets mechanically. Successful exec control wrappers are removed automatically; oversized read-only output remains exact-capsule recoverable even at normal context pressure.
• For automatic-context-compaction cost, call insight with compaction:true and the session id/path. It returns aggregates, the latest event, and context_pressure with occupancy, predicted runway, locally reported quota pressure, compaction churn, retained-image risk, reasons, and the active policy. Request bounded rows only with compaction_events:true. Treat pre/post input as observable exposure because hidden compactor generation is not exposed.
• Treat gain as contract-valid local exposure telemetry, not billing. Legacy hook projections are excluded unless include_unverified:true is explicitly requested.
• Start view_image at high; use original only for pixel-level evidence.
• Give subagents self-contained messages. Automatic policy uses none for independent tasks, a small recent window for deictic continuations, and full history only for explicit whole-conversation dependencies.
• Subagent work is task-bounded: keep the default total to 16 (never above 20), use one agent per UI/i18n/performance/release lane, pause after 10 independent agents for one compact {decisions,evidence,files,blockers} summary, and reuse that digest instead of opening another review wave. Return no raw screenshots, base64 media, full diffs, or transcripts; retain exact recovery in Capsule.
• After each grouped mutation and decisive verification, call stats once and gain once. Treat both as local exposure evidence, not billing; do not rerun exploratory reads just to obtain a measurement.
• Return the smallest complete answer; explicit detail requirements win.
• Never sacrifice accuracy for compression.
• Skill apply|restore requires an explicit request and confirm:true.
• See capsule-contract.md for exact behavior and recovery.