Priority: P1 (HIGH)

Workflow

1. Inspect host capabilities; do not assume agents can rewrite history, prompts, or memory.
2. Project outputs after consuming them: retain decisions, evidence, errors, and required values; point to complete logs/artifacts by stable reference.
3. Bound context at host-supported boundaries. Carry goal, active slice, authority, decisions, blockers, evidence links, and next action; preserve a stable append-only prefix where supported.
4. Compact only when useful and supported. Keep required details retrievable; never discard source evidence or rely on a fixed turn/token threshold.
5. Measure cache reads, replayed input, and total actor cost across the whole task. Separate observed usage from estimates; no invoice, savings, or efficacy claims without evidence.

Anti-Patterns

• No history rewrite assumption: Project outputs or create artifact references; use only host-supported controls.
• No fixed threshold: Compact from observed context pressure and task needs, not a universal turn/token count.
• No evidence deletion: Keep stable references to source outputs, decisions, and verification. Archive required packets, reports and receipts at durable references before cleanup.
• No omission-as-optimization: Never claim savings from suppressing required instruction bodies. Selection or body-return counts are not acknowledgment, comprehension, or enforcement.
• No partial cost claim: Include cache reads, replay, and total actor cost; label estimates. Missing coverage or rates means no savings claim.

References

• Compaction
• Masking and output projection
• Implementation and cost measurement