context-compression

name	context-compression
description	Use when an ongoing task must be compacted so another turn or agent can continue with its files, decisions, evidence, failures, and next action intact. Do not use for one-off summaries, research briefs, or domain implementation.

Context Compression

Compress active task context into a smaller, structured summary that preserves the artifact trail. The goal is not minimum tokens at any cost. The goal is to keep enough state that work can continue without re-reading the same material.

Use this skill when a conversation or task has grown large enough that the next step would otherwise risk losing:

• which files were read or modified
• which decisions were made and why
• exact error messages or identifiers
• what remains unresolved
• the next action that should happen first

Do Not Use For

• Short sessions that still fit comfortably in context
• Generic conversation summarization without task continuity needs
• Research briefs that need external sourcing rather than compression
• Test design, code review, or implementation work owned by another skill
• Rewriting a single message when no ongoing task state must be preserved

Workflow

1. Identify the minimum state that must survive compression:
   • task intent
   • changed files
   • read-only files worth remembering
   • decisions already made
   • exact failures, warnings, or constraints
   • open questions and next steps
2. Preserve the artifact trail explicitly:
   • use full file paths for files that matter
   • keep symbol names, config keys, error strings, and branch names
   • note dates or versions when facts may change
   • label important statements as observed, user-provided, inferred, or still unverified
   • retain the last verification time for volatile external state and active processes
3. Compress only the newest or least useful detail:
   • keep durable facts
   • drop repetition, chatter, and dead ends
   • do not regenerate the entire summary if only one span changed
4. Write the compressed state in a structured form:
   • Session Intent
   • Files Modified
   • Files Read
   • Decisions Made
   • Harness State
   • Current State
   • Next Steps
   • Open Questions
   • Stop Or Escalation Conditions
   • Freshness And Provenance
   • Superseded State
5. Check the result for continuity:
   • could the task continue without re-reading the original thread
   • are the key files and decisions still named explicitly
   • are unresolved issues still visible instead of flattened away
   • can a reader distinguish current state from an earlier result that has been replaced
6. If compression would remove a critical fact, keep the fact and compress elsewhere instead:
   • prefer a slightly larger summary over a misleadingly small one
   • never invent continuity that is not actually present
7. Update incrementally when a prior compressed state exists:
   • preserve still-valid decisions instead of regenerating them from memory
   • move replaced commands, hypotheses, and status into Superseded State
   • update verification timestamps only for facts that were actually rechecked

Output Expectations

Return a compact task summary that lets the next agent or future turn continue safely. The summary should include:

• the current task goal
• the important file paths or identifiers
• the last known decisions
• commands run, validation results, active sessions or servers, failed attempts, and remaining checks
• the active blocker or uncertainty
• the next concrete step
• when to continue, retry, ask the user, or stop
• provenance and freshness for state whose validity can change
• explicit superseded items when an earlier summary is being updated

Guardrails

• Do not discard file paths just to save tokens.
• Do not replace unresolved issues with vague language.
• Do not merge distinct decisions into one blurred summary.
• Do not invent a clean state if the task is still messy.
• Do not use this skill as a substitute for a domain-specific brief or a code review.
• Do not present inferred or stale state as freshly observed.
• Do not silently drop an earlier blocker merely because a newer summary is shorter.

Stop Conditions

• Stop and keep a layered or longer summary when the requested compression would make current and superseded state indistinguishable.
• Stop and mark state unverified when the original evidence is unavailable; do not reconstruct it from plausibility.