mhylle/claude-skills-collection/strategic-compact

Strategic Compact

<!-- contents: generated by tests/skill_structure.py --write; edit headings, not this block -->

Contents

• Why strategic > automatic
• When to use
• The core heuristic
• Suggest when…
• Don't suggest when…
• Threshold
• Preserve before compact
• Suggestion format
• Integration with other skills
• Core principles (keep these in mind)
• References (loaded on demand)

Reference files (open one when a step points to it):

• references/boundary-rules.md: Boundary rules — when to suggest, when not to
• references/hook-setup.md: Hook setup and configuration

<!-- /contents -->

Suggest context compaction at logical boundaries — not arbitrary token thresholds. Auto-compaction is a safety net; strategic compaction is a workflow optimization that makes compaction feel like a natural pause instead of an interruption.

Why strategic > automatic

Auto-compaction fires on simple thresholds (token count, time). That creates real problems:

• Mid-task disruption — auto-compact might fire while you're debugging, losing the mental model built up over many interactions.
• Context fragmentation — arbitrary cutoffs create artificial boundaries in session history.
• Lost continuity — "why" decisions get compacted away at random points rather than preserved at meaningful boundaries.
• Incomplete state — mid-implementation compaction can lose partially completed work.

Strategic compaction instead:

• Monitors for logical boundaries (task completion, phase transitions, successful commits).
• Considers semantic state (actively debugging vs. between tasks).
• Provides suggestions — never forces compaction.
• Integrates with context-saver to preserve critical state first.
• Respects workflow rhythm rather than imposing arbitrary limits.

Goal: compaction that feels like a natural pause, not an interruption.

When to use

Passive monitoring — via PreToolUse hook that attaches to every tool call, tracks counters, and evaluates boundaries. Hook setup → references/hook-setup.md.

Direct invocation:

• /strategic-compact — check current session state and get a recommendation.
• /strategic-compact status — view tool counts and threshold proximity.
• /strategic-compact now — force a suggestion (auto-invokes context-saver first).

Trigger phrases: "should I compact", "when should I clear context", "getting long", "checkpoint", "context is getting big".

The core heuristic

Two questions:

1. Has enough work happened that compaction is worthwhile? (weighted tool-call count vs. threshold)
2. Is this moment a good place to compact? (logical boundary vs. mid-task)

Both must be yes before suggesting. The full rule set is in references/boundary-rules.md. Summary:

Suggest when…

Don't suggest when…

• Active debugging — recent errors, investigation-pattern tool sequence, or task name contains "debug/fix/investigate".
• Active implementation — uncommitted changes, recent edits, unresolved test failures.
• User recently deferred — respect the last "not now".
• Critical operation in progress — rollback, migration, destructive commands.
• Last suggestion was <10 min ago — avoid nagging.

Threshold

Default: 50 weighted tool calls between logical boundaries.

"Significant" means the weighted count — raw tool calls multiplied by the weight table below. A tool call with weight 1.0 contributes 1 significant call; a Read with weight 0.5 contributes 0.5. Weights:

• Edit / Write / Bash / NotebookEdit / TaskUpdate — weight 1.0 (state-modifying or complex)
• Read — 0.5 (inspection, lower cognitive load)
• Glob / Grep — 0.3 (searches, low cost)
• WebFetch / WebSearch — 0.5 (external lookups)

Scenario-tuned thresholds (standard dev 50, debugging 75, quick fixes 30, large refactor 40, docs 60). Full config + dynamic adjustment rules → references/hook-setup.md.

Preserve before compact

When a suggestion is accepted, auto-invoke context-saver before compaction (via integrate_context_saver: true in hook config — default). This produces a handoff doc capturing:

• What was being worked on
• Files mid-edit with line numbers
• Decisions made
• Next steps

The compacted session can be resumed from that file. See context-saver/SKILL.md for the doc format.

Suggestion format

Keep suggestions informative but non-intrusive:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Context Compaction Suggested
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 
 Session Metrics:
   Tool calls:     72 significant (threshold: 50)
   Session time:   45 minutes
   Last boundary:  Task completed 3 min ago
 
 Boundary Detected:
   Task "Add user authentication" marked complete
   Git commit: "feat: implement JWT auth flow"
   No pending changes or active debugging
 
 Recommendation: GOOD TIME TO COMPACT
 
 Options:
   [1] Compact now (will invoke context-saver first)
   [2] Defer for 15 minutes
   [3] Defer until next boundary
   [4] Dismiss (don't suggest again this session)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Always show why now — the specific boundary and conditions that triggered the suggestion. Users trust the heuristic faster when they can see its reasoning.

Minimal mode (for experienced users):

[Compact? 72/50 calls | Task done | Clean state] Y/n/defer

Integration with other skills

• `context-saver` — auto-invoked before compact. Captures in-flight state as a handoff doc.
• `implement-phase` — emits phase_complete boundary events.
• `create-plan` — plan creation emits a boundary.
• `code-review` — review completion emits a boundary.

Full integration specs → references/hook-setup.md.

Core principles (keep these in mind)

1. Suggest, don't force — respect the user's workflow.
2. Boundary-aware — align with natural pause points, not token counts.
3. Context-preserving — always integrate with context-saver before compact.
4. Configurable — adapt thresholds per project and scenario.
5. Non-intrusive — clear but minimal suggestions; include the "why."

References (loaded on demand)

• references/hook-setup.md — PreToolUse hook config, state tracking, tool weights, dynamic thresholds, per-project configuration, integration specs, troubleshooting
• references/boundary-rules.md — Full suggest-when / suppress-when rule specifications with YAML patterns, plus accept/defer heuristics for users