Stats

Actions

Tags

context-optimizer | claude-agent-dev

Skill

context-optimizer

From claude-agent-dev

Diagnoses conversation bloat and prunes active context using token diagnostics, rolling summaries, and script-based compaction. Generates a flat key-value state summary file before executing the `/clear` command to reduce active token usage. Trigger on: 'optimize context', 'compress context', 'prune memory', 'reduce tokens', 'context size too large', 'out of tokens', 'compact context', 'clear history', 'cleanup context'. Also triggers when managing long development sessions, handling large file modifications, or resolving memory drift in nested subagent workflows. Always prefer this optimizer over running `/clear` directly when critical task progress and decisions need to be preserved.

Invocation

How this skill is triggered — by the user, by Claude, or both

Slash command

/claude-agent-dev:context-optimizer

User invocable

Model invocable

Inline context

Default effort

Context Preview

The summary Claude sees in its skill listing — used to decide when to auto-load this skill

Optimize and prune active conversation context to prevent token limits and context drift. This skill provides systematic diagnostics to sniff context bloat (unignored directories, lockfiles, large source files) and pruning strategies (JSON compaction, traceback log filtering, and rolling summaries) to preserve agent reasoning efficiency in long developer sessions.

Supporting Files

references/context-pruning-guidelines.mdscripts/diagnose_context_bloat.pyscripts/prune_context.pytests/test_diagnose_context_bloat.pytests/test_prune_context.py

SKILL.md

61 lines · ~996 tokens

Stats

LanguagePython

Stars0

MaintenanceExcellent

Last CommitJun 24, 2026

Actions

View Source View Plugin View on GitHub View README

context-optimizer

Optimize and prune active conversation context to prevent token limits and context drift. This skill provides systematic diagnostics to sniff context bloat (unignored directories, lockfiles, large source files) and pruning strategies (JSON compaction, traceback log filtering, and rolling summaries) to preserve agent reasoning efficiency in long developer sessions.

Process Flow

Phase 1: Diagnose (Run scripts/diagnose_context_bloat.py and /context)
  -> Phase 2: Select Strategy (Choose KV compaction, log filtering, or rolling summary)
  -> Phase 3: Action (Run scripts/prune_context.py or specify StartLine/EndLine slices)
  -> Phase 4: Reset & Reload (Save summary, run /clear, reload summary)

Step 1: Diagnose Bloat

Check for workspace dirtiness and large files: python skills/context-optimizer/scripts/diagnose_context_bloat.py (Use $CLAUDE_PLUGIN_ROOT/ prefix if required).
Run /context to inspect token usage.
Verify CLAUDE.md and GEMINI.md link to AGENTS.md instead of full rules text.

Step 2: Apply Strategy

MANDATORY PREREQUISITE: Before executing Strategy A, B, or Step 3, read references/context-pruning-guidelines.md completely from start to finish (it is under 100 lines, so the global 300-line range-limit rule below does not apply to it).

Select ONE strategy based on diagnostics:

Strategy A: KV Compaction (JSON/Configs) python skills/context-optimizer/scripts/prune_context.py --to-kv < data.json
Strategy B: Log Truncation (Test Failures/Traces) python skills/context-optimizer/scripts/prune_context.py --logs < test_output.log
Strategy C: Line Slicing (Large Files) Find target lines using grep_search, then read ONLY the required lines using StartLine and EndLine.
Strategy D: Subagent Isolation (Complex Tasks) Delegate to a subagent to protect the main context.

Step 3: Clear and Resume (For Bloated Sessions)

Write a flat Key-Value status, including which skill or gate was active so resume doesn't re-derive routing from scratch: python skills/context-optimizer/scripts/prune_context.py --summary --timestamp "$(date -Iseconds)" --current-skill "<skill-name or gate>" --done "completed items" --blocking "blockers" --next-step "next actions" --decisions "key decisions"
Tell the user: "Clearing history. Please type 'resume'."
Run /clear.
Reload .claude/rolling_summary.md and resume the current_skill recorded there directly — do not re-enter using-agent-dev-skills from Gate 0.

STRICT PROHIBITIONS

NEVER read files larger than 300 lines without specifying StartLine and EndLine. Run grep_search first.
NEVER run /clear without writing .claude/rolling_summary.md first. Progress will be lost.
NEVER paste full JSON structures or verbose test/build stdout. Always use Strategy A or B to compact them.

Next Skills:

planning: Trigger if context reveals major specification gaps.
Whichever skill was recorded as current_skill in the rolling summary: resume it directly.
using-agent-dev-skills: Only if no current_skill was recorded (e.g. optimizing before any route was chosen).