Token Saver Context Compression

Use this skill to reduce token usage while preserving grounded evidence. This integrates:

• pnpm search:code (hybrid retrieval)
• token-saver Python compression scripts
• MemoryRecord persistence into framework memory
• spawn prompt evidence injection ([mem:*] / [rag:*])

Activation

The token-saver skill can be invoked in two ways:

Manual Invocation (always available)

Skill({ skill: 'token-saver-context-compression' });

Use this when context pressure is high, pnpm search:tokens shows a file/directory exceeds 32K tokens, or you need query-targeted compression.

Auto-enforcement via compression-reminder.txt (requires AUTO_COMPRESSION_PHASE_3=1)

Set AUTO_COMPRESSION_PHASE_3=1 in .env to enable the compression-reminder.txt trigger:

# In .env
AUTO_COMPRESSION_PHASE_3=1

When enabled, compression-trigger.cjs writes .claude/context/runtime/compression-reminder.txt whenever a compression event fires. The router reads this file and spawns context-compressor automatically.

Without this env var: compression events are logged to .claude/context/compression-stats.jsonl but no compression-reminder.txt is written, so the router does not auto-spawn compression. The skill must be invoked manually.

Token thresholds enforced by the router (from CLAUDE.md Section 8):

• 80K tokens — spawn context-compressor proactively
• 120K tokens — compression mandatory before new spawns
• 150K tokens — no new agent spawns until compression completes

Note: These thresholds are router behavioral guidelines checked in CLAUDE.md Section 8. The compression-trigger.cjs triggers are separate heuristics (budget >90%, reads >10KB, fetches >5KB, periodic every 10 ops). There is no automated hook enforcing the 80K/120K/150K thresholds — they rely on the router reading compression-reminder.txt.

When to Use

• pnpm search:tokens shows a file/directory exceeds 32K tokens
• Context is large or expensive and you need a compressed summary
• You need query-targeted compression before synthesis
• You need hard evidence sufficiency gating before persisting memory
• You're building a prompt and search:code results alone aren't enough context

Iron Law

Do not persist compressed content directly to memory files from a subprocess. Emit MemoryRecord payloads and let framework hooks process sync/indexing.

Workflow

1. Retrieve candidate context (pnpm search:code "<query>").

Step 0.5: Check Actual Token Usage + Cost (ccusage-adapter)

Before compressing, query actual API token usage and cost for today via ccusage-adapter. This lets you make data-driven compression decisions and report accurate cost savings.

// Attempt to read actual token usage (graceful degradation — never blocks compression)
let usageData = null;
let costs = null;
try {
  const ccusage = require('.claude/lib/utils/ccusage-adapter.cjs');
  usageData = ccusage.getTodayTotals();
  if (usageData) {
    costs = ccusage.calculateCost(usageData, process.env.CCUSAGE_MODEL || 'opus');
  }
} catch (_err) {
  // ccusage not installed or unavailable — fall back to heuristic estimation
}

if (usageData && costs) {
  console.log('[token-saver] Usage today:', {
    total: usageData.inputTokens + usageData.outputTokens,
    cost: `$${costs.actualCost.toFixed(4)}`,
    cacheSaved: `$${costs.cacheSavings.toFixed(4)}`,
  });

  // Use actual counts to decide compression aggressiveness
  const totalTokens = usageData.inputTokens + usageData.outputTokens;
  if (totalTokens > 120_000) {
    console.log('[token-saver] HIGH pressure (>120K tokens) — aggressive compression mode');
  } else if (totalTokens > 80_000) {
    console.log('[token-saver] MODERATE pressure (>80K tokens) — standard compression mode');
  } else {
    console.log('[token-saver] LOW pressure (<80K tokens) — light compression');
  }
} else {
  // ccusage unavailable — fall through to heuristic estimation from compression-trigger.cjs
  console.log('[token-saver] ccusage unavailable — using heuristic token estimation');
}

Fallback behavior: when getTodayTotals() returns null (ccusage not installed, timeout, or CCUSAGE_DISABLED=1), the workflow continues using existing heuristic thresholds from compression-trigger.cjs. The step never blocks compression.

Status file: the ccusage-statusline hook writes a live status to .claude/context/runtime/ccusage-status.txt on every prompt. Read it for a quick human-readable summary without calling the adapter directly.

Pricing Table

Canonical reference — these rates are used by ccusage-adapter.cjs → calculateCost(). When skill-updater refreshes this skill, it must verify these values via Exa search and update both this table and PRICING in .claude/lib/utils/ccusage-adapter.cjs.

Last verified: March 2026 (sources: Silicon Data, IntuitionLabs, DevTk.AI)

Model	Input	Output	Cache Write (1.25× input)	Cache Read (0.10× input)
opus → Claude Opus 4.6	$5.00/M	$25.00/M	$6.25/M	$0.50/M
sonnet → Claude Sonnet 4.6	$3.00/M	$15.00/M	$3.75/M	$0.30/M
haiku → Claude Haiku 4.5	$1.00/M	$5.00/M	$1.25/M	$0.10/M

Cache pricing rules (Anthropic standard):

• Cache write = 1.25× base input price (25% premium to store the prefix)
• Cache read = 0.10× base input price (90% discount on repeated context)

Set CCUSAGE_MODEL=sonnet or CCUSAGE_MODEL=haiku to match your active model. Default is opus (conservative — highest rates, largest savings shown).

Cost Tracking Components

The cost tracking pipeline consists of three files:

File	Role
.claude/lib/utils/ccusage-adapter.cjs	Parses JSONL session logs, sums tokens, calculates cost via PRICING table
.claude/hooks/monitoring/ccusage-statusline.cjs	UserPromptSubmit hook — runs adapter each prompt, writes status to runtime file
.claude/context/runtime/ccusage-status.txt	Live status file — read by router for pipeline summaries

Status output format (three savings layers):

[tokens]      <N> today (in: <N> / out: <N>) | Cost: $<X>
[cache]       $<X> saved | <N> reads, <N> writes
[compression] <N> events | <X>KB freed (~<N> tokens) | ~$<X> saved

• [tokens] — actual API spend
• [cache] — server-side savings (Anthropic prompt caching, 90% discount on cache reads)
• [compression] — client-side savings (tokens this skill avoided sending entirely)

Maintenance Instructions (for skill-updater)

When skill-updater refreshes this skill, execute this additional step after the standard research protocol:

Pricing Verification Step

1. Run an Exa search to confirm current Claude API pricing:

   mcp__Exa__web_search_exa({
     query: 'Anthropic Claude API pricing per million tokens Opus Sonnet Haiku current',
     numResults: 4,
   });
   

2. Cross-reference at least two sources (e.g., Anthropic docs, Silicon Data, IntuitionLabs, DevTk.AI).

3. Compare found rates against the Pricing Table above.

4. If any rate has changed:

   • Update the table in this SKILL.md
   • Update the PRICING constant in .claude/lib/utils/ccusage-adapter.cjs
   • Update the Last verified date
   • Record the change in .claude/context/memory/decisions.md

5. If no rates changed, update only the Last verified date.

Why this matters: stale pricing shows inflated savings ($86 showing as $258 with old Opus 4.0 rates). Accurate pricing is core to this skill's cost-reporting value.

1. Compress using token-saver in JSON mode (run_skill_workflow.py --output-format json).
2. If evidence is insufficient and fail gate is on, stop.
3. Map distilled insights into MemoryRecord-ready payloads.
4. Persist through MemoryRecord so .claude/hooks/memory/sync-memory-index.cjs runs.

Mapping Rule (Deterministic)

• gotchas.json:
  • text contains gotcha|pitfall|anti-pattern|risk|warning|failure
• issues.md:
  • text contains issue|bug|error|incident|defect|gap
• decisions.md:
  • text contains decision|tradeoff|choose|selected|rationale
• patterns.json:
  • default fallback for all remaining distilled evidence

Tooling Commands

Preferred wrapper entrypoint:

node .claude/skills/token-saver-context-compression/scripts/main.cjs --query "<question>" --mode evidence_aware --limit 20 --fail-on-insufficient-evidence

Direct Python engine (advanced):

python .claude/skills/token-saver-context-compression/scripts/run_skill_workflow.py --file <path> --mode evidence_aware --query "<question>" --output-format json --fail-on-insufficient-evidence

Output Contract

• Wrapper emits JSON with:
  • search summary
  • compression summary
  • memoryRecords grouped by target (patterns, gotchas, issues, decisions)
  • evidence sufficiency status

Workflow References

• Skill workflow: .claude/workflows/token-saver-context-compression-skill-workflow.md
• Companion tool: .claude/tools/token-saver-context-compression/token-saver-context-compression.cjs
• Command surface: .claude/commands/token-saver-context-compression.md
• Citation format is unchanged:
  • memory entries become [mem:xxxxxxxx]
  • RAG entries remain [rag:xxxxxxxx]

Integration with search:tokens

Use pnpm search:tokens to decide when to invoke this skill:

# Check if you need compression
pnpm search:tokens .claude/lib/memory
# Output: 60 files, 500KB, ~128K tokens ⚠ OVER CONTEXT

# Then compress with a targeted query
node .claude/skills/token-saver-context-compression/scripts/main.cjs \
  --query "how does memory persistence work" --mode evidence_aware --limit 10

The tool reads actual file content from search results (not just file paths), compresses via the Python engine, and extracts memory records classified by type (patterns, gotchas, issues, decisions).

Adaptive Compression

Adaptive compression (adjusting compression ratio based on corpus size) is automatic and requires no env var configuration. When the input corpus is small, compression is lighter; when it is large, compression is more aggressive. This is controlled internally by the Python engine based on token counts.

Requirements

• Node.js 18+
• Python 3.10+

Iron Laws

1. ALWAYS run hybrid search (pnpm search:code) before compressing to retrieve grounded evidence for the distilled output
2. NEVER compress context that still has open uncertainties — resolve ambiguities before compressing
3. ALWAYS persist distilled learnings via MemoryRecord immediately after compression
4. NEVER discard evidence that contradicts the current working hypothesis during compression
5. ALWAYS inject [mem:*] and [rag:*] citations in the compressed output for downstream spawn prompt grounding

Anti-Patterns

Anti-Pattern	Why It Fails	Correct Approach
Compressing without prior hybrid search	Output lacks grounded evidence, hallucination risk	Run pnpm search:code first, embed citations
Discarding contradicting evidence	Creates false confidence in distilled output	Preserve all conflicting signals in summary
No MemoryRecord after compression	Learnings lost on next context reset	Persist key findings immediately via MemoryRecord
Compressing too late (past 80K tokens)	Severe accuracy degradation before compression	Trigger compression at 80K tokens, not at limit
Skipping [mem:*] / [rag:*] citations	Downstream agents cannot verify claims	Always annotate evidence sources in output

Memory Protocol (MANDATORY)

Before work:

cat .claude/context/memory/learnings.md

After work:

• Add integration learnings to .claude/context/memory/learnings.md
• Add integration risks to .claude/context/memory/issues.md