context-audit — Token Hygiene for the Always-Loaded Context Stack

Convention: see conventions/brain-first.md — before running a fresh audit, check the brain for prior audit reports (gbrain recall "context audit report") so you can compute token DRIFT since the last run and avoid re-flagging findings the user already declined.

Convention: see conventions/quality.md — every finding cites its file and evidence; no unsourced claims.

What this is

Every file that loads on every turn is a per-turn tax: tokens, latency, and — past a point — instruction-following quality. Always-loaded files accrete (append-only release notes, promoted memory blocks nobody re-reads, rules restated in three files that drift into contradiction). This skill audits the whole always-loaded stack at once and returns a ranked, evidence-cited action list sorted by token savings.

It is an auditor, not a surgeon. It measures, finds, ranks, and recommends. The user (or a skill the user explicitly invokes afterward) applies changes.

Scope: what counts as "always-loaded"

Enumerate what THIS harness actually loads every turn — do not assume a fixed list. Typical stack:

File	Role	Fix belongs in
project CLAUDE.md / AGENTS.md	orientation, routing, invariants	the file itself (source-editable)
user-global CLAUDE.md	cross-project instructions	the file itself (source-editable)
auto-memory MEMORY.md	promoted memory blocks	the memory store (demote/expire)
SOUL.md, USER.md, ACCESS_POLICY.md, HEARTBEAT.md, rendered AGENTS.md	bootstrap-rendered identity files	the interview answer bank / templates — NEVER the rendered file
harness system-prompt fragments (identity/tools files)	per-harness	wherever that harness sources them

Skills, reference docs, and anything loaded on demand are OUT of scope as audit subjects — but they are the DESTINATION for skill-extraction findings (content that only matters for one workflow should move out of the always-loaded stack into a skill).

Contract

This skill guarantees:

• Report-only. No audited file is edited, no page is written, nothing is auto-fixed — including 🟢 zero-risk findings. The output is a recommendation list the user applies deliberately.
• Rendered-file safety. Any recommendation touching a bootstrap-rendered file is expressed as an answer-bank or template change (gbrain bootstrap interview --set KEY "..." then gbrain bootstrap render --only <FILE> --force), never as a direct edit. See skills/soul-audit/SKILL.md for the mechanics.
• Estimated with a stated basis, never invented. Token figures come from the deterministic pre-pass (wc -c bytes/2.8 — Claude-family tokenizers run ~2.6-3.5 bytes/token on markdown dense with paths and code spans; 2.8 is the calibrated midpoint of measured always-loaded markdown, see #4988). The report prints the divisor so a reader can re-derive every number. If the host client reports an exact per-category context breakdown (e.g. Claude Code /context), that figure outranks the estimate — quote it and use it for the stack total.
• Native judging. The draft report is quality-gated through gbrain eval cross-modal — no raw model API calls, no hardcoded model IDs.
• Cost line. Default judging is ONE cheap model (the user's utility-tier model, all three slots, --cycles 1 — a few cents). The full three-provider frontier panel runs only when the user explicitly asks for a "full" or "multi-model" audit (~3x+ the cost per cycle).

Procedure

1. Enumerate the stack (deterministic)

List the always-loaded files for this harness and measure each:

# bytes/2.8 (calibrated for Claude-family tokenizers on markdown, #4988); integer ceil: (n*10+27)/28
for f in CLAUDE.md AGENTS.md SOUL.md USER.md ACCESS_POLICY.md HEARTBEAT.md MEMORY.md; do
  [ -f "$f" ] && echo "$f: $(wc -c < "$f") bytes (~$(( ( $(wc -c < "$f") * 10 + 27 ) / 28 )) tokens)"
done

Record the total. If a prior audit report exists in the brain, compute drift (net tokens grown/shrunk since last run, which files moved).

2. Read and analyze (the agent does this — no model calls yet)

Read every file in the stack in full. Evaluate against six dimensions:

1. Token efficiency — tokens spent per unit of behavioral value
2. Redundancy — the same rule/fact stated in more than one file
3. Contradictions — conflicting rules, numbers, or policies across files
4. Skill-worthiness — content that only matters for a specific workflow (extraction candidate: move to a skill, load on demand)
5. Staleness — outdated facts, references to removed features, promoted memory blocks that no longer earn their slot
6. Clarity — instructions compressible without behavior change, or ambiguous enough to misfire

3. Classify every finding by risk

• 🟢 Zero risk — pure deletion of exact redundancy or dead content
• 🟡 Low risk — compression or skill extraction with a clear trigger
• 🔴 Medium risk — changes that could shift edge-case behavior

All three classes are recommendations. The risk class tells the user how much care to apply — it does not authorize this skill to act.

4. Judge the draft through the native eval runner

Write the draft report to a temp file, then gate it:

# Resolve the cheap judge from the user's model tiers — never hardcode an ID.
# (`gbrain models` shows all resolved tiers if the config key is unset.)
JUDGE=$(gbrain config get models.tier.utility)

gbrain eval cross-modal \
  --task "Context-stack token-hygiene audit: every finding cites file + quoted evidence; savings are estimated (bytes/2.8, divisor stated in the report), never invented; findings ranked by token savings; every rendered-file recommendation targets the interview answer bank or template, never a direct edit; risk class on every row" \
  --output /tmp/context-audit-draft.md \
  --slug context-audit-report \
  --cycles 1 \
  --slot-a-model "$JUDGE" --slot-b-model "$JUDGE" --slot-c-model "$JUDGE"

Full multi-model panel (explicit opt-in only — the user asked for a "full" / "multi-model" audit): omit the --slot-*-model overrides so the runner's native three-provider defaults apply.

Exit codes: 0 PASS — deliver. 1 FAIL — fix the flagged weaknesses in the draft (usually: an unquoted claim or a rendered-file edit recommendation) and re-judge. 2 INCONCLUSIVE (provider/key trouble) — deliver the report but label it "unjudged" prominently.

5. Deliver

Print the report in the conversation (see Output Format). If the user wants it persisted, hand off to the brain-ops skill to file it under openclaw/ (agent-state notes) — this skill does not write pages itself.

Re-running after major edits to the stack, or on a schedule, is a harness-routing convention the user can set up (see the cron-scheduler skill) — nothing here runs automatically or guarantees a cadence.

Output Format

# Context Audit — YYYY-MM-DD

Stack total: ~NN,NNN tokens across N files (drift since last audit: +/-N,NNN)
Estimate basis: bytes/2.8 | host-reported exact total (e.g. `/context`): NN,NNN or n/a
Findings: N (~NN,NNN tokens recoverable) | Contradictions: N
Judge verdict: PASS (single-model, utility tier) | receipt: <path>

| # | Save (tok) | Risk | File | Finding | Evidence | Recommended fix (and WHERE it lives) |
|---|-----------|------|------|---------|----------|--------------------------------------|
| 1 | ~2,400    | 🟢   | ...  | redundancy: X restated | "quoted line" | delete from A; canonical copy stays in B |
| 2 | ~1,100    | 🟡   | SOUL.md | stale: ... | "quoted line" | update answer bank key VOICE_REGISTER, re-render — NOT a SOUL.md edit |
...

## Contradictions (fix these first, savings aside)
- FILE-A says "..." but FILE-B says "..." — resolve toward <one>, delete the other.

## Skill-extraction candidates
- <content> only matters when <workflow> — extract via skill-creator, load on demand.

Sorted by token savings, descending — except contradictions, which are called out first regardless of size (they cost correctness, not just tokens). Every row carries evidence (a quote or line reference) and names WHERE the fix belongs: source file, answer bank/template, memory store, or a new skill.

When it fails

Follow the agent operator protocol for any gbrain error code, exit code, [AGENT] block or notice block. Specific to this skill:

• gbrain eval cross-modal exits 1 (FAIL): fix the flagged weaknesses and re-run; do not deliver a failed audit as passing.
• No judge model or provider key is configured: say the audit ran structure-only and name the missing key; do not fabricate scores.
• A paid multi-model audit hits no_pricing or a cost cap: fall back to the default single cheap judge and tell the user why.

Anti-Patterns

• Editing any audited file. Report-only — even 🟢 zero-risk deletions are recommendations, not actions. "Auto-fix" promises contradict the rendered-file guard and are out of contract.
• Recommending a direct edit to a rendered file. SOUL.md / USER.md / ACCESS_POLICY.md / HEARTBEAT.md edits are overwritten by the next gbrain bootstrap render. Target the answer bank or template, then re-render.
• Raw model API calls for judging. The eval runner owns provider config, receipts, and verdict aggregation — route through gbrain eval cross-modal.
• Hardcoding model IDs. Resolve the judge from the user's model tiers; model names in a skill body rot.
• Running the full multi-model panel by default. It is an explicit opt-in; the single-cheap-model pass is the default for cost reasons.
• Auditing on-demand content as if always-loaded. Skills and reference docs don't pay the per-turn tax; flagging them inflates savings numbers.
• Inventing token counts. Run the pre-pass; estimates are labeled ~N with the divisor stated, and a host-reported exact figure always wins.
• Rewriting identity content yourself. If a finding is about WHAT an identity file says (wrong persona, outdated profile), route to soul-audit — the interview is the only author of that content.

Dedup

• soul-audit — identity CONTENT via interview: what SOUL.md/USER.md should SAY, sourced from the user's own words. context-audit is token/structure hygiene: what the stack COSTS per turn, where it repeats or contradicts itself. A finding like "USER.md's profile is outdated" hands off to soul-audit; "USER.md restates 800 tokens already in SOUL.md" stays here. Both respect the same rendered-file rule.
• skill-optimizer — tunes ONE skill's body against a benchmark and can mutate it. context-audit never mutates and looks only at always-loaded files; skills appear only as extraction destinations.
• functional-area-resolver — the compression TECHNIQUE for oversized routing tables (>=12KB). context-audit may cite it as the recommended fix when a routing section is the finding; it never applies it.
• skillpack-check — install/runtime health (DB, worker, migrations), not context size or prompt content.
• cross-modal-review — general second-opinion gate on arbitrary work products. context-audit uses the same underlying runner but as its own fixed judging step with audit-specific pass criteria; asking for "a second opinion on this code" routes there, not here.