/audit-style

description: "Internal facet of the audit family — invoke via '/audit style'. Mines session-log history for revealed working-style rules; receipt-gated (at least 2 distinct sessions of dated verbatim quotes)."

/audit style — Working-Style Sub-Audit

Canonical invocation: **`/audit style`** (args ride the umbrella: `/audit style recent`). The direct `/audit-style` form is retained for compatibility and is not advertised.

Mine your own past session-log corpus for working-style rules you have already revealed through action — not rules you're asked to invent, rules extracted from what you actually said and did across real sessions. Evidence-gated at every step: a candidate that cannot show its receipts does not ship, no matter how plausible it sounds.

Step 0: Config + Corpus Locate

Read `~/.gemini/antigravity/aria-knowledge.local.md` and extract the three style-audit config keys (parsed by `config.sh` as `KT_STYLE_LOOKBACK_DAYS`, `KT_STYLE_MAX_SESSIONS`, `KT_STYLE_AUDIT_LOG`). These keys are **bare-assigned** — the config file may have no value at all for any of them, in which case the shell variable is an empty string. Apply these defaults yourself in the skill body whenever the corresponding value is empty:

• `KT_STYLE_LOOKBACK_DAYS` empty → default to **90** days.
• `KT_STYLE_MAX_SESSIONS` empty → default to **50** sessions.
• `KT_STYLE_AUDIT_LOG` empty → default to `{knowledge_folder}/logs/style-audit-log.md`.

If `~/.gemini/antigravity/aria-knowledge.local.md` doesn't exist, stop: "aria-knowledge is not configured. Run /setup to get started."

**Locate the corpus.** The session-log corpus for the current project lives at `~/.gemini/antigravity/transcripts/<cwd-encoded>/*.jsonl`, where `<cwd-encoded>` is the current working directory with `/` replaced by `-` (Claude Code's standard transcript-directory encoding). Glob that directory for `*.jsonl` files — each file is one session transcript.

**Skip subagent sessions (hard pre-filter).** Exclude any transcript whose filename begins with `agent-` (e.g. `agent-a1b2c3….jsonl`). These are subagent worker transcripts spawned by the parent session (Task/Agent dispatches, workflow workers, review agents) — they are the *agent's* execution logs, not the user's authored prose, and mining them would (a) attribute agent-authored dispatch text to the user and (b) massively inflate the corpus on any session that fanned out workers. The `agent-` filename prefix is a clean, unambiguous discriminator (Claude Code names all subagent transcripts this way). This filter runs BEFORE the count in Step 1b, so the over-cap gate sees only genuine user sessions. (Validated on a live run: a 43-file delta was 29 subagent transcripts + 14 user sessions — skipping the `agent-*` set is what made the delta tractable.)

**Read the style-audit log for incremental scope.** Read `KT_STYLE_AUDIT_LOG` (resolved path per the default above). If it exists, its most recent timestamp entry marks the boundary of the last mining pass — this run only needs to consider sessions modified/created after that timestamp (incremental scope, keeps repeat runs cheap). **If the log doesn't exist yet (first run)**, there is no prior boundary: window the initial scan to the last `KT_STYLE_LOOKBACK_DAYS` (default 90) days of session files, by file mtime or the transcript's own embedded timestamps.

**Reuse a prior external mine as a first-run boundary (avoid re-mining what's already mined).** If there is no audit-log yet BUT a prior full mine of this same corpus exists on disk — most notably a ditto run archived under `{knowledge_folder}/references/` (its `you-corpus.txt` mtime marks when it ran, and its `stats.json` records the session count/date-range it covered) — treat that prior mine's timestamp as the incremental boundary instead of doing a full-lookback re-scan. Then only the **delta** (sessions newer than that mine, after the `agent-*` skip above) needs fresh extraction; fold the delta's evidence in with the prior mine's already-reduced result rather than re-processing the whole corpus. This turns an all-time first run from a multi-million-token re-scan into a small delta mine. State clearly in the report which portion was reused vs freshly mined. (Validated on a live run: reused 1,697 prior-mined sessions + freshly mined a 10-session user delta, instead of re-scanning ~8.5M tokens.) If no such prior mine exists, fall back to the lookback window as above.

Step 1: Extract User Prose (multi-stage filter)

The reference implementation of this filter is `extract-user-prose.py`, which lives alongside this SKILL.md at `plugin-claude-code/skills/audit-style/extract-user-prose.py`. **That script is the canonical algorithm** — this section documents the same stages in prose so the mining logic is auditable without reading Python, but the script is the executable source of truth; when in doubt about an edge case, defer to what the script actually does, or invoke it directly (`python3 .../extract-user-prose.py <session.jsonl>`) rather than re-deriving the filter by hand.

Each `.jsonl` line is one JSON object (a transcript event). The filter applies, in order:

1. **Stage 1 — role + block-type gate.** Keep only objects where `type == "user"` AND the nested `message.role == "user"`, and only the **text** content (a plain string, or list blocks with `type == "text"`). This is the load-bearing exclusion: it drops every `tool_result` block outright — tool outputs are not the user's voice, no matter how much prose they contain. 2. **Stage 2 — strip command/tool wrapper noise.** Even after Stage 1, some `role: user` events are Claude Code's own wrapper markup rather than something the user typed — local-command invocations and their echoed output/errors. Drop any text block containing `<local-command-*>`, `<command-name>`, `<command-args>`, `<local-command-stdout>`, `<local-command-caveat>`, or similar `<command-*>` tags. 3. **Stage 3 — drop skill-injection preambles and resume scaf