/woz-kb

description: The WOZCODE knowledge base behind the /woz review reviewer. Subcommands — `tune` (end-to-end reviewer tuning), `backtest` (tuning building blocks), `architecture-doc-fetch` (fetch woz/architecture.md from the KB), `cross-repo` (cross-repo planning brief), and knowledge-base ops (`status`, `query`, `note`, `suppress`, `boost`, `ingest`, `refresh`, `ops`).
argument-hint: tune | backtest | architecture-doc-fetch | cross-repo | status | query | note | suppress | boost | ingest | refresh | ops
arguments: subcommand

/woz-kb — the reviewer's knowledge base

One skill, these subcommands:

• **`woz-kb tune`** — the start-to-finish orchestrator: distill → backtest → learn (autotuner new-personas + per-PR missed-fixes) → re-measure the lift. Dry-run by default; `--apply` writes to the KB. Use this to onboard/tune a new repo or a whole org.
• **`woz-kb backtest`** — the building blocks for power users: the raw backtest run plus `--tune`, `--missed-report`, `--org-tune`, `--ab-compare` (unchanged).
• **`woz-kb architecture-doc-fetch`** — fetch the current repo's COMPANY-scope `woz/architecture.md` manifest from the KB into the working tree (no-op when already up to date).
• **`woz-kb cross-repo <feature>`** — print a cross-repo planning brief from the org's architecture manifests; `--ground` additionally appends a two-pass code/PR grounding search on top of the manifest brief. Flags: `--repos owner/a,owner/b`, `--topk <n>`.
• **knowledge-base ops** (`status`, `query`, `note`, `unnote`, `suppress`, `unsuppress`, `boost`, `unboost`, `ingest`, `refresh`, `ops`) — inspect and customize the knowledge base directly, e.g. `/woz-kb status`, `/woz-kb note "..."`, `/woz-kb query <text>`.

The first positional arg selects the subcommand; if omitted it defaults to `backtest` (back-compat). `tune`, `backtest`, `architecture-doc-fetch`, and `cross-repo` run through `woz-kb.cjs`; the knowledge-base ops run through `woz-knowledge.cjs` (see that section).

When to use

TRIGGER for tune/backtest: "tune the reviewer", "tune this repo", "onboard a repo", "tune the org", "backtest the reviewer", "how well does the reviewer do", or `/woz-kb`.

TRIGGER for architecture-doc-fetch/cross-repo: "sync the architecture doc", "fetch the architecture manifest", "cross-repo plan", "plan this across repos".

TRIGGER for knowledge-base ops: "knowledge base status", "what's in the knowledge base", "search the knowledge base for …", "add a knowledge-base note", "remember that X", "suppress this" / "stop the reviewer from citing X", "boost this", "ingest <file> into the knowledge base", "refresh the knowledge base", or `/woz-kb status` / `/woz-kb note` / `/woz-kb query`.

DO NOT use for: reviewing the current branch (`/woz review`) or past-session recall (`/woz-recall` — searches Claude Code session transcripts, not the knowledge base). General-purpose code search belongs to `mcp__plugin_woz_code__Search`, not the KB.

---

`woz-kb tune` — start-to-finish

node --no-warnings=ExperimentalWarning "${CLAUDE_PLUGIN_ROOT}/scripts/woz-kb.cjs" tune \
  --repo with-woz/wozcode \
  --anthropic-api-key-file ~/.woz/.anthropic-backtest-key

Pipeline per repo: **(1)** trigger a KB refresh → **(2)** distill human PR comments into baseline persona-hints → **(3)** training backtest (20 PRs × 3 rounds, or reuse a cached run) → **(4)** learn: the autotuner contributes **new personas** (its unique cross-PR synthesis), then missed-fixes contributes the **per-PR durable persona-hints** → **(5)** re-measure on the same PRs and print the recall/precision lift.

Flags:

• `--repo <owner/name>` — tune one repo. **Mutually exclusive with `--org`.**
• `--org <orgId>` — tune every repo the org has indexed, then run a final company-scope org-tune.
• `--apply` — write to the KB. **Without it, `tune` is a dry-run**: it computes + reports proposed hints/personas and counts but writes nothing (and skips the re-measure, since nothing changed).
• `--reuse-run <runId>` — reuse an existing backtest run's PRs + baseline instead of running a fresh (expensive) reviewer pass. `tune` also auto-reuses the newest cached run for the repo when present.
• `--count <n>` / `--rounds <n|all>` — training-backtest size (defaults 20 / 3) when not reusing a run.
• `--skip-distill` / `--skip-refresh` / `--skip-remeasure` — drop individual phases.
• `--anthropic-api-key-file <path>` — use API pricing for the reviewer/judge subprocesses.
• `--judge-model` / `--reviewer-model` / `--reviewer-via` / `--source` — as in `backtest`.

Note: after an `--apply`, the reviewer cache is invalidated (KB changed), so the re-measure re-runs the reviewer on the (small) PR set — a real cost, not a cache hit.

---

`woz-kb backtest` — building blocks

node --no-warnings=ExperimentalWarning "${CLAUDE_PLUGIN_ROOT}/scripts/woz-kb.cjs" backtest \
  --repo with-woz/wozcode --count 3 --rounds 1

Runs the deep reviewer against a sample of historical merged PRs and scores how close it gets — the reviewer never sees the human comments or merged diff, so the score is a real recall measure. Key flags:

• `--repo <owner/name>` (required), `--source <path>` (default cwd), `--count <n>` (default 3), `--rounds <n|all>` (per-review-round scoring), `--prs <n,n,n>` (explicit PRs; reuse a prior run's set for a comparable measurement).
• `--tune <runId> [--tune-apply] [--min-apply-ratio <n>]` — autotuner over a finished run.
• `--missed-report <runId> [--tune-apply]` — per-PR missed→suggested-fixes; writes `missed-fixes.json` + `.md`.
• `--org-tune <orgId> [--org-tune-all | --org-tune-repos <list>] [--org-tune-apply]` — cross-repo → company scope.
• `--ab-compare <baselineRunId> <newRunId> [--auto-rollback]`, `--personas <ids>`, `--no-apply`, `--timeout-min <n>`.

Safety contract (backtest runs)

• A fresh clone per PR at `<repo>/.wozcode/backtests/<runId>/pr-<n>/clone/`; `.wozcode/` is gitignored.
• The clone's `origin` is removed and push URLs are rewritten to `unreachab