@pugi/cli 0.1.0-beta.1 → 0.1.0-beta.100

package/dist/core/compact/summarizer.js

@@ -0,0 +1,208 @@
+import { estimateTokens } from './token-counter.js';
+/**
+ * System prompt for the summarizer. Six closed sections + brand
+ * voice clamp + "no tool calls" sentinel. The headings are stable
+ * markdown so a downstream renderer can split + reformat without
+ * re-parsing the model output.
+ *
+ * Why the explicit `If a section has no content` rule: empty sections
+ * cost zero tokens but their absence is meaningful — the second model
+ * reading the memo learns "no decisions were made yet" from the empty
+ * `## Decisions` header. Without the rule we observed models silently
+ * dropping empty sections and the reader could not distinguish "no
+ * decisions" from "summarizer forgot the section".
+ */
+const SUMMARIZE_SYSTEM_PROMPT = [
+    'You are the Pugi conversation summarizer. Compress the supplied',
+    'transcript into a six-section memo. Operator picks the work back up',
+    'from this memo — accuracy and completeness matter more than brevity.',
+    '',
+    'OUTPUT FORMAT (verbatim section headings, in this order):',
+    '',
+    "## Intent",
+    '(What the operator is trying to accomplish, in one paragraph.)',
+    '',
+    "## Decisions",
+    '(Bullet list of decisions made + the reasoning. Empty section means',
+    'no decisions yet — render the heading anyway.)',
+    '',
+    "## Files",
+    '(Bullet list of file paths touched, with one-line "why".)',
+    '',
+    "## Errors",
+    '(Bullet list of errors encountered + how each was resolved. Empty',
+    'means no errors — still render the heading.)',
+    '',
+    "## Tools",
+    '(Bullet list of notable tool calls and their outcomes. Group similar',
+    'calls; the goal is to surface meaningful state changes, not log',
+    'every Read.)',
+    '',
+    "## Next",
+    '(One paragraph: the immediate next planned action.)',
+    '',
+    'RULES:',
+    '- Do not invent state. Only summarise what is in the transcript.',
+    '- Preserve file paths verbatim.',
+    '- Do not call any tools (you have none).',
+    '- Do not address the operator. Write in third person.',
+    '- No emoji. No em dashes.',
+].join('\n');
+/**
+ * Convert a slice of session events into the user message body the
+ * summarizer ingests. We keep the format simple: one event per line,
+ * prefixed with the kind, so the model can see role boundaries. Tool
+ * outputs are length-capped at 4 KB each so a single 200 KB grep result
+ * does not blow the summarizer's own context budget.
+ */
+const TOOL_PAYLOAD_CAP_BYTES = 4096;
+export function renderEventsForSummary(events) {
+    const lines = [];
+    for (const event of events) {
+        const rendered = renderOneEvent(event);
+        if (rendered !== null)
+            lines.push(rendered);
+    }
+    return lines.join('\n');
+}
+function renderOneEvent(event) {
+    const payload = (event.payload ?? null);
+    switch (event.kind) {
+        case 'user': {
+            const text = stringField(payload, 'brief') ?? stringField(payload, 'text') ?? '';
+            if (text.length === 0)
+                return null;
+            return `[operator] ${text}`;
+        }
+        case 'persona': {
+            const text = stringField(payload, 'text') ?? '';
+            if (text.length === 0)
+                return null;
+            const slug = stringField(payload, 'personaSlug') ?? 'persona';
+            return `[${slug}] ${text}`;
+        }
+        case 'system': {
+            const text = stringField(payload, 'text') ?? '';
+            if (text.length === 0)
+                return null;
+            return `[system] ${text}`;
+        }
+        case 'tool.start': {
+            const toolName = stringField(payload, 'toolName') ?? 'unknown';
+            const args = stringField(payload, 'args') ?? '';
+            return `[tool.start ${toolName}] ${truncate(args, TOOL_PAYLOAD_CAP_BYTES)}`;
+        }
+        case 'tool.result': {
+            const toolName = stringField(payload, 'toolName') ?? 'unknown';
+            const result = stringField(payload, 'result') ?? '';
+            return `[tool.result ${toolName}] ${truncate(result, TOOL_PAYLOAD_CAP_BYTES)}`;
+        }
+        case 'agent.spawned': {
+            const slug = stringField(payload, 'personaSlug') ?? 'unknown';
+            return `[agent.spawned ${slug}]`;
+        }
+        case 'agent.completed': {
+            const slug = stringField(payload, 'personaSlug') ?? 'unknown';
+            return `[agent.completed ${slug}]`;
+        }
+        case 'compaction': {
+            // Pre-existing compact marker — its `summary` payload IS the
+            // condensed form of older history. We pass it through verbatim so
+            // the summarizer treats it as already-summarised state and folds
+            // it into the new memo (rather than re-summarising garbage).
+            const summary = stringField(payload, 'summary') ?? '';
+            if (summary.length === 0)
+                return null;
+            return `[prior compaction]\n${summary}`;
+        }
+        case 'rewind-marker': {
+            // L9 : rewind tombstones are infrastructure rows that
+            // describe a transcript edit, not conversation content. They
+            // never carry user-visible prose; the summariser drops them so
+            // the produced memo focuses on actual operator/persona turns.
+            // The masked range (events the marker covers) is already elided
+            // by `applyRewindMask` before this function is reached, so the
+            // summariser only sees a marker when it sits OUTSIDE every active
+            // rewind range — in which case the marker has already been
+            // cancelled by an undo-rewind and contributes nothing.
+            return null;
+        }
+        default: {
+            // Forward-compat: unknown kinds get a structural fingerprint so
+            // the summarizer can still represent them.
+            const exhaustive = event.kind;
+            void exhaustive;
+            return null;
+        }
+    }
+}
+function stringField(payload, key) {
+    if (!payload)
+        return undefined;
+    const value = payload[key];
+    return typeof value === 'string' ? value : undefined;
+}
+function truncate(s, capBytes) {
+    if (Buffer.byteLength(s, 'utf8') <= capBytes)
+        return s;
+    // Naive cut on UTF-16 chars — good enough for the summarizer; we
+    // append a marker so the model knows content was elided.
+    return `${s.slice(0, Math.floor(capBytes / 2))}\n... [truncated]`;
+}
+/**
+ * Run one summarisation round. Throws on transport error (the caller
+ * surfaces a one-line message to the operator and aborts the compact).
+ * Returns the structured result on success.
+ */
+export async function summarizeEvents(input) {
+    if (input.events.length === 0) {
+        throw new SummarizerError('refusing-to-summarize-empty-slice', 'No events to summarize.');
+    }
+    const userBody = renderEventsForSummary(input.events);
+    if (userBody.length === 0) {
+        throw new SummarizerError('refusing-to-summarize-empty-slice', 'All events rendered as empty.');
+    }
+    const messages = [
+        { role: 'system', content: SUMMARIZE_SYSTEM_PROMPT },
+        { role: 'user', content: userBody },
+    ];
+    const response = await input.client.send(messages, [], {
+        personaSlug: input.personaSlug,
+        tag: { tag: 'summarize' },
+        maxTokens: 2048,
+        temperature: 0.1,
+        ...(input.model !== undefined ? { model: input.model } : {}),
+        ...(input.signal !== undefined ? { signal: input.signal } : {}),
+    });
+    if (response.stop === 'error') {
+        throw new SummarizerError(response.code, `Summarizer transport failed: ${response.message}`);
+    }
+    if (response.stop === 'tool_use') {
+        // Sanity guard. We pass tools: [] so the model should not be able
+        // to invoke any; if Anvil's prompt template ever leaks tool defs
+        // we want a hard failure rather than a silent dropped summary.
+        throw new SummarizerError('unexpected-tool-call', 'Summarizer returned tool_use despite tools: []. Treating as failure.');
+    }
+    const summary = response.content.trim();
+    if (summary.length === 0) {
+        throw new SummarizerError('empty-summary', 'Summarizer returned an empty body.');
+    }
+    return {
+        summary,
+        tokensSummarised: estimateTokens(userBody),
+        eventsSummarised: input.events.length,
+    };
+}
+/**
+ * Error class so the caller can branch on `error instanceof
+ * SummarizerError` and surface `error.code` to the operator.
+ */
+export class SummarizerError extends Error {
+    code;
+    constructor(code, message) {
+        super(message);
+        this.name = 'SummarizerError';
+        this.code = code;
+    }
+}
+//# sourceMappingURL=summarizer.js.map

package/dist/core/compact/token-counter.js

@@ -0,0 +1,108 @@
+/**
+ * Approximate token counter for the `/compact` auto-trigger gate.
+ *
+ * Pugi does not bundle tiktoken — adding a 3 MB native module for a
+ * single heuristic check after every turn is the wrong trade. We instead
+ * use the OpenAI rule-of-thumb that 1 token ≈ 4 characters of English
+ * (≈ ¾ of a word). For mixed-language conversations the approximation
+ * skews high by 10-20% which is the safe direction — the auto-compact
+ * threshold trips slightly EARLY, never LATE. A late trigger would
+ * exceed the model's actual context window and crash the next turn.
+ *
+ * Per-model context windows are exported from `MODEL_CONTEXT_WINDOW` so
+ * the caller can resolve the budget from the active model slug without
+ * round-tripping to the server. New models added in Anvil should grow
+ * this table in lockstep; the fall-back is the conservative 32k window.
+ *
+ * Override hook: when `PUGI_TOKEN_COUNTER_OVERRIDE` is set the function
+ * parses it as a JSON object `{"text": <n>}` (number of tokens to claim
+ * per char). Used only by the integration spec — never by production.
+ */
+/** Default chars-per-token ratio for the OpenAI rule-of-thumb. */
+const DEFAULT_CHARS_PER_TOKEN = 4;
+/**
+ * Conservative fallback window when the model slug is unknown. 32k is
+ * the smallest window any current Anvil-served model exposes; using it
+ * as a fall-back guarantees we never overestimate budget and skip a
+ * compaction that should have fired.
+ */
+const FALLBACK_WINDOW = 32_000;
+/**
+ * Known context windows for models Anvil exposes today. Keep in sync
+ * with `apps/admin-api/src/pugi/model-registry.ts` — when a new model
+ * lands there, add it here. The table is intentionally narrow: only
+ * models we have actually validated.
+ */
+export const MODEL_CONTEXT_WINDOW = Object.freeze({
+    'sonnet-4.6': 200_000,
+    'sonnet-4.5': 200_000,
+    'opus-4.7': 1_000_000,
+    'opus-4.6': 200_000,
+    'haiku-4.5': 200_000,
+    'deepseek-chat-v3.1': 128_000,
+    'gpt-5': 200_000,
+    'gemini-2.5-pro': 1_000_000,
+});
+/**
+ * Estimate the token count of a single string. Returns a positive
+ * integer for non-empty input and zero for empty input. The function is
+ * pure — same input always yields the same output.
+ *
+ * The approximation is `ceil(byteLength / chars-per-token)`. We use
+ * `Buffer.byteLength` so multi-byte UTF-8 sequences count proportionally
+ * to their on-the-wire size, not their char count — this matches the
+ * tokenizer's behaviour on CJK / cyrillic / emoji where one char often
+ * eats 3-4 tokens.
+ */
+export function estimateTokens(text) {
+    if (text.length === 0)
+        return 0;
+    const ratio = readCharsPerTokenOverride() ?? DEFAULT_CHARS_PER_TOKEN;
+    const bytes = Buffer.byteLength(text, 'utf8');
+    return Math.max(1, Math.ceil(bytes / ratio));
+}
+/**
+ * Resolve the context window in tokens for the given model slug. Falls
+ * back to the conservative 32k window when the slug is unknown.
+ */
+export function contextWindowForModel(model) {
+    if (!model)
+        return FALLBACK_WINDOW;
+    const known = MODEL_CONTEXT_WINDOW[model];
+    return known ?? FALLBACK_WINDOW;
+}
+/**
+ * Sum the token estimates of an arbitrary list of strings. Convenience
+ * for the auto-trigger which has to count across N transcript turns.
+ */
+export function estimateTokensInMany(parts) {
+    let total = 0;
+    for (const part of parts) {
+        total += estimateTokens(part);
+    }
+    return total;
+}
+/**
+ * Read the test override env. Returns the parsed chars-per-token ratio
+ * when set, or undefined when absent / malformed. Never throws.
+ */
+function readCharsPerTokenOverride() {
+    const raw = process.env['PUGI_TOKEN_COUNTER_OVERRIDE'];
+    if (!raw)
+        return undefined;
+    try {
+        const parsed = JSON.parse(raw);
+        if (typeof parsed === 'object'
+            && parsed !== null
+            && typeof parsed.charsPerToken === 'number') {
+            const ratio = parsed.charsPerToken;
+            if (Number.isFinite(ratio) && ratio > 0)
+                return ratio;
+        }
+    }
+    catch {
+        /* malformed env, fall through */
+    }
+    return undefined;
+}
+//# sourceMappingURL=token-counter.js.map

package/dist/core/consensus/anvil-fanout.js

@@ -1,21 +1,21 @@
 /**
- * Anvil fan-out — `pugi review --consensus` (α6.7).
+ * Anvil fan-out — `pugi review --consensus` .
  *
  * Posts the captured diff to Anvil's consensus endpoint and consumes the
  * SSE stream that interleaves per-reviewer events (`type:"verdict"`) and
  * the final consensus event (`type:"consensus"`).
  *
- * Endpoint contract (admin-api side, ships as α6.7.1 follow-up):
+ * Endpoint contract (admin-api side, ships as follow-up):
  *
- *   POST {apiUrl}/api/pugi/review-consensus
- *   Authorization: Bearer {apiKey}
- *   Content-Type: application/json
- *   Body: { diff, context: { branch, commit, title } }
- *   Response: text/event-stream
- *     data: { reviewer: "codex"|"claude"|"deepseek", type:"started" }
- *     data: { reviewer, type:"verdict", severity:"P0|P1|P2|P3|CLEAN",
- *             rawContent:"<reviewer text>", latencyMs, error? }
- *     data: { type:"consensus", rubric_verdict, reasoning }
+ *  POST {apiUrl}/api/pugi/review-consensus
+ *  Authorization: Bearer {apiKey}
+ *  Content-Type: application/json
+ *  Body: { diff, context: { branch, commit, title } }
+ *  Response: text/event-stream
+ *    data: { reviewer: "codex"|"claude"|"deepseek", type:"started" }
+ *    data: { reviewer, type:"verdict", severity:"P0|P1|P2|P3|CLEAN",
+ *            rawContent:"<reviewer text>", latencyMs, error? }
+ *    data: { type:"consensus", rubric_verdict, reasoning }
  *
  * The CLI side does NOT trust the server's `rubric_verdict` — we recompute
  * it locally from the per-reviewer verdicts so a malformed / forged server
@@ -24,15 +24,15 @@
  *
  * Graceful degradation:
  *
- *   - 404 from runtime  → "endpoint_missing" (admin-api endpoint pending,
- *                         α6.7 ships CLI-only). Caller falls back to the
- *                         legacy `pugi review --triple --remote` flow OR
- *                         prints a "backend not deployed" notice depending
- *                         on the operator's invocation.
- *   - 401/403 / 429     → matching status with an actionable message.
- *   - 5xx / timeout     → "failed" with the truncated body.
+ *  - 404 from runtime → "endpoint_missing" (admin-api endpoint pending,
+ *                         ships CLI-only). Caller falls back to the
+ *                        legacy `pugi review --triple --remote` flow OR
+ *                        prints a "backend not deployed" notice depending
+ *                        on the operator's invocation.
+ *  - 401/403 / 429    → matching status with an actionable message.
+ *  - 5xx / timeout    → "failed" with the truncated body.
  *
- * Local-first contract (ADR-0037): this module never touches the disk,
+ * Local-first contract : this module never touches the disk,
  * never logs the diff payload, and never retries on transient errors.
  */
 /**
@@ -75,7 +75,7 @@ export async function dispatchConsensus(config, request, sink) {
             return {
                 status: 'endpoint_missing',
                 code,
-                message: 'POST /api/pugi/review-consensus not deployed on this runtime (α6.7.1 follow-up).',
+                message: 'POST /api/pugi/review-consensus not deployed on this runtime .',
             };
         }
         if (code === 401 || code === 403) {
@@ -143,11 +143,11 @@ export async function dispatchConsensus(config, request, sink) {
  * Async-iterable SSE parser. Spec'd against the
  * [HTML SSE spec](https://html.spec.whatwg.org/multipage/server-sent-events.html):
  *
- *   - Events are delimited by a blank line.
- *   - Each line is `field:value` (whitespace after `:` stripped).
- *   - Multiple `data:` lines in one event concatenate with `\n`.
- *   - We only care about `data:` payloads carrying JSON; everything
- *     else (event:, id:, retry:) is ignored.
+ *  - Events are delimited by a blank line.
+ *  - Each line is `field:value` (whitespace after `:` stripped).
+ *  - Multiple `data:` lines in one event concatenate with `\n`.
+ *  - We only care about `data:` payloads carrying JSON; everything
+ *    else (event:, id:, retry:) is ignored.
  *
  * The parser tolerates JSON-parse failures by dropping the malformed
  * event and continuing; a single bad event must not block the consensus

package/dist/core/consensus/diff-capture.js

@@ -1,15 +1,19 @@
 /**
- * Diff capture — `pugi review --consensus` (α6.7).
+ * Diff capture — `pugi review --consensus` .
  *
  * Captures the diff that the consensus fan-out will send to Anvil. Four
  * supported source kinds (in order of precedence):
  *
- *   1. `--pr <number>`    — uses `gh pr diff <num>` (gh CLI required).
- *   2. `--commit <sha>`   — diff of that commit vs its first parent.
- *   3. `--branch <name>`  — diff of HEAD vs `origin/<name>` merge-base.
- *   4. (default)          — diff of HEAD vs `origin/main` merge-base
- *                           covering BOTH committed-since-base AND
- *                           uncommitted (staged + working tree) edits.
+ *  1. `--pr <number>`   — uses `gh pr diff <num>` (gh CLI required).
+ *  2. `--commit <sha>`  — diff of that commit vs its first parent.
+ *                          When `--base <ref>` is ALSO provided, the
+ *                          diff is the range `<base>..<commit>` instead
+ *                          (mirrors `git diff base..commit` — covers the
+ *                          full PR-style payload, not just the tip).
+ *  3. `--branch <name>` — diff of HEAD vs `origin/<name>` merge-base.
+ *  4. (default)         — diff of HEAD vs `origin/main` merge-base
+ *                          covering BOTH committed-since-base AND
+ *                          uncommitted (staged + working tree) edits.
  *
  * The shape mirrors the existing `performRemoteTripleReview` flow:
  * uncommitted edits are deliberately included by computing the diff
@@ -95,6 +99,16 @@ export function captureDiff(spec) {
         return captureFromPr(cwd, spec.pr);
     }
     if (typeof spec.commit === 'string' && spec.commit.length > 0) {
+        // When `--base` is supplied alongside `--commit`, callers want the
+        // full PR-style range diff (`base..commit`), not just the tip
+        // commit's parent diff. This matches the convention used by
+        // `git diff <base>..<commit>` everywhere else in the toolchain and
+        // is the verified-correct mode for reviewing a PR head ref. Without
+        // this branch, `--base` was silently ignored when `--commit` was
+        // present — see feedback_pugi_review_use_range_diff_not_worktree.
+        if (typeof spec.baseRef === 'string' && spec.baseRef.length > 0) {
+            return captureFromRange(cwd, spec.baseRef, spec.commit);
+        }
         return captureFromCommit(cwd, spec.commit);
     }
     if (typeof spec.branch === 'string' && spec.branch.length > 0) {
@@ -114,7 +128,18 @@ function captureFromPr(cwd, pr) {
     // Fetch the PR head into a private ref so we have local objects to
     // diff against. `pull/<num>/head` is GitHub's special refspec exposed
     // to anyone with read access on the repo.
-    safeExec(cwd, 'git', ['fetch', 'origin', `pull/${pr}/head:${tempRef}`]);
+    //
+    // The leading `+` (force-update) is REQUIRED: if a prior invocation
+    // died before reaching the `finally` cleanup (SIGKILL, host crash,
+    // operator Ctrl-C inside a `try` block in a wrapping caller, hung
+    // gh CLI), the tempRef survives on disk. Without `+`, the next
+    // `fetch <pr>/head:<tempRef>` aborts with
+    //  `! [rejected] pull/N/head -> refs/pugi/consensus-pr-N (non-fast-forward)`
+    // and the entire consensus run fails for an operator who did nothing
+    // wrong. `+` semantics: replace the ref unconditionally — exactly
+    // the recovery behavior we want for a sandboxed `refs/pugi/...` slot
+    // that no human ever reads.
+    safeExec(cwd, 'git', ['fetch', 'origin', `+pull/${pr}/head:${tempRef}`]);
     try {
         // Resolve the base ref to diff against. Prefer the PR's declared
         // base; fall back to `origin/main`. We compute the merge-base so a
@@ -149,8 +174,10 @@ function captureFromPr(cwd, pr) {
             safeExec(cwd, 'git', ['update-ref', '-d', tempRef]);
         }
         catch {
-            // Swallow: a leftover ref under refs/pugi/ is harmless. The next
-            // run will overwrite it via `fetch ... :ref` anyway.
+            // Swallow: a leftover ref under refs/pugi/ is harmless. The
+            // next run force-overwrites it via `fetch +pull/<n>/head:<ref>`
+            // (note the leading `+` in the refspec above), so a stale tempRef
+            // never blocks a future invocation even if cleanup never runs.
         }
     }
 }
@@ -192,6 +219,88 @@ function captureFromCommit(cwd, commit) {
         },
     };
 }
+/**
+ * Range capture for `--commit <X> --base <Y>` — diff equivalent to
+ * `git diff <base>..<commit>`. Used when the operator names BOTH endpoints
+ * (typical PR review against a remote head SHA).
+ *
+ * Critical: this MUST be a pure read-only range diff against named refs.
+ * The previous behavior fell through to `captureFromCommit` which only
+ * showed the tip commit (`commit~1..commit`) — fine for single-commit
+ * review, wrong for multi-commit PRs. Worse, a stale fallback path was
+ * sending the working tree diff (`git diff` with no args), which caused
+ * every review on to surface identical noise from uncommitted
+ * `.gitignore` edits instead of the actual PR contents.
+ *
+ * Working tree integrity: only `git diff <ref>..<ref>` and metadata
+ * `log` / `rev-parse` / `name-rev` are used — none of these touch the
+ * index, working tree, or HEAD.
+ */
+function captureFromRange(cwd, baseRef, commit) {
+    // Resolve both endpoints up front so an unknown ref errors with a
+    // clear message before the diff invocation. `rev-parse` is read-only.
+    const fullCommit = safeExec(cwd, 'git', ['rev-parse', commit]).trim();
+    if (!fullCommit)
+        throw new Error(`Unknown commit ref: ${commit}`);
+    // Task #63 — refresh the base ref via `git fetch` BEFORE
+    // resolving. Previously a stale local `main` (last fetched days ago)
+    // would silently produce a diff containing every commit that landed
+    // upstream after the fetch, swamping the model's review с unrelated
+    // changes. The fetch is opportunistic: failures (offline, auth) are
+    // swallowed so the existing rev-parse path stays the safety net и
+    // returns the clear "Unknown base ref" error if the stale local copy
+    // is also missing.
+    //
+    // We fetch ONLY the base ref (not all refs) so the overhead is
+    // bounded — typical PR-style review payload, base = main, one ref.
+    // `--no-tags --quiet` keeps the network footprint minimal.
+    const remoteCandidate = baseRef.includes('/') ? baseRef : `origin/${baseRef}`;
+    if (remoteCandidate.startsWith('origin/')) {
+        const bareRef = remoteCandidate.slice('origin/'.length);
+        safeExecOptional(cwd, 'git', [
+            'fetch',
+            '--no-tags',
+            '--quiet',
+            'origin',
+            bareRef,
+        ]);
+    }
+    // Resolve the base: accept already-qualified refs (`origin/main`,
+    // `refs/heads/foo`) and bare branch names. If the bare name isn't
+    // locally resolvable, retry against `origin/<name>` — the common
+    // CI shape where local main is absent but the remote tracking ref is.
+    let resolvedBase = safeExecOptional(cwd, 'git', ['rev-parse', baseRef]).trim();
+    let effectiveBase = baseRef;
+    if (!resolvedBase && !baseRef.includes('/')) {
+        const remoteBase = `origin/${baseRef}`;
+        resolvedBase = safeExecOptional(cwd, 'git', ['rev-parse', remoteBase]).trim();
+        if (resolvedBase)
+            effectiveBase = remoteBase;
+    }
+    if (!resolvedBase)
+        throw new Error(`Unknown base ref: ${baseRef}`);
+    const diff = safeExec(cwd, 'git', [
+        'diff',
+        `${resolvedBase}..${fullCommit}`,
+        '--',
+        '.',
+        ...PROTECTED_PATHSPEC_EXCLUDES,
+    ]);
+    const cappedDiff = capDiff(diff);
+    const subject = safeExec(cwd, 'git', ['log', '-1', '--pretty=%s', fullCommit]).trim();
+    const branch = safeExec(cwd, 'git', ['name-rev', '--name-only', fullCommit]).trim() || 'detached';
+    const stats = computeStats(cappedDiff);
+    return {
+        diff: cappedDiff,
+        context: {
+            branch,
+            commit: shortSha(fullCommit),
+            title: subject || `commit ${shortSha(fullCommit)}`,
+            ref: `range:${effectiveBase}..${shortSha(fullCommit)}`,
+            stats,
+        },
+    };
+}
 function captureFromBranch(cwd, branch, baseRef) {
     const remoteRef = branch.includes('/') ? branch : `origin/${branch}`;
     const mergeBase = safeExec(cwd, 'git', ['merge-base', baseRef, remoteRef]).trim();
@@ -227,8 +336,8 @@ function captureFromBase(cwd, baseRef) {
     const mergeBase = safeExecOptional(cwd, 'git', ['merge-base', baseRef, 'HEAD']).trim();
     if (mergeBase) {
         // Two parts (non-overlapping):
-        //   1. Committed since base:  `<base>..HEAD`
-        //   2. Uncommitted (staged + working tree as a single union): `git diff HEAD`
+        //  1. Committed since base: `<base>..HEAD`
+        //  2. Uncommitted (staged + working tree as a single union): `git diff HEAD`
         // `git diff HEAD` already reports BOTH staged AND working-tree
         // changes relative to HEAD, so we MUST NOT add a separate
         // `--cached` invocation: doing so emits the same staged hunks

package/dist/core/consensus/rubric.js

@@ -1,5 +1,5 @@
 /**
- * Consensus rubric — `pugi review --consensus` (α6.7).
+ * Consensus rubric — `pugi review --consensus` .
  *
  * Three independent reviewers (Codex / Claude / DeepSeek) produce findings
  * tagged `[P0]` / `[P1]` / `[P2]` / `[P3]`. The rubric translates the per-
@@ -7,16 +7,16 @@
  *
  * Rubric (verbatim from /triple-review skill + admin-api OES MCP triple_review):
  *
- *   any reviewer reports [P0]              -> BLOCK
- *   two or more reviewers report [P1]      -> BLOCK   (consensus)
- *   exactly one reviewer reports [P1]      -> WARN    (asymmetric)
- *   no reviewer reports [P0] or [P1]       -> PASS    (P2/P3 only)
- *   every reviewer errored                 -> BLOCK   (no signal)
+ *  any reviewer reports [P0]             -> BLOCK
+ *  two or more reviewers report [P1]     -> BLOCK  (consensus)
+ *  exactly one reviewer reports [P1]     -> WARN   (asymmetric)
+ *  no reviewer reports [P0] or [P1]      -> PASS   (P2/P3 only)
+ *  every reviewer errored                -> BLOCK  (no signal)
  *
  * The rubric never reads model text beyond the severity markers; the
  * reviewer-side narrative is shown to the operator unchanged. Keeping the
  * verdict deterministic + LLM-free is the entire point of the gate (CEO
- * directive 2026-05-19): a model that disagrees with the rubric can be
+ * directive): a model that disagrees with the rubric can be
  * audited, a model that produces the verdict cannot.
  */
 /**
@@ -35,13 +35,13 @@ const SEVERITY_TOKEN = /\[\s*[Pp]([0-3])\s*\]/g;
  * Heuristics (intentionally permissive — different models format very
  * differently and a strict parser would drop signal):
  *
- *   1. Split the text on `[Px]` tokens, preserving the marker.
- *   2. Each marker starts a new finding. The summary is the rest of the
- *      same line (and the next line if the first is `:` or empty after
- *      stripping whitespace).
- *   3. Empty / whitespace-only summaries are dropped — a bare `[P1]`
- *      with no context cannot be acted on, and treating it as a finding
- *      would falsely trigger consensus.
+ *  1. Split the text on `[Px]` tokens, preserving the marker.
+ *  2. Each marker starts a new finding. The summary is the rest of the
+ *     same line (and the next line if the first is `:` or empty after
+ *     stripping whitespace).
+ *  3. Empty / whitespace-only summaries are dropped — a bare `[P1]`
+ *     with no context cannot be acted on, and treating it as a finding
+ *     would falsely trigger consensus.
  */
 export function parseFindings(raw) {
     if (typeof raw !== 'string' || raw.length === 0)
@@ -88,8 +88,8 @@ function extractSummary(slice) {
 /**
  * Compute the highest BLOCKING severity from a finding list. Returns
  * `null` when the reviewer is clean for gating purposes, i.e. either:
- *   - no findings at all, OR
- *   - only P2 / P3 findings (informational, non-blocking by rubric).
+ *  - no findings at all, OR
+ *  - only P2 / P3 findings (informational, non-blocking by rubric).
  *
  * `null` is the right contract for downstream tooling that gates on
  * "did this reviewer flag anything that should block ship?" - the
@@ -143,7 +143,7 @@ export function aggregate(verdicts) {
     const erroredReviewers = verdicts.filter((v) => v.errored).length;
     // Zero reviewers = zero signal. Falling through to the no-P0/no-P1
     // branch would emit a false PASS — exactly the regression flagged by
-    // Codex during PR #370 review. Treat empty input as BLOCK so the gate
+    // Codex during PR review. Treat empty input as BLOCK so the gate
     // fails closed when the backend returns no events at all (5xx that
     // somehow drained the SSE, server-side bug, dispatcher misconfig).
     if (totalReviewers === 0) {
@@ -214,11 +214,11 @@ export function aggregate(verdicts) {
 }
 /**
  * Map a rubric verdict to the conventional exit code Pugi CLI uses for
- * gates (spec α6.7):
+ * gates (spec ):
  *
- *   PASS  -> 0
- *   WARN  -> 1
- *   BLOCK -> 2
+ *  PASS -> 0
+ *  WARN -> 1
+ *  BLOCK -> 2
  *
  * The non-zero codes are distinct so a shell script can branch on the
  * exact outcome without re-parsing stdout.