codehere 1.0.1 → 1.0.2

package/dist/application/use-cases/generate-narration.js

@@ -1,130 +0,0 @@
-/**
- * PM Narration — v0.6 M3 mid-task summary.
- *
- * While a task is in flight, periodically generates a one-paragraph
- * plain-English summary of what the agent has done so far. Fires on
- * step-count or elapsed-time thresholds (caller decides when).
- *
- * Per claim_user_persona HIGH: non-technical users primary. They don't
- * read tool calls; they read narration. Mid-task narration tells them
- * "the agent is currently editing your auth file and ran the tests"
- * instead of forcing them to interpret a list of Bash/Edit/Read steps.
- *
- * Per feedback_overconfidence_guard.md: narration is an LLM call.
- * Hallucination is possible. Surface MUST frame it as "narrator says…"
- * not authoritative truth. The PROMPT_TEMPLATE includes the same
- * anti-hallucination rule as generate-final-report (ground every claim
- * in the structured data).
- *
- * Provider strategy: env override, else claude (the dead claude-free
- * first, fall back to claude). Returns null if neither is installed.
- */
-import { spawn } from 'child_process';
-import { loadTemplate, renderTemplate } from './narrator-prompt-template.js';
-import { isMockMode } from '../../infrastructure/orchestration/mocks/resolve-agent-spawn.js';
-import { mockNarration } from '../../infrastructure/orchestration/mocks/mock-synthesis.js';
-/** v0.6 S6 — prompt template lifted into ~/.codehere/narrator-prompt.md so
- *  the user can edit the agent's voice. loadTemplate() falls back to the
- *  baked-in default on missing file or missing required placeholders, so
- *  the prompt is never silently broken. See narrator-prompt-template.ts. */
-function buildPrompt(input) {
-    const { template } = loadTemplate();
-    return renderTemplate(template, input);
-}
-function selectProviderBinary() {
-    // MED #27 (atomic audit) · the old chain led with 'claude-free' (the dead
-    // FCC/NIM proxy final-report.ts abandoned) — every narration burned its
-    // timeout against a down proxy first. Env override → claude, that's it.
-    const override = process.env.CODEHERE_NARRATOR_BINARY?.trim();
-    if (override)
-        return [override];
-    return ['claude'];
-}
-export async function generateNarration(input) {
-    if (isMockMode()) {
-        const m = mockNarration({ steps: input.steps });
-        return {
-            summary: m.summary,
-            generatedAt: Date.now(),
-            durationMs: m.durationMs,
-            stepCountAtSnapshot: input.steps.length,
-        };
-    }
-    const prompt = buildPrompt(input);
-    const candidates = selectProviderBinary();
-    for (const binary of candidates) {
-        const result = await tryProvider(binary, prompt, input.sessionId, input.steps.length);
-        if (result)
-            return result;
-    }
-    return null;
-}
-function tryProvider(binary, prompt, sessionId, stepCount) {
-    return new Promise((resolve) => {
-        const startedAt = Date.now();
-        let stdout = '';
-        let stderr = '';
-        let resolved = false;
-        const finish = (val) => {
-            if (resolved)
-                return;
-            resolved = true;
-            resolve(val);
-        };
-        let child;
-        try {
-            // Same shape as generate-final-report's tryProvider — claude --bare -p.
-            child = spawn(binary, ['--bare', '-p', prompt], {
-                stdio: ['ignore', 'pipe', 'pipe'],
-            });
-        }
-        catch {
-            finish(null);
-            return;
-        }
-        // 20s timeout — narration should be SNAPPIER than Final Report
-        // (which is 60s). Mid-task = the user is actively waiting.
-        const timeout = setTimeout(() => {
-            try {
-                child.kill();
-            }
-            catch {
-                /* ignore */
-            }
-            finish(null);
-        }, 20_000);
-        child.stdout?.on('data', (chunk) => {
-            stdout += chunk.toString('utf-8');
-        });
-        child.stderr?.on('data', (chunk) => {
-            stderr += chunk.toString('utf-8');
-        });
-        child.on('close', (code) => {
-            clearTimeout(timeout);
-            if (code !== 0) {
-                // Non-zero usually means binary missing OR the model refused. Either
-                // way silent fail; caller treats null as "no narration this cycle".
-                void stderr;
-                void sessionId;
-                finish(null);
-                return;
-            }
-            const summary = stdout.trim();
-            if (!summary || summary.length < 10) {
-                finish(null);
-                return;
-            }
-            finish({
-                summary,
-                generatedAt: Date.now(),
-                durationMs: Date.now() - startedAt,
-                stepCountAtSnapshot: stepCount,
-            });
-        });
-        child.on('error', () => {
-            clearTimeout(timeout);
-            finish(null);
-        });
-    });
-}
-//# sourceMappingURL=generate-narration.js.map

package/dist/application/use-cases/generate-thread-title.js

@@ -1,181 +0,0 @@
-/**
- * Hidden Title agent — v0.6.0 M9.
- *
- * Generates a 4-6 word title for a project thread from its first user prompt.
- * Hidden = users don't pick it from any picker; fires automatically when the
- * project thread gets its first task entry. The title surfaces in the browser
- * tab + future thread switcher UI.
- *
- * Companion to generate-final-report.ts:
- *   - Provider: env override → claude → graceful skip (dead claude-free hop removed)
- *   - Same single-shot, no streaming, returns string
- *   - Different prompt (terse title generation, NOT 3-paragraph narrative)
- *   - Cheaper budget (~50 tokens per title vs 200 per Final Report paragraph)
- *
- * Triggering: project-thread-mirror invokes this when seeding the project-meta
- * header for a NEW project (i.e., this cwd's thread.jsonl didn't exist before).
- * Idempotent: subsequent task entries do NOT regenerate the title.
- *
- * Cost: ~1 LLM call per NEW project (not per task). At a typical 1-2 new
- * projects per week per user, the cost is negligible. FCC-routed when
- * available; degrades silently to "(untitled)" when no provider responds.
- */
-import { spawn } from 'child_process';
-import { isMockMode } from '../../infrastructure/orchestration/mocks/resolve-agent-spawn.js';
-import { mockTitle } from '../../infrastructure/orchestration/mocks/mock-synthesis.js';
-const TITLE_TIMEOUT_MS = 15_000; // titles are short; abort fast if stuck
-const PROMPT_TEMPLATE = (firstPrompt, cwd) => `Generate a 4-6 word project title from this first user prompt.
-
-Project working directory: ${cwd}
-First prompt: ${firstPrompt}
-
-Rules:
-- 4 to 6 words MAX
-- No quotes, no punctuation, no trailing period
-- Title-case (e.g. "Auth Bug Fix Sprint")
-- Ground the title in the prompt + cwd ONLY. Do NOT invent project subject
-  matter you can't derive from the data above (e.g. don't infer "Thread
-  Interface Updates" from a "list files" prompt — say what the user
-  actually asked, not what you imagine the project is about).
-- If the prompt is too generic to title meaningfully, prefer naming the
-  cwd's last path segment (e.g. "Taste V0.6 Workspace") over invention.
-- Output ONLY the title — no preamble, no explanation
-
-Title:`;
-function selectProviderBinary() {
-    // MED #28 (atomic audit) · same dead 'claude-free' first hop as narration —
-    // removed so a title either comes from claude or degrades fast, not after
-    // a proxy timeout.
-    const override = process.env.CODEHERE_TITLE_BINARY;
-    if (override)
-        return [override];
-    return ['claude'];
-}
-function tryProvider(binary, prompt) {
-    return new Promise((resolve) => {
-        let stdout = '';
-        let settled = false;
-        const finish = (result) => {
-            if (settled)
-                return;
-            settled = true;
-            resolve(result);
-        };
-        const child = spawn(binary, ['--bare', '-p', prompt], {
-            stdio: ['ignore', 'pipe', 'pipe'],
-        });
-        const timeout = setTimeout(() => {
-            child.kill('SIGTERM');
-            if (process.env.CODEHERE_DEBUG === '1') {
-                console.error(`[title-agent] ${binary} timeout after ${TITLE_TIMEOUT_MS}ms`);
-            }
-            finish(null);
-        }, TITLE_TIMEOUT_MS);
-        child.on('error', () => {
-            clearTimeout(timeout);
-            finish(null); // ENOENT etc. — silent fallback to next provider
-        });
-        child.stdout?.on('data', (chunk) => {
-            stdout += chunk.toString('utf-8');
-        });
-        child.on('exit', (code) => {
-            clearTimeout(timeout);
-            if (code !== 0) {
-                finish(null);
-                return;
-            }
-            const title = sanitizeTitle(stdout);
-            if (!title) {
-                finish(null);
-                return;
-            }
-            finish(title);
-        });
-    });
-}
-/**
- * Strip whitespace / quotes / trailing punctuation. Cap at 80 chars.
- *
- * Garbage rejection (added 2026-05-06 after dogfood Bug A4): the underlying
- * `claude --bare` call can return error-shaped text on stdout with exit 0
- * (rate-limit messages, auth errors, refusals). Pre-fix those flowed straight
- * into the project sidebar as titles like "Request ID: req_xyz...". We now
- * detect obvious error / refusal patterns and return '' so the caller falls
- * back to the cwd-derived title.
- */
-export function sanitizeTitle(raw) {
-    const lines = raw
-        .split('\n')
-        .map((l) => l.trim())
-        .filter(Boolean);
-    if (lines.length === 0)
-        return '';
-    let title = lines[lines.length - 1]; // last non-empty line is the answer
-    // Strip surrounding quotes
-    title = title.replace(/^["'`]+|["'`]+$/g, '');
-    // Strip trailing punctuation
-    title = title.replace(/[.!?,;:]+$/, '');
-    // Cap at 80 chars
-    if (title.length > 80)
-        title = title.slice(0, 77) + '...';
-    title = title.trim();
-    if (!title)
-        return '';
-    // Garbage rejection — patterns that indicate an error response, NOT a title.
-    // Each pattern matches a real failure mode observed in dogfood:
-    //   - "Request ID: req_..."         → API error envelope leaking through
-    //   - "Error:" / "ERROR" / "error " → generic error prefix
-    //   - "I cannot" / "I can't"        → model refusal
-    //   - "I apologize" / "I'm sorry"   → model refusal
-    //   - "Rate limit" / "rate-limited" → quota error
-    //   - "Authentication" / "Unauthorized" / "401" / "403" → auth failure
-    //   - Long sentences (>10 words)    → not a title; model ignored the rules
-    //   - Contains http:// or https://  → URL leakage from error envelope
-    const lower = title.toLowerCase();
-    const garbageMarkers = [
-        /^request\s+id\s*[:=]/i,
-        /^error\b/i,
-        /\berror[:\s]/i,
-        /^i\s+(cannot|can'?t|am\s+unable|won'?t)\b/i,
-        /^i\s+(apologi[sz]e|am\s+sorry)\b/i,
-        /^i'?m\s+sorry\b/i,
-        /\brate\s*[-_]?\s*limit/i,
-        /\b(authentication|unauthori[sz]ed|forbidden)\b/i,
-        /\b(401|403|429|500|502|503)\b/i,
-        /https?:\/\//i,
-        /api[_-]?key/i,
-    ];
-    if (garbageMarkers.some((re) => re.test(lower))) {
-        if (process.env.CODEHERE_DEBUG === '1') {
-            console.error(`[title-agent] rejected garbage title: ${title.slice(0, 60)}`);
-        }
-        return '';
-    }
-    // Length sanity: a real title is 4-6 words per the prompt; allow up to 10
-    // for slack. Anything longer is the model ignoring instructions.
-    const wordCount = title.split(/\s+/).filter(Boolean).length;
-    if (wordCount > 10) {
-        if (process.env.CODEHERE_DEBUG === '1') {
-            console.error(`[title-agent] rejected long title (${wordCount} words): ${title.slice(0, 60)}`);
-        }
-        return '';
-    }
-    return title;
-}
-/**
- * Generate a thread title from the first user prompt + cwd context.
- * Returns the title string, or null if no provider is reachable.
- * Never throws — failure modes return null.
- */
-export async function generateThreadTitle(firstPrompt, cwd) {
-    if (isMockMode())
-        return sanitizeTitle(mockTitle(firstPrompt, cwd)) || null;
-    const prompt = PROMPT_TEMPLATE(firstPrompt, cwd);
-    for (const binary of selectProviderBinary()) {
-        const result = await tryProvider(binary, prompt);
-        if (result)
-            return result;
-    }
-    return null;
-}
-//# sourceMappingURL=generate-thread-title.js.map

package/dist/application/use-cases/narrator-prompt-template.js

@@ -1,158 +0,0 @@
-/**
- * v0.6 S6 — editable PM narrator prompt template.
- *
- * The mid-task narration (M3) used to ship a hard-coded prompt baked into
- * generate-narration.ts. That made the agent's "voice" un-tunable: a user
- * who wanted shorter / longer / more-jargony / less-jargony narration
- * had no lever short of forking the source.
- *
- * S6 lifts the template into ~/.codehere/narrator-prompt.md, a plain
- * Markdown file the user edits with whatever editor they like. The file
- * uses `{{placeholder}}` substitution for the runtime values:
- *   {{prompt}}            — the user's original task ask (truncated 500ch)
- *   {{agent}}             — which adapter is running ("claude-code" etc.)
- *   {{elapsed_seconds}}   — how long the task has been running
- *   {{step_count}}        — how many steps completed/in-progress
- *   {{steps}}             — last 15 steps, "  N. [state] label"
- *
- * If the file is missing OR the user-edited file is missing required
- * placeholders, we fall back to the baked-in DEFAULT_TEMPLATE so a typo
- * never breaks narration silently.
- *
- * Per claim_user_persona HIGH: the narrator IS the product voice for
- * non-technical users. Letting them control it is a v1 promise.
- *
- * Per feedback_overconfidence_guard.md: the DEFAULT_TEMPLATE keeps the
- * anti-hallucination rule + "frame as narrator not authority" cue. Users
- * editing the template can drop those rules at their own risk; the file
- * comment warns about it.
- */
-import { existsSync, mkdirSync, readFileSync, unlinkSync, writeFileSync } from 'fs';
-import { dirname, join } from 'path';
-import { homedir } from 'os';
-/** The required placeholders. If a user-edited template is missing ANY of
- *  these we fall back to the default so the rendered prompt is never
- *  silently broken. {{step_count}} is omitted from the required list — it's
- *  derivable from {{steps}} and a custom template might intentionally drop it. */
-const REQUIRED_PLACEHOLDERS = ['{{prompt}}', '{{agent}}', '{{elapsed_seconds}}', '{{steps}}'];
-export const DEFAULT_TEMPLATE = `<!--
-  codehere narrator prompt — drives the mid-task summary the user sees as
-  "narrator says…" while their AI coding agent is working. Edit freely.
-
-  Available placeholders (substituted at render time):
-    {{prompt}}            user's original task ask
-    {{agent}}             which adapter is running
-    {{elapsed_seconds}}   how long the task has been running
-    {{step_count}}        number of steps completed/in-progress
-    {{steps}}             last 15 steps, formatted as "  N. [state] label"
-
-  REQUIRED: {{prompt}}, {{agent}}, {{elapsed_seconds}}, {{steps}}.
-  Missing any of those → codehere falls back to its baked-in default.
-
-  Reset to default any time with: codehere narrator reset
--->
-
-You are a PM giving a 1-sentence-or-2 mid-task update to a non-technical user
-about what their AI coding agent is doing RIGHT NOW. Write ONE short paragraph
-(2-3 sentences). No headers, no bullet lists, no markdown.
-
-CRITICAL anti-hallucination rule: ground every claim in the structured data
-below. Do not invent file names, status, or actions not in the steps list.
-If you don't know what something does, say "the agent is working on …" with
-the literal target. Better to under-claim than to over-claim.
-
-Tone: like a colleague checking in. "It's reading the auth file and ran the
-tests — looks like 2 failed so far." Not robotic, not breathless.
-
-Reference frame for the user (not for the LLM): the user did NOT see this
-list directly; you're translating it for them. The user only sees your
-paragraph.
-
-────────── Structured data (the only source of truth) ──────────
-Task: "{{prompt}}"
-Agent: {{agent}}
-Elapsed: {{elapsed_seconds}}s
-Steps so far ({{step_count}}):
-{{steps}}
-
-Write the paragraph now. Do not preface with "Update:" or anything similar.
-Plain prose. No "I'm currently…" — narrate the AGENT, not yourself.
-`;
-function templateFile() {
-    return join(process.env.CODEHERE_HOME || homedir(), '.codehere', 'narrator-prompt.md');
-}
-/** Read the user-customized template if present and valid. Falls back to
- *  baked-in default on missing file, parse failure, or missing required
- *  placeholders (so a typo never silently breaks narration). */
-export function loadTemplate() {
-    const path = templateFile();
-    if (!existsSync(path))
-        return { source: 'default', template: DEFAULT_TEMPLATE };
-    let raw;
-    try {
-        raw = readFileSync(path, 'utf-8');
-    }
-    catch {
-        return { source: 'default', template: DEFAULT_TEMPLATE };
-    }
-    for (const ph of REQUIRED_PLACEHOLDERS) {
-        if (!raw.includes(ph)) {
-            return { source: 'default', template: DEFAULT_TEMPLATE };
-        }
-    }
-    return { source: 'user', template: raw };
-}
-/** Persist the default template to disk so the user has a starting point
- *  to edit. Used by `codehere narrator edit`. Doesn't overwrite an existing
- *  user-edited file unless `force` is set. */
-export function writeDefaultTemplate(force = false) {
-    const path = templateFile();
-    if (!force && existsSync(path))
-        return { wrote: false, path };
-    const dir = dirname(path);
-    if (!existsSync(dir))
-        mkdirSync(dir, { recursive: true });
-    writeFileSync(path, DEFAULT_TEMPLATE, 'utf-8');
-    return { wrote: true, path };
-}
-/** Delete the user-customized template. Future loads use the baked-in
- *  default. Idempotent: missing file is a no-op success. */
-export function resetTemplate() {
-    const path = templateFile();
-    if (!existsSync(path))
-        return { removed: false, path };
-    try {
-        unlinkSync(path);
-        return { removed: true, path };
-    }
-    catch {
-        return { removed: false, path };
-    }
-}
-/** Substitute the runtime values into the template. Same {{placeholder}}
- *  protocol as the template-file comments document. */
-export function renderTemplate(template, input) {
-    const stepsBlock = input.steps
-        .slice(-15)
-        .map((s, idx) => `  ${idx + 1}. [${s.state}] ${(s.label ?? '').slice(0, 120)}`)
-        .join('\n');
-    const replacements = {
-        '{{prompt}}': input.prompt.slice(0, 500),
-        '{{agent}}': input.agent,
-        '{{elapsed_seconds}}': String(Math.round(input.elapsedMs / 1000)),
-        '{{step_count}}': String(input.steps.length),
-        '{{steps}}': stepsBlock,
-    };
-    let out = template;
-    for (const [k, v] of Object.entries(replacements)) {
-        // Replace ALL occurrences. Use split/join to avoid regex escaping
-        // pain on the curly braces.
-        out = out.split(k).join(v);
-    }
-    return out;
-}
-/** Path to the user-customized template (for CLI status display). */
-export function templatePath() {
-    return templateFile();
-}
-//# sourceMappingURL=narrator-prompt-template.js.map

package/dist/application/use-cases/plan-auto-review.js

@@ -1,57 +0,0 @@
-/**
- * Tier 2.1 · slice 2b — turn the eligibility decision into a concrete,
- * loop-safe plan for the auto cross-vendor review spawn. Still pure: it decides
- * WHAT to spawn (vendor, prompt, cwd, tag), not how — the live spawn + the
- * real-spend verification are the next slice. Kept pure so the loop guard and
- * the refute framing are unit-testable without a daemon or spend.
- *
- * THE LOOP GUARD is the load-bearing safety here: an auto-review is itself a
- * task that completes and flows through the same completion handler. Without a
- * guard it would trigger its own auto-review, recursing + burning spend forever.
- * A reviewer session is tagged; planAutoReview returns null for tagged sessions.
- *
- * The reviewer runs in REFUTE mode and READ-ONLY: codehere proves the PROCESS
- * (a different vendor independently reviewed + recorded findings), it never
- * adjudicates code correctness (assess-cross-vendor honesty note). The prompt
- * asks the reviewer to try to disprove "done", not to fix anything.
- */
-import { shouldAutoReview } from './should-auto-review.js';
-/** Tag stamped on an auto-review reviewer session so it never triggers its own
- *  auto-review (the loop guard) and so surfaces can label it. */
-export const AUTO_REVIEW_TAG = 'auto-cross-vendor-review';
-/** The refute-mode brief: read-only, adversarial, names the writer's files. */
-function buildRefutePrompt(writerAgent, files) {
-    const fileList = files.length <= 10 ? files.join(', ') : `${files.slice(0, 10).join(', ')} (+${files.length - 10} more)`;
-    return (`A different agent (${writerAgent}) just changed these files in this repo: ${fileList}. ` +
-        `Independently review that work READ-ONLY — do not edit anything. ` +
-        `Try to disprove that the task is correctly done: look for bugs, missed edge cases, ` +
-        `or claims the diff doesn't actually support. Report what you find (or that it holds up). ` +
-        `You are the second vendor in a cross-vendor check; your job is to refute, not to fix.`);
-}
-/**
- * Build the auto-review spawn plan, or null if it shouldn't / can't run.
- * Returns null (with no side effects) for: the loop guard, or any failed
- * eligibility gate. Pure.
- */
-export function planAutoReview(facts, thresholds) {
-    // Loop guard FIRST — a reviewer's own completion must never spawn a review.
-    if (facts.isAutoReviewSession)
-        return null;
-    const inputs = {
-        writerAgent: facts.writerAgent,
-        installedAgents: facts.installedAgents,
-        filesChanged: facts.writerFiles.length,
-        budgetRemainingUsd: facts.budgetRemainingUsd,
-    };
-    const decision = shouldAutoReview(inputs, thresholds);
-    if (!decision.eligible || !decision.reviewerVendor)
-        return null;
-    return {
-        reviewerVendor: decision.reviewerVendor,
-        prompt: buildRefutePrompt(facts.writerAgent, facts.writerFiles),
-        cwd: facts.cwd,
-        tag: AUTO_REVIEW_TAG,
-        reason: decision.reason,
-    };
-}
-//# sourceMappingURL=plan-auto-review.js.map

package/dist/application/use-cases/reconcile-audit.js

@@ -1,100 +0,0 @@
-/**
- * Tier 1.2 (verdict-trust punch list, 2026-06-09) · reconcile the AUDIT against
- * the real on-disk effect.
- *
- * The gap this closes: an audit entry records the agent's DECLARED tool call
- * (PreToolUse intent for claude, a post-hoc mirror for the others) — not an
- * independent observation that the write actually landed. An agent (or adapter)
- * that emits a file-write tool_use which never really changed anything lands that
- * fiction in the integrity-checked chain as a first-class entry. The chain then
- * faithfully shows the CLAIM was never casually edited — it never checks the claim
- * against reality.
- *
- * This reconciles the audit's file-write tool calls against the SAME git-diff
- * ground truth verify-claim already uses for file claims. It does NOT mutate the
- * append-only, hash-chained audit (that would break tamper-evidence) — it's a
- * read-only layer that produces an honest verdict the surface can show:
- *   - consistent   · writes recorded AND files changed on disk
- *   - unconfirmed  · writes recorded but codehere couldn't tie them to a tracked
- *                    git change. Honest, NOT an accusation: the writes may have
- *                    targeted gitignored / scratch / .codehere paths (which the
- *                    git-diff ground truth excludes), been reverted within the
- *                    run, or — the case worth a look — never really happened.
- *                    "couldn't confirm", never "it's fake" (refuter 2026-06-09:
- *                    a false fabrication accusation on honest work is the worst
- *                    error class on a proof surface).
- *   - no-writes    · nothing claimed, nothing to reconcile
- *   - unobservable · not a git repo, so codehere can't confirm
- *
- * Count-based by design: toolInput is truncated/heterogeneous across adapters, so
- * per-file path matching is unreliable; the file-write COUNT vs the git-change
- * count is robust. Bash/shell tools can also write files but codehere can't tell,
- * so they're deliberately NOT counted as declared writes (conservative — never
- * fabricates a "declared-not-observed" the user can't act on).
- */
-/** Tool names that explicitly mutate a file on disk, across every adapter.
- *  Derived from the real audit's toolName distribution (Write/Edit/write_file/
- *  edit_file observed) plus the standard variants. Shell tools are excluded — a
- *  bash `echo > f` can't be distinguished from a read, so counting it would
- *  manufacture false "declared-not-observed" verdicts. */
-// Matched case-INSENSITIVELY (adapters emit varied casings: claude 'Edit',
-// codex 'Delete', gemini raw tool_name like 'write_file'/'replace'). A DELETE is
-// a real file mutation that shows in git — omitting it made a codex delete-only
-// run falsely read "no writes recorded" (self-audit 2026-06-09). Keep this broad;
-// over-counting a non-write is far less harmful than a false "nothing happened".
-const WRITE_TOOLS = new Set(
-// SPECIFIC file-mutation tool names only. Bare generic verbs ('replace',
-// 'patch') are deliberately EXCLUDED — a non-file tool named 'replace' (an
-// in-memory string replace) would false-count and produce a false 'consistent'
-// (worse than under-counting, which is an honest 'unconfirmed'). A real write
-// the set misses → 'unconfirmed' (honest), never a fabricated-write laundered
-// into green (self-audit round 2, 2026-06-09).
-[
-    'Write',
-    'Edit',
-    'MultiEdit',
-    'NotebookEdit',
-    'Delete',
-    'write_file',
-    'edit_file',
-    'create_file',
-    'delete_file',
-    'str_replace',
-    'str_replace_editor',
-    'apply_patch',
-].map((t) => t.toLowerCase()));
-export function reconcileAudit(entries, artifacts) {
-    const writeToolCalls = entries.filter((e) => WRITE_TOOLS.has((e.toolName || '').toLowerCase())).length;
-    const filesObserved = artifacts?.files?.length ?? 0;
-    if (artifacts?.noGitTracking) {
-        return {
-            writeToolCalls,
-            filesObserved: 0,
-            verdict: 'unobservable',
-            note: `${writeToolCalls} file-write tool call${writeToolCalls === 1 ? '' : 's'} recorded, but this folder isn't a git repo — codehere can't confirm them against disk.`,
-        };
-    }
-    if (writeToolCalls === 0) {
-        return {
-            writeToolCalls,
-            filesObserved,
-            verdict: 'no-writes',
-            note: 'No file-write tool calls recorded — nothing to reconcile against disk.',
-        };
-    }
-    if (filesObserved === 0) {
-        return {
-            writeToolCalls,
-            filesObserved,
-            verdict: 'unconfirmed',
-            note: `${writeToolCalls} file-write tool call${writeToolCalls === 1 ? '' : 's'} recorded, but codehere couldn't confirm them against tracked git changes — they may have targeted gitignored/scratch paths or been reverted. Worth a look, not proof of a problem.`,
-        };
-    }
-    return {
-        writeToolCalls,
-        filesObserved,
-        verdict: 'consistent',
-        note: `${writeToolCalls} file-write tool call${writeToolCalls === 1 ? '' : 's'} recorded · ${filesObserved} file${filesObserved === 1 ? '' : 's'} confirmed changed on disk.`,
-    };
-}
-//# sourceMappingURL=reconcile-audit.js.map