codehere 0.8.0 → 1.0.2

package/dist/application/use-cases/domain-trust.js

@@ -1,110 +0,0 @@
-/**
- * Per-domain trust grade for W-PASS write targets · v0.9 must-ship.
- *
- * Demo v12 + locked v1 tier scope: "linear.app A · *.atlassian.net B ·
- * unknown F · first-write confirmation per domain."
- *
- * Two pieces of state:
- *   1. WELL_KNOWN_GRADES · hardcoded · the platforms we've vetted ourselves
- *      ship with grade A. Subdomain wildcards get B. Everything else
- *      defaults to F (forces explicit user confirmation on first write).
- *   2. USER_TRUSTED_DOMAINS · ~/.codehere/keys/trusted-domains.json ·
- *      domains the user has explicitly trusted after a first-write prompt.
- *      Domains here promote from F → B.
- *
- * Pure logic. CLI + HTTP wrap it.
- */
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'fs';
-import { homedir } from 'os';
-import { dirname, join } from 'path';
-const WELL_KNOWN_GRADES = [
-    // Grade A · platforms codehere has shipped+tested adapters for
-    { pattern: /^api\.linear\.app$/i, grade: 'A', reason: 'official Linear API · codehere ships a vetted adapter' },
-    { pattern: /^api\.github\.com$/i, grade: 'A', reason: 'official GitHub API · codehere ships a vetted adapter' },
-    { pattern: /^api\.figma\.com$/i, grade: 'A', reason: 'official Figma API · codehere ships a vetted adapter' },
-    // Grade B · same vendor, surrounding subdomains
-    { pattern: /^[a-z0-9-]+\.linear\.app$/i, grade: 'B', reason: 'Linear subdomain · adjacent to the vetted API' },
-    { pattern: /^[a-z0-9-]+\.github\.com$/i, grade: 'B', reason: 'GitHub subdomain · adjacent to the vetted API' },
-    { pattern: /^[a-z0-9-]+\.figma\.com$/i, grade: 'B', reason: 'Figma subdomain · adjacent to the vetted API' },
-    {
-        pattern: /^[a-z0-9-]+\.atlassian\.net$/i,
-        grade: 'B',
-        reason: 'Atlassian-hosted instance · vendor known but instance not vetted',
-    },
-    // Grade C · public hosts known to be commonly used
-    { pattern: /^raw\.githubusercontent\.com$/i, grade: 'C', reason: 'GitHub-hosted raw content · static, read-mostly' },
-];
-function trustedDomainsPath() {
-    const home = process.env.CODEHERE_HOME || homedir();
-    return join(home, '.codehere', 'keys', 'trusted-domains.json');
-}
-export function loadTrustedDomains() {
-    const path = trustedDomainsPath();
-    if (!existsSync(path))
-        return { schema: 1, domains: {} };
-    try {
-        const raw = JSON.parse(readFileSync(path, 'utf-8'));
-        if (raw && raw.schema === 1 && typeof raw.domains === 'object')
-            return raw;
-    }
-    catch {
-        /* fall through · empty trust store */
-    }
-    return { schema: 1, domains: {} };
-}
-export function trustDomain(host, note) {
-    const file = loadTrustedDomains();
-    file.domains[host.toLowerCase()] = { addedAt: Date.now(), note };
-    mkdirSync(dirname(trustedDomainsPath()), { recursive: true });
-    writeFileSync(trustedDomainsPath(), JSON.stringify(file, null, 2) + '\n', 'utf-8');
-}
-export function untrustDomain(host) {
-    const file = loadTrustedDomains();
-    const key = host.toLowerCase();
-    if (!file.domains[key])
-        return false;
-    delete file.domains[key];
-    writeFileSync(trustedDomainsPath(), JSON.stringify(file, null, 2) + '\n', 'utf-8');
-    return true;
-}
-export function gradeDomain(host) {
-    const userFile = loadTrustedDomains();
-    const userTrusted = !!userFile.domains[host.toLowerCase()];
-    for (const rule of WELL_KNOWN_GRADES) {
-        if (rule.pattern.test(host)) {
-            return {
-                host,
-                grade: rule.grade,
-                reason: rule.reason,
-                userTrusted,
-                requiresFirstWriteConfirmation: false,
-            };
-        }
-    }
-    if (userTrusted) {
-        return {
-            host,
-            grade: 'B',
-            reason: 'previously trusted by you · stored in trusted-domains.json',
-            userTrusted: true,
-            requiresFirstWriteConfirmation: false,
-        };
-    }
-    return {
-        host,
-        grade: 'F',
-        reason: 'unknown domain · not in well-known list · not previously trusted',
-        userTrusted: false,
-        requiresFirstWriteConfirmation: true,
-    };
-}
-/** Extract the host from a fully-resolved HTTPS URL. */
-export function hostOf(url) {
-    try {
-        return new URL(url).host;
-    }
-    catch {
-        return '';
-    }
-}
-//# sourceMappingURL=domain-trust.js.map

package/dist/application/use-cases/generate-final-report.js

@@ -1,134 +0,0 @@
-/**
- * Final Report — the headline deliverable. At task exit, produce a plain-English
- * summary of what the agent did.
- *
- * Engine (locked 2026-06-03, after the real dogfood found NO report on a real
- * conductor run): REUSE THE AGENT SESSION. The agent already narrates what it
- * did (kind:'say' steps — e.g. "Added subtract(a,b) to utils.js:5"); we use that
- * directly. No fresh summarizer spawn.
- *
- * Why the old path was replaced: it spawned a SEPARATE provider binary chain
- * (`claude-free` → `claude`) to summarize. `claude-free` (the FCC/NIM proxy
- * binary) is on PATH but its proxy was down → 30s hang → null; the `claude`
- * fallback also returned null in the conductor exit-handler. Net: a completed
- * real task showed "⚠ Final Report unavailable" — the headline deliverable
- * missing. Reusing the agent's own words is cheaper, has full context, and can't
- * be taken down by an external proxy.
- *
- * Fallback (locked): when the agent left no usable summary (or it looks like a
- * provider failure), build a DETERMINISTIC report from the recorded steps +
- * artifacts. It states ONLY verified facts, says plainly that no written summary
- * was available, and NEVER implies success — so a completed task always shows a
- * deliverable, and it's honest about which kind it is (see `source`).
- */
-const MAX_SUMMARY_CHARS = 1600;
-/**
- * UAT-003 — patterns indicating the agent's own text is actually a network /
- * auth / quota failure, not a real summary. Such "summaries" must NOT be shown
- * as the report; they fall through to the deterministic fallback.
- */
-const NETWORK_FAILURE_PATTERNS = /(?:all connection attempts failed|ECONNREFUSED|ENOTFOUND|fetch failed|network error|request id:\s*req_|connection timed out|getaddrinfo)/i;
-function isLikelyNetworkFailureSummary(text) {
-    return NETWORK_FAILURE_PATTERNS.test(text);
-}
-/**
- * ENGINE: the agent's own narration. kind:'say' steps are the words the agent
- * already emitted (reasoning + wrap-up). Joined newest-context-last, they ARE
- * the summary — no extra spawn. Returns null if the agent said nothing usable.
- */
-function extractAgentSummary(steps) {
-    const says = steps
-        .filter((s) => s.kind === 'say' && typeof s.label === 'string' && s.label.trim().length > 0)
-        .map((s) => s.label.trim());
-    if (!says.length)
-        return null;
-    const text = says.join('\n\n').slice(0, MAX_SUMMARY_CHARS);
-    // Too short to be a real summary (e.g. a one-word ack) → let the fallback give facts.
-    return text.length >= 12 ? text : null;
-}
-/**
- * FALLBACK: deterministic, honest. Facts only, from steps + artifacts. Says
- * plainly that no written summary was available. Never implies success.
- */
-function buildDeterministicReport(input) {
-    const files = input.artifacts?.files ?? [];
-    const toolSteps = input.steps.filter((s) => s.kind !== 'say');
-    const noGit = input.artifacts?.noGitTracking;
-    const agentName = input.agent === 'conductor' ? 'the agent codehere chose' : input.agent;
-    const parts = [
-        `No plain-English summary was available for this run — ${agentName} didn't leave a written summary, and codehere doesn't fabricate one. Here are the recorded facts.`,
-        `The session ran ${input.steps.length} step${input.steps.length === 1 ? '' : 's'}` +
-            (toolSteps.length
-                ? ` (${toolSteps.slice(0, 5).map((s) => s.label).join(', ')}${toolSteps.length > 5 ? ', …' : ''})`
-                : '') +
-            `.`,
-    ];
-    if (noGit) {
-        parts.push(`This folder isn't a git repository, so codehere couldn't track which files changed — check the working tree yourself.`);
-    }
-    else if (files.length) {
-        parts.push(`Files changed: ` +
-            files
-                .map((f) => `${f.path} (+${f.added}/-${f.removed}${f.trust !== undefined ? `, trust ${f.trust}` : ''})`)
-                .join('; ') +
-            `.`);
-    }
-    else {
-        parts.push(`No files were changed.`);
-    }
-    return { summary: parts.join('\n\n'), generatedAt: Date.now(), durationMs: 0, source: 'deterministic' };
-}
-function buildErrorRecoveryReport(input, durationMs) {
-    const agentName = input.agent === 'conductor' ? 'the agent codehere chose' : input.agent;
-    return {
-        summary: `The ${agentName} session ended without completing the task. ` +
-            `This usually means a network hiccup, an expired login, or a rate limit on the provider.\n\n` +
-            `What to try: check your internet, run \`codehere health\` to confirm reachability + auth, ` +
-            `or wait a minute and rerun the same prompt. If it keeps happening, try a different agent ` +
-            `from the dropdown — codehere is vendor-neutral on purpose.\n\n` +
-            `codehere couldn't confirm any file changes from this attempt.`,
-        generatedAt: Date.now(),
-        durationMs,
-        source: 'error-recovery',
-    };
-}
-/**
- * Generate a Final Report. Reuses the agent's own summary; falls back to a
- * deterministic, honest facts report. A completed task ALWAYS gets a report
- * (never null for non-error terminal states) — no external dependency.
- */
-export async function generateFinalReport(input) {
-    // No mock short-circuit: under CODEHERE_MOCK the SAME engine runs, so the
-    // narration gap is reproduced faithfully — claude/aider (emit `say`) get a
-    // rich agent-summary; codex/gemini/opencode (no `say`) get the honest
-    // deterministic report, exactly as on a real run. (A prior mock-only
-    // short-circuit synthesized a report from steps for every agent, hiding the
-    // cross-vendor gap — you couldn't dogfood "codex's report is thin" under mock.)
-    // Errored session — deterministic recovery hint, no summarization of nothing.
-    if (input.status === 'error') {
-        return buildErrorRecoveryReport(input, 0);
-    }
-    // ENGINE: reuse the agent's own narration (no fresh spawn).
-    const agentSummary = extractAgentSummary(input.steps);
-    if (agentSummary && !isLikelyNetworkFailureSummary(agentSummary)) {
-        return { summary: agentSummary, generatedAt: Date.now(), durationMs: 0, source: 'agent-summary' };
-    }
-    // FALLBACK: deterministic, honest — always a deliverable, never a fake ✓.
-    return buildDeterministicReport(input);
-}
-/**
- * Pretty-print to terminal. Used by session-manager when a spawned task exits.
- * Wrapped in try/catch upstream — never let a print failure surface.
- */
-export function printFinalReportToTerminal(input, output) {
-    const banner = '─'.repeat(60);
-    process.stdout.write('\n');
-    process.stdout.write(`${banner}\n`);
-    process.stdout.write(`Final Report · ${input.agent} · ${input.sessionId.slice(0, 8)}\n`);
-    process.stdout.write(`${banner}\n`);
-    process.stdout.write(`${output.summary}\n`);
-    process.stdout.write(`${banner}\n`);
-    process.stdout.write(`(generated in ${output.durationMs}ms)\n`);
-    process.stdout.write('\n');
-}
-//# sourceMappingURL=generate-final-report.js.map

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

@@ -1,128 +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: same as generate-final-report (try 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() {
-    // Same precedence as generate-final-report: env override → claude-free → claude.
-    const override = process.env.CODEHERE_NARRATOR_BINARY?.trim();
-    if (override)
-        return [override];
-    return ['claude-free', '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

@@ -1,178 +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:
- *   - Same provider chain pattern (claude-free → claude → graceful skip)
- *   - 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() {
-    const override = process.env.CODEHERE_TITLE_BINARY;
-    if (override)
-        return [override];
-    return ['claude-free', '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