@hone-ai/cli 1.39.0 → 1.40.0

package/lib/branch-protection.js

@@ -17,14 +17,24 @@
  *  mergeContexts(existing, added)
  *      Pure. Dedupes context names and preserves first-seen order so a
  *      pre-existing custom check (e.g. 'lint') is preserved when Hone
- *      appends 'ai-code-review'.
+ *      appends the checks it requires.
+ *
+ *  pruneLegacyChecks(contexts, { reportedChecks, keep })
+ *      Pure. Removes LEGACY_REQUIRED_CHECK ('ai-code-review'), which earlier
+ *      setups required and no served workflow reports (HC-GF-348), unless a
+ *      workflow in the repository reports it or the caller requires it.
+ *      Removes nothing else.
  *
  *  parseGhError(stderr, exitCode)
  *      Pure. Classifies a `gh` invocation failure into a stable kind:
  *      no-gh / no-auth / no-admin / plan-restricted / org-ruleset / unknown.
  *
  *  installBranchProtection({ exec?, log?, opts? })
- *      Orchestrator. Returns { status, reason?, branch?, manualCommand? }
+ *      Orchestrator. opts.requiredChecks is the list `hone setup` reads from
+ *      the installed workflow (cli/lib/required-checks.js) and may be empty;
+ *      opts.reportedChecks is every check name the repository's workflows
+ *      report (null = unknown, which keeps the legacy check).
+ *      Returns { status, reason?, branch?, manualCommand? }
  *      where status ∈ 'installed' | 'skipped' | 'failed'. Never throws,
  *      never exits — degrades gracefully so `hone setup` always returns 0.
  *
@@ -34,7 +44,14 @@
  * invocations during testing.
  */
 
-const DEFAULT_REQUIRED_CHECKS = ['ai-code-review'];
+// HC-GF-348: no check is required by default. Setup passes the checks the repository's own workflow
+// reports (cli/lib/required-checks.js); an empty list is valid — the PUT still requires a review and
+// sets enforce_admins. This used to be ['ai-code-review'], a check nothing reports, which held every PR.
+const DEFAULT_REQUIRED_CHECKS = [];
+
+// The one context an earlier `hone setup` installed that no served workflow reports. Re-running setup
+// removes it (pruneLegacyChecks), so a repository already held by it is released.
+const LEGACY_REQUIRED_CHECK = 'ai-code-review';
 
 // ─────────────────────────────────────────────────────────────────────
 // defaultExec — production exec, wraps spawnSync
@@ -59,7 +76,8 @@ function buildPayload({ contexts = [], requireReviews } = {}) {
   for (const c of contexts) {
     if (!seen.has(c)) { seen.add(c); dedupedContexts.push(c); }
   }
-  const reviewCount = typeof requireReviews === 'number' && requireReviews > 0
+  // 0 is a real answer since HC-GF-353 (a repository with one writer); anything that is not a count is 1.
+  const reviewCount = Number.isInteger(requireReviews) && requireReviews >= 0
     ? requireReviews
     : 1;
   return {
@@ -90,13 +108,27 @@ function mergeContexts(existing, added) {
   return out;
 }
 
+// ─────────────────────────────────────────────────────────────────────
+// pruneLegacyChecks — pure (HC-GF-348)
+// ─────────────────────────────────────────────────────────────────────
+function pruneLegacyChecks(contexts, { reportedChecks = [], keep = [] } = {}) {
+  const arr = Array.isArray(contexts) ? contexts : [];
+  // null: the repository's workflows could not all be read, so whether one reports it is unknown.
+  if (reportedChecks == null) return arr.slice();
+  const reported = Array.isArray(reportedChecks) ? reportedChecks : [];
+  const kept = Array.isArray(keep) ? keep : [];
+  if (reported.includes(LEGACY_REQUIRED_CHECK) || kept.includes(LEGACY_REQUIRED_CHECK)) return arr.slice();
+  return arr.filter((c) => c !== LEGACY_REQUIRED_CHECK);
+}
+
 // ─────────────────────────────────────────────────────────────────────
 // parseGhError — pure
 // ─────────────────────────────────────────────────────────────────────
 function parseGhError(stderr = '', exitCode = 0) {
   const s = String(stderr || '');
   // gh binary missing entirely
-  if (exitCode === 127 || /command not found|gh:.*not found/i.test(s)) {
+  // Not GitHub's own `gh: Not Found (HTTP 404)`, which the shell's wording also matched.
+  if (exitCode === 127 || /command not found/i.test(s) || (/gh:.*not found/i.test(s) && !/\(HTTP \d{3}\)/.test(s))) {
     return { kind: 'no-gh', message: 'gh CLI not installed (or not on PATH)' };
   }
   // gh installed but not authenticated
@@ -128,28 +160,67 @@ function parseGhError(stderr = '', exitCode = 0) {
 // ─────────────────────────────────────────────────────────────────────
 // buildManualCommand — produces a copy-paste shell snippet for the user
 // ─────────────────────────────────────────────────────────────────────
-function buildManualCommand({ contexts = DEFAULT_REQUIRED_CHECKS, owner = '<owner>', repo = '<repo>', branch = 'main' } = {}) {
-  const contextLines = contexts
-    .map(c => `  -f 'required_status_checks[contexts][]=${c}' \\`)
-    .join('\n');
+function buildManualCommand({ contexts = DEFAULT_REQUIRED_CHECKS, owner = '<owner>', repo = '<repo>', branch = 'main', requireReviews } = {}) {
+  // HC-GF-348: the exact payload setup PUTs, sent as JSON. `-f` fields cannot carry an empty contexts
+  // list (now the default) or a null `restrictions` — `-f 'restrictions='` sent the empty string.
   return [
-    `gh api --method PUT repos/${owner}/${repo}/branches/${branch}/protection \\`,
-    `  -f 'required_status_checks[strict]=true' \\`,
-    contextLines,
-    `  -f 'enforce_admins=true' \\`,
-    `  -f 'required_pull_request_reviews[required_approving_review_count]=1' \\`,
-    `  -f 'restrictions='`,
+    `gh api --method PUT repos/${owner}/${repo}/branches/${branch}/protection --input - <<'JSON'`,
+    JSON.stringify(buildPayload({ contexts, requireReviews }), null, 2),
+    'JSON',
   ].join('\n');
 }
 
+// ─────────────────────────────────────────────────────────────────────
+// resolveChecks — the required list, which may depend on the branch (HC-GF-354)
+// ─────────────────────────────────────────────────────────────────────
+// Setup's list depends on the protected branch, which only `gh repo view` names. On the paths that return
+// before that, the printed manual command is a template, so the list is resolved with branch null. A
+// resolver that throws requires nothing rather than failing the install.
+function resolveChecks(requiredChecks, branch) {
+  if (typeof requiredChecks === 'function') {
+    try {
+      const r = requiredChecks(branch);
+      return Array.isArray(r) ? r : [];
+    } catch {
+      return [];
+    }
+  }
+  return Array.isArray(requiredChecks) ? requiredChecks : DEFAULT_REQUIRED_CHECKS;
+}
+
+// ─────────────────────────────────────────────────────────────────────
+// countWriters — how many people can push (HC-GF-353)
+// ─────────────────────────────────────────────────────────────────────
+function countWriters(exec, owner, repo) {
+  let r;
+  try {
+    r = exec('gh', ['api', '--paginate', `repos/${owner}/${repo}/collaborators?per_page=100`]);
+  } catch (e) {
+    return { count: null, reason: `gh could not be run: ${e && e.message}` };
+  }
+  if (r.status !== 0) return { count: null, reason: parseGhError(r.stderr, r.status).message };
+  let list;
+  try {
+    // --paginate concatenates one JSON array per page: `[...][...]`.
+    const pages = JSON.parse(`[${String(r.stdout || '').trim().replace(/\]\s*\[/g, '],[')}]`);
+    list = pages.length && pages.every(Array.isArray) ? pages.flat() : null;
+  } catch {
+    list = null;
+  }
+  if (!Array.isArray(list) || list.length === 0) return { count: null, reason: 'the collaborator list could not be read' };
+  const n = list.filter((c) => c && c.permissions && (c.permissions.push === true || c.permissions.admin === true)).length;
+  // Nobody with a readable push or admin permission is an answer that could not be read, not "zero
+  // writers": it must never remove the review.
+  if (n === 0) return { count: null, reason: 'no collaborator in the list has a readable push or admin permission' };
+  return { count: n };
+}
+
 // ─────────────────────────────────────────────────────────────────────
 // installBranchProtection — orchestrator
 // ─────────────────────────────────────────────────────────────────────
 async function installBranchProtection({ exec = defaultExec, log = () => {}, opts = {} } = {}) {
-  const requiredChecks = (opts.requiredChecks && opts.requiredChecks.length)
-    ? opts.requiredChecks
-    : DEFAULT_REQUIRED_CHECKS;
-  const requireReviews = opts.requireReviews || 1;
+  // Not given: unknown, which keeps the legacy check. Only a caller that read the workflows can prune it.
+  const reportedChecks = opts.reportedChecks === undefined ? null : opts.reportedChecks;
 
   // 1. Honor opt-out — short-circuit before any subprocess call
   if (opts.skip) {
@@ -159,6 +230,12 @@ async function installBranchProtection({ exec = defaultExec, log = () => {}, opt
     };
   }
 
+  // An explicit list — empty included — is what the caller requires (HC-GF-348); a function of the
+  // protected branch (HC-GF-354); none given: the default. `requiredChecks` is the answer before the
+  // branch is known: for a function that is nothing, since a check it would name might never report on
+  // the real branch, and the manual command printed on those paths is a template the user edits.
+  const requiredChecks = typeof opts.requiredChecks === 'function' ? [] : resolveChecks(opts.requiredChecks, null);
+
   // 2. Resolve git remote
   const remote = exec('git', ['remote', 'get-url', 'origin']);
   if (remote.status !== 0) {
@@ -206,10 +283,23 @@ async function installBranchProtection({ exec = defaultExec, log = () => {}, opt
     };
   }
 
+  const branchChecks = resolveChecks(opts.requiredChecks, branch);
+
   // 4. GET existing protection (404 = none, 200 = merge)
   const protPath = `repos/${owner}/${repo}/branches/${branch}/protection`;
   const existing = exec('gh', ['api', protPath]);
   let existingContexts = [];
+  // HC-GF-348: only "Branch not protected" (HTTP 404) means there is nothing to merge. Any other failure
+  // to read the protection must not be taken for "none": the PUT replaces the contexts, so it would drop
+  // every check the repository already requires — and with this story the computed list can be empty.
+  if (existing.status !== 0 && !/Branch not protected/i.test(`${existing.stderr || ''}${existing.stdout || ''}`)) {
+    const err = parseGhError(existing.stderr, existing.status);
+    return {
+      status: 'failed',
+      reason: `could not read the existing branch protection, so it was left unchanged: ${err.message}`,
+      manualCommand: buildManualCommand({ contexts: branchChecks, owner, repo, branch }),
+    };
+  }
   if (existing.status === 0) {
     try {
       const e = JSON.parse(existing.stdout || '{}');
@@ -220,11 +310,32 @@ async function installBranchProtection({ exec = defaultExec, log = () => {}, opt
       existingContexts = [];
     }
   }
-  // (non-zero exit is treated as "no existing protection" — the PUT below will succeed if the user has permissions)
 
-  // 5. Build merged payload
-  const mergedContexts = mergeContexts(existingContexts, requiredChecks);
-  const payload = buildPayload({ contexts: mergedContexts, requireReviews });
+  // 5. Build merged payload; drop the legacy check an earlier setup installed (HC-GF-348)
+  const mergedContexts = pruneLegacyChecks(mergeContexts(existingContexts, branchChecks), {
+    reportedChecks, keep: branchChecks,
+  });
+
+  // HC-GF-353: GitHub does not count a PR author's approval of their own PR, so one required review on a
+  // repository with a single writer makes every PR unmergeable. An explicit opts.requireReviews wins. A
+  // count that could not be established keeps 1: failing open on human review would be the regression.
+  let reviewCount = 1;
+  let reviewReason;
+  if (Number.isInteger(opts.requireReviews) && opts.requireReviews >= 0) {
+    reviewCount = opts.requireReviews;
+    reviewReason = 'set by the caller';
+  } else {
+    const w = countWriters(exec, owner, repo);
+    if (w.count == null) {
+      reviewReason = `the number of people who can push could not be counted (${w.reason}), so one review is still required`;
+    } else if (w.count === 1) {
+      reviewCount = 0;
+      reviewReason = 'one person can push, and GitHub does not count a pull request author\'s own approval; a pull request is still required';
+    } else {
+      reviewReason = `${w.count} people can push`;
+    }
+  }
+  const payload = buildPayload({ contexts: mergedContexts, requireReviews: reviewCount });
 
   // 6. PUT
   const put = exec(
@@ -233,7 +344,7 @@ async function installBranchProtection({ exec = defaultExec, log = () => {}, opt
     { input: JSON.stringify(payload) }
   );
   if (put.status === 0) {
-    return { status: 'installed', branch };
+    return { status: 'installed', branch, requiredChecks: mergedContexts, reviewCount, reviewReason };
   }
 
   // 7. PUT failed — classify and return graceful failure
@@ -241,16 +352,21 @@ async function installBranchProtection({ exec = defaultExec, log = () => {}, opt
   return {
     status: 'failed',
     reason: err.message,
-    manualCommand: buildManualCommand({ contexts: mergedContexts, owner, repo, branch }),
+    manualCommand: buildManualCommand({ contexts: mergedContexts, owner, repo, branch, requireReviews: reviewCount }),
   };
 }
 
 module.exports = {
   buildPayload,
   mergeContexts,
+  pruneLegacyChecks,
+  resolveChecks,
+  countWriters,
   parseGhError,
   installBranchProtection,
   // Internal helpers exposed for inspection / future reuse
   buildManualCommand,
+  defaultExec,
   DEFAULT_REQUIRED_CHECKS,
+  LEGACY_REQUIRED_CHECK,
 };

package/lib/bundle-paths.js

@@ -256,8 +256,20 @@ function extractPathRefs(text) {
   return out;
 }
 
-/** Backticked spans in a story, longest first — the longest is the most specific anchor. */
-const BACKTICKED_RE = /`([^`\n]{3,200})`/g;
+/**
+ * Backticked spans in a story, PAIRED FIRST and length-unbounded; `storyTermsOf` applies the
+ * 3-to-200 rule to each paired span. HC-GF-188-followup-2.
+ *
+ * A BOUND INSIDE THE MATCH DECIDES WHICH BACKTICKS PAIR. At `` `x`, and `foo` `` the opening
+ * backtick of `x` could not start a 3-character match, so its CLOSING backtick did: the span ran
+ * to `foo`'s OPENING backtick, the "term" was `, and`, and `foo` was lost because no backtick was
+ * left to open it. A span over 200 characters, and an empty pair, shift the pairing the same way.
+ *
+ * The pairing is the one `maskCodeSpans` uses below: one rule for what a code span is, and the
+ * filter decides which spans can anchor. Fences are NOT masked here, which would widen the change
+ * past what the row measured.
+ */
+const BACKTICKED_RE = /`([^`\n]*)`/g;
 
 /**
  * The terms a story quotes that could locate code, longest first — one copy of the filter.
@@ -269,7 +281,8 @@ const BACKTICKED_RE = /`([^`\n]{3,200})`/g;
 function storyTermsOf(storyText) {
   return [...String(storyText || '').matchAll(BACKTICKED_RE)]
     .map((m) => m[1].trim())
-    .filter((c) => c.length >= 3
+    // Both bounds read the TRIMMED span; the ceiling used to live in the regex, on the raw span.
+    .filter((c) => c.length >= 3 && c.length <= 200
       && !new RegExp(`^[\\w.\\-/]+\\.(?:${FILE_EXT})$`).test(c)
       && !new RegExp(`^(?:${EXTENSIONLESS})$`).test(c))
     .sort((a, b) => b.length - a.length);
@@ -924,6 +937,10 @@ const LABEL_RE = /\*\*([A-Z][^*\n]{0,80}?):\*\*/g;
  * Fences are masked FIRST so a backtick inside a fence cannot pair with one
  * outside it.
  *
+ * HC-GF-188-followup-2: `BACKTICKED_RE` (read by `storyTermsOf`) pairs backticks the same way,
+ * `[^`\n]*` between two of them. Change the pairing in one place and change it in the other, or
+ * the two will disagree about what a code span is.
+ *
  * @param {string} text
  * @returns {string} same length as the input
  */

package/lib/judge-provider.js

@@ -16,6 +16,8 @@
  * the unshipped `scripts/` tree.
  */
 
+const { readAnthropicText } = require('./anthropic-response');
+
 /** Said by every command that used to default to GitHub Models, so the cause is one sentence. */
 const GH_MODELS_RETIRED =
   'GitHub Models (--provider gh-models), the free default, was retired by GitHub on 2026-07-30: '
@@ -38,25 +40,32 @@ function buildJudgeCallLLM(provider, { axios }) {
       return { error: 'ANTHROPIC_API_KEY required for --provider claude. Set it: export ANTHROPIC_API_KEY=sk-ant-...' };
     }
     // MODEL CHOICE (feedback_model_choice_cost_amplifier): the judge is NOT on the
-    // OPUS_AGENTS allowlist — paid mode uses Sonnet (`claude-sonnet-5`), never Opus.
+    // OPUS_AGENTS allowlist — paid mode uses Sonnet (`claude-sonnet-5-5`), never Opus.
+    //
+    // HC-GF-337f: Sonnet 5.5 answers HTTP 400 to `thinking: {type: 'disabled'}`. Thinking and effort are
+    // TWO sibling fields, never one: `thinking: {type: 'adaptive'}` and `output_config: {effort: 'low'}`
+    // (a `thinking` object with an `effort` key is wrong). Effort is sent because its default is `high`
+    // and a judge wants `low`. Thinking tokens share `max_tokens`, hence 8192. No `temperature`, `top_p`,
+    // `top_k` or `tool_choice` is sent, and that is load-bearing now: a non-default sampling value or a
+    // forced tool choice is a 400 on this model.
     const callLLM = async (systemPrompt, userPrompt) => {
       const { data } = await axios.post('https://api.anthropic.com/v1/messages', {
-        model: 'claude-sonnet-5',
-        thinking: { type: 'disabled' },
-        max_tokens: 2048,
+        model: 'claude-sonnet-5-5',
+        thinking: { type: 'adaptive' },
+        output_config: { effort: 'low' },
+        max_tokens: 8192,
         system: systemPrompt,
         messages: [{ role: 'user', content: userPrompt }],
       }, {
         headers: { 'x-api-key': apiKey, 'anthropic-version': '2023-06-01', 'content-type': 'application/json' },
-        timeout: 60000,
+        timeout: 120000,
       });
-      const text = (data.content || []).find((b) => b?.type === 'text')?.text;
       // A provider that answered with nothing is a provider failure, not a failed criterion:
-      // returning '' here made every criterion read "could not parse response".
-      if (typeof text !== 'string' || !text.trim()) {
-        throw new Error(`the claude provider returned no text (stop_reason: ${data.stop_reason || 'unknown'})`);
-      }
-      return text;
+      // returning '' here made every criterion read "could not parse response". A refusal (HTTP 200,
+      // stop_reason "refusal") names its category so it is not read as a quota or parse problem.
+      const r = readAnthropicText(data, { label: 'the claude provider', allowPartial: true }); // a verdict cut off defaults to NO
+      if (!r.ok) throw new Error(r.message);
+      return r.text;
     };
     return { callLLM, provider: 'claude', paid: true };
   }

package/lib/release-review-config.js

@@ -67,8 +67,9 @@ function resolveBaseRef(base) {
  * documentation later confirms the cap is TOTAL (input + output) instead
  * of input-only, drop this number AND `max_tokens` together.
  *
- * Anthropic Opus `claude-opus-4-8`: 1M context window (upgraded from
- * the legacy 4-20250514's 200K). The 100K char cap is well below the
+ * Anthropic Opus `claude-opus-5-5`: 1M context window (upgraded from
+ * the legacy 4-20250514's 200K; HC-GF-337f moved the id from claude-opus-4-8; the 1M figure is the
+ * Anthropic models table's context window for claude-opus-5-5, read 2026-10-10). The 100K char cap is well below the
  * model's headroom but kept conservative — the prompt + diff together
  * comfortably fit, and bumping the cap is a separate decision.
  *