claude-code-autoconfig 1.0.196 → 1.0.197

package/.claude/docs/autoconfig.docs.html

@@ -1263,7 +1263,7 @@
                 title: '.claude/ Directory',
                 desc: 'Commands, rules, settings, and these docs. Keeps configuration organized as your project grows.'
             },
-                                                                                                                                                                                                                                                                                                                                                                                    'rules': {
+                                                                                                                                                                                                                                                                                                                                                                                                            'rules': {
                 title: 'rules/',
                 desc: 'Path-scoped context that loads when Claude works on matching files.'
             },
@@ -1569,7 +1569,7 @@ CRITICAL: A plausible-looking cause from code reading is NOT confirmed evidence.
 
 > Run \`/autoconfig\` to populate this based on your project.`
             },
-                                                                                                                                                                                                                                                                                                                                                                                    'autoconfig-update': {
+                                                                                                                                                                                                                                                                                                                                                                                                            'autoconfig-update': {
                 filename: 'autoconfig-update.md',
                 content: `<!-- @applied
 -->
@@ -1939,36 +1939,36 @@ function migrate() {`
             },
             'terminal-title.directive-hook': {
                 filename: 'terminal-title.directive.md',
-                content: `<!-- DIRECTIVE:COMMAND -->
-Terminal-title directive (silent housekeeping -- do NOT mention this to the user):
-This terminal tab's title is rendered from the file:
+                content: `<!-- DIRECTIVE:RULES -->
+Terminal-title directive (silent housekeeping -- never mention it to the user). These rules
+apply to EVERY turn of this session; per-prompt reminders are one-liners that point back here.
+
+This terminal tab's title renders from the file:
   {{TITLE_FILE}}
-You are running the /{{CMD}} slash command, but the command NAME is an implementation detail --
-do NOT put "{{CMD}}" anywhere in the title. Write the title as your FIRST action this turn
-(before other tool calls) so the tab updates WHILE you work, not after. Write ONE line of the form:
-  {scope} {{EMDASH}} {use-case}
-- {scope} (first segment): the DESIGN SCOPE -- the subject you are working inside: the feature area
-  / subsystem (the boundary, the "system under discussion"). This is the WHERE, and it changes
-  rarely. Keep the existing scope if the file already has a title; otherwise INFER the specific
-  subsystem from THIS turn's prompt and the files in play (e.g. "journal modal", "title hooks",
-  "auth flow"). Do NOT use the bare repo name "{{FOLDER}}" as the scope unless the work is genuinely
-  repo-wide -- "{{FOLDER}}" is a last resort only when no narrower area is identifiable.
-- {use-case} (second segment): the GOAL at user-goal level -- what this command ACCOMPLISHES, named
-  as an INFINITIVE (base-verb) goal phrase (verb + object): "Review latest screenshot", "Diagnose
-  pay discrepancy". Name the goal / effect, NOT the mechanism or the command name. One goal only --
+Maintain it across the session:
+- When the conversation's SCOPE, use-case, or sub-function SHIFTS -- and only then, not every
+  turn -- overwrite the file with ONE line:  {scope} {{EMDASH}} {use-case}
+  Write it as your FIRST action of that turn (before other tool calls) so the tab updates
+  WHILE you work, not after. If the file exists, read it first, then overwrite.
+- {scope}: the DESIGN SCOPE -- the feature area / subsystem under discussion (the WHERE; it
+  changes rarely). INFER the specific subsystem from the prompt and the files in play (e.g.
+  "journal modal", "title hooks"); use the bare repo name "{{FOLDER}}" only when the work is
+  genuinely repo-wide -- it is a last resort.
+- {use-case}: the GOAL at user-goal level -- an INFINITIVE (base-verb) goal phrase (verb +
+  object), e.g. "Refine title taxonomy". Name the goal, not the mechanism. One goal only --
   never "and"-join two use cases.
-e.g. for /{{CMD}}, write "{{FOLDER}} {{EMDASH}} <verb the goal>" -- NOT "{{FOLDER}} {{EMDASH}} {{CMD}}".
-Use ' {{EMDASH}} ' as the separator; write only the BARE title (a state indicator is prepended
-automatically). If the file already exists, read it first then overwrite.
-<!-- /DIRECTIVE:COMMAND -->
-
-<!-- DIRECTIVE:SHIFT -->
-Terminal-title directive (silent housekeeping -- do NOT mention this to the user):
-This terminal tab's title is rendered from the file:
-  {{TITLE_FILE}}
-When the conversation's SCOPE, use-case, or sub-function SHIFTS -- and only then, not every turn --
-write ONE line to that file. When it DOES shift, write it as your FIRST action that turn (before
-other tool calls) so the tab updates WHILE you work, not after. The line's form:`
+- Add a third segment ( {{EMDASH}} {sub-function} ) only when the work goes a level deeper --
+  a step beneath the user goal.
+- Use ' {{EMDASH}} ' (space, em-dash, space) as the separator; keep segments short (a few
+  words); write only the BARE title -- a state glyph is prepended automatically. This is a
+  compass, not a log: change it rarely.
+- If no title has been set yet, treat your first turn as the baseline: infer the title from
+  that turn's prompt and write it immediately, even though nothing has "shifted".
+- Slash-command turns: the command NAME is an implementation detail -- never put it in the
+  title. Name the goal the command ACCOMPLISHES ("{{FOLDER}} {{EMDASH}} <verb the goal>",
+  never "{{FOLDER}} {{EMDASH}} <command-name>").
+
+Pending-question signal: when you END a turn on a question the user must answer before you`
             },
             'terminal-title-hook': {
                 filename: 'terminal-title.js',
@@ -1976,13 +1976,20 @@ other tool calls) so the tab updates WHILE you work, not after. The line's form:
 /**
  * Terminal Title — distributable plugin hook (installed to <project>/.claude/hooks/terminal-title.js).
  * ONE self-dispatching hook for five events (keyed on hook_event_name):
- *   UserPromptSubmit -> ⬤ working  + inject the title directive (incl. the {sid}.ask path) +
- *                       clear any stale {sid}.ask flag so it reflects only the turn about to run
+ *   UserPromptSubmit -> ⬤ working  + inject the per-prompt REMINDER one-liner (+BASELINE while no
+ *                       title exists, +COMMAND on /slash turns) + clear any stale {sid}.ask flag
  *   PostToolUse      -> ⬤ working  (refresh, so a mid-turn title flip shows live + clears a stale ◐)
  *   Notification     -> ◐ awaiting your approval (permission_prompt matcher only)
  *   Stop             -> ✻ idle / done — OR ◐ awaiting (+ a 2nd BEL = gold tab) when the turn ended
  *                       on a question (last visible response text ends in '?', or a {sid}.ask flag)
  *   SessionStart     -> ✻ idle "Claude Code — New session" (or an existing title on resume/compact)
+ *                       + inject the FULL RULES block — once per session instead of every prompt
+ *                       (~90% less directive overhead); resume/compact re-inject so a squeezed
+ *                       context re-learns the rules.
+ *
+ * Dedupe: when this file is the USER-LEVEL copy (~/.claude/hooks) and the project ships its own
+ * managed copy (<cwd>/.claude/hooks/terminal-title.js), this copy stands down entirely — otherwise
+ * both hooks inject directives naming two different title files (double tokens, double Writes).
  *
  * The title is set TWO ways on every event: \`process.title\` (= SetConsoleTitleW on Windows — the ONLY
  * mechanism that flips the tab on UserPromptSubmit, where Claude Code drops \`terminalSequence\`) AND
@@ -1994,13 +2001,7 @@ other tool calls) so the tab updates WHILE you work, not after. The line's form:
  * ~512KB-capped, written to .titles/_debug.log — so it never ships a growing log.
  *
  * Requires \`env.CLAUDE_CODE_DISABLE_TERMINAL_TITLE = "1"\` (set by plugin.json), or CC's own writer races.
- */
-const fs = require('fs');
-const path = require('path');
-
-const ESC = '\\x1b';
-const BEL = '\\x07';
-const EMDASH = String.fromCodePoint(0x2014);`
+ */`
             },
             'feedback-feedback': {
                 filename: 'FEEDBACK.md',

package/.claude/hooks/terminal-title.directive.md

@@ -1,82 +1,86 @@
-<!--
-  terminal-title.directive.md — the instruction terminal-title.js injects into the
-  model's context on every prompt. THIS is the tunable "product" surface: reword it
-  freely without touching the hook's logic.
-
-  Blocks below:
-    COMMAND — used when the turn starts with a /slash-command
-    SHIFT   — used on a normal turn
-    PENDING — appended to WHICHEVER of the above was selected (shared tail)
-  terminal-title.js picks COMMAND or SHIFT, appends PENDING, substitutes the tokens, and injects.
-
-  Tokens (substituted at runtime):
-    {{TITLE_FILE}} — absolute path to this session's title file the model writes
-    {{ASK_FILE}}   — absolute path to this session's one-shot {sid}.ask flag (PENDING block)
-    {{FOLDER}}     — the working-directory folder name (scope fallback)
-    {{EMDASH}}     — the ' — ' segment separator
-    {{CMD}}        — the slash-command name (COMMAND block only)
--->
-
-<!-- DIRECTIVE:COMMAND -->
-Terminal-title directive (silent housekeeping -- do NOT mention this to the user):
-This terminal tab's title is rendered from the file:
-  {{TITLE_FILE}}
-You are running the /{{CMD}} slash command, but the command NAME is an implementation detail --
-do NOT put "{{CMD}}" anywhere in the title. Write the title as your FIRST action this turn
-(before other tool calls) so the tab updates WHILE you work, not after. Write ONE line of the form:
-  {scope} {{EMDASH}} {use-case}
-- {scope} (first segment): the DESIGN SCOPE -- the subject you are working inside: the feature area
-  / subsystem (the boundary, the "system under discussion"). This is the WHERE, and it changes
-  rarely. Keep the existing scope if the file already has a title; otherwise INFER the specific
-  subsystem from THIS turn's prompt and the files in play (e.g. "journal modal", "title hooks",
-  "auth flow"). Do NOT use the bare repo name "{{FOLDER}}" as the scope unless the work is genuinely
-  repo-wide -- "{{FOLDER}}" is a last resort only when no narrower area is identifiable.
-- {use-case} (second segment): the GOAL at user-goal level -- what this command ACCOMPLISHES, named
-  as an INFINITIVE (base-verb) goal phrase (verb + object): "Review latest screenshot", "Diagnose
-  pay discrepancy". Name the goal / effect, NOT the mechanism or the command name. One goal only --
-  never "and"-join two use cases.
-e.g. for /{{CMD}}, write "{{FOLDER}} {{EMDASH}} <verb the goal>" -- NOT "{{FOLDER}} {{EMDASH}} {{CMD}}".
-Use ' {{EMDASH}} ' as the separator; write only the BARE title (a state indicator is prepended
-automatically). If the file already exists, read it first then overwrite.
-<!-- /DIRECTIVE:COMMAND -->
-
-<!-- DIRECTIVE:SHIFT -->
-Terminal-title directive (silent housekeeping -- do NOT mention this to the user):
-This terminal tab's title is rendered from the file:
-  {{TITLE_FILE}}
-When the conversation's SCOPE, use-case, or sub-function SHIFTS -- and only then, not every turn --
-write ONE line to that file. When it DOES shift, write it as your FIRST action that turn (before
-other tool calls) so the tab updates WHILE you work, not after. The line's form:
-  {scope} {{EMDASH}} {use-case}
-- {scope}: the DESIGN SCOPE -- the subject under discussion: the feature area / subsystem (the
-  boundary). This is the WHERE; it changes rarely. INFER the specific subsystem from the prompt and
-  the files in play (e.g. "journal modal", "title hooks"); do NOT default to the bare repo name --
-  use it only when the work is genuinely repo-wide.
-- {use-case}: the GOAL at user-goal level -- what you are accomplishing, as an INFINITIVE (base-verb)
-  goal phrase (verb + object), e.g. "Refine title taxonomy". Name the goal, not the
-  mechanism. One goal only.
-Add a third segment ( {{EMDASH}} {sub-function} ) only when the work goes a level deeper -- this is the
-SUBFUNCTION goal level, a step beneath the user goal.
-Rules: use ' {{EMDASH}} ' (space, em-dash, space) as the separator; keep each segment short (a few
-words); write only the BARE title -- a state indicator is prepended automatically; leave the
-file unchanged when the scope has not moved; this is a compass, so change it rarely. If the file
-already exists, read it first then overwrite.
-<!-- /DIRECTIVE:SHIFT -->
-
-<!-- DIRECTIVE:PENDING -->
-Pending-question signal: when you END this turn on a question the user must answer before you can
-proceed, do BOTH of these as near-final actions so the tab flips to the AWAITING half-circle
-(instead of the idle asterisk):
-  1. Write the flag file {{ASK_FILE}} (any short content, e.g. "1"). This is the RELIABLE trigger --
-     it is on disk before the turn ends, so it never misses (the transcript-text check in 2 can race
-     the turn-end write and silently miss, or miss on phrasing -- see below). The flag is one-shot: it
-     is consumed at turn end and auto-cleared next turn, so write it ONLY on a turn you are genuinely
-     blocked on an answer. Write it WHENEVER the turn genuinely needs the user's answer to proceed --
-     the flag doesn't parse your text, so it is correct even when your closing question is wrapped in
-     parens, phrased as a parenthetical aside, or is not the literal final character of the message.
-  2. Phrase your FINAL line so it ends with a question mark ('?') -- a backup signal, and good UX. A
-     single trailing parenthetical aside after the '?' is fine ("...option A or B? (I lean B.)"), but
-     do not rely on phrasing for the signal -- the flag in step 1 is what reliably flips the tab.
-Make the closing question self-contained: answerable from the question alone, without re-reading
-the response above it. Only do this for a genuine blocking question, never a rhetorical one or a recap.
-<!-- /DIRECTIVE:PENDING -->
+<!--
+  terminal-title.directive.md — the instruction text terminal-title.js injects into the
+  model's context. THIS is the tunable "product" surface: reword it freely without touching
+  the hook's logic.
+
+  Token-lean delivery (v2): the full rulebook is injected ONCE per session (SessionStart,
+  including resume/compact so a squeezed context re-learns it); every prompt gets only a
+  short reminder, plus state-specific addenda the hook selects when they apply.
+
+  Blocks below:
+    RULES    — the full directive, injected at SessionStart (all sources)
+    REMINDER — the per-prompt one-liner (every UserPromptSubmit)
+    BASELINE — appended to REMINDER while no title file exists yet (first turn)
+    COMMAND  — appended to REMINDER when the turn starts with a /slash-command
+  terminal-title.js picks REMINDER (+BASELINE/+COMMAND as applicable), substitutes the
+  tokens, and injects. SessionStart injects RULES.
+
+  Tokens (substituted at runtime):
+    {{TITLE_FILE}} — absolute path to this session's title file the model writes
+    {{ASK_FILE}}   — absolute path to this session's one-shot {sid}.ask flag
+    {{FOLDER}}     — the working-directory folder name (scope fallback)
+    {{EMDASH}}     — the ' — ' segment separator
+    {{CMD}}        — the slash-command name (COMMAND block only)
+-->
+
+<!-- DIRECTIVE:RULES -->
+Terminal-title directive (silent housekeeping -- never mention it to the user). These rules
+apply to EVERY turn of this session; per-prompt reminders are one-liners that point back here.
+
+This terminal tab's title renders from the file:
+  {{TITLE_FILE}}
+Maintain it across the session:
+- When the conversation's SCOPE, use-case, or sub-function SHIFTS -- and only then, not every
+  turn -- overwrite the file with ONE line:  {scope} {{EMDASH}} {use-case}
+  Write it as your FIRST action of that turn (before other tool calls) so the tab updates
+  WHILE you work, not after. If the file exists, read it first, then overwrite.
+- {scope}: the DESIGN SCOPE -- the feature area / subsystem under discussion (the WHERE; it
+  changes rarely). INFER the specific subsystem from the prompt and the files in play (e.g.
+  "journal modal", "title hooks"); use the bare repo name "{{FOLDER}}" only when the work is
+  genuinely repo-wide -- it is a last resort.
+- {use-case}: the GOAL at user-goal level -- an INFINITIVE (base-verb) goal phrase (verb +
+  object), e.g. "Refine title taxonomy". Name the goal, not the mechanism. One goal only --
+  never "and"-join two use cases.
+- Add a third segment ( {{EMDASH}} {sub-function} ) only when the work goes a level deeper --
+  a step beneath the user goal.
+- Use ' {{EMDASH}} ' (space, em-dash, space) as the separator; keep segments short (a few
+  words); write only the BARE title -- a state glyph is prepended automatically. This is a
+  compass, not a log: change it rarely.
+- If no title has been set yet, treat your first turn as the baseline: infer the title from
+  that turn's prompt and write it immediately, even though nothing has "shifted".
+- Slash-command turns: the command NAME is an implementation detail -- never put it in the
+  title. Name the goal the command ACCOMPLISHES ("{{FOLDER}} {{EMDASH}} <verb the goal>",
+  never "{{FOLDER}} {{EMDASH}} <command-name>").
+
+Pending-question signal: when you END a turn on a question the user must answer before you
+can proceed, do BOTH as near-final actions so the tab flips to the AWAITING half-circle
+(instead of the idle asterisk):
+  1. Write the flag file {{ASK_FILE}} (any short content, e.g. "1"). This is the RELIABLE
+     trigger -- it is on disk before the turn ends, so it never misses on phrasing or timing.
+     The flag is one-shot (consumed at turn end, auto-cleared next turn), so write it ONLY on
+     a turn genuinely blocked on an answer -- but ALWAYS then, even when the closing question
+     is wrapped in parens or is not the literal final character of the message.
+  2. Phrase your FINAL line to end with a question mark ('?') -- the backup signal, and good
+     UX. A single trailing parenthetical aside after the '?' is fine.
+Make the closing question self-contained: answerable without re-reading the response above
+it. Never signal for a rhetorical question or a recap.
+<!-- /DIRECTIVE:RULES -->
+
+<!-- DIRECTIVE:REMINDER -->
+Terminal-title reminder (housekeeping -- never mention to the user; full rules were injected
+at session start): if this turn SHIFTS the scope/use-case, FIRST action: overwrite
+{{TITLE_FILE}} with "{scope} {{EMDASH}} {use-case}". If you END this turn blocked on a
+question, write the flag file {{ASK_FILE}} and end your final line with '?'.
+<!-- /DIRECTIVE:REMINDER -->
+
+<!-- DIRECTIVE:BASELINE -->
+No title is set yet -- treat THIS turn as the baseline: infer {scope} {{EMDASH}} {use-case}
+from this prompt and the files in play, and write {{TITLE_FILE}} NOW (first action). Do not
+default to the bare folder name "{{FOLDER}}" unless the work is genuinely repo-wide.
+<!-- /DIRECTIVE:BASELINE -->
+
+<!-- DIRECTIVE:COMMAND -->
+This turn runs the /{{CMD}} slash command -- the command NAME is an implementation detail:
+never put "{{CMD}}" in the title. If the scope shifted, name the goal the command
+ACCOMPLISHES (e.g. "<area> {{EMDASH}} <verb the goal>").
+<!-- /DIRECTIVE:COMMAND -->

package/.claude/hooks/terminal-title.js

@@ -2,13 +2,20 @@
 /**
  * Terminal Title — distributable plugin hook (installed to <project>/.claude/hooks/terminal-title.js).
  * ONE self-dispatching hook for five events (keyed on hook_event_name):
- *   UserPromptSubmit -> ⬤ working  + inject the title directive (incl. the {sid}.ask path) +
- *                       clear any stale {sid}.ask flag so it reflects only the turn about to run
+ *   UserPromptSubmit -> ⬤ working  + inject the per-prompt REMINDER one-liner (+BASELINE while no
+ *                       title exists, +COMMAND on /slash turns) + clear any stale {sid}.ask flag
  *   PostToolUse      -> ⬤ working  (refresh, so a mid-turn title flip shows live + clears a stale ◐)
  *   Notification     -> ◐ awaiting your approval (permission_prompt matcher only)
  *   Stop             -> ✻ idle / done — OR ◐ awaiting (+ a 2nd BEL = gold tab) when the turn ended
  *                       on a question (last visible response text ends in '?', or a {sid}.ask flag)
  *   SessionStart     -> ✻ idle "Claude Code — New session" (or an existing title on resume/compact)
+ *                       + inject the FULL RULES block — once per session instead of every prompt
+ *                       (~90% less directive overhead); resume/compact re-inject so a squeezed
+ *                       context re-learns the rules.
+ *
+ * Dedupe: when this file is the USER-LEVEL copy (~/.claude/hooks) and the project ships its own
+ * managed copy (<cwd>/.claude/hooks/terminal-title.js), this copy stands down entirely — otherwise
+ * both hooks inject directives naming two different title files (double tokens, double Writes).
  *
  * The title is set TWO ways on every event: `process.title` (= SetConsoleTitleW on Windows — the ONLY
  * mechanism that flips the tab on UserPromptSubmit, where Claude Code drops `terminalSequence`) AND
@@ -23,6 +30,7 @@
  */
 const fs = require('fs');
 const path = require('path');
+const os = require('os');
 
 const ESC = '\x1b';
 const BEL = '\x07';
@@ -61,6 +69,10 @@ async function handle(data) {
   const event = data.hook_event_name || '';
   const sid = data.session_id || '';
   const cwd = data.cwd || process.cwd();
+  // User-level copy stands down when a project-level managed copy exists (see header).
+  if (shouldDefer(cwd, __dirname, __filename, path.join(os.homedir(), '.claude', 'hooks'))) {
+    process.exit(0);
+  }
   // PROJECT-SCOPED title dir (cwd-relative) — the plugin installs per-project. (The live ~/.claude
   // variant uses os.homedir() instead; that is the only difference between the two.)
   const dir = path.join(cwd, '.claude', 'hooks', '.titles');
@@ -96,7 +108,14 @@ async function handle(data) {
   if (event === 'SessionStart') {
     // Fresh-session placeholder; on resume/compact an existing model-authored title is preferred.
     const title = normalize(readTitle(file) || 'Claude Code - New session');
-    emit(setTitle(GLYPH.idle, title));
+    const out = setTitle(GLYPH.idle, title);
+    // Inject the FULL rulebook here — once per session — instead of on every prompt. All sources
+    // get it: startup/clear teach a fresh context, resume/compact re-teach a squeezed one.
+    const rules = buildBlocks(['RULES'], file, cwd, '');
+    if (rules) {
+      out.hookSpecificOutput = { hookEventName: 'SessionStart', additionalContext: rules };
+    }
+    emit(out);
     return;
   }
 
@@ -281,28 +300,47 @@ function emit(obj) {
   process.exit(0);
 }
 
-// Build the injected directive from terminal-title.directive.md (COMMAND block for a /slash-command
-// turn, SHIFT otherwise; PENDING tail appended). Wording lives in the .md so it tunes without code edits.
+// Build the per-prompt injection: the REMINDER one-liner, plus BASELINE while no title file
+// exists yet (first turn), plus COMMAND on a /slash-command turn. The full rulebook (RULES) is
+// injected by SessionStart, not here. Wording lives in the .md so it tunes without code edits.
 function buildDirective(data, file, cwd) {
+  const prompt = typeof data.prompt === 'string' ? data.prompt : '';
+  const m = prompt.match(/^\s*\/([A-Za-z0-9][\w:.-]*)/);
+  const blocks = ['REMINDER'];
+  if (!fileExists(file)) blocks.push('BASELINE');
+  if (m) blocks.push('COMMAND');
+  return buildBlocks(blocks, file, cwd, m ? m[1] : '');
+}
+
+// Extract the named blocks from terminal-title.directive.md, join them, substitute the tokens.
+function buildBlocks(names, file, cwd, cmd) {
   let tpl = '';
   try {
     tpl = fs.readFileSync(path.join(__dirname, 'terminal-title.directive.md'), 'utf8');
   } catch (_) {
     return '';
   }
-  const prompt = typeof data.prompt === 'string' ? data.prompt : '';
-  const m = prompt.match(/^\s*\/([A-Za-z0-9][\w:.-]*)/);
-  const cmd = m ? m[1] : '';
-  const block = extractBlock(tpl, cmd ? 'COMMAND' : 'SHIFT');
-  if (!block) return '';
-  const pending = extractBlock(tpl, 'PENDING');
-  const combined = pending ? `${block}\n\n${pending}` : block;
-  return combined
+  const parts = names.map(n => extractBlock(tpl, n)).filter(Boolean);
+  if (!parts.length) return '';
+  return parts.join('\n')
     .split('{{TITLE_FILE}}').join(file)
     .split('{{ASK_FILE}}').join(file.replace(/\.txt$/, '.ask'))
     .split('{{FOLDER}}').join(folderName(cwd))
     .split('{{EMDASH}}').join(EMDASH)
-    .split('{{CMD}}').join(cmd);
+    .split('{{CMD}}').join(cmd || '');
+}
+
+// Should THIS copy of the hook stand down? True only for the user-level copy (~/.claude/hooks)
+// running inside a project that ships its own managed copy — the project copy wins, so a session
+// gets ONE directive and ONE title file. Parameterized (no ambient __dirname/homedir) for tests.
+function shouldDefer(cwd, selfDir, selfFile, homeHooksDir) {
+  try {
+    if (path.resolve(selfDir) !== path.resolve(homeHooksDir)) return false; // we ARE the project copy
+    const projectCopy = path.join(cwd, '.claude', 'hooks', 'terminal-title.js');
+    return fs.existsSync(projectCopy) && path.resolve(projectCopy) !== path.resolve(selfFile);
+  } catch (_) {
+    return false;
+  }
 }
 
 function extractBlock(tpl, name) {
@@ -314,4 +352,4 @@ function extractBlock(tpl, name) {
 }
 
 // Exported for tests (require()'d when require.main !== module). The hook itself never reads these.
-module.exports = { inspectLastResponse, normalize, GLYPH };
+module.exports = { inspectLastResponse, normalize, GLYPH, shouldDefer };

package/CHANGELOG.md

@@ -1,5 +1,8 @@
 # Changelog
 
+## v1.0.197
+- perf(terminal-title): deliver the directive once per session; dedupe user/project hooks
+
 ## v1.0.196
 - fix(terminal-title): stop at the turn boundary when grading the closing '?' (stale-awaiting race)
 - fix(cli): require non-TTY stdout for inside-Claude block (env var alone false-positives in inherited terminals)
@@ -145,6 +148,3 @@
 ## v1.0.148
 - fix: refine inside-Claude error wording
 
-## v1.0.147
-- fix: make inside-Claude error message more specific
-

package/CLAUDE.md

@@ -42,17 +42,13 @@ See `.claude/feedback/` for corrections and guidance from the team.
 ## Discoveries
 <!-- Claude: append project-specific learnings, gotchas, and context below. This section persists across /autoconfig runs. -->
 
-### Publishing to npm — web-auth (passkey) flow, run by the human
+### Publishing to npm — bypass-2FA token flow (web-auth publish is a DEAD END)
 
-`npm publish` for this package is a **two-actor** flow; a bare in-agent `npm publish` will FAIL. (Supersedes the auto-generated "## Publishing" one-liner above.)
+(Supersedes the auto-generated "## Publishing" one-liner above.)
 
-- The npm account has **2FA**. A sandboxed/agent shell can't open a browser, so `npm publish` there dies with `npm error code EOTP`. Don't retry it in-agent; don't default to asking for a typed OTP.
-- **Working flow:** the agent runs `npm test` + `npm version patch` (commit + tag), then the **human runs the publish in their own terminal** (needs a browser):
-  ```
-  cd C:\CODE\claude-code-autoconfig && npm login --auth-type=web && npm publish
-  ```
-  → browser opens → pick the Google passkey → close it → `+ claude-code-autoconfig@<version>`.
-- **After** the success line, the agent pushes: `git push origin main --follow-tags`. Never double-bump if the version is already bumped. Full procedure: `.claude/commands/deploy-to-npmjs.md` (`/deploy-to-npmjs`).
+- **Primary flow (agent can do it end-to-end):** `npm test` + `npm version patch` (commit + tag) → `npm whoami` to verify the Bypass-2FA granular token in `~/.npmrc` still authenticates → plain `npm publish` (works headless, no OTP) → after the `+ claude-code-autoconfig@<version>` line, `git push origin main --follow-tags`. Never double-bump if the version is already bumped.
+- **If `npm whoami` returns E401**, the token is dead/expired (90-day cap). The human rotates it (npmjs.com → Access Tokens → Granular → "Bypass 2FA", read-write, scoped to this package only) and pastes it into `~/.npmrc` in an editor — never via a command. Stopgap without a token: `npm publish --otp=<recovery-code>` in the human's own terminal.
+- **Do NOT use `npm login --auth-type=web && npm publish`:** login succeeds but publish still dies `EOTP` — the npm CLI can't route publish-time 2FA to the browser on this webauthn-only account (confirmed 2026-07-03, npm 10.9.2). Don't retry it in-agent, and don't ask for a typed TOTP (none exists). Full procedure: `.claude/commands/deploy-to-npmjs.md` (`/deploy-to-npmjs`).
 
 ## Debugging Methodology — Evidence Before Solutions
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-code-autoconfig",
-  "version": "1.0.196",
+  "version": "1.0.197",
   "description": "Intelligent, self-configuring setup for Claude Code. One command analyzes your project, configures Claude, and shows you what it did.",
   "author": "ADAC 1001 <info@adac1001.com>",
   "license": "MIT",
@@ -23,14 +23,14 @@
   ],
   "scripts": {
     "test": "node test/box-alignment.test.js && node test/cli-install.test.js && node test/update-system.test.js && node test/plugin-system.test.js && node test/terminal-title.test.js",
-    "pub": "npm login --auth-type=web && npm publish",
+    "pub": "npm whoami && npm publish",
     "test:box": "node test/box-alignment.test.js",
     "test:install": "node test/cli-install.test.js",
     "test:update": "node test/update-system.test.js",
     "postversion": "node scripts/generate-changelog.js --postversion"
   },
   "bin": {
-    "claude-code-autoconfig": "./bin/cli.js"
+    "claude-code-autoconfig": "bin/cli.js"
   },
   "files": [
     "bin",