flowviant 0.97.0 → 0.99.0

package/bin/lib/authproxy.mjs

@@ -116,6 +116,10 @@ const MAX_SOURCES = 500;
 
 const digest = (s) => createHash('sha256').update(String(s)).digest();
 
+/** An Authorization header in the Basic scheme — the only shape the gate's own
+ *  password ever arrives in, and so the only one it grades. */
+const isBasic = (v) => typeof v === 'string' && /^\s*basic(\s|$)/i.test(v);
+
 /** Constant-time over sha256 digests, so length never leaks and a missing
  *  header costs the same as a wrong one. */
 function sameSecret(a, b) {
@@ -137,7 +141,18 @@ function sameSecret(a, b) {
  * URL under attack. A single source is blocked on its own, at MAX_FAILED,
  * without ending anybody else's share.
  */
-export function startAuthProxy({ targetPort, log, onAbuse, grantSecret, shareId, authorizeUrl }) {
+export function startAuthProxy({
+  targetPort,
+  // The address that reaches the ATTRIBUTED socket (listeners.mjs `originFor`)
+  // — `::1` for a dev server bound to the IPv6 loopback. Defaults to the IPv4
+  // loopback it always dialled, for a caller that measured nothing.
+  targetHost = '127.0.0.1',
+  log,
+  onAbuse,
+  grantSecret,
+  shareId,
+  authorizeUrl,
+}) {
   // ALL THREE OR NONE. Two of the three is a gate that cannot bounce anybody:
   // a secret with no authorize URL has nowhere to send them, an authorize URL
   // with no secret cannot verify what comes back. An older SERVER sends none of
@@ -217,7 +232,36 @@ export function startAuthProxy({ targetPort, log, onAbuse, grantSecret, shareId,
     // whole thing down. An expired-but-validly-signed grant must never count
     // either, or a viewer who left a tab open overnight closes the share on
     // their own reload.
-    if (req.headers['authorization']) {
+    //
+    // AND ONLY A *BASIC* HEADER IS A PASSWORD ATTEMPT (audit 2026-09-24). Any
+    // Authorization header used to be graded against the gate password, so a
+    // previewed SPA calling its own `/api` with `Authorization: Bearer …` was
+    // answered 401 with a Basic challenge — over a valid grant cookie — and 25
+    // such calls blocked that viewer, eight viewers tore the share down, and
+    // the origin never saw its own header. A Bearer (or any other scheme) is
+    // the APP's credential: it is not ours to grade and it rides through.
+    //
+    // A VALID GRANT COOKIE IS ASKED FIRST, for the same reason: an app that
+    // itself speaks Basic sends a header that is not our password, and a
+    // viewer the app has vouched for must not be graded on it. A grant is
+    // HMAC-verified, so asking it first opens no brute-force path.
+    let grantVerdict = 'none';
+    if (grants) {
+      // EVERY value for our name, not the first — a duplicate must not shadow.
+      for (const raw of cookieValues(req.headers.cookie, GRANT_COOKIE)) {
+        const r = verifyGrant(raw, { secret: grantSecret, shareId });
+        if (r.ok) return 'ok';
+        if (r.reason === 'exp' && grantVerdict === 'none') {
+          lastExpired = r.payload;
+          grantVerdict = 'expired';
+        }
+      }
+      if (grantVerdict === 'none' && String(req.headers.cookie ?? '').includes(GRANT_COOKIE)) {
+        grantVerdict = 'forged';
+      }
+    }
+
+    if (isBasic(req.headers['authorization'])) {
       // A blocked source is refused BEFORE the comparison — a block that
       // still grades guesses would let the brute force run to a correct hit.
       if (sourceBlocked(req)) return 'blocked';
@@ -230,18 +274,7 @@ export function startAuthProxy({ targetPort, log, onAbuse, grantSecret, shareId,
       noteFailure(req);
       return 'badpass';
     }
-
-    if (!grants) return 'none';
-    // EVERY value for our name, not the first — a duplicate must not shadow.
-    for (const raw of cookieValues(req.headers.cookie, GRANT_COOKIE)) {
-      const r = verifyGrant(raw, { secret: grantSecret, shareId });
-      if (r.ok) return 'ok';
-      if (r.reason === 'exp') {
-        lastExpired = r.payload;
-        return 'expired';
-      }
-    }
-    return String(req.headers.cookie ?? '').includes(GRANT_COOKIE) ? 'forged' : 'none';
+    return grantVerdict;
   };
 
   /**
@@ -338,7 +371,10 @@ export function startAuthProxy({ targetPort, log, onAbuse, grantSecret, shareId,
   // its request would be us editing their app's input.
   const forwardOpts = (req) => {
     const headers = { ...req.headers };
-    delete headers.authorization;
+    // Only when it IS the gate credential. Anything else in Authorization is
+    // the previewed app's own (a Bearer to its API, its own Basic realm) and
+    // the app breaks without it; ours must still never reach branch code.
+    if (sameSecret(headers.authorization, expected)) delete headers.authorization;
     delete headers['proxy-authorization'];
     // ONLY OURS. The driver's app owns its own cookies and breaks without them;
     // our grant is a signed bearer token and must not reach branch code.
@@ -346,7 +382,7 @@ export function startAuthProxy({ targetPort, log, onAbuse, grantSecret, shareId,
     if (rest) headers.cookie = rest;
     else delete headers.cookie; // never send a bare empty `cookie:`
     return {
-      host: '127.0.0.1',
+      host: targetHost,
       port: targetPort,
       method: req.method,
       path: req.url,

package/bin/lib/childEnv.mjs

@@ -223,9 +223,41 @@ export const DEPLOY_KEEP_NAMES = [...DEPLOY_KEEP];
  * never leave the box either way; names would, and a name is enough to tell
  * somebody which credential to go looking for.
  */
+/**
+ * …AND THE DAEMON'S OWN CREDENTIALS, REDACTED AND NEVER ADMITTED (2026-09-24,
+ * the audit).
+ *
+ * The list above was written about the deploy lane, and it left out every
+ * credential the DAEMON itself runs on: the machine credential
+ * (`FLOWVIANT_MACHINE_TOKEN`, or its permanent `FLOWVIANT_FLEET` fallback on a
+ * headless box), and whatever signs the operator's CLIs and `gh` in. Those are
+ * the values a CLI turn can most easily reach — a turn inherits this process's
+ * environment, and `env` is one Bash call — and a turn trace, an answer and an
+ * artifact are all posted to the server. Redacting a value we never pass costs
+ * nothing (the `CF_*` argument above), so they join the REDACTION list only;
+ * `DEPLOY_KEEP` is untouched and `childEnv` still hands none of them to anybody.
+ *
+ * The machine credential is ALSO caught by shape in `env.mjs` (`fva_…`),
+ * because the copy that matters most is not in any environment variable at all:
+ * it is every project's `fleetToken` in `~/.flowviant/credentials.json`, plus
+ * the per-turn work/capture tokens the server mints. The names here are belt.
+ */
+const REDACT_ONLY = [
+  'CF_API_TOKEN',
+  'CF_ACCOUNT_ID',
+  'FLOWVIANT_MACHINE_TOKEN',
+  'FLOWVIANT_FLEET',
+  'ANTHROPIC_API_KEY',
+  'ANTHROPIC_AUTH_TOKEN',
+  'CLAUDE_CODE_OAUTH_TOKEN',
+  'OPENAI_API_KEY',
+  'GH_TOKEN',
+  'GITHUB_TOKEN',
+];
+
 export function processEnvSecrets() {
   const out = [];
-  for (const name of [...DEPLOY_KEEP, 'CF_API_TOKEN', 'CF_ACCOUNT_ID']) {
+  for (const name of [...DEPLOY_KEEP, ...REDACT_ONLY]) {
     const value = process.env[name];
     if (typeof value === 'string' && value) out.push({ name, value });
   }

package/bin/lib/claude.mjs

@@ -61,10 +61,66 @@ export const READ_GUARD_SETTINGS = JSON.stringify({
  *  `--settings` pair must stand before it or be swallowed as a tool name. */
 const READ_GUARD = ['--settings', READ_GUARD_SETTINGS];
 
+/** A path a permission rule can name: absolute POSIX, and free of the glob and
+ *  rule-syntax characters that would change what the rule means. Anything else
+ *  gets no rule — `--add-dir` still admits the directory (measured). */
+const ruleSafeAbsolute = (p) =>
+  typeof p === 'string' && p.startsWith('/') && !/[*?[\]{}()\\\n]/.test(p) ? p.replace(/\/+$/, '') : null;
+
+/**
+ * THE READ FENCE — the worktree, plus the knowledge library by its absolute
+ * path, and nothing else on the box (2026-09-24, the audit).
+ *
+ * Every curated posture used to allow bare `Read`, `Grep` and `Glob`, and a
+ * bare allow is not "read the repo": measured on Claude Code 2.1.281, bare
+ * `Read` read a file outside the working directory and bare `Grep` searched
+ * one. So a design, consult or capture turn — each steered by words anyone who
+ * can file a card can write — could read `~/.flowviant/credentials.json`, which
+ * holds the MACHINE CREDENTIAL of every project connected on this box, and put
+ * it in an artifact, a plan note or a transcript that leaves the machine. That
+ * is a cross-project leak from a posture this file calls read-only.
+ *
+ * The fence is research's, measured the same day in the design shape:
+ * `Read(./**)` and `Glob(./**)` read the worktree; the knowledge library reads
+ * through `--add-dir` plus its own `//` rule; a file outside both was refused
+ * ("Claude requested permissions to read from …, but you haven't granted it
+ * yet"). GREP CARRIES NO ALLOW RULE AT ALL, on purpose: with none, Grep in the
+ * working directory (and in the added knowledge directory) ran, and Grep of a
+ * path outside them was refused. The Bash readers these postures keep are
+ * path-validated by the CLI on its own (see CONSULT_BASH).
+ */
+function fencedReads(knowledgeDir) {
+  const kd = ruleSafeAbsolute(knowledgeDir);
+  return [
+    'Read(./**)',
+    ...(kd ? [`Read(/${kd}/**)`] : []),
+    'Glob(./**)',
+    ...(kd ? [`Glob(/${kd}/**)`] : []),
+  ];
+}
+
+/** The shell readers CONSULT and PLAN keep. Each one is PATH-VALIDATED by the
+ *  CLI itself — measured on 2.1.281 (2026-09-24): with `Bash(cat:*)`,
+ *  `Bash(head:*)` and `Bash(ls:*)` allowed, `cat`/`ls`/`head` on a file outside
+ *  the working directories were each refused ("Claude Code may only
+ *  concatenate files from the allowed working directories") — so it is the
+ *  Read/Grep/Glob TOOLS, not these, that needed the fence above. */
+const CONSULT_BASH = [
+  'Bash(ls:*)',
+  'Bash(wc:*)',
+  'Bash(head:*)',
+  'Bash(cat:*)',
+  'Bash(git log:*)',
+  'Bash(git show:*)',
+  'Bash(git diff:*)',
+  'Bash(git rev-parse:*)',
+];
+
 /**
  * PLAN — read the repo, write the plan, never the code.
  *
- * The read half is CONSULT_PERM verbatim: this turn's prompt is steered by
+ * The read half is `consultPermFor`'s verbatim (the read fence and the
+ * path-validated shell readers): this turn's prompt is steered by
  * anything a project editor can type, so the same threat applies and the same
  * allowlist answers it. What is added is the control plane and NOTHING else —
  * `mcp__flowviant` is the plan principal's token, whose entire tool set is the
@@ -77,22 +133,15 @@ const READ_GUARD = ['--settings', READ_GUARD_SETTINGS];
  * through an API, and there is no file on this machine it has any business
  * touching.
  */
-const PLAN_PERM = [
-  ...READ_GUARD,
-  '--allowedTools',
-  'mcp__flowviant',
-  'Read',
-  'Grep',
-  'Glob',
-  'Bash(ls:*)',
-  'Bash(wc:*)',
-  'Bash(head:*)',
-  'Bash(cat:*)',
-  'Bash(git log:*)',
-  'Bash(git show:*)',
-  'Bash(git diff:*)',
-  'Bash(git rev-parse:*)',
-];
+export function planPermFor(knowledgeDir) {
+  return [
+    ...READ_GUARD,
+    '--allowedTools',
+    'mcp__flowviant',
+    ...fencedReads(knowledgeDir),
+    ...CONSULT_BASH,
+  ];
+}
 
 // Unattended (default) skips prompts so the agent never stalls with no terminal;
 // FLOWVIANT_SAFE=1 restricts to a curated toolset instead.
@@ -154,21 +203,9 @@ const WIKI_PERM = [
 // someone else's machine. The permission list is the enforcement; the prompt's
 // "do not change anything" is only an instruction, and instructions are exactly
 // what an injected question competes with.
-const CONSULT_PERM = [
-  ...READ_GUARD,
-  '--allowedTools',
-  'Read',
-  'Grep',
-  'Glob',
-  'Bash(ls:*)',
-  'Bash(wc:*)',
-  'Bash(head:*)',
-  'Bash(cat:*)',
-  'Bash(git log:*)',
-  'Bash(git show:*)',
-  'Bash(git diff:*)',
-  'Bash(git rev-parse:*)',
-];
+export function consultPermFor(knowledgeDir) {
+  return [...READ_GUARD, '--allowedTools', ...fencedReads(knowledgeDir), ...CONSULT_BASH];
+}
 
 /**
  * THE TWO NON-CODE POSTURES (0.97.0) — a design card and a research card may
@@ -195,8 +232,10 @@ const CONSULT_PERM = [
  *
  * THEY ARE NOT THE SAME READER, and the difference is the network:
  *
- *  · DESIGN draws THIS product, so it reads the repo broadly — Read, Grep,
- *    Glob and `ls` — and has no web at all (a mockup that went looking on the
+ *  · DESIGN draws THIS product, so it reads the repo — the worktree and the
+ *    knowledge library through the read fence (`fencedReads`, 2026-09-24:
+ *    it was bare Read/Grep/Glob, which reach the whole box), Grep with no
+ *    allow rule, `ls`, and `.env*` denied by name — and has no web at all (a mockup that went looking on the
  *    web would be a mockup of somebody else's product). No git readers and no
  *    cat/head/wc: Read covers every one of them, and each git reader was an
  *    `--output` away from a write (see READ_GUARD, which it still carries for
@@ -233,9 +272,10 @@ const CONSULT_PERM = [
  * STATED, not closed: research may still read any file in the worktree and
  * name it in a WebFetch URL — the repository is what a research card is about,
  * and a turn that could not read it could not answer. Design may read any
- * file the Read tool reaches and put it in an artifact the server stores;
- * text artifacts are scrubbed of the machine's known secret values on the way
- * out (artifacts.mjs), and binary ones are withheld when they carry one.
+ * file in its worktree and put it in an artifact the server stores; text
+ * artifacts are scrubbed of the machine's known secret values — and of any
+ * Flowviant credential, by shape — on the way out (artifacts.mjs, env.mjs), and
+ * binary ones are withheld when they carry one.
  * Design still runs a plain `git log` on the CLI's own read-only classifier
  * though no git reader is allowed (measured) — the guard is what refuses its
  * `--output`. And the Agent tool is offered with no allow rule on this CLI:
@@ -274,21 +314,22 @@ const ARTIFACT_WRITE = 'Edit(.flowviant/artifacts/**)';
  *    plane whose every call the CLI refuses. See work.mjs.
  */
 export const PLAN_MODE_PERM = ['--permission-mode', 'plan'];
-export const DESIGN_PERM = [
-  ...READ_GUARD,
-  '--allowedTools',
-  'Read',
-  'Grep',
-  'Glob',
-  'Bash(ls:*)',
-  ARTIFACT_WRITE,
-];
-
-/** A path a permission rule can name: absolute POSIX, and free of the glob and
- *  rule-syntax characters that would change what the rule means. Anything else
- *  gets no rule — `--add-dir` still admits the directory (measured). */
-const ruleSafeAbsolute = (p) =>
-  typeof p === 'string' && p.startsWith('/') && !/[*?[\]{}()\\\n]/.test(p) ? p.replace(/\/+$/, '') : null;
+export function designPermFor(knowledgeDir) {
+  return [
+    ...READ_GUARD,
+    '--allowedTools',
+    ...fencedReads(knowledgeDir),
+    'Bash(ls:*)',
+    ARTIFACT_WRITE,
+    // A mockup is uploaded and runs scripts in a frame that may navigate
+    // itself, so the checkout's own secrets are denied by name too — the same
+    // pair research carries, measured load-bearing there.
+    '--disallowedTools',
+    'Read(./.env*)',
+    'Read(./**/.env*)',
+  ];
+}
+export const DESIGN_PERM = designPermFor(null);
 
 /**
  * RESEARCH, built at spawn — the knowledge library is the one directory
@@ -296,14 +337,10 @@ const ruleSafeAbsolute = (p) =>
  * `//` is the CLI's own spelling for an absolute path in a rule.
  */
 export function researchPerm(knowledgeDir) {
-  const kd = ruleSafeAbsolute(knowledgeDir);
   return [
     ...READ_GUARD,
     '--allowedTools',
-    'Read(./**)',
-    ...(kd ? [`Read(/${kd}/**)`] : []),
-    'Glob(./**)',
-    ...(kd ? [`Glob(/${kd}/**)`] : []),
+    ...fencedReads(knowledgeDir),
     'WebSearch',
     'WebFetch',
     ARTIFACT_WRITE,
@@ -322,10 +359,26 @@ export const sleep = (s) => new Promise((r) => setTimeout(r, s * 1000));
 // matching falsely fired when an agent merely *mentioned* the word in prose
 // (e.g. "I won't fabricate a BLOCKED:<id> line"), trapping the worker in a fake
 // blocked loop. Anchor to a full line instead.
-export const sawSentinel = (out, name) => new RegExp(`^\\s*${name}\\s*$`, 'm').test(out);
+//
+// LINE-SPLIT, NEVER A `^...$` REGEX WITH A `\s*` ON BOTH ENDS (audit
+// 2026-09-24). `new RegExp('^\\s*NAME\\s*$', 'm')` and
+// `/^\s*BLOCKED:(\S+)\s*$/m` are QUADRATIC against a long run of
+// whitespace on one line — the same shape `taskIdsFromMessage` already had
+// to fix, measured ~0.5s at 40KB here — and a CLI's own stdout is exactly
+// the kind of text a stray control sequence or a pasted blob can grow past
+// that. A length cap first, then plain string work, same as there.
+const SENTINEL_MAX_LINE = 2000;
+export const sawSentinel = (out, name) =>
+  String(out ?? '')
+    .split('\n')
+    .some((l) => l.length < SENTINEL_MAX_LINE && l.trim() === name);
 export const blockedId = (out) => {
-  const m = out.match(/^\s*BLOCKED:(\S+)\s*$/m);
-  return m ? m[1] : null;
+  for (const l of String(out ?? '').split('\n')) {
+    if (l.length > SENTINEL_MAX_LINE) continue;
+    const t = l.trim();
+    if (t.startsWith('BLOCKED:') && /^BLOCKED:\S+$/.test(t)) return t.slice(8);
+  }
+  return null;
 };
 
 /**
@@ -539,6 +592,27 @@ export function handleStreamLine(line, { cwd, emit, onActivity, onToolEvent, app
   }
 }
 
+/**
+ * THE ENVIRONMENT A CLI TURN IS SPAWNED WITH — the daemon's own, plus the
+ * turn's MCP token, MINUS the machine credential (2026-09-24, the audit).
+ *
+ * A turn cannot be handed a curated environment the way `childEnv` builds one
+ * for a deploy: the CLI's own sign-in lives in this environment, and stripping
+ * it signs the CLI out. But on a headless box started as
+ * `FLOWVIANT_MACHINE_TOKEN=… npx flowviant` the MACHINE CREDENTIAL lives here
+ * too, and every turn — and every dev server and script a turn starts, which
+ * inherit it — could read it with one `env`. Nothing a CLI or its children do
+ * needs it: the daemon authenticates to the server itself, and a turn that
+ * needs the project's tools gets its own per-turn token through `mcpEnv` or
+ * the MCP config file. So those two names, and only those, are removed.
+ */
+const MACHINE_CREDENTIAL_ENV = ['FLOWVIANT_MACHINE_TOKEN', 'FLOWVIANT_FLEET'];
+export function cliEnv(mcpEnv) {
+  const env = { ...process.env, ...(mcpEnv ?? {}) };
+  for (const k of MACHINE_CREDENTIAL_ENV) delete env[k];
+  return env;
+}
+
 // One Claude Code turn. Output is captured (for sentinel detection) and streamed
 // through, line-prefixed with the worker label so a fleet stays legible.
 //
@@ -548,7 +622,7 @@ export function handleStreamLine(line, { cwd, emit, onActivity, onToolEvent, app
 // returned string for sentinel detection, and each activity is handed to
 // `onActivity` so the caller can forward progress. Build-agent turns leave it
 // off and keep the raw text passthrough + line sentinels.
-export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEnv, runtime = 'claude', label, onSpawn, streamJson, answerFromResult, onActivity, onToolEvent, onInit, onUsage, onThreadId, wikiPerm, readOnly, planPerm, planMode, posture, vaultDir, knowledgeDir, resultSchemaArgs, model, effort, adoptResumeId, resumeThreadId, resumeConversationId }) {
+export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEnv, agentTools, runtime = 'claude', label, onSpawn, streamJson, answerFromResult, onActivity, onToolEvent, onInit, onUsage, onThreadId, onAnswer, wikiPerm, readOnly, planPerm, planMode, posture, vaultDir, knowledgeDir, resultSchemaArgs, model, effort, adoptResumeId, resumeThreadId, resumeConversationId }) {
   return new Promise((resolve) => {
     const rt = runtimeById(runtime);
     // PLAN MODE IS CLAUDE'S (0.97.0). The other adapters build their argv
@@ -622,6 +696,7 @@ export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEn
       resume,
       streamJson,
       profile,
+      agentTools,
       // Adopting a terminal session (work.mjs): Claude turns it into
       // `--resume <id> --fork-session` (a FORK — the original is untouched);
       // agy turns it into `--conversation <id>` (a MOVE — agy has no fork, the
@@ -652,15 +727,19 @@ export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEn
       perm: planMode
         ? PLAN_MODE_PERM
         : profile === 'design'
-          ? DESIGN_PERM
+          ? designPermFor(knowledgeDir)
           : profile === 'research'
             ? researchPerm(knowledgeDir)
-            : planPerm ? PLAN_PERM : readOnly ? CONSULT_PERM : wikiPerm ? WIKI_PERM : PERM,
+            : planPerm
+              ? planPermFor(knowledgeDir)
+              : readOnly
+                ? consultPermFor(knowledgeDir)
+                : wikiPerm ? WIKI_PERM : PERM,
       // Handed to the adapter rather than appended here, because WHERE these go
       // is a property of the CLI: Codex reads its prompt as a trailing
       // positional, so a flag after it is a flag in the wrong place.
       // Wiki-vault turns are pure file work and pass neither — no MCP at all.
-      mcp: mcpConfig ? ['--mcp-config', mcpConfig] : (mcpArgs ?? []),
+      mcp: planMode ? ['--strict-mcp-config'] : mcpConfig ? ['--mcp-config', mcpConfig] : (mcpArgs ?? []),
       // Resuming a SPECIFIC held conversation by its own id (work.mjs, codex
       // sessions). Runtimes without a by-id resume ignore it and keep their
       // `resume` behavior unchanged.
@@ -706,8 +785,9 @@ export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEn
       detached: true,
       // Only ADDS to the environment (the worker token, for runtimes that read
       // it from there). Never replaces it: the CLI's own credentials live in
-      // this environment, and handing it a curated one signs it out.
-      ...(mcpEnv ? { env: { ...process.env, ...mcpEnv } } : {}),
+      // this environment, and handing it a curated one signs it out. The one
+      // thing it REMOVES is this daemon's machine credential — see `cliEnv`.
+      env: cliEnv(mcpEnv),
     });
     onSpawn?.(child);
     let out = '';
@@ -734,14 +814,28 @@ export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEn
         // thread.started). Purely additive: callers that pass no onThreadId —
         // every dispatch path — see zero behavior change.
         if (ev.threadId) onThreadId?.(ev.threadId);
+        // The runtime's own token count (codex's `turn.completed`) — the same
+        // `onUsage` Claude's result event feeds, so a caller charges a turn
+        // identically whichever CLI ran it.
+        if (ev.usage) onUsage?.(ev.usage);
+        // One agent MESSAGE, alone — see parseCodexLine's `answer`. Last call
+        // wins at the caller, which is what "the final answer" means.
+        if (typeof ev.answer === 'string') onAnswer?.(ev.answer);
         if (ev.text) appendText(ev.text);
         if (ev.activity) {
           emit(`${ev.activity.label}\n`);
           onActivity?.(ev.activity);
         }
       };
+      // DECODED AS A STREAM, never chunk by chunk (2026-09-24, the audit): a
+      // pipe read can end inside a multi-byte character — the em dash models
+      // write constantly — and `d.toString()` per chunk turned each half into
+      // U+FFFD, which JSON.parse accepts and the reply then carried verbatim.
+      // `setEncoding` holds the partial bytes over to the next chunk.
+      child.stdout.setEncoding('utf8');
+      child.stderr.setEncoding('utf8');
       child.stdout.on('data', (d) => {
-        buf += d.toString();
+        buf += d;
         let nl;
         while ((nl = buf.indexOf('\n')) >= 0) {
           const line = buf.slice(0, nl);
@@ -750,8 +844,7 @@ export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEn
         }
       });
       // stderr is not JSON (warnings/errors) — pass through and keep for sentinels.
-      child.stderr.on('data', (d) => {
-        const s = d.toString();
+      child.stderr.on('data', (s) => {
         out += s;
         emit(s);
       });
@@ -780,11 +873,13 @@ export function runTurn({ prompt, resume, system, cwd, mcpConfig, mcpArgs, mcpEn
       return;
     }
 
-    const onChunk = (d) => {
-      const s = d.toString();
+    const onChunk = (s) => {
       out += s;
       emit(s);
     };
+    // Stream-decoded for the same reason as the line-parsed path above.
+    child.stdout.setEncoding('utf8');
+    child.stderr.setEncoding('utf8');
     child.stdout.on('data', onChunk);
     child.stderr.on('data', onChunk);
     child.on('error', (e) => {

package/bin/lib/config.mjs

@@ -327,7 +327,30 @@ export let FLEET_TOKEN =
  * A frozen copy read at import would salt a picked project's report with null
  * and report nothing at all, for the whole life of the process.
  */
-export let PROJECT_ID = CREDENTIAL.entry?.projectId ?? null;
+export let PROJECT_ID =
+  CREDENTIAL.entry?.fleetToken && CREDENTIAL.entry.fleetToken === FLEET_TOKEN
+    ? (CREDENTIAL.entry.projectId ?? null)
+    : null;
+
+/**
+ * THE STORE'S PROJECT IS THIS DAEMON'S ONLY WHEN IT IS THIS DAEMON'S TOKEN.
+ * `FLEET_TOKEN` prefers `--fleet` and the environment over the store, and the
+ * id above used to come from the store unconditionally — so a box holding one
+ * unbound project A, started with project B's token in the environment, salted
+ * B's env report with A's id (every variable then read `differs` beside B's
+ * other boxes, the confident wrong answer the salt exists to prevent) and
+ * printed "serves · A". It starts null in that case, and the ROSTER's own word
+ * settles it here on the first poll — the "names no project until the roster
+ * does" the paragraph above always promised and nothing ever did.
+ */
+export function learnProjectId(id) {
+  if (typeof id === 'string' && id) PROJECT_ID = id;
+}
+
+/** Whether the stored resolution describes the token actually in use. */
+export function storedCredentialInUse() {
+  return Boolean(CREDENTIAL.entry?.fleetToken) && CREDENTIAL.entry.fleetToken === FLEET_TOKEN;
+}
 
 /** cli.mjs's picker chose. Must run BEFORE runFleetDaemon — nothing here
  *  re-authenticates a connection already made. */

package/bin/lib/credentials.mjs

@@ -48,6 +48,7 @@ import { execFileSync } from 'node:child_process';
 import { mkdirSync, readFileSync, realpathSync, renameSync, writeFileSync } from 'node:fs';
 import { homedir } from 'node:os';
 import { join } from 'node:path';
+import { printable } from './printable.mjs';
 
 const CRED_DIR = join(homedir(), '.flowviant');
 const CRED_FILE = join(CRED_DIR, 'credentials.json');
@@ -133,7 +134,28 @@ export function listStoredProjects() {
 
 /** What a project is CALLED on a terminal: its name, or an id you can grep. */
 export function projectLabel(e) {
-  return e?.name ?? (e?.projectId ? `project ${e.projectId.slice(0, 8)}…` : 'an unnamed project');
+  return safeName(e?.name) ?? (e?.projectId ? `project ${e.projectId.slice(0, 8)}…` : 'an unnamed project');
+}
+
+/**
+ * A SERVER-SUPPLIED NAME, SAFE TO PRINT ON A TERMINAL — or null.
+ *
+ * A project's name is whatever an owner typed in the app, and the server's
+ * schema caps its length and nothing else. Printed raw it reaches every
+ * terminal that connects a machine to the project — `login`, the banner, the
+ * picker, `projects`, `machines` — and a name carrying an escape sequence can
+ * rewrite the clipboard (OSC 52 in kitty, WezTerm, Windows Terminal) or hide
+ * lines of the very listing whose job is to say which project is which. The
+ * scrub is `printable()` (C0, DEL, C1 — the server's own `parseBoxName` makes
+ * the same cut for hostnames — plus the bidi/format controls that reorder a
+ * line, added 2026-09-24 rather than re-derived: see printable.mjs), and what
+ * is left is trimmed; nothing left is null, so the caller's id fallback
+ * speaks instead of an empty label.
+ */
+export function safeName(name) {
+  if (typeof name !== 'string') return null;
+  const clean = printable(name).trim();
+  return clean || null;
 }
 
 /**
@@ -316,7 +338,7 @@ export function saveLogin({ fleetToken, projectId, mcpUrl, name, repoRoot }) {
       ...(f.projects[projectId] ?? {}),
       fleetToken,
       mcpUrl: mcpUrl ?? null,
-      ...(name ? { name } : {}),
+      ...(safeName(name) ? { name: safeName(name) } : {}),
       ...(repoRoot ? { repoRoot } : {}),
       savedAt: new Date().toISOString(),
     };
@@ -391,8 +413,9 @@ export function bindStoredRepo(projectId, repoRoot) {
 
 /** The roster names the project on every poll; remember it so the picker and
  *  `flowviant projects` can say a NAME instead of an id. */
-export function setStoredProjectName(projectId, name) {
-  if (!projectId || typeof name !== 'string' || !name) return;
+export function setStoredProjectName(projectId, raw) {
+  const name = safeName(raw);
+  if (!projectId || !name) return;
   const entries = listStoredProjects();
   const e = entries.find((x) => x.projectId === projectId);
   if (!e || e.name === name) return;