Merge pull request 'fix(mcplocal): stream chat SSE through the proxy instead of buffering it' (#109) from fix/chat-sse-streaming into main

`mcpctl chat reviewer` showed nothing until the turn finished, then dumped
the whole answer at once. mcpd streams token deltas and the CLI renders
them incrementally — the sole buffering point was mcplocal's catch-all
/api/v1/* proxy reading the whole SSE body via res.text() before replying.
The previous commit (cherry-picked from feat/agentic-teams) pipes the body
through instead; this one adds the cover that was missing:

- proxy-long-running.test.ts: a progressive-delivery test in which the
  stand-in mcpd withholds its final frame until the client has observed the
  first delta through the proxy. A buffering proxy cannot satisfy that
  ordering — verified: the test fails in 3s (no hang) against a res.text()
  proxy and passes against the piped one. The existing SSE test used
  inject(), which collects the whole body and so passes either way.

- agent-chat.smoke.test.ts: a live smoke that posts through mcplocal on
  localhost:3200 (the path `mcpctl chat` actually takes — every other chat
  smoke uses --direct and bypasses the proxy entirely) and asserts delta
  frames arrive spread across the generation window, not in one burst at
  stream end. Uses its own agent: the shared smoke agent pins replies to a
  single token, too short to tell live streaming from a buffer dump.

Also settles the strict-boolean-expressions lint on the auth-header check
the streaming split touched.

Local: workspace 2546 passed, proxy suite 10/10, smoke file loads + self-skips.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dNpnBqyyz9GxfznVcX2sP

docs/claude-integration.md

@@ -101,13 +101,48 @@ to open, and re-scoping it would silently change which skills sync into it.
   ⏵⏵ bypass permissions on · ← for agents
 ```
 
-`mcpctl statusline` resolves the project from a directory-scoped `.mcp.json`
-first (a repo that pinned itself wins), then the user-scope entry in
-`.claude.json`, then a `.mcpctl-project` marker up the tree so a checkout that is scoped but not yet
-wired still reports. It reads the directory from the JSON Claude Code pipes in,
-so it follows `/cwd` rather than reporting wherever the binary was launched, and
-prints **nothing** when no project is active — an empty status line beats one
-saying "none" on every unrelated repo.
+`mcpctl statusline` reads the directory from the JSON Claude Code pipes in, so it
+follows `/cwd` rather than reporting wherever the binary was launched, and prints
+**nothing** when no project is active — an empty status line beats one saying
+"none" on every unrelated repo.
+
+It then takes the project from the most deliberate source that names one:
+
+1. a canonical `mcpctl` entry in that directory's `.mcp.json` — a repo that
+   pinned itself wins, and it is the scope Claude Code itself prefers when both
+   define that server name;
+2. the user-scope entry in `.claude.json` — what `config claude --project`
+   writes, so a switch takes effect everywhere it is not overridden;
+3. a **legacy** project-named entry in `.mcp.json` (`homeautomation`,
+   `docmost`, …), left by an mcpctl older than the constant server name;
+4. a `.mcpctl-project` marker up the tree, so a checkout that is scoped but not
+   yet wired still reports.
+
+> **Legacy entries rank below user scope on purpose.** They used to outrank it,
+> which made switching look broken: a user-scope switch never rewrites a
+> checkout's `.mcp.json`, so the leftover kept naming the old project for good.
+> A pin is a decision; residue is not.
+
+A server Claude Code has switched off for that directory (`disabledMcpServers` /
+`disabledMcpjsonServers`) is skipped at every step — a disabled server is not
+mounted, so naming its project would be a lie. A `.mcp.json` server that is in
+neither list is still awaiting its approval prompt and does count, since blanking
+the status line on a fresh checkout is the more confusing failure.
+
+### When a directory contradicts a switch
+
+Claude Code merges the two scopes rather than picking one, so switching in user
+scope cannot clean up what a directory declares. `config claude` says so rather
+than reporting plain success:
+
+```
+Warning: /path/to/repo/.mcp.json still registers 'homeautomation' for this
+directory — mounted alongside 'sre', not replaced by it.
+  Re-run with --scope project to retire it, or delete the entry by hand.
+```
+
+A canonical entry pinned to another project gets the stronger wording — it
+*overrides* the switch in that directory rather than sitting beside it.
 
 ### It is never installed over yours
 

docs/pi-extension.md

@@ -101,9 +101,34 @@ src/pi-ext/
   mcp-http.ts           # vendored Streamable-HTTP JSON-RPC client (no deps)
 ```
 
-The extension imports only from pi-bundled packages
-(`@earendil-works/pi-coding-agent`, `@earendil-works/pi-ai`, `typebox`), so it
-loads standalone.
+The extension imports only from pi-bundled packages, so it loads standalone.
+
+### `typebox` is the only bare runtime import
+
+pi does not let an extension resolve modules the ordinary way: it hands jiti a
+hard-coded alias table built from its *own* dependencies, and that table is not
+the same across pi distributions. The newer `@earendil-works/pi-coding-agent`
+aliases both the `@earendil-works/*` and the legacy `@mariozechner/*` names;
+older `@mariozechner/pi-coding-agent` installs (0.73.x and earlier) alias only
+the `@mariozechner/*` ones. Neither resolves the other's namespace.
+
+So an import of anything outside the intersection kills the *whole* extension on
+someone else's pi — every tool, the `/mcpctl` command, the status line — with:
+
+```
+Failed to load extension ".../mcpctl-pi.ts": Cannot find module '@earendil-works/pi-ai'
+```
+
+which is exactly what `import { StringEnum } from "@earendil-works/pi-ai"` did.
+`typebox` is aliased by every published pi, so it is the only bare specifier
+allowed at runtime. Everything else must be a `node:` builtin, a relative path,
+an `import type` (erased before jiti resolves anything), or inlined — pi-ai's
+`StringEnum` is now a six-line local `stringEnum`. The
+`tests/config/pi-extension-embed.test.ts` guard fails the build on a reintroduced
+runtime import.
+
+If a user does hit this error, check `type -a pi`: two installs on `$PATH` is the
+usual cause, and the extension has to load under whichever one wins.
 
 ## Typechecking
 

docs/project-summary.md

@@ -884,6 +884,39 @@ All pushed to `mysources.co.uk/michal/` registry.
 source .env && bash scripts/release.sh
 ```
 
+**The build refuses to run from a branch that is behind `main`.** Everyone
+branches off main, so a stale branch still builds and installs cleanly — it just
+ships a binary missing whatever landed on main meanwhile, and `rpm -U --force`
+overwrites the good one with it. That happened on 2026-08-10: a build from a
+stale checkout replaced `/usr/bin/mcpctl` with one that had no `statusline`
+command, months after the status line landed. `scripts/check-main-sync.sh`
+fetches `main`, compares, and fails before any work happens, listing the commits
+you are missing. It gates every path that produces something others consume:
+`build-rpm.sh`, `build-deb.sh`, `build-mcpd.sh` (each is also run standalone, so
+none can rely on another having checked) and `deploy-k8s.sh` — where a stale
+branch would pin its sha in Pulumi and make it the cluster's source of truth.
+`deploy-k8s.sh --dry-run` skips the check: it builds and cuts over nothing, and
+blocking a read-only inspection only teaches people to export the escape hatch
+permanently, disabling the gate for real deploys too.
+
+```bash
+git merge main                                   # the fix
+MCPCTL_ALLOW_BEHIND_MAIN=1 bash scripts/release.sh  # deliberate old-tree build
+MCPCTL_BASE_BRANCH=release-2.x bash scripts/build-rpm.sh  # compare to another branch
+```
+
+Offline it falls back to the last fetched `origin/main`, then to a local `main`,
+and says which it used; outside a git checkout it skips entirely.
+
+**A failing smoke run fails the release.** It used to print
+`WARNING: Smoke tests failed!` and exit 0 — which is exactly how four broken
+readiness probes shipped unnoticed (see `docs/reliability.md`): the warning
+scrolled past and the release reported success. Note what the gate does and does
+not do — smoke runs *last*, against the installed binary, so the package is
+already published and installed by the time it fails. It reports the breakage
+rather than preventing it, so investigate the fleet rather than assuming the
+artifact is bad. Override with `MCPCTL_ALLOW_SMOKE_FAILURE=1`.
+
 Installs via nfpm:
 - `/usr/bin/mcpctl` — CLI binary (bun compiled)
 - `/usr/bin/mcpctl-local` — Local proxy binary (bun compiled)

scripts/build-deb.sh

@@ -19,6 +19,11 @@ source "$SCRIPT_DIR/arch-helper.sh"
 resolve_arch "${MCPCTL_TARGET_ARCH:-}"
 # Sets: NFPM_ARCH, BUN_TARGET, ARCH_SUFFIX
 
+# Same guard as build-rpm.sh: this script is also run on its own, so it cannot
+# rely on that one having checked.
+source "$SCRIPT_DIR/check-main-sync.sh"
+check_main_sync
+
 # Check and install missing build dependencies
 source "$SCRIPT_DIR/ensure-deps.sh"
 ensure_build_deps

scripts/build-mcpd.sh

@@ -16,6 +16,11 @@ if [ -f .env ]; then
   set -a; source .env; set +a
 fi
 
+# This pushes an image to the registry, so the same staleness gate as the package
+# builds applies. Run standalone as well as from deploy-k8s.sh, hence its own copy.
+source "$SCRIPT_DIR/check-main-sync.sh"
+check_main_sync
+
 # Push directly to internal address (external proxy has body size limit)
 REGISTRY="10.0.0.194:3012"
 IMAGE="mcpd"

scripts/build-rpm.sh

@@ -19,6 +19,11 @@ source "$SCRIPT_DIR/arch-helper.sh"
 resolve_arch "${MCPCTL_TARGET_ARCH:-}"
 # Sets: NFPM_ARCH, BUN_TARGET, ARCH_SUFFIX
 
+# Before anything expensive: a branch behind main packages a binary missing
+# whatever landed there, and installing it silently downgrades the machine.
+source "$SCRIPT_DIR/check-main-sync.sh"
+check_main_sync
+
 # Check and install missing build dependencies
 source "$SCRIPT_DIR/ensure-deps.sh"
 ensure_build_deps

scripts/check-main-sync.sh

@@ -0,0 +1,101 @@
+#!/bin/bash
+# Refuse to build a package from a branch that main has already moved past.
+#
+# WHY
+#
+# Everyone branches off main and builds from their own branch. A branch that is
+# behind main still builds and installs perfectly — it just quietly ships a
+# binary missing whatever landed on main in the meantime, and `rpm -U --force`
+# overwrites the good one with it.
+#
+# That is not hypothetical: on 2026-08-10 a build from a stale checkout replaced
+# /usr/bin/mcpctl with one that had no `statusline` command at all, months after
+# the status line landed on main. Nothing reported an error — the release
+# succeeded, the feature just vanished from the installed CLI.
+#
+# So this is a hard failure rather than a warning. A warning scrolls past in a
+# build log; the whole point is to stop before the artifact exists.
+#
+# ESCAPE HATCH
+#
+#   MCPCTL_ALLOW_BEHIND_MAIN=1   build anyway (deliberate build of an old tree)
+#   MCPCTL_BASE_BRANCH=<name>    compare against something other than main
+#
+# Skips itself entirely outside a git checkout, so tarball builds still work.
+
+# Resolve the ref to compare against, echoing it on stdout. Prefers a fresh
+# fetch; falls back to whatever is already on disk so an offline build is
+# degraded rather than blocked. Returns 1 when there is nothing to compare to.
+_main_sync_ref() {
+  local base="$1" remote="$2"
+
+  if [ -n "$remote" ] && git fetch --quiet "$remote" "$base" 2>/dev/null; then
+    # FETCH_HEAD rather than refs/remotes/<remote>/<base>: it is what this fetch
+    # just wrote, so it cannot be a stale opportunistic update.
+    echo "FETCH_HEAD"
+    return 0
+  fi
+
+  if [ -n "$remote" ] && git rev-parse --verify --quiet "refs/remotes/$remote/$base" >/dev/null; then
+    echo "  (could not reach $remote — comparing against the last fetched $remote/$base)" >&2
+    echo "refs/remotes/$remote/$base"
+    return 0
+  fi
+
+  if git rev-parse --verify --quiet "refs/heads/$base" >/dev/null; then
+    echo "  (no reachable remote — comparing against local $base)" >&2
+    echo "refs/heads/$base"
+    return 0
+  fi
+
+  return 1
+}
+
+check_main_sync() {
+  local base="${MCPCTL_BASE_BRANCH:-main}"
+
+  if ! git rev-parse --git-dir >/dev/null 2>&1; then
+    return 0  # not a checkout; nothing to be behind
+  fi
+
+  if [ "${MCPCTL_ALLOW_BEHIND_MAIN:-}" = "1" ]; then
+    echo "==> Skipping the '$base' sync check (MCPCTL_ALLOW_BEHIND_MAIN=1)"
+    return 0
+  fi
+
+  echo "==> Checking this branch is not behind '$base'..."
+
+  local remote ref
+  remote="$(git remote | head -1)"
+  if ! ref="$(_main_sync_ref "$base" "$remote")"; then
+    echo "  (no '$base' branch found anywhere — skipping)"
+    return 0
+  fi
+
+  local behind
+  behind="$(git rev-list --count "HEAD..$ref" 2>/dev/null || echo 0)"
+  if [ "$behind" -eq 0 ]; then
+    echo "  up to date with $base"
+    return 0
+  fi
+
+  local branch
+  branch="$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo HEAD)"
+  echo "" >&2
+  echo "ERROR: '$branch' is $behind commit(s) behind $base — refusing to build." >&2
+  echo "" >&2
+  # Deliberately artifact-agnostic: the same helper gates RPM/DEB packages, the
+  # mcpd image, and the k8s deploy.
+  echo "  Building now would produce an artifact without these, and shipping it" >&2
+  echo "  would replace a good one with a version missing them:" >&2
+  echo "" >&2
+  git log --oneline --no-decorate "HEAD..$ref" | head -15 | sed 's/^/    /' >&2
+  if [ "$behind" -gt 15 ]; then
+    echo "    … and $((behind - 15)) more" >&2
+  fi
+  echo "" >&2
+  echo "  Fix it:  git merge $base        # or: git rebase $base" >&2
+  echo "  Anyway:  MCPCTL_ALLOW_BEHIND_MAIN=1 $0" >&2
+  echo "" >&2
+  return 1
+}

scripts/deploy-k8s.sh

@@ -80,6 +80,20 @@ cat <<EOF
 EOF
 [ -f "$PULUMI_YAML" ] || die "Pulumi config not found: $PULUMI_YAML"
 
+# ── 0. Staleness gate ──
+# Same hazard as the RPM build, with the cluster on the receiving end: a branch
+# behind main deploys images missing whatever landed there, and the sha pinned in
+# Pulumi makes that the new source of truth. Skipped for --dry-run, which builds
+# and cuts over nothing — blocking a read-only inspection only teaches people to
+# export MCPCTL_ALLOW_BEHIND_MAIN=1 permanently, which would disable the gate for
+# the real deploys too.
+if [ "$DRY_RUN" = true ]; then
+  warn "dry-run: skip the main-sync check"
+else
+  source "$SCRIPT_DIR/check-main-sync.sh"
+  check_main_sync || die "branch is behind main — merge it before deploying"
+fi
+
 # ── 1. Test gate ──
 if [ "$SKIP_TESTS" = true ]; then warn "skipping unit tests (--skip-tests)"; else
   say "1/7  Unit tests (pnpm test:run)"

scripts/release.sh

@@ -75,9 +75,28 @@ echo "==> Running smoke tests..."
 export PATH="$HOME/.npm-global/bin:$PATH"
 if pnpm test:smoke; then
   echo "==> Smoke tests passed!"
+elif [ "${MCPCTL_ALLOW_SMOKE_FAILURE:-}" = "1" ]; then
+  echo "==> WARNING: Smoke tests failed, continuing (MCPCTL_ALLOW_SMOKE_FAILURE=1)."
 else
-  echo "==> WARNING: Smoke tests failed! Check mcplocal/mcpd are running."
-  echo "    Continuing anyway — deployment is complete, but verify manually."
+  # This used to print a warning and exit 0. That is how four broken readiness
+  # probes shipped unnoticed on 2026-08-10: the warning scrolled past in the
+  # build log and the release reported success. A failing smoke run means
+  # something in the live fleet is genuinely broken — say so in the exit code.
+  #
+  # Note what this does and does not do: smoke runs LAST, against the installed
+  # binary, so the package is already published and installed by now. Failing
+  # here reports the breakage, it does not prevent it — investigate, do not
+  # assume the artifact is bad.
+  echo "" >&2
+  echo "ERROR: smoke tests failed — the release is published and installed, but" >&2
+  echo "       something in the live fleet is broken. Investigate before relying" >&2
+  echo "       on this build; do not just re-run." >&2
+  echo "" >&2
+  echo "  Common causes: mcplocal/mcpd not running, a readiness probe pointing at" >&2
+  echo "  a tool the upstream renamed, or an expired credential." >&2
+  echo "  Override:      MCPCTL_ALLOW_SMOKE_FAILURE=1 $0" >&2
+  echo "" >&2
+  exit 1
 fi
 echo ""
 

src/cli/src/commands/config.ts

@@ -46,6 +46,8 @@ import {
   mergeUserScopeServer,
   userScopeProject,
   activeProjectIn,
+  canonicalProjectIn,
+  legacyEntriesIn,
   claudeJsonPath,
   type McpJson,
   type ClaudeJson,
@@ -104,11 +106,51 @@ function readMcpJson(path: string): McpJson | null {
   }
 }
 
+/**
+ * Warnings about a `.mcp.json` in `dir` that contradicts a user-scope switch to
+ * `project`.
+ *
+ * Claude Code merges the two scopes rather than picking one, so a
+ * directory-scoped entry does not go away when you switch globally:
+ *   - a canonical `mcpctl` entry shares the name, and project scope wins — the
+ *     switch has no effect in this directory at all;
+ *   - a legacy project-named entry has a *different* name, so it is simply
+ *     mounted alongside and the old project keeps answering here.
+ * Either way the user is owed the file path, because nothing else will tell
+ * them. Exported for tests.
+ */
+export function shadowWarnings(dir: string, project: string | undefined): string[] {
+  if (project === undefined || project === '') return [];
+  const path = join(dir, '.mcp.json');
+  const parsed = readMcpJson(path);
+  if (parsed === null) return [];
+
+  const pinned = canonicalProjectIn(parsed);
+  if (pinned !== null && pinned !== project) {
+    return [
+      `Warning: ${path} pins '${MCPCTL_SERVER_NAME}' to '${pinned}' for this directory, which overrides the switch here.`,
+      `  Re-run with --scope project to repoint it, or delete the '${MCPCTL_SERVER_NAME}' entry to follow the user-scope project.`,
+    ];
+  }
+
+  const stale = legacyEntriesIn(parsed).filter((e) => e.project !== project);
+  if (stale.length > 0) {
+    const names = stale.map((e) => `'${e.server}'`).join(', ');
+    return [
+      `Warning: ${path} still registers ${names} for this directory — mounted alongside '${project}', not replaced by it.`,
+      `  Re-run with --scope project to retire ${stale.length === 1 ? 'it' : 'them'}, or delete the ${stale.length === 1 ? 'entry' : 'entries'} by hand.`,
+    ];
+  }
+  return [];
+}
+
 export interface ConfigCommandDeps {
   configDeps: Partial<ConfigLoaderDeps>;
   log: (...args: string[]) => void;
   /** API client for the skills sync side-effect of `config claude --project`. Optional so existing call sites work; without it we skip the sync step. */
   apiClient?: ApiClient;
+  /** Working directory to check for a shadowing `.mcp.json`. Injectable so tests need not chdir. */
+  cwd?: () => string;
 }
 
 export interface ConfigApiDeps {
@@ -124,6 +166,7 @@ const defaultDeps: ConfigCommandDeps = {
 
 export function createConfigCommand(deps?: Partial<ConfigCommandDeps>, apiDeps?: ConfigApiDeps): Command {
   const { configDeps, log } = { ...defaultDeps, ...deps };
+  const cwd = deps?.cwd ?? ((): string => process.cwd());
   // PR-5: api client used by `mcpctl config claude --project` to run the
   // initial skills sync after wiring the .mcp.json. Threaded through from
   // index.ts; falls back to apiDeps.client when not explicitly passed (the
@@ -406,6 +449,11 @@ export function createConfigCommand(deps?: Partial<ConfigCommandDeps>, apiDeps?:
         if (userScope) {
           // The whole point of user scope: you do this once, not per checkout.
           log('This applies in every directory — no need to re-run it per repo.');
+          // ...except where a directory-scoped entry contradicts it. That file
+          // is never rewritten by a user-scope switch, so staying silent is how
+          // a switch ends up looking like it did nothing: the status line keeps
+          // naming the old project, and its server keeps answering here.
+          for (const line of shadowWarnings(cwd(), opts.project)) log(line);
         }
 
         // PR-5: write project marker, run initial skills sync, install

src/cli/src/commands/statusline.ts

@@ -2,7 +2,16 @@ import { Command } from 'commander';
 import { readFileSync } from 'node:fs';
 import { join, resolve } from 'node:path';
 import { homedir } from 'node:os';
-import { activeProjectIn, claudeJsonPath, userScopeProject, type McpJson, type ClaudeJson } from '../config/claude-mcp.js';
+import {
+  MCPCTL_SERVER_NAME,
+  canonicalProjectIn,
+  claudeJsonPath,
+  disabledServersFor,
+  legacyEntriesIn,
+  userScopeProject,
+  type McpJson,
+  type ClaudeJson,
+} from '../config/claude-mcp.js';
 import { findProjectMarker } from '../utils/project-marker.js';
 
 /**
@@ -16,9 +25,10 @@ import { findProjectMarker } from '../utils/project-marker.js';
  * `setStatus`.
  *
  * Claude Code pipes a JSON blob in on stdin (session id, model, workspace). We
- * only need the directory — the project is whatever `.mcp.json` there mounts,
- * falling back to a `.mcpctl-project` marker up the tree so a checkout that is
- * scoped but not yet wired still reports.
+ * only need the directory; the project is then resolved from the most
+ * deliberate source that names one — see the ranking in the action below.
+ * Whatever it reports has to be a project that is genuinely mounted, so a
+ * server Claude Code has switched off for that directory is skipped.
  *
  * Prints nothing at all when no project is active: an empty status line is
  * better than one that says "none" on every unrelated repo you open.
@@ -53,25 +63,47 @@ export function resolveDirectory(input: StatusLineInput, fallback: string): stri
   return input.workspace?.current_dir ?? input.workspace?.project_dir ?? input.cwd ?? fallback;
 }
 
-/** The project Claude Code's user-scope config mounts, or null. */
-export function projectFromUserScope(path: string): string | null {
+/** Claude Code's user-scope config, or null if it is missing or unreadable. */
+export function readClaudeJson(path: string): ClaudeJson | null {
   try {
-    return userScopeProject(JSON.parse(readFileSync(path, 'utf-8')) as ClaudeJson);
+    return JSON.parse(readFileSync(path, 'utf-8')) as ClaudeJson;
   } catch {
     return null;
   }
 }
 
-/** The project `.mcp.json` in `dir` mounts, or null. */
-export function projectFromMcpJson(dir: string): string | null {
+/** The project Claude Code's user-scope config mounts, or null. */
+export function projectFromUserScope(doc: ClaudeJson | null): string | null {
+  return userScopeProject(doc);
+}
+
+/** The `.mcp.json` in `dir`, or null if there isn't a readable one. */
+export function readDirMcpJson(dir: string): McpJson | null {
   try {
-    const parsed = JSON.parse(readFileSync(join(dir, '.mcp.json'), 'utf-8')) as McpJson;
-    return activeProjectIn(parsed);
+    return JSON.parse(readFileSync(join(dir, '.mcp.json'), 'utf-8')) as McpJson;
   } catch {
     return null;
   }
 }
 
+/**
+ * The project the canonical `mcpctl` entry in `dir`'s `.mcp.json` pins, or null
+ * — skipped when Claude Code has that server switched off for `dir`.
+ */
+export function projectFromDirPin(mcpJson: McpJson | null, disabled: Set<string>): string | null {
+  if (disabled.has(MCPCTL_SERVER_NAME)) return null;
+  return canonicalProjectIn(mcpJson);
+}
+
+/**
+ * The project a *legacy* project-named entry in `dir`'s `.mcp.json` mounts, or
+ * null. Disabled entries are skipped, so a leftover the user already turned off
+ * in `/mcp` stops being reported.
+ */
+export function projectFromDirLegacy(mcpJson: McpJson | null, disabled: Set<string>): string | null {
+  return legacyEntriesIn(mcpJson).find((e) => !disabled.has(e.server))?.project ?? null;
+}
+
 /** Format for the status line. Empty string means "render nothing". */
 export function formatStatus(project: string | null, prefix: string): string {
   return project !== null && project !== '' ? `${prefix}${project}` : '';
@@ -96,12 +128,29 @@ export function createStatuslineCommand(deps?: Partial<StatuslineDeps>): Command
       const input = opts.directory !== undefined ? {} : await readStdinJson();
       const dir = opts.directory !== undefined ? resolve(opts.directory) : resolveDirectory(input, cwd());
 
-      // Directory-scoped wiring wins: a repo with its own .mcp.json entry has
-      // deliberately pinned itself, and that beats the global default.
-      let project = projectFromMcpJson(dir) ?? projectFromUserScope(claudeJsonPath());
+      const claudeJson = readClaudeJson(claudeJsonPath());
+      const mcpJson = readDirMcpJson(dir);
+      const disabled = disabledServersFor(claudeJson, dir);
+
+      // Ranked by how deliberate each source is, because a switch has to be
+      // able to win:
+      //   1. a canonical `mcpctl` entry in this directory's .mcp.json — a
+      //      deliberate pin, and the scope Claude Code itself prefers when both
+      //      define the same server name;
+      //   2. user scope — what `config claude --project` writes, so switching
+      //      projects must beat anything less deliberate than a pin;
+      //   3. a *legacy* project-named entry in .mcp.json. This used to outrank
+      //      user scope, which made a switch look like it had done nothing: the
+      //      residue an older mcpctl left in a checkout is not a pin, and never
+      //      gets rewritten by a user-scope switch, so it reported the old
+      //      project forever;
+      //   4. the .mcpctl-project marker — the other thing `config claude`
+      //      writes, and what skills sync already trusts.
+      let project =
+        projectFromDirPin(mcpJson, disabled)
+        ?? projectFromUserScope(claudeJson)
+        ?? projectFromDirLegacy(mcpJson, disabled);
       if (project === null) {
-        // Not wired here (or wired above this directory) — the marker is the
-        // other thing `config claude` writes, and skills sync already trusts it.
         const marker = await findProjectMarker(dir, homeDir()).catch(() => null);
         project = marker?.project ?? null;
       }

src/cli/src/config/claude-mcp.ts

@@ -100,16 +100,30 @@ export function isLegacyMcpctlEntry(name: string, entry: unknown): boolean {
   return projectOfEntry(entry) === name;
 }
 
+/** The project the canonical `mcpctl` entry mounts, or null if there isn't one. */
+export function canonicalProjectIn(config: Pick<McpJson, 'mcpServers'> | null | undefined): string | null {
+  return projectOfEntry(config?.mcpServers?.[MCPCTL_SERVER_NAME]);
+}
+
+/**
+ * Legacy project-named entries still present, in file order.
+ *
+ * Kept separate from the canonical entry because the two mean different things
+ * to a reader: the canonical entry is a deliberate pin, a legacy entry is
+ * residue from an older mcpctl that nothing has cleaned up yet. Callers that
+ * rank sources (the status line) must be able to tell them apart.
+ */
+export function legacyEntriesIn(config: Pick<McpJson, 'mcpServers'> | null | undefined): { server: string; project: string }[] {
+  const servers = config?.mcpServers;
+  if (!servers) return [];
+  return Object.entries(servers)
+    .filter(([name, entry]) => isLegacyMcpctlEntry(name, entry))
+    .map(([name]) => ({ server: name, project: name }));
+}
+
 /** The project currently mounted by `.mcp.json`, preferring the canonical entry. */
 export function activeProjectIn(config: Pick<McpJson, 'mcpServers'> | null | undefined): string | null {
-  const servers = config?.mcpServers;
-  if (!servers) return null;
-  const canonical = projectOfEntry(servers[MCPCTL_SERVER_NAME]);
-  if (canonical !== null) return canonical;
-  for (const [name, entry] of Object.entries(servers)) {
-    if (isLegacyMcpctlEntry(name, entry)) return name;
-  }
-  return null;
+  return canonicalProjectIn(config) ?? legacyEntriesIn(config)[0]?.project ?? null;
 }
 
 export interface MergeResult {
@@ -169,12 +183,39 @@ export function claudeJsonPath(env: NodeJS.ProcessEnv = process.env, homeDir?: s
     : join(home, '.claude.json');
 }
 
+/** Per-directory state Claude Code keeps in `.claude.json`'s `projects` map. */
+export interface ClaudeProjectEntry {
+  /** Servers switched off for this directory, whatever scope they came from. */
+  disabledMcpServers?: string[];
+  /** `.mcp.json` servers declined at the approval prompt. */
+  disabledMcpjsonServers?: string[];
+  [key: string]: unknown;
+}
+
 /** Shape of the bits of `.claude.json` we touch. Everything else is preserved. */
 export interface ClaudeJson {
   mcpServers?: Record<string, McpServerEntry>;
+  projects?: Record<string, ClaudeProjectEntry>;
   [key: string]: unknown;
 }
 
+/**
+ * Server names Claude Code has switched off in `dir`.
+ *
+ * A disabled server is not mounted, so naming its project as "active" is a
+ * plain lie — this is what lets the status line skip one. Only *explicit*
+ * disables count: a `.mcp.json` server in neither list is pending its approval
+ * prompt, and treating pending as off would blank the status line on a fresh
+ * checkout, which is the more confusing failure.
+ */
+export function disabledServersFor(doc: ClaudeJson | null | undefined, dir: string): Set<string> {
+  const entry = doc?.projects?.[dir];
+  return new Set([
+    ...(Array.isArray(entry?.disabledMcpServers) ? entry.disabledMcpServers : []),
+    ...(Array.isArray(entry?.disabledMcpjsonServers) ? entry.disabledMcpjsonServers : []),
+  ]);
+}
+
 /**
  * Set the user-scope entry, returning the new document and any legacy
  * project-named entries retired from it.

src/cli/tests/commands/claude.test.ts

@@ -359,4 +359,46 @@ describe('config claude — user scope', () => {
     process.exitCode = prevExit;
     expect(output.join('\n')).toContain("unknown --scope 'global'");
   });
+
+  // A user-scope switch never rewrites a directory's .mcp.json, so anything of
+  // ours left in one keeps answering in that directory. Saying so is the only
+  // way the user finds out — the switch otherwise reports plain success.
+  describe('warns when the working directory contradicts the switch', () => {
+    const switchTo = async (project: string): Promise<string> => {
+      await createConfigCommand({ configDeps: {}, log, cwd: () => tmpDir })
+        .parseAsync(['claude', '--project', project, '--skip-skills', '--skip-ui'], { from: 'user' });
+      return output.join('\n');
+    };
+
+    it('names a legacy entry that stays mounted alongside the new project', async () => {
+      writeFileSync(join(tmpDir, '.mcp.json'), JSON.stringify({
+        mcpServers: { homeautomation: { command: 'mcpctl', args: ['mcp', '-p', 'homeautomation'] } },
+      }));
+      const out = await switchTo('sre');
+      expect(out).toContain(join(tmpDir, '.mcp.json'));
+      expect(out).toContain("'homeautomation'");
+      expect(out).toContain('mounted alongside');
+    });
+
+    it('says a canonical pin overrides the switch in that directory', async () => {
+      writeFileSync(join(tmpDir, '.mcp.json'), JSON.stringify({
+        mcpServers: { mcpctl: { command: 'mcpctl', args: ['mcp', '-p', 'docmost'] } },
+      }));
+      expect(await switchTo('sre')).toContain('overrides the switch here');
+    });
+
+    it('stays quiet when the directory already agrees, or wires nothing of ours', async () => {
+      writeFileSync(join(tmpDir, '.mcp.json'), JSON.stringify({
+        mcpServers: {
+          mcpctl: { command: 'mcpctl', args: ['mcp', '-p', 'sre'] },
+          'their-server': { command: 'docker', args: ['run', 'x'] },
+        },
+      }));
+      expect(await switchTo('sre')).not.toContain('Warning:');
+    });
+
+    it('stays quiet when there is no .mcp.json at all', async () => {
+      expect(await switchTo('sre')).not.toContain('Warning:');
+    });
+  });
 });

src/cli/tests/commands/statusline.test.ts

@@ -0,0 +1,121 @@
+import { describe, it, expect, beforeEach, afterEach } from 'vitest';
+import { mkdtempSync, rmSync, writeFileSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { createStatuslineCommand } from '../../src/commands/statusline.js';
+
+/**
+ * The status line is what tells you which project you are in, so the property
+ * under test throughout is: after a switch, does it name the project you
+ * switched to?
+ *
+ * These drive the real command rather than the resolution helpers, because the
+ * bug they cover was in the *ranking* of sources, not in any one source.
+ */
+
+const bridge = (project: string): Record<string, unknown> => ({
+  command: 'mcpctl',
+  args: ['mcp', '-p', project],
+});
+
+let home: string;
+let dir: string;
+
+/** Claude Code's user-scope config, at the path `claudeJsonPath()` resolves. */
+function writeClaudeJson(doc: unknown): void {
+  writeFileSync(join(home, '.claude.json'), JSON.stringify(doc));
+}
+
+function writeMcpJson(doc: unknown): void {
+  writeFileSync(join(dir, '.mcp.json'), JSON.stringify(doc));
+}
+
+/** Run `statusline` for `dir` and return exactly what it printed. */
+async function statusline(): Promise<string> {
+  const out: string[] = [];
+  const cmd = createStatuslineCommand({ log: (l) => out.push(l), cwd: () => dir, homeDir: () => home });
+  await cmd.parseAsync(['--directory', dir], { from: 'user' });
+  return out.join('');
+}
+
+beforeEach(() => {
+  home = mkdtempSync(join(tmpdir(), 'mcpctl-statusline-home-'));
+  dir = mkdtempSync(join(tmpdir(), 'mcpctl-statusline-dir-'));
+  // Point claudeJsonPath() at the fake home; the CLI reads $CLAUDE_CONFIG_DIR
+  // first, which keeps this off the developer's real ~/.claude.json.
+  process.env['CLAUDE_CONFIG_DIR'] = home;
+});
+
+afterEach(() => {
+  delete process.env['CLAUDE_CONFIG_DIR'];
+  rmSync(home, { recursive: true, force: true });
+  rmSync(dir, { recursive: true, force: true });
+});
+
+describe('mcpctl statusline', () => {
+  it('reports the user-scope project when the directory wires nothing', async () => {
+    writeClaudeJson({ mcpServers: { mcpctl: bridge('sre') } });
+    expect(await statusline()).toBe('mcpctl:sre');
+  });
+
+  it('prints nothing at all when no project is active', async () => {
+    writeClaudeJson({ mcpServers: {} });
+    expect(await statusline()).toBe('');
+  });
+
+  it('lets a canonical .mcp.json pin override the user-scope project', async () => {
+    // Same server name in both scopes: Claude Code prefers project scope, so a
+    // deliberate pin is genuinely what is mounted here.
+    writeClaudeJson({ mcpServers: { mcpctl: bridge('sre') } });
+    writeMcpJson({ mcpServers: { mcpctl: bridge('docmost') } });
+    expect(await statusline()).toBe('mcpctl:docmost');
+  });
+
+  it('does not let a legacy project-named entry outrank a user-scope switch', async () => {
+    // The regression: an older mcpctl wrote `homeautomation` into a checkout,
+    // and a user-scope switch never rewrites that file — so the status line
+    // reported the old project forever and the switch looked like a no-op.
+    writeClaudeJson({ mcpServers: { mcpctl: bridge('sre') } });
+    writeMcpJson({ mcpServers: { homeautomation: bridge('homeautomation') } });
+    expect(await statusline()).toBe('mcpctl:sre');
+  });
+
+  it('still reports a legacy entry when nothing more deliberate names a project', async () => {
+    writeClaudeJson({ mcpServers: {} });
+    writeMcpJson({ mcpServers: { homeautomation: bridge('homeautomation') } });
+    expect(await statusline()).toBe('mcpctl:homeautomation');
+  });
+
+  it('skips a directory server Claude Code has switched off', async () => {
+    // A disabled server is not mounted, so naming its project is a lie.
+    writeClaudeJson({
+      mcpServers: {},
+      projects: { [dir]: { disabledMcpServers: ['homeautomation'] } },
+    });
+    writeMcpJson({ mcpServers: { homeautomation: bridge('homeautomation') } });
+    expect(await statusline()).toBe('');
+  });
+
+  it('skips a disabled pin and falls through to the user-scope project', async () => {
+    writeClaudeJson({
+      mcpServers: { mcpctl: bridge('sre') },
+      projects: { [dir]: { disabledMcpjsonServers: ['mcpctl'] } },
+    });
+    writeMcpJson({ mcpServers: { mcpctl: bridge('docmost') } });
+    expect(await statusline()).toBe('mcpctl:sre');
+  });
+
+  it('falls back to a .mcpctl-project marker when nothing is wired', async () => {
+    writeClaudeJson({ mcpServers: {} });
+    writeFileSync(join(dir, '.mcpctl-project'), 'lab\n');
+    expect(await statusline()).toBe('mcpctl:lab');
+  });
+
+  it('honours a custom prefix', async () => {
+    writeClaudeJson({ mcpServers: { mcpctl: bridge('sre') } });
+    const out: string[] = [];
+    const cmd = createStatuslineCommand({ log: (l) => out.push(l), cwd: () => dir, homeDir: () => home });
+    await cmd.parseAsync(['--directory', dir, '--prefix', 'proj '], { from: 'user' });
+    expect(out.join('')).toBe('proj sre');
+  });
+});

src/cli/tests/config/claude-mcp.test.ts

@@ -5,6 +5,9 @@ import {
   projectOfEntry,
   isLegacyMcpctlEntry,
   activeProjectIn,
+  canonicalProjectIn,
+  legacyEntriesIn,
+  disabledServersFor,
 } from '../../src/config/claude-mcp.js';
 
 const bridge = (project: string): Record<string, unknown> => ({
@@ -62,6 +65,50 @@ describe('activeProjectIn', () => {
   });
 });
 
+describe('canonicalProjectIn / legacyEntriesIn', () => {
+  it('tells a deliberate pin apart from pre-migration residue', () => {
+    const config = { mcpServers: { [MCPCTL_SERVER_NAME]: bridge('sre'), homeautomation: bridge('homeautomation') } };
+    expect(canonicalProjectIn(config)).toBe('sre');
+    expect(legacyEntriesIn(config)).toEqual([{ server: 'homeautomation', project: 'homeautomation' }]);
+  });
+
+  it('reports no canonical entry when only legacy ones are present', () => {
+    const config = { mcpServers: { docmost: bridge('docmost') } };
+    expect(canonicalProjectIn(config)).toBeNull();
+    expect(legacyEntriesIn(config)).toEqual([{ server: 'docmost', project: 'docmost' }]);
+  });
+
+  it('leaves servers that are not ours out of both', () => {
+    const config = { mcpServers: { other: { command: 'echo' } } };
+    expect(canonicalProjectIn(config)).toBeNull();
+    expect(legacyEntriesIn(config)).toEqual([]);
+    expect(legacyEntriesIn(null)).toEqual([]);
+  });
+});
+
+describe('disabledServersFor', () => {
+  const doc = {
+    projects: {
+      '/repo': { disabledMcpServers: ['homeautomation'], disabledMcpjsonServers: ['mcpctl'] },
+      '/other': { disabledMcpServers: ['sre'] },
+    },
+  };
+
+  it('unions both of Claude Code\'s disable lists for that directory', () => {
+    expect([...disabledServersFor(doc, '/repo')].sort()).toEqual(['homeautomation', 'mcpctl']);
+  });
+
+  it('is scoped to the directory asked about', () => {
+    expect([...disabledServersFor(doc, '/other')]).toEqual(['sre']);
+    expect([...disabledServersFor(doc, '/unknown')]).toEqual([]);
+    expect([...disabledServersFor(null, '/repo')]).toEqual([]);
+  });
+
+  it('survives a malformed entry rather than throwing on the status line', () => {
+    expect([...disabledServersFor({ projects: { '/repo': { disabledMcpServers: 'nope' } } }, '/repo')]).toEqual([]);
+  });
+});
+
 describe('mergeMcpctlServers', () => {
   it('writes one constant entry regardless of project', () => {
     const { config } = mergeMcpctlServers(null, { project: 'my-fancy-project' });

src/cli/tests/config/pi-extension-embed.test.ts

@@ -29,6 +29,34 @@ describe('embedded pi extension', () => {
     expect(PI_EXTENSION_FILES['mcpctl-pi.ts']).toContain('./mcp-http.js');
   });
 
+  /**
+   * pi resolves an extension's bare specifiers through a hard-coded alias table
+   * in its own loader, and that table is not the same across pi distributions:
+   * `@earendil-works/*` exists only in the newer packages, `@mariozechner/*`
+   * installs alias only the old names, and neither resolves the other. An
+   * import of a package outside the intersection makes the whole extension fail
+   * to load with `Cannot find module` — every tool gone, on someone else's pi.
+   *
+   * `typebox` is aliased by every published pi, so it is the only safe bare
+   * runtime import. Type-only imports are erased before jiti resolves anything,
+   * so they may name whatever they like.
+   */
+  it('imports nothing at runtime that some pi build cannot resolve', () => {
+    // `import x from "s"` / `import {..} from "s"` (but not `import type`),
+    // plus the side-effect form `import "s"`.
+    const runtimeImport =
+      /^\s*import\s+(?!type\s)[^;]*?from\s*["']([^"']+)["']|^\s*import\s*["']([^"']+)["']/gm;
+    const allowed = /^(node:|\.\/|\.\.\/|typebox$|typebox\/)/;
+
+    for (const name of PI_EXTENSION_FILENAMES) {
+      const src = PI_EXTENSION_FILES[name] ?? '';
+      for (const match of src.matchAll(runtimeImport)) {
+        const specifier = match[1] ?? match[2] ?? '';
+        expect(specifier, `${name} runtime-imports ${specifier}`).toMatch(allowed);
+      }
+    }
+  });
+
   it('carries the fixes the pi API requires', () => {
     const main = PI_EXTENSION_FILES['mcpctl-pi.ts'] ?? '';
     // ctx.ui.select takes string[] and returns the chosen string.

src/mcplocal/src/http/index.ts

@@ -2,7 +2,7 @@ export { createHttpServer } from './server.js';
 export type { HttpServerDeps } from './server.js';
 export { loadHttpConfig } from './config.js';
 export type { HttpConfig } from './config.js';
-export { McpdClient, AuthenticationError, ConnectionError } from './mcpd-client.js';
+export { McpdClient, AuthenticationError, ConnectionError, UpstreamTimeoutError } from './mcpd-client.js';
 export { registerProxyRoutes } from './routes/proxy.js';
 export { registerMcpEndpoint } from './mcp-endpoint.js';
 export { registerProjectMcpEndpoint } from './project-mcp-endpoint.js';

src/mcplocal/src/http/mcpd-client.ts

@@ -20,9 +20,41 @@ export class ConnectionError extends Error {
   }
 }
 
+/**
+ * Thrown when mcpd was reachable but did not finish in time.
+ *
+ * Deliberately NOT a ConnectionError. Folding timeouts into "cannot connect"
+ * is what made this class of failure so expensive to diagnose: mcpd answered
+ * /healthz in 32ms while the proxy insisted the daemon was down. A timeout and
+ * an unreachable daemon need different messages and different status codes.
+ */
+export class UpstreamTimeoutError extends Error {
+  constructor(readonly url: string, readonly timeoutMs: number) {
+    super(`mcpd did not respond within ${String(timeoutMs)}ms: ${url}`);
+    this.name = 'UpstreamTimeoutError';
+  }
+}
+
+/** True when `err` is an AbortSignal.timeout() firing. */
+function isTimeout(err: unknown): boolean {
+  return err instanceof DOMException && err.name === 'TimeoutError';
+}
+
 /** Default timeout for mcpd requests (ms). Prevents indefinite hangs on slow upstream tool calls. */
 export const DEFAULT_TIMEOUT_MS = 30_000;
 
+/**
+ * Budget for routes that are *expected* to run long: agent/project chat and
+ * raw inference. An agent turn is a multi-turn tool-use loop and legitimately
+ * runs for minutes, so the 30s default is not a safety net there — it is a
+ * guaranteed failure. Matches `STREAM_TIMEOUT_MS` in the CLI's chat command
+ * (src/cli/src/commands/chat.ts), which already allowed 10 minutes; mcplocal
+ * sitting in the middle with 30s was the binding constraint.
+ *
+ * Override with `MCPLOCAL_LONG_TIMEOUT_MS`.
+ */
+export const LONG_RUNNING_TIMEOUT_MS = Number(process.env['MCPLOCAL_LONG_TIMEOUT_MS']) || 600_000;
+
 /**
  * Discovery-class operations (tools/list, resources/list, prompts/list) should not share
  * the full tool-call timeout budget — a single dead upstream would stall session init for
@@ -121,9 +153,7 @@ export class McpdClient {
     try {
       res = await fetch(url, init);
     } catch (err: unknown) {
-      if (err instanceof DOMException && err.name === 'TimeoutError') {
-        throw new ConnectionError(this.baseUrl, new Error(`Request timed out after ${this.timeoutMs}ms`));
-      }
+      if (isTimeout(err)) throw new UpstreamTimeoutError(this.baseUrl, this.timeoutMs);
       throw new ConnectionError(this.baseUrl, err);
     }
 
@@ -131,7 +161,18 @@ export class McpdClient {
       throw new AuthenticationError();
     }
 
-    const text = await res.text();
+    // The body read MUST be inside a try. mcpd writes SSE headers immediately
+    // on chat routes, so fetch() resolves long before the turn finishes and the
+    // abort lands here instead — previously escaping as a raw DOMException and
+    // surfacing to the user as an opaque `500 code:23`.
+    let text: string;
+    try {
+      text = await res.text();
+    } catch (err: unknown) {
+      if (isTimeout(err)) throw new UpstreamTimeoutError(this.baseUrl, this.timeoutMs);
+      throw new ConnectionError(this.baseUrl, err);
+    }
+
     let parsed: unknown;
     try {
       parsed = JSON.parse(text);
@@ -142,6 +183,51 @@ export class McpdClient {
     return { status: res.status, body: parsed };
   }
 
+  /**
+   * Forward a request and hand back the raw Response, body unread.
+   *
+   * `forward()` buffers through `res.text()`, which is fine for CRUD but
+   * defeats streaming entirely: an SSE chat arrives at the client as one blob
+   * after the turn ends, so the token-by-token output the CLI draws never
+   * appears. Streaming routes use this instead and pipe the body straight
+   * through.
+   */
+  async forwardStream(
+    method: string,
+    path: string,
+    query: string,
+    body: unknown | undefined,
+    authOverride?: string,
+  ): Promise<Response> {
+    const url = `${this.baseUrl}${path}${query ? `?${query}` : ''}`;
+    const headers: Record<string, string> = {
+      ...this.extraHeaders,
+      'Authorization': `Bearer ${authOverride ?? this.token}`,
+      // Accept both: mcpd picks SSE or JSON based on the request's `stream` flag.
+      'Accept': 'text/event-stream, application/json',
+    };
+
+    const init: RequestInit = {
+      method,
+      headers,
+      signal: AbortSignal.timeout(this.timeoutMs),
+    };
+    if (body !== undefined && body !== null && method !== 'GET' && method !== 'HEAD') {
+      headers['Content-Type'] = 'application/json';
+      init.body = JSON.stringify(body);
+    }
+
+    try {
+      const res = await fetch(url, init);
+      if (res.status === 401) throw new AuthenticationError();
+      return res;
+    } catch (err: unknown) {
+      if (err instanceof AuthenticationError) throw err;
+      if (isTimeout(err)) throw new UpstreamTimeoutError(this.baseUrl, this.timeoutMs);
+      throw new ConnectionError(this.baseUrl, err);
+    }
+  }
+
   private async request<T>(method: string, path: string, body?: unknown): Promise<T> {
     const result = await this.forward(method, path, '', body);
 

src/mcplocal/src/http/routes/proxy.ts

@@ -1,10 +1,62 @@
 /**
  * Catch-all proxy route that forwards /api/v1/* requests to mcpd.
  */
-import type { FastifyInstance } from 'fastify';
-import { AuthenticationError, ConnectionError } from '../mcpd-client.js';
+import { Readable } from 'node:stream';
+
+import type { FastifyInstance, FastifyReply } from 'fastify';
+
+import { AuthenticationError, ConnectionError, UpstreamTimeoutError, LONG_RUNNING_TIMEOUT_MS } from '../mcpd-client.js';
 import type { McpdClient } from '../mcpd-client.js';
 
+/**
+ * Routes that are expected to run long and/or stream.
+ *
+ * An agent turn is a multi-turn tool-use loop — minutes, not seconds — so the
+ * 30s default budget guarantees failure rather than guarding against it. These
+ * also stream SSE, which must be piped rather than buffered or the client sees
+ * one blob at the end instead of live output.
+ */
+const LONG_RUNNING = [
+  /^\/api\/v1\/agents\/[^/]+\/chat\b/,
+  /^\/api\/v1\/projects\/[^/]+\/chat\b/,
+  /^\/api\/v1\/llms\/[^/]+\/infer\b/,
+  /^\/api\/v1\/inference-tasks\/[^/]+\/stream\b/,
+];
+
+function isLongRunning(path: string): boolean {
+  return LONG_RUNNING.some((re) => re.test(path));
+}
+
+/** Headers worth preserving from mcpd; everything else is re-derived by Fastify. */
+const PASSTHROUGH_HEADERS = ['content-type', 'cache-control', 'x-accel-buffering'];
+
+function sendUpstreamError(reply: FastifyReply, err: unknown): FastifyReply | undefined {
+  if (err instanceof AuthenticationError) {
+    return reply.code(401).send({
+      error: 'unauthorized',
+      message: 'Authentication with mcpd failed. Run `mcpctl login` to refresh your token.',
+    });
+  }
+  if (err instanceof UpstreamTimeoutError) {
+    // 504, not 503 — mcpd was reachable, it just did not finish. Reporting this
+    // as "cannot reach mcpd" sent a previous debugging session chasing a
+    // network fault while /healthz answered in 32ms.
+    return reply.code(504).send({
+      error: 'upstream_timeout',
+      message:
+        `mcpd did not respond within ${String(err.timeoutMs)}ms. The daemon is reachable — the ` +
+        'request itself ran long. Raise MCPLOCAL_LONG_TIMEOUT_MS if this is a legitimately slow turn.',
+    });
+  }
+  if (err instanceof ConnectionError) {
+    return reply.code(503).send({
+      error: 'service_unavailable',
+      message: 'Cannot reach mcpd daemon. Is it running?',
+    });
+  }
+  return undefined;
+}
+
 export function registerProxyRoutes(app: FastifyInstance, client: McpdClient): void {
   app.all('/api/v1/*', async (request, reply) => {
     const path = (request.url.split('?')[0]) ?? '/';
@@ -19,25 +71,78 @@ export function registerProxyRoutes(app: FastifyInstance, client: McpdClient): v
     // Forward the user's auth token to mcpd so RBAC applies per-user.
     // If no user token is present, mcpd will use its auth hook to reject.
     const authHeader = request.headers['authorization'] as string | undefined;
-    const userToken = authHeader?.startsWith('Bearer ') ? authHeader.slice(7) : undefined;
+    const userToken = authHeader !== undefined && authHeader.startsWith('Bearer ')
+      ? authHeader.slice(7)
+      : undefined;
+
+    if (isLongRunning(path)) {
+      return proxyStreaming(reply, client, request.method, path, querystring, body, userToken);
+    }
 
     try {
       const result = await client.forward(request.method, path, querystring, body, userToken);
       return reply.code(result.status).send(result.body);
     } catch (err: unknown) {
-      if (err instanceof AuthenticationError) {
-        return reply.code(401).send({
-          error: 'unauthorized',
-          message: 'Authentication with mcpd failed. Run `mcpctl login` to refresh your token.',
-        });
-      }
-      if (err instanceof ConnectionError) {
-        return reply.code(503).send({
-          error: 'service_unavailable',
-          message: 'Cannot reach mcpd daemon. Is it running?',
-        });
-      }
+      const handled = sendUpstreamError(reply, err);
+      if (handled) return handled;
       throw err;
     }
   });
 }
+
+/**
+ * Pipe a long-running response straight through, headers and all.
+ *
+ * Hijacks the reply so Fastify does not try to serialize a stream, then copies
+ * mcpd's status and content-type before piping. `x-accel-buffering` matters:
+ * mcpd sets it to `no` so intermediaries don't buffer SSE, and dropping it here
+ * would reintroduce the exact stall we are fixing.
+ */
+async function proxyStreaming(
+  reply: FastifyReply,
+  client: McpdClient,
+  method: string,
+  path: string,
+  querystring: string,
+  body: unknown,
+  userToken: string | undefined,
+): Promise<void> {
+  const longClient = client.withTimeout(LONG_RUNNING_TIMEOUT_MS);
+
+  let res: Response;
+  try {
+    res = await longClient.forwardStream(method, path, querystring, body, userToken);
+  } catch (err: unknown) {
+    const handled = sendUpstreamError(reply, err);
+    if (handled) return;
+    throw err;
+  }
+
+  const headers: Record<string, string> = {};
+  for (const name of PASSTHROUGH_HEADERS) {
+    const value = res.headers.get(name);
+    if (value !== null) headers[name] = value;
+  }
+
+  reply.hijack();
+  reply.raw.writeHead(res.status, headers);
+
+  if (res.body === null) {
+    reply.raw.end();
+    return;
+  }
+
+  try {
+    // Node's Readable.fromWeb bridges the fetch ReadableStream onto the socket.
+    await new Promise<void>((resolve, reject) => {
+      const upstream = Readable.fromWeb(res.body as Parameters<typeof Readable.fromWeb>[0]);
+      upstream.on('error', reject);
+      reply.raw.on('close', () => { upstream.destroy(); resolve(); });
+      upstream.pipe(reply.raw).on('finish', resolve).on('error', reject);
+    });
+  } catch {
+    // Headers are already on the wire, so there is no status left to change.
+    // Close the socket; the client surfaces the truncated stream.
+    if (!reply.raw.writableEnded) reply.raw.end();
+  }
+}

src/mcplocal/src/index.ts

@@ -11,7 +11,7 @@ export type { MainResult } from './main.js';
 export { ProviderRegistry } from './providers/index.js';
 export type { LlmProvider, CompletionOptions, CompletionResult, ChatMessage } from './providers/index.js';
 export { OpenAiProvider, AnthropicProvider, OllamaProvider, GeminiCliProvider, DeepSeekProvider } from './providers/index.js';
-export { createHttpServer, loadHttpConfig, McpdClient, AuthenticationError, ConnectionError, registerProxyRoutes } from './http/index.js';
+export { createHttpServer, loadHttpConfig, McpdClient, AuthenticationError, ConnectionError, UpstreamTimeoutError, registerProxyRoutes } from './http/index.js';
 export type { HttpConfig, HttpServerDeps } from './http/index.js';
 export type {
   JsonRpcRequest,

src/mcplocal/tests/mcpd-client.test.ts

@@ -1,6 +1,6 @@
 import { describe, it, expect, afterAll, afterEach } from 'vitest';
 import http from 'node:http';
-import { McpdClient, ConnectionError } from '../src/http/mcpd-client.js';
+import { McpdClient, ConnectionError, UpstreamTimeoutError } from '../src/http/mcpd-client.js';
 
 /**
  * Create a local HTTP server for testing McpdClient behavior.
@@ -85,7 +85,7 @@ describe('McpdClient', () => {
 
   // ── Timeout behavior ──
 
-  it('times out on slow responses and throws ConnectionError', async () => {
+  it('times out on slow responses and throws UpstreamTimeoutError', async () => {
     const { server, url } = await createTestServer((_req, _res) => {
       // Never respond — simulates a hanging upstream tool call
     });
@@ -96,7 +96,7 @@ describe('McpdClient', () => {
 
     const start = Date.now();
     await expect(client.post('/api/v1/mcp/proxy', { serverId: 's1' })).rejects.toThrow(
-      /timed out/,
+      /did not respond within/,
     );
     const elapsed = Date.now() - start;
 
@@ -105,7 +105,7 @@ describe('McpdClient', () => {
     expect(elapsed).toBeLessThan(3000);
   });
 
-  it('timeout error is a ConnectionError with descriptive message', async () => {
+  it('timeout is NOT a ConnectionError — a slow daemon is not an absent one', async () => {
     const { server, url } = await createTestServer((_req, _res) => {
       // Never respond
     });
@@ -117,8 +117,12 @@ describe('McpdClient', () => {
       await client.get('/test');
       expect.unreachable('Should have thrown');
     } catch (err) {
-      expect(err).toBeInstanceOf(ConnectionError);
-      expect((err as Error).message).toContain('Request timed out after 200ms');
+      // Reporting a timeout as "cannot connect" is what sent a previous
+      // debugging session chasing a network fault that did not exist.
+      expect(err).toBeInstanceOf(UpstreamTimeoutError);
+      expect(err).not.toBeInstanceOf(ConnectionError);
+      expect((err as UpstreamTimeoutError).timeoutMs).toBe(200);
+      expect((err as Error).message).toContain('did not respond within 200ms');
     }
   });
 
@@ -146,7 +150,7 @@ describe('McpdClient', () => {
     const derived = client.withHeaders({ 'X-Custom': 'val' });
 
     const start = Date.now();
-    await expect(derived.get('/test')).rejects.toThrow(/timed out/);
+    await expect(derived.get('/test')).rejects.toThrow(/did not respond within/);
     const elapsed = Date.now() - start;
     expect(elapsed).toBeLessThan(2000);
   });

src/mcplocal/tests/proxy-long-running.test.ts

@@ -0,0 +1,255 @@
+import http from 'node:http';
+
+import Fastify, { type FastifyInstance } from 'fastify';
+import { describe, it, expect, afterEach } from 'vitest';
+
+import {
+  McpdClient,
+  UpstreamTimeoutError,
+  ConnectionError,
+  LONG_RUNNING_TIMEOUT_MS,
+  DEFAULT_TIMEOUT_MS,
+} from '../src/http/mcpd-client.js';
+import { registerProxyRoutes } from '../src/http/routes/proxy.js';
+
+/**
+ * Regression cover for the 30s proxy timeout that made `mcpctl chat` fail with
+ * a misleading "Cannot reach mcpd daemon" 503 while mcpd was answering
+ * /healthz in 32ms.
+ *
+ * Three separate defects are pinned here:
+ *   1. chat routes inherited the 30s CRUD budget, so any turn longer than 30s
+ *      failed — and an agent turn is a tool-use loop that routinely exceeds it;
+ *   2. a timeout was reported as a connection failure, sending diagnosis after
+ *      a network fault that did not exist;
+ *   3. SSE was buffered through res.text(), so streaming never reached the
+ *      client even when the turn finished in time.
+ */
+let app: FastifyInstance | null = null;
+let upstream: FastifyInstance | null = null;
+
+afterEach(async () => {
+  if (app) { await app.close(); app = null; }
+  if (upstream) { await upstream.close(); upstream = null; }
+});
+
+/** A stand-in mcpd. Returns its base URL. */
+async function startUpstream(register: (a: FastifyInstance) => void): Promise<string> {
+  upstream = Fastify();
+  register(upstream);
+  await upstream.listen({ port: 0, host: '127.0.0.1' });
+  const addr = upstream.server.address();
+  if (addr === null || typeof addr === 'string') throw new Error('no address');
+  return `http://127.0.0.1:${String(addr.port)}`;
+}
+
+async function startProxy(baseUrl: string, timeoutMs?: number): Promise<FastifyInstance> {
+  app = Fastify();
+  registerProxyRoutes(app, new McpdClient(baseUrl, 'test-token', {}, timeoutMs));
+  await app.ready();
+  return app;
+}
+
+describe('proxy — long-running route budget', () => {
+  it('gives chat routes the long budget, not the 30s CRUD default', () => {
+    // The constants themselves are the contract: a 30s cap on an agent turn is
+    // a guaranteed failure, not a safety net.
+    expect(DEFAULT_TIMEOUT_MS).toBe(30_000);
+    expect(LONG_RUNNING_TIMEOUT_MS).toBeGreaterThanOrEqual(600_000);
+  });
+
+  it('does not abort an agent chat that outlives the CRUD budget', async () => {
+    const base = await startUpstream((a) => {
+      a.post('/api/v1/agents/:name/chat', async () => {
+        // Longer than the (deliberately tiny) CRUD budget below. Before the
+        // fix this inherited that budget and 503'd.
+        await new Promise((r) => setTimeout(r, 250));
+        return { answer: 'pong' };
+      });
+    });
+    // CRUD budget of 50ms — a chat route must NOT inherit it.
+    const proxy = await startProxy(base, 50);
+
+    const res = await proxy.inject({
+      method: 'POST',
+      url: '/api/v1/agents/reviewer/chat',
+      payload: { message: 'hi' },
+    });
+
+    expect(res.statusCode).toBe(200);
+    expect(res.json()).toEqual({ answer: 'pong' });
+  });
+
+  it('still applies the short budget to ordinary CRUD routes', async () => {
+    const base = await startUpstream((a) => {
+      a.get('/api/v1/servers', async () => {
+        await new Promise((r) => setTimeout(r, 300));
+        return [];
+      });
+    });
+    const proxy = await startProxy(base, 50);
+
+    const res = await proxy.inject({ method: 'GET', url: '/api/v1/servers' });
+    // Times out — and is now reported honestly as a timeout, not a connection fault.
+    expect(res.statusCode).toBe(504);
+    expect(res.json().error).toBe('upstream_timeout');
+  });
+
+  it('reports a timeout as 504, never as "cannot reach mcpd"', async () => {
+    const base = await startUpstream((a) => {
+      a.get('/api/v1/servers', async () => {
+        await new Promise((r) => setTimeout(r, 300));
+        return [];
+      });
+    });
+    const proxy = await startProxy(base, 50);
+
+    const res = await proxy.inject({ method: 'GET', url: '/api/v1/servers' });
+    const body = res.json();
+    expect(body.message).toMatch(/did not respond within/);
+    expect(body.message).not.toMatch(/Cannot reach mcpd/);
+    expect(body.message).toMatch(/reachable/);
+  });
+
+  it('streams SSE through instead of buffering it', async () => {
+    const base = await startUpstream((a) => {
+      a.post('/api/v1/agents/:name/chat', async (_req, reply) => {
+        reply.raw.writeHead(200, {
+          'Content-Type': 'text/event-stream',
+          'Cache-Control': 'no-cache',
+          'X-Accel-Buffering': 'no',
+        });
+        reply.raw.write('data: {"type":"text","delta":"po"}\n\n');
+        reply.raw.write('data: {"type":"text","delta":"ng"}\n\n');
+        reply.raw.write('data: [DONE]\n\n');
+        reply.raw.end();
+        return reply;
+      });
+    });
+    const proxy = await startProxy(base, 50);
+
+    const res = await proxy.inject({
+      method: 'POST',
+      url: '/api/v1/agents/reviewer/chat',
+      payload: { message: 'hi', stream: true },
+    });
+
+    expect(res.statusCode).toBe(200);
+    // Content-type must survive — a client that gets application/json will not
+    // parse the event stream.
+    expect(res.headers['content-type']).toMatch(/text\/event-stream/);
+    // x-accel-buffering=no must survive too, or intermediaries re-buffer the
+    // stream and reintroduce the stall.
+    expect(res.headers['x-accel-buffering']).toBe('no');
+    expect(res.body).toContain('"delta":"po"');
+    expect(res.body).toContain('"delta":"ng"');
+    expect(res.body).toContain('[DONE]');
+  });
+
+  it('delivers each SSE frame while the upstream is still generating', async () => {
+    // The buffering regression is invisible to the pass-through test above:
+    // `inject()` collects the whole body, so a proxy that buffers via
+    // res.text() still passes it. This test proves *progressive* delivery by
+    // making the upstream withhold its final frame until the client has
+    // observed the first one. A buffering proxy can never satisfy that
+    // ordering — the 3s guard resolves the gate so the run fails cleanly
+    // instead of deadlocking.
+    let openGate: (seen: boolean) => void = () => {};
+    const clientSawFirstFrame = new Promise<boolean>((r) => { openGate = r; });
+    const guard = setTimeout(() => openGate(false), 3_000);
+
+    const base = await startUpstream((a) => {
+      a.post('/api/v1/agents/:name/chat', async (_req, reply) => {
+        reply.raw.writeHead(200, { 'Content-Type': 'text/event-stream' });
+        reply.raw.write('data: {"type":"text","delta":"live"}\n\n');
+        await clientSawFirstFrame;
+        reply.raw.write('data: {"type":"final"}\n\n');
+        reply.raw.write('data: [DONE]\n\n');
+        reply.raw.end();
+        return reply;
+      });
+    });
+    const proxy = await startProxy(base, 50);
+    await proxy.listen({ port: 0, host: '127.0.0.1' });
+    const addr = proxy.server.address();
+    if (addr === null || typeof addr === 'string') throw new Error('no address');
+
+    const body = await new Promise<string>((resolve, reject) => {
+      const req = http.request({
+        hostname: '127.0.0.1',
+        port: addr.port,
+        path: '/api/v1/agents/reviewer/chat',
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+      }, (res) => {
+        let acc = '';
+        res.setEncoding('utf-8');
+        res.on('data', (chunk: string) => {
+          acc += chunk;
+          if (acc.includes('"delta":"live"')) openGate(true);
+        });
+        res.on('end', () => resolve(acc));
+        res.on('error', reject);
+      });
+      req.on('error', reject);
+      req.end(JSON.stringify({ message: 'hi', stream: true }));
+    });
+    clearTimeout(guard);
+
+    // The ordering proof: the first frame reached the client while the
+    // upstream was still holding the stream open.
+    await expect(clientSawFirstFrame).resolves.toBe(true);
+    expect(body).toContain('"type":"final"');
+    expect(body).toContain('[DONE]');
+  });
+
+  it('relays a non-200 status from a streaming route', async () => {
+    const base = await startUpstream((a) => {
+      a.post('/api/v1/agents/:name/chat', async (_req, reply) => {
+        return reply.code(404).send({ error: 'Agent not found' });
+      });
+    });
+    const proxy = await startProxy(base, 50);
+
+    const res = await proxy.inject({
+      method: 'POST',
+      url: '/api/v1/agents/ghost/chat',
+      payload: { message: 'hi' },
+    });
+    expect(res.statusCode).toBe(404);
+    expect(res.body).toContain('Agent not found');
+  });
+
+  it('still reports a genuinely unreachable daemon as 503', async () => {
+    // Port 1 is reserved and refuses instantly.
+    const proxy = await startProxy('http://127.0.0.1:1', 500);
+    const res = await proxy.inject({ method: 'GET', url: '/api/v1/servers' });
+
+    expect(res.statusCode).toBe(503);
+    expect(res.json().error).toBe('service_unavailable');
+  });
+
+  it('propagates 401 from a streaming route so login guidance still fires', async () => {
+    const base = await startUpstream((a) => {
+      a.post('/api/v1/agents/:name/chat', async (_req, reply) => reply.code(401).send({}));
+    });
+    const proxy = await startProxy(base, 50);
+
+    const res = await proxy.inject({
+      method: 'POST',
+      url: '/api/v1/agents/reviewer/chat',
+      payload: { message: 'hi' },
+    });
+    expect(res.statusCode).toBe(401);
+    expect(res.json().message).toMatch(/mcpctl login/);
+  });
+});
+
+describe('error taxonomy', () => {
+  it('keeps timeout and unreachable as distinct types', () => {
+    const timeout = new UpstreamTimeoutError('http://mcpd', 30_000);
+    expect(timeout).not.toBeInstanceOf(ConnectionError);
+    expect(timeout.timeoutMs).toBe(30_000);
+    expect(timeout.message).toMatch(/did not respond within 30000ms/);
+  });
+});

src/mcplocal/tests/smoke/agent-chat.smoke.test.ts

@@ -18,8 +18,12 @@ import { describe, it, expect, beforeAll, afterAll } from 'vitest';
 import http from 'node:http';
 import https from 'node:https';
 import { spawnSync, execSync } from 'node:child_process';
+import { existsSync, readFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { homedir } from 'node:os';
 
 const MCPD_URL = process.env.MCPD_URL ?? 'https://mcpctl.ad.itaz.eu';
+const MCPLOCAL_URL = process.env.MCPLOCAL_URL ?? 'http://localhost:3200';
 const LLM_URL = process.env.MCPCTL_SMOKE_LLM_URL;
 const LLM_MODEL = process.env.MCPCTL_SMOKE_LLM_MODEL ?? 'qwen3-thinking';
 const LLM_KEY = process.env.MCPCTL_SMOKE_LLM_KEY;
@@ -27,6 +31,10 @@ const SUFFIX = Date.now().toString(36);
 const SECRET_NAME = `smoke-chat-sec-${SUFFIX}`;
 const LLM_NAME = `smoke-chat-llm-${SUFFIX}`;
 const AGENT_NAME = `smoke-chat-agent-${SUFFIX}`;
+// Dedicated agent for the streaming-timing test: the shared agent's system
+// prompt pins the reply to a single token, which is too short to distinguish
+// live streaming from an end-of-turn buffer dump.
+const STREAM_AGENT_NAME = `smoke-stream-agent-${SUFFIX}`;
 
 interface CliResult { code: number; stdout: string; stderr: string }
 
@@ -99,6 +107,7 @@ describe('agent chat smoke (live LLM)', () => {
   afterAll(() => {
     if (!liveLlmConfigured || !mcpdUp) return;
     run(`delete agent ${AGENT_NAME}`);
+    run(`delete agent ${STREAM_AGENT_NAME}`);
     run(`delete llm ${LLM_NAME}`);
     run(`delete secret ${SECRET_NAME}`);
   });
@@ -139,6 +148,92 @@ describe('agent chat smoke (live LLM)', () => {
     expect(result.stderr).toMatch(/thread:\s+c[a-z0-9]+/);
   });
 
+  it('streams progressively THROUGH mcplocal — frames arrive during generation, not in one burst', async () => {
+    if (!liveLlmConfigured || !mcpdUp) return;
+    // The regression this pins: mcplocal's /api/v1/* proxy buffered SSE via
+    // res.text(), so the CLI showed nothing until the turn finished and then
+    // dumped the whole answer at once. The --direct tests above bypass
+    // mcplocal entirely and cannot catch that. This one posts to the local
+    // proxy (the path `mcpctl chat` actually takes) and asserts frames are
+    // spread across the generation window: with buffering, everything lands
+    // within a few ms of stream end.
+    if (!(await healthz(MCPLOCAL_URL))) {
+      // eslint-disable-next-line no-console
+      console.warn(`\n  ○ mcplocal streaming smoke: skipped — ${MCPLOCAL_URL}/healthz unreachable.\n`);
+      return;
+    }
+    let token = '';
+    try {
+      const credsPath = join(homedir(), '.mcpctl', 'credentials');
+      if (existsSync(credsPath)) {
+        const creds = JSON.parse(readFileSync(credsPath, 'utf-8')) as { token?: string };
+        if (creds.token !== undefined) token = creds.token;
+      }
+    } catch { /* unauthenticated — the request will 401 and fail loudly */ }
+
+    run(`delete agent ${STREAM_AGENT_NAME}`);
+    const agent = run([
+      `create agent ${STREAM_AGENT_NAME}`,
+      `--llm ${LLM_NAME}`,
+      `--description "mcplocal streaming smoke"`,
+      `--system-prompt "You are a smoke test. Follow the user's instructions exactly."`,
+      '--default-temperature 0',
+      '--default-max-tokens 512',
+    ].join(' '));
+    expect(agent.code, agent.stderr).toBe(0);
+
+    const url = new URL(`${MCPLOCAL_URL.replace(/\/$/, '')}/api/v1/agents/${STREAM_AGENT_NAME}/chat`);
+    const deltaTimes: number[] = [];
+    let endTime = 0;
+    let status = 0;
+    let raw = '';
+
+    await new Promise<void>((resolve, reject) => {
+      const req = http.request({
+        hostname: url.hostname,
+        port: url.port || 80,
+        path: url.pathname,
+        method: 'POST',
+        timeout: 120_000,
+        headers: {
+          'Content-Type': 'application/json',
+          ...(token !== '' ? { Authorization: `Bearer ${token}` } : {}),
+        },
+      }, (res) => {
+        status = res.statusCode ?? 0;
+        res.setEncoding('utf-8');
+        let buf = '';
+        res.on('data', (chunk: string) => {
+          raw += chunk;
+          buf += chunk;
+          let nl: number;
+          while ((nl = buf.indexOf('\n\n')) !== -1) {
+            const frame = buf.slice(0, nl);
+            buf = buf.slice(nl + 2);
+            if (/"type":"(text|thinking)"/.test(frame)) deltaTimes.push(Date.now());
+          }
+        });
+        res.on('end', () => { endTime = Date.now(); resolve(); });
+        res.on('error', reject);
+      });
+      req.on('error', reject);
+      req.on('timeout', () => { req.destroy(); reject(new Error('stream timed out')); });
+      req.end(JSON.stringify({
+        message: 'Count from 1 to 40, one number per line. No other text.',
+        stream: true,
+        max_tokens: 400,
+      }));
+    });
+
+    expect(status, raw.slice(0, 500)).toBe(200);
+    expect(deltaTimes.length).toBeGreaterThanOrEqual(2);
+    // The buffering signature: every frame lands in the same final burst as
+    // stream end. Live streaming puts the first delta well before the end —
+    // a 40-line generation spans seconds; 300ms is a conservative floor.
+    const firstDelta = deltaTimes[0]!;
+    expect(endTime - firstDelta).toBeGreaterThanOrEqual(300);
+  }, 150_000);
+
   it('streaming `mcpctl chat` emits text deltas', () => {
     if (!liveLlmConfigured || !mcpdUp) return;
     // Default mode is streaming. Pipe stdout/stderr separately.

src/mcplocal/tests/smoke/fixtures/smoke-data.yaml

@@ -12,6 +12,19 @@ servers:
     env:
       - name: FASTMCP_LOG_LEVEL
         value: "ERROR"
+    # Mirrors the production `aws-docs` probe. Without it this fixture is a
+    # RUNNING server with no readiness probe, so it fails the very assertion in
+    # health-readiness.smoke.test.ts that the fixture exists to support — the
+    # suite reporting its own scaffolding as a fleet regression.
+    # `search_documentation` needs a phrase; the 300s interval matches aws-docs,
+    # since the call leaves the cluster.
+    healthCheck:
+      tool: search_documentation
+      arguments:
+        search_phrase: "s3 bucket"
+      timeoutSeconds: 20
+      intervalSeconds: 300
+      failureThreshold: 3
 
 projects:
   - name: smoke-data

src/mcplocal/tests/smoke/security.test.ts

@@ -32,6 +32,18 @@ function httpRequest(opts: {
   headers?: Record<string, string>;
   body?: string;
   timeout?: number;
+  /**
+   * Resolve as soon as the response headers arrive, then hang up, instead of
+   * waiting for the body to end.
+   *
+   * Required for a streaming endpoint: SSE responses never end, so the normal
+   * path can only settle via the socket's *inactivity* timeout — which never
+   * fires while the stream is busy. `/inspect` relays every project's MCP
+   * traffic, so during a full smoke run it is never idle, and the request hung
+   * until vitest killed the test. Alone it looked flaky; under load it failed
+   * every time. Reading the status does not need the body anyway.
+   */
+  headersOnly?: boolean;
 }): Promise<{ status: number; headers: http.IncomingHttpHeaders; body: string }> {
   return new Promise((resolve, reject) => {
     const parsed = new URL(opts.url);
@@ -46,6 +58,12 @@ function httpRequest(opts: {
         timeout: opts.timeout ?? 10_000,
       },
       (res) => {
+        if (opts.headersOnly === true) {
+          resolve({ status: res.statusCode ?? 0, headers: res.headers, body: '' });
+          res.destroy();
+          req.destroy();
+          return;
+        }
         const chunks: Buffer[] = [];
         res.on('data', (chunk: Buffer) => chunks.push(chunk));
         res.on('end', () => {
@@ -93,17 +111,15 @@ describe('Smoke: Security — mcplocal unauthenticated endpoints', () => {
 
     // /inspect streams ALL MCP traffic (tool calls, arguments, responses)
     // for ALL projects to any unauthenticated local client
+    // headersOnly: the stream never ends, and waiting for it to go idle is what
+    // made this hang whenever other suites were generating traffic. The status
+    // line is all this assertion needs.
     const res = await httpRequest({
       url: `${MCPLOCAL_URL}/inspect`,
       method: 'GET',
       headers: { 'Accept': 'text/event-stream' },
       timeout: 3_000,
-    }).catch((err) => {
-      // Timeout is expected (SSE keeps connection open) — still means endpoint is accessible
-      if ((err as Error).message.includes('timed out')) {
-        return { status: 200, headers: {} as http.IncomingHttpHeaders, body: '' };
-      }
-      throw err;
+      headersOnly: true,
     });
 
     // Should be accessible without auth (documenting the vulnerability)

src/pi-ext/mcpctl-pi.ts

@@ -20,9 +20,16 @@
  * or via settings:  "extensions": ["/abs/path/to/mcpctl-pi.ts"]
  *
  * Only imports pi-bundled packages — no @mcpctl/*, no ~/.claude.
+ *
+ * RUNTIME IMPORTS ARE LOAD-BEARING: pi resolves an extension's bare specifiers
+ * through a fixed alias table in its own loader, and that table differs between
+ * pi distributions — `@earendil-works/*` exists only in the newer packages,
+ * while `@mariozechner/*` installs alias only the old names. `typebox` is the
+ * one specifier every published pi aliases, so it is the ONLY runtime import
+ * allowed here. Anything else must be `import type` (erased before jiti runs)
+ * or inlined — see `stringEnum` below.
  */
 import { Type, type TSchema } from "typebox";
-import { StringEnum } from "@earendil-works/pi-ai";
 import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
 import {
   McpHttpSession,
@@ -110,6 +117,23 @@ async function listProjects(mcplocalUrl: string, token?: string): Promise<string
 }
 
 // ── JSON Schema → TypeBox ────────────────────────────────────────────────────
+
+/**
+ * `{ type: "string", enum: [...] }` rather than a union of literals: Google's
+ * API (and other providers that reject anyOf/const) only accept the flat form.
+ *
+ * Inlined from pi-ai's `StringEnum` on purpose — importing it dragged in
+ * `@earendil-works/pi-ai`, which older pi installs cannot resolve, and the
+ * whole extension then failed to load. See the import note at the top.
+ */
+function stringEnum(values: string[], description?: string): TSchema {
+  return Type.Unsafe<string>({
+    type: "string",
+    enum: values,
+    ...(description ? { description } : {}),
+  });
+}
+
 function convertSchema(inputSchema: unknown): TSchema {
   if (!inputSchema || typeof inputSchema !== "object") {
     return Type.Object({});
@@ -147,7 +171,7 @@ function convertProp(raw: unknown): TSchema {
   const enumVals = Array.isArray(s.enum) && s.enum.length > 0 ? s.enum : undefined;
 
   if (enumVals && enumVals.every((v) => typeof v === "string")) {
-    return StringEnum(enumVals as string[]);
+    return stringEnum(enumVals as string[], desc);
   }
   if (enumVals && enumVals.every((v) => typeof v === "number")) {
     const literals = enumVals.map((v) => Type.Literal(v));