sync: defer 51 commits (agent-loop-core fix + new gateway relay family + 1 unported tui_gateway)

.gitignore

@@ -5,6 +5,7 @@
 *.pyc*
 __pycache__/
 .venv/
+.venv
 .vscode/
 .env
 .env.local

Dockerfile

@@ -11,8 +11,11 @@ FROM ghcr.io/astral-sh/uv:0.11.6-python3.13-trixie@sha256:b3c543b6c4f23a5f2df228
 FROM node:26-bookworm-slim@sha256:79723b41edbedf595f62e943a9f8b0ba9af5b1e61045c5f8f59c2c02c1212a16 AS node_source
 FROM debian:13.4
 
-# Disable Python stdout buffering to ensure logs are printed immediately
+# Disable Python stdout buffering to ensure logs are printed immediately.
+# Do not write .pyc files at runtime: /opt/hermes is immutable in the
+# published container and writable state belongs under /opt/data.
 ENV PYTHONUNBUFFERED=1
+ENV PYTHONDONTWRITEBYTECODE=1
 
 # Store Playwright browsers outside the volume mount so the build-time
 # install survives the /opt/data volume overlay at runtime.
@@ -196,36 +199,31 @@ RUN cd web && npm run build && \
 
 # ---------- Source code ----------
 # .dockerignore excludes node_modules, so the installs above survive.
-COPY --chown=hermes:hermes . .
+COPY . .
 
 # ---------- Permissions ----------
-# Make install dir world-readable so any HERMES_UID can read it at runtime.
-# The venv needs to be traversable too.
-# node_modules trees additionally need to be writable by the hermes user
-# so the runtime `npm install` triggered by _tui_need_npm_install() in
-# hermes_cli/main.py succeeds (see #18800). /opt/hermes/web is build-time
-# only (HERMES_WEB_DIST points at hermes_cli/web_dist) and is intentionally
-# not chowned here.
-# /opt/hermes/gateway is runtime-writable: Python may create __pycache__ and
-# gateway state artifacts beneath the package after services drop privileges,
-# especially when the hermes UID is remapped at boot (#27221).
-# The .venv MUST remain hermes-writable so lazy_deps.py can install
-# remaining optional platform packages and future pin bumps at first use.
-# Without this, `uv pip install` fails with EACCES and adapters silently
-# fail to load.  See tools/lazy_deps.py.
+# Link hermes-agent itself (editable). Deps are already installed in the
+# cached layer above; `--no-deps` makes this a fast egg-link creation with no
+# resolution or downloads.
+RUN uv pip install --no-cache-dir --no-deps -e "."
+
+# Keep /opt/hermes immutable for the runtime hermes user. Hosted/container
+# instances must not be able to self-edit the installed source or venv; user
+# data, skills, plugins, config, logs, and dashboard uploads live under
+# /opt/data instead. Root can still repair the image during build/boot, but
+# supervised Hermes processes drop to the non-root hermes user.
 USER root
-RUN chmod -R a+rX /opt/hermes && \
-    chown -R hermes:hermes /opt/hermes/.venv /opt/hermes/ui-tui /opt/hermes/gateway /opt/hermes/node_modules
+RUN mkdir -p /opt/hermes/bin && \
+    cp /opt/hermes/docker/hermes-exec-shim.sh /opt/hermes/bin/hermes && \
+    chmod 0755 /opt/hermes/bin/hermes && \
+    chown -R root:root /opt/hermes && \
+    chmod -R a+rX /opt/hermes && \
+    chmod -R a-w /opt/hermes
 # Start as root so the s6-overlay stage2 hook can usermod/groupmod and chown
 # the data volume. Each supervised service then drops to the hermes user via
 # `s6-setuidgid hermes` in its run script. If HERMES_UID is unset, services
 # run as the default hermes user (UID 10000).
 
-# ---------- Link hermes-agent itself (editable) ----------
-# Deps are already installed in the cached layer above; `--no-deps` makes
-# this a fast (~1s) egg-link creation with no resolution or downloads.
-RUN uv pip install --no-cache-dir --no-deps -e "."
-
 # ---------- Bake build-time git revision ----------
 # .dockerignore excludes .git, so `git rev-parse HEAD` from inside the
 # container always returns nothing — meaning `hermes dump` reports
@@ -245,8 +243,9 @@ RUN uv pip install --no-cache-dir --no-deps -e "."
 # every published image has it.
 ARG HERMES_GIT_SHA=
 RUN if [ -n "${HERMES_GIT_SHA}" ]; then \
+        chmod u+w /opt/hermes && \
         printf '%s\n' "${HERMES_GIT_SHA}" > /opt/hermes/.hermes_build_sha && \
-        chown hermes:hermes /opt/hermes/.hermes_build_sha; \
+        chmod a-w /opt/hermes /opt/hermes/.hermes_build_sha; \
     fi
 
 # ---------- s6-overlay service wiring ----------
@@ -292,6 +291,8 @@ ENV HERMES_WEB_DIST=/opt/hermes/hermes_cli/web_dist
 # check. (A separate launcher hardening is tracked independently.)
 ENV HERMES_TUI_DIR=/opt/hermes/ui-tui
 ENV HERMES_HOME=/opt/data
+ENV HERMES_WRITE_SAFE_ROOT=/opt/data
+ENV HERMES_DISABLE_LAZY_INSTALLS=1
 
 # `docker exec` privilege-drop shim. When operators run
 # `docker exec <c> hermes ...` they default to root, and any file the
@@ -304,7 +305,6 @@ ENV HERMES_HOME=/opt/data
 # Recursion is impossible because the shim exec's the venv binary by
 # absolute path (/opt/hermes/.venv/bin/hermes). See the shim source for
 # the opt-out env var (HERMES_DOCKER_EXEC_AS_ROOT=1).
-COPY --chmod=0755 docker/hermes-exec-shim.sh /opt/hermes/bin/hermes
 
 # Pre-s6 entrypoint.sh did `source .venv/bin/activate` which exported
 # the venv bin onto PATH; Architecture B's main-wrapper.sh does the

agent/codex_responses_adapter.py

@@ -262,6 +262,26 @@ def _responses_tools(tools: Optional[List[Dict[str, Any]]] = None) -> Optional[L
     return converted or None
 
 
+# Provider-executed built-in tool *declaration* types accepted on the
+# Responses ``tools`` array.  These are declared by ``type`` alone (no
+# client-side name/parameters schema) and run server-side — the provider
+# owns the implementation and reports progress via the matching ``*_call``
+# output items.  Hermes injects xAI's native ``web_search`` for the xAI
+# transport (see agent/transports/codex.py); the rest are listed so the
+# preflight validator passes them through rather than rejecting them as
+# "unsupported type".  Mirrors the ``*_call`` item-type set used in
+# _normalize_codex_response.
+_RESPONSES_BUILTIN_TOOL_TYPES = {
+    "web_search",
+    "web_search_preview",
+    "file_search",
+    "code_interpreter",
+    "image_generation",
+    "computer_use_preview",
+    "local_shell",
+}
+
+
 # ---------------------------------------------------------------------------
 # Message format conversion
 # ---------------------------------------------------------------------------
@@ -802,7 +822,22 @@ def _preflight_codex_api_kwargs(
         for idx, tool in enumerate(tools):
             if not isinstance(tool, dict):
                 raise ValueError(f"Codex Responses tools[{idx}] must be an object.")
-            if tool.get("type") != "function":
+
+            tool_type = tool.get("type")
+
+            # Provider-executed built-in tools (xAI native web_search, code
+            # interpreter, etc.) are declared by ``type`` alone and carry no
+            # ``name``/``parameters`` schema — the provider owns the
+            # implementation.  Pass them through verbatim instead of forcing
+            # them through the function-tool validation below (which would
+            # otherwise reject them with "unsupported type").  See
+            # agent/transports/codex.py for where xAI's native web_search is
+            # injected.
+            if tool_type in _RESPONSES_BUILTIN_TOOL_TYPES:
+                normalized_tools.append(dict(tool))
+                continue
+
+            if tool_type != "function":
                 raise ValueError(f"Codex Responses tools[{idx}] has unsupported type {tool.get('type')!r}.")
 
             name = tool.get("name")
@@ -1086,6 +1121,33 @@ def _normalize_codex_response(
     saw_final_answer_phase = False
     saw_reasoning_item = False
 
+    # Server-side built-in tool calls (xAI's native web_search, code
+    # interpreter, etc.) are executed by the provider and reported as
+    # discrete ``*_call`` output items.  xAI's /v1/responses surface
+    # (e.g. grok-composer-2.5-fast on SuperGrok OAuth) routinely leaves
+    # these items at ``status="in_progress"`` even when the overall
+    # ``response.status == "completed"`` — the search ran to completion
+    # server-side, the per-item status simply isn't reconciled.  These
+    # are NOT a signal that the model's turn is unfinished, so they must
+    # not flip ``has_incomplete_items``.  Only the response-level status
+    # and genuine model output items (message/reasoning/function_call)
+    # govern the incomplete verdict.  Without this guard, any turn where
+    # grok-composer invokes server-side search is misclassified as
+    # ``finish_reason="incomplete"`` and burns 3 fruitless continuation
+    # retries before failing with "Codex response remained incomplete
+    # after 3 continuation attempts".  client-side function/custom tool
+    # calls keep their own in_progress handling below (they are skipped,
+    # not awaited).
+    _SERVER_SIDE_TOOL_CALL_TYPES = {
+        "web_search_call",
+        "file_search_call",
+        "code_interpreter_call",
+        "image_generation_call",
+        "computer_call",
+        "local_shell_call",
+        "mcp_call",
+    }
+
     for item in output:
         item_type = getattr(item, "type", None)
         item_status = getattr(item, "status", None)
@@ -1094,7 +1156,10 @@ def _normalize_codex_response(
         else:
             item_status = None
 
-        if item_status in {"queued", "in_progress", "incomplete"}:
+        if (
+            item_status in {"queued", "in_progress", "incomplete"}
+            and item_type not in _SERVER_SIDE_TOOL_CALL_TYPES
+        ):
             has_incomplete_items = True
             saw_streaming_or_item_incomplete = True
 

agent/conversation_loop.py

@@ -3756,8 +3756,30 @@ def _perform_api_call(next_api_kwargs):
                     assistant_msg = agent._build_assistant_message(assistant_message, finish_reason)
                     messages.append(assistant_msg)
                     for tc in assistant_message.tool_calls:
-                        if tc.function.name not in agent.valid_tool_names:
-                            content = f"Tool '{tc.function.name}' does not exist. Available tools: {available}"
+                        _tc_name = tc.function.name
+                        if _tc_name not in agent.valid_tool_names:
+                            # A blank/whitespace-only name is not a typo the
+                            # model can fuzzy-correct toward a real tool — it is
+                            # almost always a weak open model echoing tool-call
+                            # XML/JSON it saw in file or tool output (#47967:
+                            # <tool_call>/<invoke name=...> payloads in a file
+                            # prime mimo/nemotron-class models to emit empty
+                            # structured calls). Dumping the full tool catalog
+                            # in that case feeds the priming loop more names to
+                            # mimic and inflates context 3-4x across retries, so
+                            # send a terse error that tells the model in-context
+                            # tool-call syntax is DATA, not a call to make.
+                            if not (_tc_name or "").strip():
+                                content = (
+                                    "Tool call rejected: the tool name was empty. "
+                                    "If tool-call XML or JSON appeared in file "
+                                    "contents or tool output, that is data — do "
+                                    "not re-emit it as a tool call. To call a "
+                                    "tool, use a valid name from your tool list; "
+                                    "otherwise reply in plain text."
+                                )
+                            else:
+                                content = f"Tool '{_tc_name}' does not exist. Available tools: {available}"
                         else:
                             content = "Skipped: another tool call in this turn used an invalid name. Please retry this tool call."
                         messages.append({

agent/model_metadata.py

@@ -275,6 +275,10 @@ def _save_model_metadata_disk_cache(data: Dict[str, Dict[str, Any]]) -> None:
     # via a custom provider. Values sourced from models.dev (2026-04).
     # Keys use substring matching (longest-first), so e.g. "grok-4.20"
     # matches "grok-4.20-0309-reasoning" / "-non-reasoning" / "-multi-agent-0309".
+    # OAuth-only slug; absent from GET /v1/models. xAI publishes a 200k
+    # usable context window for Composer 2.5 on Grok Build (SuperGrok /
+    # Premium+); /v1/responses additionally enforces a ~262144 input+output
+    # budget, but the usable context (what we track here) is 200k.
     "grok-composer": 200000,    # grok-composer-2.5-fast (Grok Build CLI)
     "grok-build": 256000,       # grok-build-0.1
     "grok-code-fast": 256000,   # grok-code-fast-1

agent/transports/codex.py

@@ -128,6 +128,65 @@ def build_kwargs(
         reasoning_effort = _effort_clamp.get(reasoning_effort, reasoning_effort)
 
         response_tools = _responses_tools(tools)
+
+        # xAI server-side web search.
+        #
+        # grok models on xAI's /v1/responses surface (notably
+        # grok-composer-2.5-fast on SuperGrok OAuth) have a *native*,
+        # server-executed web search.  When the model is handed a
+        # client-side function literally named ``web_search``, it routes
+        # the intent to that native engine — but because the tool is
+        # declared as a plain ``function`` rather than xAI's first-class
+        # ``{"type": "web_search"}`` built-in, the server-side search is
+        # dispatched but never reconciled: the response streams reasoning
+        # + ``web_search_call`` progress items, the searches never reach
+        # ``status="completed"`` in the assembled output, no final
+        # message is emitted, and ``_normalize_codex_response`` correctly
+        # sees reasoning-with-no-answer and reports ``incomplete``.  The
+        # turn then burns 3 continuation retries and fails with "Codex
+        # response remained incomplete after 3 continuation attempts".
+        # Verified live against grok-composer-2.5-fast (2026-06).
+        #
+        # Fix: when the agent HAS a client-side ``web_search`` function (i.e.
+        # the user enabled the web toolset), declare xAI's native
+        # ``web_search`` built-in instead so the search actually runs to
+        # completion server-side and the model streams a real answer.  The
+        # Responses API rejects two tools sharing the name ``web_search``
+        # (HTTP 400 "Duplicate tool names"), so we drop the client-side
+        # ``web_search`` function for the xAI path and let the native tool
+        # satisfy it.  All other client-side tools (read_file, terminal,
+        # web_extract, MCP tools, …) are untouched and continue to dispatch
+        # through Hermes's agent loop.
+        #
+        # Scope: we ONLY swap in the native built-in when the client
+        # ``web_search`` was actually present.  We do NOT force-enable Grok
+        # server-side search on turns where the user never had web enabled —
+        # that would silently route around Hermes's web-provider config and
+        # tool-trace/citation plumbing for every xai-oauth turn.  The swap is
+        # a 1:1 replacement of an already-requested capability, not an
+        # additive grant.
+        #
+        # NOTE: for the swapped case this routes ``web_search`` to Grok's
+        # native search engine for xAI sessions instead of Hermes's
+        # configured web provider (Tavily/etc.), and those results bypass
+        # Hermes's tool-trace / citation plumbing (they arrive baked into the
+        # model's answer rather than as a tool result the loop observes).
+        # Scoped to ``is_xai_responses`` deliberately; narrow to specific
+        # models if a future grok variant should keep the client-side
+        # function.
+        if is_xai_responses and response_tools:
+            has_client_web_search = any(
+                isinstance(t, dict) and t.get("name") == "web_search"
+                for t in response_tools
+            )
+            if has_client_web_search:
+                filtered = [
+                    t for t in response_tools
+                    if not (isinstance(t, dict) and t.get("name") == "web_search")
+                ]
+                filtered.append({"type": "web_search"})
+                response_tools = filtered
+
         # ``tools`` MUST be omitted entirely when there are no functions to
         # expose: the openai SDK's ``responses.stream()`` / ``responses.parse()``
         # eagerly call ``_make_tools(tools)`` which does ``for tool in tools``

apps/bootstrap-installer/src-tauri/src/update.rs

@@ -286,7 +286,7 @@ async fn run_update(app: AppHandle) -> Result<()> {
     emit_stage(&app, "rebuild", StageState::Running, None, None);
     let started = Instant::now();
     let rebuild_args: Vec<String> = vec!["desktop".into(), "--build-only".into()];
-    let rebuild = run_streamed(
+    let mut rebuild = run_streamed(
         &app,
         &hermes,
         &rebuild_args,
@@ -295,6 +295,33 @@ async fn run_update(app: AppHandle) -> Result<()> {
         Some("rebuild"),
     )
     .await?;
+
+    // Retry-once: the first `--build-only` can return nonzero on a still-settling
+    // post-update tree or a network-blocked Electron fetch that our self-heal
+    // repaired mid-run. A second attempt then builds clean off the healed dist
+    // (the content-hash stamp makes it a near-no-op when the first actually
+    // succeeded). Without this the updater bails here and never reaches the
+    // relaunch below — the app updates but doesn't restart. Matches the
+    // retry-once `hermes update` already does above, and `hermes update`'s own
+    // desktop rebuild in cmd_update.
+    if rebuild_needs_retry(rebuild.exit_code) {
+        emit_log(
+            &app,
+            Some("rebuild"),
+            LogStream::Stdout,
+            "[rebuild] first desktop rebuild failed; retrying once (a self-healed \
+             Electron download builds clean on the second run)…",
+        );
+        rebuild = run_streamed(
+            &app,
+            &hermes,
+            &rebuild_args,
+            &install_root,
+            &child_env,
+            Some("rebuild"),
+        )
+        .await?;
+    }
     let rebuild_ms = started.elapsed().as_millis() as u64;
 
     if rebuild.exit_code != Some(0) {
@@ -533,6 +560,14 @@ fn is_locked(path: &Path) -> bool {
     }
 }
 
+/// Whether the `desktop --build-only` rebuild should be retried once. Any
+/// non-success exit qualifies: the common cause is a transient first-attempt
+/// failure (still-settling tree / self-healed Electron download) that a clean
+/// second run resolves.
+fn rebuild_needs_retry(exit_code: Option<i32>) -> bool {
+    exit_code != Some(0)
+}
+
 /// Spawn `hermes <args>` from `cwd`, stream stdout/stderr as Log events on the
 /// bootstrap channel, and return the exit code. Mirrors powershell::run_script
 /// but for an arbitrary command (no install.ps1 -File wrapping).
@@ -970,6 +1005,16 @@ mod tests {
         assert_eq!(update_branch_from_args(["--update"]), None);
     }
 
+    #[test]
+    fn rebuild_retries_only_on_failure() {
+        assert!(!rebuild_needs_retry(Some(0)), "a clean rebuild must not retry");
+        assert!(rebuild_needs_retry(Some(1)), "a failed rebuild retries once");
+        assert!(
+            rebuild_needs_retry(None),
+            "a killed/signalled rebuild (no exit code) retries once"
+        );
+    }
+
     #[test]
     fn parses_only_app_targets() {
         assert_eq!(