codex-autorunner 0.1.2__tar.gz → 1.0.0__tar.gz

{codex_autorunner-0.1.2/src/codex_autorunner.egg-info → codex_autorunner-1.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codex-autorunner
-Version: 0.1.2
+Version: 1.0.0
 Summary: Codex autorunner CLI per DESIGN-V1
 Author: Codex
 License: MIT License
@@ -65,11 +65,17 @@ Dynamic: license-file
 # codex-autorunner
 [![PyPI](https://img.shields.io/pypi/v/codex-autorunner.svg)](https://pypi.org/project/codex-autorunner/)
 
-An opinionated autorunner that uses the Codex CLI to work on large tasks via a simple loop. On each loop we feed the Codex instance the last one's final output along with core documents.
-1. TODO - Tracks long-horizon tasks
-2. PROGRESS - High level overview of what's been done already that may be relevant for future agents
-3. OPINIONS - Guidelines for how we should approach implementation
-4. SPEC - Source-of-truth requirements and scope for large features/projects
+An opinionated autorunner that uses the Codex app-server as the primary execution backend with OpenCode support to work on large tasks via a simple loop. On each loop we feed the Codex instance the last one's final output along with core documents.
+
+In the current model, the primary work surface is **tickets**:
+
+- `.codex-autorunner/tickets/TICKET-###.md`
+
+Optionally, you can maintain lightweight **workspace docs** (auto-created on write; missing is OK):
+
+- `.codex-autorunner/workspace/active_context.md`
+- `.codex-autorunner/workspace/decisions.md`
+- `.codex-autorunner/workspace/spec.md`
 
 ## Sneak Peak
 Run multiple agents on many repositories, with git worktree support
@@ -81,8 +87,7 @@ See the progress of your long running tasks with a high level overview
 Dive deep into specific agent execution with a rich but readable log
 ![Desktop logs](docs/screenshots/car-desktop-logs.png)
 
-All memory and opinions are markdown files! Edit them directly or chat with the document!
-![Desktop TODO](docs/screenshots/car-desktop-todo.png)
+Tickets and workspace docs are markdown files. Edit them directly or use the web UI’s file chat to iterate with the agent.
 
 Use codex CLI directly for multi-shot problem solving or `/review`
 ![Desktop terminal](docs/screenshots/car-desktop-terminal.png)
@@ -92,11 +97,10 @@ Mobile-first experience, code on the go with Whisper support (BYOK)
 
 ## What it does
 - Initializes a repo with Codex-friendly docs and config.
-- Runs Codex in a loop against the repo, streaming logs.
+- Runs Codex app-server in a loop against the repo, streaming logs via OpenCode runtime.
 - Tracks state, logs, and config under `.codex-autorunner/`.
-- Exposes a power-user HTTP API and web UI for docs, logs, runner control, and a Codex TUI terminal.
+- Exposes a power-user HTTP API and web UI for tickets, workspace docs, file chat, logs, runner control, and a Codex TUI terminal.
 - Optionally runs a Telegram bot for interactive, user-in-the-loop Codex sessions.
-- Generates a pasteable repo snapshot (`.codex-autorunner/SNAPSHOT.md`) for sharing with other LLM chats.
 
 CLI commands are available as `codex-autorunner` or the shorter `car`.
 
@@ -166,7 +170,7 @@ If you set a base path, prefix all checks with it.
 
 ## Quick start
 1) Install (editable): `pip install -e .`
-2) Initialize (hub + repo): `codex-autorunner init --git-init` (or `car init --git-init` if you prefer short). This creates the hub config at `.codex-autorunner/config.yml`, plus state/log files and the docs under `.codex-autorunner/`.
+2) Initialize (hub + repo): `codex-autorunner init --git-init` (or `car init --git-init` if you prefer short). This creates the hub config at `.codex-autorunner/config.yml`, plus state/log files and starter content under `.codex-autorunner/` (tickets and optional workspace docs).
 3) Run once: `codex-autorunner once` / `car once`
 4) Continuous loop: `codex-autorunner run` / `car run`
 5) If stuck: `codex-autorunner kill` then `codex-autorunner resume` (or the `car` equivalents)
@@ -231,19 +235,12 @@ If you set `server.auth_token_env`, the API requires `Authorization: Bearer <tok
 - `run` / `once` — run the loop (continuous or single iteration).
 - `resume` — clear stale lock/state and restart; `--once` for a single run.
 - `kill` — SIGTERM the running loop and mark state error.
-- `status` — show current state and outstanding TODO count.
+- `status` — show current state.
 - `sessions` — list terminal sessions (server-backed when available).
 - `stop-session` — stop a terminal session by repo (`--repo`) or id (`--session`).
 - `log` — view logs (tail or specific run).
-- `edit` — open TODO/PROGRESS/OPINIONS/SPEC in `$EDITOR`.
-- `ingest-spec` — generate TODO/PROGRESS/OPINIONS from SPEC using Codex (use `--force` to overwrite).
-- `clear-docs` — reset TODO/PROGRESS/OPINIONS to empty templates (type CLEAR to confirm).
-- `snapshot` — generate/update `.codex-autorunner/SNAPSHOT.md` (incremental by default when one exists; use `--from-scratch` to regenerate).
+- `edit` — open `active_context|decisions|spec` in `$EDITOR`.
 - `serve` — start the HTTP API (FastAPI) on host/port from config (defaults 127.0.0.1:4173).
 
-## Snapshot (repo briefing)
-- Web UI: open the Snapshot tab. If no snapshot exists, you’ll see “Generate snapshot”; otherwise you’ll see “Update snapshot (incremental)” and “Regenerate snapshot (from scratch)”, plus “Copy to clipboard”.
-- CLI: `codex-autorunner snapshot` (or `car snapshot`) writes `.codex-autorunner/SNAPSHOT.md` and `.codex-autorunner/snapshot_state.json`.
-
 ## Star history
 [![Star History Chart](https://api.star-history.com/svg?repos=Git-on-my-level/codex-autorunner&type=Date)](https://star-history.com/#Git-on-my-level/codex-autorunner&Date)

{codex_autorunner-0.1.2 → codex_autorunner-1.0.0}/README.md

@@ -1,11 +1,17 @@
 # codex-autorunner
 [![PyPI](https://img.shields.io/pypi/v/codex-autorunner.svg)](https://pypi.org/project/codex-autorunner/)
 
-An opinionated autorunner that uses the Codex CLI to work on large tasks via a simple loop. On each loop we feed the Codex instance the last one's final output along with core documents.
-1. TODO - Tracks long-horizon tasks
-2. PROGRESS - High level overview of what's been done already that may be relevant for future agents
-3. OPINIONS - Guidelines for how we should approach implementation
-4. SPEC - Source-of-truth requirements and scope for large features/projects
+An opinionated autorunner that uses the Codex app-server as the primary execution backend with OpenCode support to work on large tasks via a simple loop. On each loop we feed the Codex instance the last one's final output along with core documents.
+
+In the current model, the primary work surface is **tickets**:
+
+- `.codex-autorunner/tickets/TICKET-###.md`
+
+Optionally, you can maintain lightweight **workspace docs** (auto-created on write; missing is OK):
+
+- `.codex-autorunner/workspace/active_context.md`
+- `.codex-autorunner/workspace/decisions.md`
+- `.codex-autorunner/workspace/spec.md`
 
 ## Sneak Peak
 Run multiple agents on many repositories, with git worktree support
@@ -17,8 +23,7 @@ See the progress of your long running tasks with a high level overview
 Dive deep into specific agent execution with a rich but readable log
 ![Desktop logs](docs/screenshots/car-desktop-logs.png)
 
-All memory and opinions are markdown files! Edit them directly or chat with the document!
-![Desktop TODO](docs/screenshots/car-desktop-todo.png)
+Tickets and workspace docs are markdown files. Edit them directly or use the web UI’s file chat to iterate with the agent.
 
 Use codex CLI directly for multi-shot problem solving or `/review`
 ![Desktop terminal](docs/screenshots/car-desktop-terminal.png)
@@ -28,11 +33,10 @@ Mobile-first experience, code on the go with Whisper support (BYOK)
 
 ## What it does
 - Initializes a repo with Codex-friendly docs and config.
-- Runs Codex in a loop against the repo, streaming logs.
+- Runs Codex app-server in a loop against the repo, streaming logs via OpenCode runtime.
 - Tracks state, logs, and config under `.codex-autorunner/`.
-- Exposes a power-user HTTP API and web UI for docs, logs, runner control, and a Codex TUI terminal.
+- Exposes a power-user HTTP API and web UI for tickets, workspace docs, file chat, logs, runner control, and a Codex TUI terminal.
 - Optionally runs a Telegram bot for interactive, user-in-the-loop Codex sessions.
-- Generates a pasteable repo snapshot (`.codex-autorunner/SNAPSHOT.md`) for sharing with other LLM chats.
 
 CLI commands are available as `codex-autorunner` or the shorter `car`.
 
@@ -102,7 +106,7 @@ If you set a base path, prefix all checks with it.
 
 ## Quick start
 1) Install (editable): `pip install -e .`
-2) Initialize (hub + repo): `codex-autorunner init --git-init` (or `car init --git-init` if you prefer short). This creates the hub config at `.codex-autorunner/config.yml`, plus state/log files and the docs under `.codex-autorunner/`.
+2) Initialize (hub + repo): `codex-autorunner init --git-init` (or `car init --git-init` if you prefer short). This creates the hub config at `.codex-autorunner/config.yml`, plus state/log files and starter content under `.codex-autorunner/` (tickets and optional workspace docs).
 3) Run once: `codex-autorunner once` / `car once`
 4) Continuous loop: `codex-autorunner run` / `car run`
 5) If stuck: `codex-autorunner kill` then `codex-autorunner resume` (or the `car` equivalents)
@@ -167,19 +171,12 @@ If you set `server.auth_token_env`, the API requires `Authorization: Bearer <tok
 - `run` / `once` — run the loop (continuous or single iteration).
 - `resume` — clear stale lock/state and restart; `--once` for a single run.
 - `kill` — SIGTERM the running loop and mark state error.
-- `status` — show current state and outstanding TODO count.
+- `status` — show current state.
 - `sessions` — list terminal sessions (server-backed when available).
 - `stop-session` — stop a terminal session by repo (`--repo`) or id (`--session`).
 - `log` — view logs (tail or specific run).
-- `edit` — open TODO/PROGRESS/OPINIONS/SPEC in `$EDITOR`.
-- `ingest-spec` — generate TODO/PROGRESS/OPINIONS from SPEC using Codex (use `--force` to overwrite).
-- `clear-docs` — reset TODO/PROGRESS/OPINIONS to empty templates (type CLEAR to confirm).
-- `snapshot` — generate/update `.codex-autorunner/SNAPSHOT.md` (incremental by default when one exists; use `--from-scratch` to regenerate).
+- `edit` — open `active_context|decisions|spec` in `$EDITOR`.
 - `serve` — start the HTTP API (FastAPI) on host/port from config (defaults 127.0.0.1:4173).
 
-## Snapshot (repo briefing)
-- Web UI: open the Snapshot tab. If no snapshot exists, you’ll see “Generate snapshot”; otherwise you’ll see “Update snapshot (incremental)” and “Regenerate snapshot (from scratch)”, plus “Copy to clipboard”.
-- CLI: `codex-autorunner snapshot` (or `car snapshot`) writes `.codex-autorunner/SNAPSHOT.md` and `.codex-autorunner/snapshot_state.json`.
-
 ## Star history
 [![Star History Chart](https://api.star-history.com/svg?repos=Git-on-my-level/codex-autorunner&type=Date)](https://star-history.com/#Git-on-my-level/codex-autorunner&Date)

{codex_autorunner-0.1.2 → codex_autorunner-1.0.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "codex-autorunner"
-version = "0.1.2"
+version = "1.0.0"
 description = "Codex autorunner CLI per DESIGN-V1"
 readme = "README.md"
 license = {file = "LICENSE"}
@@ -86,6 +86,11 @@ markers = [
     "integration: marks tests as integration tests that require external dependencies (deselect with '-m \"not integration\"')",
     "timeout(seconds): per-test timeout (provided by pytest-timeout)",
 ]
+filterwarnings = [
+    # Cancelled auto-restart tasks can emit noisy unraisable warnings when the event
+    # loop shuts down; we explicitly clean these up, so silence the residual notice.
+    "ignore::pytest.PytestUnraisableExceptionWarning",
+]
 
 [tool.ruff]
 line-length = 88

codex_autorunner-1.0.0/src/codex_autorunner/__main__.py

@@ -0,0 +1,4 @@
+from .cli import app
+
+if __name__ == "__main__":
+    app()

{codex_autorunner-0.1.2 → codex_autorunner-1.0.0}/src/codex_autorunner/agents/opencode/client.py

@@ -4,6 +4,7 @@ import asyncio
 import dataclasses
 import json
 import logging
+import re
 from typing import Any, AsyncIterator, Iterable, Optional
 
 import httpx
@@ -48,7 +49,18 @@ def _normalize_sse_event(event: SSEEvent) -> SSEEvent:
         payload_obj = None
 
     if isinstance(payload_obj, dict) and isinstance(payload_obj.get("payload"), dict):
-        payload_obj = payload_obj["payload"]
+        outer = payload_obj
+        inner = dict(outer.get("payload") or {})
+        if "type" not in inner and isinstance(outer.get("type"), str):
+            inner["type"] = outer["type"]
+        for key in ("sessionID", "sessionId", "session_id"):
+            if key in outer and key not in inner:
+                inner[key] = outer[key]
+        if "session" in outer and "session" not in inner:
+            inner["session"] = outer["session"]
+        if "properties" in outer and "properties" not in inner:
+            inner["properties"] = outer["properties"]
+        payload_obj = inner
 
     if isinstance(payload_obj, dict):
         payload_type = payload_obj.get("type")
@@ -505,49 +517,70 @@ class OpenCodeClient:
 
     async def fetch_openapi_spec(self) -> dict[str, Any]:
         """Fetch OpenAPI spec from /doc endpoint for capability negotiation."""
-        async with self._client.stream("GET", "/doc") as response:
-            response.raise_for_status()
-            content = response.content
-            try:
-                spec = json.loads(content) if content else {}
-                log_event(
-                    self._logger,
-                    logging.INFO,
-                    "opencode.openapi.fetched",
-                    paths=len(spec.get("paths", {})) if isinstance(spec, dict) else 0,
-                    has_components=(
-                        "components" in spec if isinstance(spec, dict) else False
-                    ),
-                )
-                return spec
-            except Exception as exc:
-                log_event(
-                    self._logger,
-                    logging.WARNING,
-                    "opencode.openapi.parse_failed",
-                    exc=exc,
-                )
-                raise OpenCodeProtocolError(
-                    f"Failed to parse OpenAPI spec: {exc}",
-                    status_code=response.status_code,
-                    content_type=(
-                        response.headers.get("content-type") if response else None
-                    ),
-                ) from exc
+        response = await self._client.get("/doc")
+        response.raise_for_status()
+        content = response.content
+        try:
+            spec = json.loads(content) if content else {}
+            log_event(
+                self._logger,
+                logging.INFO,
+                "opencode.openapi.fetched",
+                paths=len(spec.get("paths", {})) if isinstance(spec, dict) else 0,
+                has_components=(
+                    "components" in spec if isinstance(spec, dict) else False
+                ),
+            )
+            return spec
+        except Exception as exc:
+            log_event(
+                self._logger,
+                logging.WARNING,
+                "opencode.openapi.parse_failed",
+                exc=exc,
+            )
+            raise OpenCodeProtocolError(
+                f"Failed to parse OpenAPI spec: {exc}",
+                status_code=response.status_code,
+                content_type=(
+                    response.headers.get("content-type") if response else None
+                ),
+            ) from exc
 
     def has_endpoint(
         self, openapi_spec: dict[str, Any], method: str, path: str
     ) -> bool:
-        """Check if endpoint is available in OpenAPI spec."""
+        """Check if endpoint is available in OpenAPI spec.
+
+        The OpenAPI spec sometimes uses different template parameter names (e.g.,
+        `{sessionID}` vs `{session_id}`). We normalize templates before matching so
+        capability detection does not depend on placeholder spelling.
+        """
         if not isinstance(openapi_spec, dict):
             return False
         paths = openapi_spec.get("paths", {})
         if not isinstance(paths, dict):
             return False
-        path_info = paths.get(path)
-        if not isinstance(path_info, dict):
-            return False
-        return method in path_info
+
+        target = _normalize_template_path(path)
+        method = method.lower()
+
+        for candidate_path, info in paths.items():
+            if not isinstance(info, dict):
+                continue
+            if _normalize_template_path(candidate_path) != target:
+                continue
+            if method in info:
+                return True
+        return False
+
+
+def _normalize_template_path(path: str) -> str:
+    """Collapse template placeholders to a canonical form.
+
+    Example: `/session/{sessionID}/prompt_async` -> `/session/{}/prompt_async`
+    """
+    return re.sub(r"{[^/]+}", "{}", path)
 
 
 __all__ = ["OpenCodeClient", "OpenCodeProtocolError", "OpenCodeApiProfile"]

{codex_autorunner-0.1.2 → codex_autorunner-1.0.0}/src/codex_autorunner/agents/opencode/logging.py

@@ -66,11 +66,6 @@ class OpenCodeEventFormatter:
             for line in complete_lines:
                 if line.strip():
                     lines.append(f"**{line.strip()}**")
-
-            remaining = coalescer.get_buffer()
-            if remaining and remaining.strip():
-                lines.append(f"**{remaining.strip()}**")
-                coalescer.clear()
         return lines
 
     def _format_tool_part(self, part: dict[str, Any]) -> list[str]:
@@ -120,6 +115,27 @@ class OpenCodeEventFormatter:
             lines.append("exec")
             lines.append(f"tool: {tool_name}")
 
+        input_preview: Optional[str] = None
+        for key in ("input", "command", "cmd", "script"):
+            value = part.get(key)
+            if isinstance(value, str) and value.strip():
+                input_preview = value.strip()
+                break
+        if input_preview is None:
+            args = part.get("args") or part.get("arguments") or part.get("params")
+            if isinstance(args, dict):
+                for key in ("command", "cmd", "script", "input"):
+                    value = args.get(key)
+                    if isinstance(value, str) and value.strip():
+                        input_preview = value.strip()
+                        break
+            elif isinstance(args, str) and args.strip():
+                input_preview = args.strip()
+        if input_preview:
+            if len(input_preview) > 240:
+                input_preview = input_preview[:240] + "…"
+            lines.append(f"input: {input_preview}")
+
         return lines
 
     def _format_patch_part(self, part: dict[str, Any]) -> list[str]:

{codex_autorunner-0.1.2 → codex_autorunner-1.0.0}/src/codex_autorunner/agents/opencode/run_prompt.py

@@ -112,6 +112,7 @@ async def run_opencode_prompt(
             client,
             session_id=session_id,
             workspace_path=config.workspace_root,
+            model_payload=model_payload,
             permission_policy=permission_policy,
             should_stop=should_stop,
             ready_event=ready_event,

{codex_autorunner-0.1.2 → codex_autorunner-1.0.0}/src/codex_autorunner/agents/opencode/runtime.py

@@ -15,6 +15,7 @@ from typing import (
     Callable,
     MutableMapping,
     Optional,
+    cast,
 )
 
 import httpx
@@ -127,17 +128,19 @@ def extract_session_id(
             return value
     properties = payload.get("properties")
     if isinstance(properties, dict):
-        value = properties.get("sessionID")
-        if isinstance(value, str) and value:
-            return value
-        part = properties.get("part")
-        if isinstance(part, dict):
-            value = part.get("sessionID")
+        for key in ("sessionID", "sessionId", "session_id"):
+            value = properties.get(key)
             if isinstance(value, str) and value:
                 return value
+        part = properties.get("part")
+        if isinstance(part, dict):
+            for key in ("sessionID", "sessionId", "session_id"):
+                value = part.get(key)
+                if isinstance(value, str) and value:
+                    return value
     session = payload.get("session")
     if isinstance(session, dict):
-        return extract_session_id(session, allow_fallback_id=allow_fallback_id)
+        return extract_session_id(session, allow_fallback_id=True)
     return None
 
 
@@ -676,7 +679,7 @@ async def opencode_missing_env(
         providers = [entry for entry in payload if isinstance(entry, dict)]
     for provider in providers:
         pid = provider.get("id") or provider.get("providerID")
-        if pid != provider_id:
+        if not pid or pid != provider_id:
             continue
         if _provider_has_auth(pid, workspace_root):
             return []
@@ -732,6 +735,7 @@ async def collect_opencode_output_from_events(
     events: Optional[AsyncIterator[SSEEvent]] = None,
     *,
     session_id: str,
+    model_payload: Optional[dict[str, str]] = None,
     progress_session_ids: Optional[set[str]] = None,
     permission_policy: str = PERMISSION_ALLOW,
     permission_handler: Optional[PermissionHandler] = None,
@@ -759,13 +763,16 @@ async def collect_opencode_output_from_events(
     last_usage_total: Optional[int] = None
     last_context_window: Optional[int] = None
     part_types: dict[str, str] = {}
-    seen_question_request_ids: set[tuple[str, str]] = set()
+    seen_question_request_ids: set[tuple[Optional[str], str]] = set()
     logged_permission_errors: set[str] = set()
     normalized_question_policy = _normalize_question_policy(question_policy)
     logger = logging.getLogger(__name__)
     providers_cache: Optional[list[dict[str, Any]]] = None
     context_window_cache: dict[str, Optional[int]] = {}
     session_model_ids: Optional[tuple[Optional[str], Optional[str]]] = None
+    default_model_ids = (
+        _extract_model_ids(model_payload) if isinstance(model_payload, dict) else None
+    )
 
     def _message_id_from_info(info: Any) -> Optional[str]:
         if not isinstance(info, dict):
@@ -843,15 +850,19 @@ async def collect_opencode_output_from_events(
         nonlocal session_model_ids
         if session_model_ids is not None:
             return session_model_ids
-        if session_fetcher is None:
-            session_model_ids = (None, None)
-            return session_model_ids
-        try:
-            payload = await session_fetcher()
-        except Exception:
-            session_model_ids = (None, None)
-            return session_model_ids
-        session_model_ids = _extract_model_ids(payload)
+        resolved_ids: Optional[tuple[Optional[str], Optional[str]]] = None
+        if session_fetcher is not None:
+            try:
+                payload = await session_fetcher()
+                resolved_ids = _extract_model_ids(payload)
+            except Exception:
+                resolved_ids = None
+        # If we failed to resolve model ids from the session (including the empty
+        # tuple case), fall back to the caller-provided model payload so we can
+        # still backfill usage metadata.
+        if not resolved_ids or all(value is None for value in resolved_ids):
+            resolved_ids = default_model_ids
+        session_model_ids = resolved_ids or (None, None)
         return session_model_ids
 
     async def _resolve_context_window_from_providers(
@@ -946,7 +957,7 @@ async def collect_opencode_output_from_events(
             await aclose()
 
     stream_iter = _new_stream().__aiter__()
-    last_event_at = time.monotonic()
+    last_relevant_event_at = time.monotonic()
     last_primary_completion_at: Optional[float] = None
     reconnect_attempts = 0
     can_reconnect = (
@@ -981,6 +992,7 @@ async def collect_opencode_output_from_events(
                             session_id=session_id,
                             exc=exc,
                         )
+                idle_seconds = now - last_relevant_event_at
                 if _status_is_idle(status_type):
                     log_event(
                         logger,
@@ -988,7 +1000,7 @@ async def collect_opencode_output_from_events(
                         "opencode.stream.stalled.session_idle",
                         session_id=session_id,
                         status_type=status_type,
-                        idle_seconds=now - last_event_at,
+                        idle_seconds=idle_seconds,
                     )
                     if not text_parts and pending_text:
                         _flush_all_pending_text()
@@ -1000,7 +1012,7 @@ async def collect_opencode_output_from_events(
                         "opencode.stream.stalled.after_completion",
                         session_id=session_id,
                         status_type=status_type,
-                        idle_seconds=now - last_event_at,
+                        idle_seconds=idle_seconds,
                     )
                 if not can_reconnect:
                     break
@@ -1015,7 +1027,7 @@ async def collect_opencode_output_from_events(
                     logging.WARNING,
                     "opencode.stream.stalled.reconnecting",
                     session_id=session_id,
-                    idle_seconds=now - last_event_at,
+                    idle_seconds=idle_seconds,
                     backoff_seconds=backoff,
                     status_type=status_type,
                     attempts=reconnect_attempts,
@@ -1023,21 +1035,86 @@ async def collect_opencode_output_from_events(
                 await _close_stream(stream_iter)
                 await asyncio.sleep(backoff)
                 stream_iter = _new_stream().__aiter__()
+                last_relevant_event_at = now
                 continue
-            last_event_at = time.monotonic()
+            now = time.monotonic()
             raw = event.data or ""
             try:
                 payload = json.loads(raw) if raw else {}
             except json.JSONDecodeError:
                 payload = {}
             event_session_id = extract_session_id(payload)
-            if not event_session_id:
-                continue
-            if progress_session_ids is None:
-                if event_session_id != session_id:
-                    continue
-            elif event_session_id not in progress_session_ids:
+            is_relevant = False
+            if event_session_id:
+                if progress_session_ids is None:
+                    is_relevant = event_session_id == session_id
+                else:
+                    is_relevant = event_session_id in progress_session_ids
+            if not is_relevant:
+                if (
+                    stall_timeout_seconds is not None
+                    and now - last_relevant_event_at > stall_timeout_seconds
+                ):
+                    idle_seconds = now - last_relevant_event_at
+                    last_relevant_event_at = now
+                    status_type = None
+                    if session_fetcher is not None:
+                        try:
+                            payload = await session_fetcher()
+                            status_type = _extract_status_type(payload)
+                        except Exception as exc:
+                            log_event(
+                                logger,
+                                logging.WARNING,
+                                "opencode.session.poll_failed",
+                                session_id=session_id,
+                                exc=exc,
+                            )
+                    if _status_is_idle(status_type):
+                        log_event(
+                            logger,
+                            logging.INFO,
+                            "opencode.stream.stalled.session_idle",
+                            session_id=session_id,
+                            status_type=status_type,
+                            idle_seconds=idle_seconds,
+                        )
+                        if not text_parts and pending_text:
+                            _flush_all_pending_text()
+                        break
+                    if last_primary_completion_at is not None:
+                        log_event(
+                            logger,
+                            logging.INFO,
+                            "opencode.stream.stalled.after_completion",
+                            session_id=session_id,
+                            status_type=status_type,
+                            idle_seconds=idle_seconds,
+                        )
+                    if not can_reconnect:
+                        break
+                    backoff_index = min(
+                        reconnect_attempts,
+                        len(_OPENCODE_STREAM_RECONNECT_BACKOFF_SECONDS) - 1,
+                    )
+                    backoff = _OPENCODE_STREAM_RECONNECT_BACKOFF_SECONDS[backoff_index]
+                    reconnect_attempts += 1
+                    log_event(
+                        logger,
+                        logging.WARNING,
+                        "opencode.stream.stalled.reconnecting",
+                        session_id=session_id,
+                        idle_seconds=idle_seconds,
+                        backoff_seconds=backoff,
+                        status_type=status_type,
+                        attempts=reconnect_attempts,
+                    )
+                    await _close_stream(stream_iter)
+                    await asyncio.sleep(backoff)
+                    stream_iter = _new_stream().__aiter__()
                 continue
+            last_relevant_event_at = now
+            reconnect_attempts = 0
             is_primary_session = event_session_id == session_id
             if event.event == "question.asked":
                 request_id, props = _extract_question_request(payload)
@@ -1419,6 +1496,7 @@ async def collect_opencode_output(
     *,
     session_id: str,
     workspace_path: str,
+    model_payload: Optional[dict[str, str]] = None,
     progress_session_ids: Optional[set[str]] = None,
     permission_policy: str = PERMISSION_ALLOW,
     permission_handler: Optional[PermissionHandler] = None,
@@ -1427,6 +1505,7 @@ async def collect_opencode_output(
     should_stop: Optional[Callable[[], bool]] = None,
     ready_event: Optional[Any] = None,
     part_handler: Optional[PartHandler] = None,
+    stall_timeout_seconds: Optional[float] = _OPENCODE_STREAM_STALL_TIMEOUT_SECONDS,
 ) -> OpenCodeTurnOutput:
     async def _respond(request_id: str, reply: str) -> None:
         await client.respond_permission(request_id=request_id, reply=reply)
@@ -1438,14 +1517,21 @@ async def collect_opencode_output(
         await client.reject_question(request_id)
 
     def _stream_factory() -> AsyncIterator[SSEEvent]:
-        return client.stream_events(directory=workspace_path, ready_event=ready_event)
+        return cast(
+            AsyncIterator[SSEEvent],
+            client.stream_events(directory=workspace_path, ready_event=ready_event),
+        )
 
     async def _fetch_session() -> Any:
         statuses = await client.session_status(directory=workspace_path)
         if isinstance(statuses, dict):
             session_status = statuses.get(session_id)
+            if session_status is None:
+                return {"status": {"type": "idle"}}
             if isinstance(session_status, dict):
                 return {"status": session_status}
+            if isinstance(session_status, str):
+                return {"status": session_status}
         return {"status": {}}
 
     async def _fetch_providers() -> Any:
@@ -1465,8 +1551,10 @@ async def collect_opencode_output(
         reject_question=_reject_question,
         part_handler=part_handler,
         event_stream_factory=_stream_factory,
+        model_payload=model_payload,
         session_fetcher=_fetch_session,
         provider_fetcher=_fetch_providers,
+        stall_timeout_seconds=stall_timeout_seconds,
     )