The container still exited immediately after the su-based fix. can
interfere with TTY/stdin handling for interactive shells. Switch to
, which is root-only, skips PAM, and preserves file descriptors so
bash stays interactive.
Also improve container failure diagnostics:
- now combines stdout and stderr
- This helps surface the real reason when a container exits with code 0
Quality gates:
- pytest tests/unit: 219 passed
- ruff: clean on changed files
- mypy: clean on changed files
compute_image_tag hashed only the manifest JSON, so cached images built
before the entrypoint fix were reused even though compile_entrypoint() now
produces a different entrypoint. This caused containers to keep using the
old (broken) entrypoint that exited immediately.
- Include a compiler_version token in the hash input so manifest compiler
changes invalidate previously built images
- This forces a fresh image build for new instances after any change to
compile_dockerfile, compile_entrypoint, or compile_compose
Quality gates:
- pytest tests/unit: 219 passed
- ruff: clean on changed files
- mypy: clean on changed files
The generated entrypoint used a brittle invocation that could
spawn a non-interactive bash and exit immediately, causing containers to
stop with exit code 0 right after startup.
- Switch to so the
container user gets a login shell and stdin/tty are preserved
- Update the unit test assertion for the new drop-privileges command
Quality gates:
- pytest tests/unit: 219 passed
- ruff: clean on changed files
- mypy: clean on changed files
Git clone with an explicit destination puts the working copy directly into
that directory; it does not create a repo-named subdirectory. The previous
change assumed the opposite, so workspaces ended up at
/data/working-copies/{workspace_id}/ with the repo contents mixed in,
causing a 500 when the expected repo-named subdirectory was missing.
- Build the target path as /data/working-copies/{workspace_id}/{repo_name}/
and pass it directly to GitService.clone
- Remove stale directory detection and fallback logic that is no longer
needed
- Keep diagnostic logging around git clone failures
Quality gates:
- pytest tests/unit: 219 passed
- mypy: clean on changed files
The workspace creation endpoint returns 500 but the actual error is not
visible. Add explicit error logging when GitService.clone fails and info
logging when git creates a directory name different from the one derived
from the remote URL.
Quality gates:
- pytest tests/unit: 219 passed
- mypy: clean on changed files
Working copies were stored as /data/working-copies/{repo_id}/{workspace_name}/,
so git clone was forced into a user-named directory. That meant the container
mount basename was the workspace name (e.g. main) instead of the repo name.
- Generate the workspace UUID before cloning and clone into
/data/working-copies/{workspace_id}/ so git creates {repo_name}/ naturally
- Set workspace.path to /data/working-copies/{workspace_id}/{repo_name}/
- Update _migrate_clone_into_workspace() to use the same layout
- _get_repository_mount_name() now prefers workspace.path basename and only
falls back to remote URL / repo.name for legacy repo-only instances
- Update unit tests to assert workspace path basename is used for mounts
Quality gates:
- pytest tests/unit: 219 passed
- ruff: clean on changed files
- mypy: clean on changed files
The pi-agent manifest still declared an explicit repo mount with
{{WORKSPACE_NAME}}, making the mount target dependent on tool config. The
instance service now synthesizes the repo mount, so the manifest no longer
needs the explicit mount.
- Add Alembic migration 2026_06_15_090500 to remove the source_type: repo
mount from the built-in pi-agent manifest
- Add _get_repository_mount_name() helper to derive the workspace directory
name from the repository remote URL (matching git clone behavior) and
fall back to the user-provided repository name
- Use the helper for WORKSPACE_NAME/REPO_NAME in manifest, legacy dockerfile,
and legacy compose template paths
- Update unit tests for the new migration and helper
Quality gates:
- pytest tests/unit: 218 passed
- ruff: clean on changed files
- mypy: clean on changed files
- alembic heads: single head
WORKSPACE_NAME was computed from os.path.basename(repo_path), so when a
workspace path ended in a directory like 'main', the container mount target
became /home/user/main instead of /home/user/{repo-name}.
- Use GitRepository.name for WORKSPACE_NAME/REPO_NAME in manifest and
legacy dockerfile flows
- Add unit test verifying prepare_manifest_instance uses repo.name even
when the workspace path basename differs
Quality gates:
- pytest tests/unit: 213 passed
- ruff: clean on changed files
- mypy: clean on changed files
Older cached images still contain a literal /home/user/{{WORKSPACE_NAME}}
directory baked in by the previous Dockerfile generation. Even though new
images no longer create it, existing images leave the placeholder folder
alongside the real repo-named mount.
- Add entrypoint cleanup that removes /{{WORKSPACE_NAME}} if it
exists before creating the real workspace target and /workspace symlink
- Update unit tests to assert the stale placeholder removal
Quality gates:
- pytest tests/unit: 212 passed
- ruff: clean on changed files
- mypy: clean on changed files
When a manifest mount target uses ~/{{WORKSPACE_NAME}}, the Dockerfile was
building a literal directory named {{WORKSPACE_NAME}} into the image and
creating a broken /workspace symlink. The runtime mount then created the
correct repo-named folder alongside the placeholder folder.
- Only create static mount target directories in the Dockerfile; skip any
target containing {{WORKSPACE_NAME}}
- Only create the /workspace compatibility symlink at image-build time when
the workspace name is known; otherwise let the entrypoint create it from
the WORKSPACE_NAME environment variable
- Update unit tests to cover both build-time workspace names and runtime
placeholders
Quality gates:
- pytest tests/unit: 211 passed
- ruff: clean on changed files
- mypy: clean on changed files
- Add final get_container_status check in start_tool_instance before
writing status=running; mark as error and return logs if container stopped
- Treat restarting as error in HealthMonitor when DB status was already
running, so crash loops are surfaced instead of preserved
- Disable auto-restart (restart: unless-stopped -> restart: no) for tool
instances in manifest compiler, legacy dockerfile path, and built-in seeds
Quality gates:
- pytest tests/unit: 210 passed
- ruff: clean on changed files
- mypy: clean on changed files
The compose file was forcing the container to run as uid 1001, so the
entrypoint could not create /workspace even with sudo configured.
- Remove Dockerfile USER directive so containers start as root
- Make compile_compose use user: 0:0 when the manifest declares a user
- Make the entrypoint drop to the container user via after setup,
preserving environment variables and command arguments
- Update unit tests to assert root startup and privilege drop
Quality gates:
- pytest tests/unit: 210 passed
- ruff: clean on changed files
- mypy: clean on changed files
The previous ordering checked SUDO before checking if the process was
already running as root. When Docker starts the container with a
non-root user, SUDO may be empty, but the real fix is that the
entrypoint should try root first (e.g. when the image is started as
root) and only then fall back to sudo.
- Reorder symlink creation logic: root first, then sudo, then best-effort
- Update unit test to assert root is checked before sudo
Quality gates:
- pytest tests/unit: 208 passed
- ruff: clean on changed files
- mypy: clean on changed files
The previous commit moved the pi-agent repo mount from /workspace to
/home/user/{repo_name}. This exposed a permission bug: the Dockerfile
creates /workspace as a root-owned symlink in the image, and the
non-root entrypoint could not replace it because / is owned by root.
- Update compile_entrypoint to recreate /workspace via sudo when running
as the container user, or directly when running as root
- Add unit test covering sudo/root symlink creation
- Update OpenSpec change docs with the additional root cause
Quality gates:
- pytest tests/unit: 208 passed
- ruff: clean on changed files
- mypy: clean on changed files
- alembic heads: single head
- Add Alembic migration to update built-in pi-agent manifest:
* repo mount target from /workspace to ~/{{WORKSPACE_NAME}}
* keep /workspace as compatibility symlink via working_dir
* update startup chown target to $HOME/$WORKSPACE_NAME
- Pass REPO_NAME and WORKSPACE_NAME to compile_compose from instance_service
- Substitute {{WORKSPACE_NAME}} in manifest mount targets and expose it as
a container env var so the entrypoint can create the /workspace symlink
- Generate entrypoint workspace symlink from runtime WORKSPACE_NAME env var
- Install npm_global packages into {home_dir}/.npm-global with PATH so the
non-root container user can update global packages
- Update manifest compiler unit tests for the new behavior
Quality gates:
- pytest tests/unit: 207 passed
- ruff: clean on changed files
- mypy: clean on changed files
- alembic heads: single head
- modify_compose_file now appends :ro when a config-profile volume entry
has readonly=true, matching the manifest compiler behavior.
- Add a guard for missing tool_type in prepare_manifest_instance.
Quality gates: python3 -m pytest (313 passed, 34 skipped)
- Switch apply_resolved_profile from per-file bind mounts to one
directory-level bind mount per ResolvedMount target.
- Stage all configured files under instance_dir/mounts/<sanitized_target>
and bind-mount that directory, so Docker no longer creates a root-owned
parent directory such as ~/.config.
- Propagate read-only mode ('ro') as the 'readonly' flag on volume entries.
- Update unit tests to expect directory-level mounts and add coverage for
readonly/writable flags.
Quality gates: python3 -m py_compile, pytest (313 passed, 34 skipped),
npm run typecheck, npm run lint.
Config-profile and git mounts staged under instance_dir were created
by the API process (root), so when bind-mounted over ~/.config the
container user could not write. Recursively chown staged sources to
the resolved container uid/gid before compose up.
Quality gates: python3 -m py_compile passed; ruff/pytest skipped
(test tooling not available in this shell, helper smoke tested
with a temporary directory).
Regenerate .pi-map*.md artifacts for source changes in:
- apps/api/src/api/tool/tool_types_validation.py
- apps/api/src/schemas/tool/tool_type.py
- apps/api/tests/integration/test_tool_types_api_extended.py
- and all affected test files from backend-frontend refactoring cleanup
- Remove clone_mode/branch from API responses and make DB columns nullable
- Remove legacy clone-mode branches from create_tool_instance
- Add WORKSPACE_PATH compose variable alongside REPO_PATH
- Add workspace migration helpers in WorkspaceManager
Remaining: POST /workspaces/:id/instances, frontend clone_mode cleanup, tests
Add generated .pi-map.md and .pi-map.index.md files across the repository
so the project navigation maps are shared and versioned. These artifacts
are maintained by project_map_init/patch/validate and must be kept in
sync with source edits.
Note: .cache/ remains ignored (added in previous commit).
Root cause: manifest-based Dockerfile created the home directory and
chowned only the home root. Files/directories copied from /etc/skel by
useradd -m (or created later by root) remained root-owned, so apps like
ranger failed when writing to ~/.config.
Changes:
- manifest_compiler.py: recursive chown of the home directory after
useradd so /etc/skel contents are owned by the container user
- Pre-create .config, .local/share, .cache and chown them to the user
so first-run apps have writable directories immediately
- Add unit test verifying the Dockerfile emits the expected user/home
setup and config directory creation
Quality gates: py_compile all backend files pass, test file compiles,
tsc --noEmit pass, npm run build pass, 82/82 web tests pass
Note: pytest not available in this shell; backend unit test was not
executed but follows existing project conventions.
Backend:
- sessions.py: include workspace_name in session response
- instance_service.py: auto-generate display names as
'Project / Workspace / Tool #N' instead of 'Workspace / Tool #N'
- instance_service.py: add rename_tool_instance() service function
- tool_instances.py: add PATCH /instances/{id} endpoint for renaming
display_name
Frontend:
- api/sessions.ts: add workspace_name to Session type, add renameInstance()
- use-instance-actions.ts: add handleRename, set document.title when opening
- session-card.tsx: click-to-edit display_name inline; always show project
context line (Project / Workspace or Repo / Tool)
- session-list.tsx: pass through onRename prop
- SessionsPage.tsx: wire handleRename to SessionCard and SessionList
- app-shell.tsx: sidebar tooltip includes workspace or repo name
- use-terminal-page.ts: set document.title based on active terminal session
Quality gates: py_compile all backend files pass, tsc --noEmit pass,
npm run build pass, 82/82 tests pass
Root causes:
1. No dedup for monitor restarts — _last_known_state is cleared on stop,
so every restart re-sent notifications for all unhealthy instances.
2. Aggressive error classification — any Docker state other than 'running'
was treated as 'error', including transient 'created' and 'restarting'.
3. Confusing metadata — when new_status == previous_status (after restart),
notifications showed previous_status equal to current status.
Fixes:
- _check_instance: when previous is None (first check) and new_status equals
the DB status, just record the snapshot and skip _handle_state_change.
This prevents duplicate events/notifications on monitor restart.
- _derive_status: only treat 'exited' and 'dead' as error. Preserve current
status for transient Docker states ('created', 'restarting').
- _derive_status: if DB says 'running' but container is 'not_found',
return 'error' instead of preserving 'running' (fixes silent failure).
- _handle_state_change: improved unhealthy message to 'Container tunnel is
unreachable' instead of generic 'Container is now unhealthy'.
Quality gates: py_compile all backend files pass, tsc --noEmit pass,
npm run build pass, 82/82 tests pass
Backend (health_monitor.py):
- Skip health checks for instances with no container_id
- Treat 'not_found' as error only when container was previously running
- Skip duplicate error notifications when already in error state
- Skip 'not_found' notifications for containers that never ran
Frontend (notification-item.tsx):
- Display notification.message (detailed error text)
- Add expandable Details section showing metadata (exit_code, previous_status, etc.)
- New CSS styles for message and metadata display
Quality gates: py_compile, tsc --noEmit, 80/80 tests pass
- Move APIRouter definition from instance_service.py back to tool_instances.py
(service files should not define FastAPI routers)
- Add missing prepare_manifest_instance import in tool_instances.py
- Guard repo.remote_url before clone_repository call
- Guard tool_type.compose_template before render_compose_template call
- Rename subprocess result variable to avoid shadowing SQLAlchemy Result
- Build error message as local string to avoid None/bool type issues
Quality gates: py_compile pass, LSP clean
From ae02e97 ('fix: tunnel URLs, session naming, git control bar placement'):
1. Tunnel URL regex: exclude api.trycloudflare.com from pattern.
Real tunnel subdomains are 10+ random chars. Prevents matching the
Cloudflare API endpoint instead of the actual tunnel URL.
2. Session auto-numbering: when user doesn't provide a display_name,
auto-generate 'project / repo / tool_type #N' where N increments
for each existing instance with the same project/repo/tool_type.
Prevents confusing duplicate display names in the sidebar.
These fixes were lost when main's merge was overwritten. Ported to
our clean dev codebase.
Quality gates: py_compile passed, ruff passed on tool_instances.py and tunnel.py.
tool-starter.tsx was hardcoding ssh_key_ids=[] and only showing a read-only
SSH key status. Users couldn't select keys when starting tools from workspaces.
Changes:
- tool-starter.tsx: add checkboxes for SSH key selection with repo key
pre-selected, pass selected keys to createInstance/startInstance
- AGENTS.md: add explicit rule forbidding docker compose commands without
user approval and proper isolation
The web container must be rebuilt to pick up the frontend changes:
docker compose up -d --build web
Quality gates: tsc clean, pytest (19 passed, 1 skipped)
When a tool container stops, the docker exec PTY reaches EOF. Previously,
the event-driven reader silently returned on EOF, leaving websockets
attached to a dead session. Input writes then failed silently.
Changes:
- _on_fd_readable: detect EOF (empty read) and call _handle_eof()
- _handle_eof: stop reading, mark process dead, close all websockets
with code 4001 to force frontend reconnection
- write_input: detect write errors and trigger EOF cleanup
Quality gates: pytest (19 passed, 1 skipped)
SSH key mounting was broken because:
1. Each selected key was mounted to a separate source dir but all targeted
the same ~/.ssh path in the container, causing Docker Compose's
last-mount-wins behavior
2. All keys were named id_ed25519, so they'd overwrite each other
Changes:
- ssh_keys.py: add key_filename param to prepare_ssh_key_files for unique
key names; add write_ssh_config for combined multi-key config
- tool_instances.py: collect all selected keys into a single ~/.ssh mount
with sanitized unique filenames (id_ed25519_<name>); generate combined
SSH config with all IdentityFile entries
- tests: add os.makedirs mock for SSH permission tests
Quality gates: pytest (19 passed, 1 skipped)
- Log REPO_PATH, SSH_PATH, EXTRA_VOLUMES, manifest mounts, and resolved
volumes in compile_compose() to trace why mounts may be missing
- Log repo_path and generated compose content in _prepare_manifest_instance()
to verify the full compose YAML at start time
- Convert WorkspaceHasInstancesError to store plain dicts instead of
SQLAlchemy ORM objects, preventing lazy-load failures outside async
session context (MissingGreenlet)
- Update both delete endpoints (top-level and nested) to use exc.instances
directly since they're already plain dicts
- Add no-cache headers for index.html in nginx.conf so browsers always
fetch new hashed JS/CSS bundles on deploy
1. Remove Docker build from create_instance for manifest types — the build
was blocking the HTTP request for several minutes, causing frontend
timeouts and retries. Image is now built lazily on start (via the
existing _prepare_manifest_instance path in start_instance).
2. Increase MAX_CONNECTIONS_PER_USER from 5 to 20 for SSE endpoint —
aggressive reconnect loops from the frontend were exhausting the limit
and causing 429 errors unrelated to tool starting.
Quality gates: ruff clean, tsc --noEmit clean, pytest workspaces (9 passed)
Bug 1 — in-container repo mounting:
- docker-compose.yml: added /data/working-copies:/data/working-copies mount
to API container so workspace dirs are visible on host filesystem
- Dockerfile: create /data/working-copies dir in image
Bug 2 — /home/user not writable:
- workspace_manager.py: chmod 777 workspace dirs + 666 files after clone
and after sync, so any container user can write
- manifest_compiler.py: explicit mkdir + chown + chmod 755 for home dir
in generated Dockerfile
Bug 3 — terminal text shifts left on typing:
- terminal.tsx: removed manual term.refresh() after fit (caused reflow)
- Track lastSentCols/lastSentRows and only send resize when dimensions
actually changed, preventing resize feedback loops
Bug 4 — ESC key captured by terminal:
- terminal.tsx: attachCustomKeyEventHandler allows ESC to propagate to
browser when not in alternate buffer (vim/tmux), so modals/navigation
work; ESC still sent to PTY when in vim/tmux alternate screen
Quality gates: ruff clean, tsc --noEmit clean, pytest workspaces (9 passed)
Developer
alex