budget_context_for_model() returns 0 when the endpoint does not prove its
context window, and compute_input_token_budget() answers 0 with the
conservative default budget (6000). The agent loop's small-local-server
clamp (`candidate_context <= 8192`) also matched 0 and cut the request to
max(1200, 0) tokens, so on any endpoint with an unreported window an open
document or a few thousand tokens of earlier context were trimmed away
before the model saw them ("5519 tokens > 1200 budget (ctx=1800)").
Apply the clamp only to a known window.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
When the chat model cannot see images, the separate vision call's description
is appended to the user's message. Nothing marked it as machine-derived, so the
chat model read it as the user's own prose and answered the description instead
of the question.
The label goes below the `[Image: <name>]` marker line because the client
strips the block by matching that exact prefix; renaming the marker would leave
the raw description rendered in the user's bubble.
Fixes#5573
Co-authored-by: Claude <noreply@anthropic.com>
`_gather_inputs()` fails soft: a source that raises leaves its inventory as
`[]`. Every probe reads an empty inventory as "not configured" and answers
`disabled`, so a broken integrations file, email store, or endpoint database is
indistinguishable from a subsystem the operator deliberately turned off. With
the remaining subsystems disabled, `GET /api/diagnostics/services` then reports
`overall: "ok"` while three of its inputs failed to load — the exceptions appear
only in DEBUG logs.
Record each failure alongside the empty value (`failed`: source -> controlled
category) and report the affected subsystem through `_source_unavailable()`
rather than probing it with an inventory that is empty for the wrong reason.
- `degraded`, not `disabled`: the configured state is unknown, not known-off.
Not `down` either — nothing reached the subsystem and failed; the inventory
could not be read. Severity is non-zero either way, so a report containing a
failed source can no longer roll up to `ok`.
- A source that loads and is genuinely empty still reports `disabled`.
- One broken source degrades one entry; the others are still probed.
- `meta` carries the source name and the controlled `_classify_error()`
category only, never exception text, credentials, or database detail.
- `inputs.get("failed")` so an inputs payload without the key still probes
normally.
Verified with the reporter's harness: settings loading with
`search_provider: "disabled"` while the other three sources raise now yields
`overall: "degraded"` with `searxng` still `disabled`; before, all five read
`disabled` and `overall` was `ok`.
Six regexes over tool output backtracked polynomially on long whitespace
runs or repeated openers: the browser snapshot generic-line filter and
post-interaction split, the task-row and skill-bullet parsers, the email
"no results" check, and the Google /sorry/ interstitial check.
Rewrite each without overlapping quantifiers, keeping the same matches for
the text these tools emit. Move the /sorry/ check into
is_google_sorry_url() so it can be tested.
Preserve complete scientific notation and signed floating-point literals across
arbitrary streaming chunk boundaries in _DegenerateStreamGuard. Buffer pending
tokens until delimiter characters arrive to avoid splitting numbers into
incomplete or arbitrary alphanumeric tokens, while maintaining the 2048-number
runaway backstop and detection of repeated phrases and word loops.
- Reconcile conflict resolution in routes/editor_agent_routes.py and src/editor_agent_model.py
- Enforce strict server-authorized endpoint resolution and reject client URL overrides
- Enforce bearer-token privilege separation: delegated API tokens are capped at non-admin scope per src/tool_security.py and cannot inspect private base URLs or probe metadata
- Preserve chat-scope authorization, missing-session 401 rejection, and owner-scoped browser bridge/SSE queues
- Preserve administrator-disabled-tool denylist enforcement introduced by PR #6357
- Fix mobile toolbar flyout hit-testing and stacking context trapping on viewports <= 700px via scoped elevation and mask suppression
- Stabilize project recovery test fixture by awaiting document draft quiescence prior to baseline capture
* fix(security): stop the agent from re-enabling tools an admin disabled
settings.json:disabled_tools is the only global tool-permission boundary
the Settings > Agent Tools panel offers, yet manage_settings exposed an
`enable_tool` action that removed entries from it. Any agent turn could
undo the admin's restriction, and because the agent reads untrusted web
pages, emails and documents, a planted instruction could re-enable bash,
email or any other tool that had been switched off.
The denylist can now only be narrowed from a turn: `disable_tool` and
`list_tools` keep working, `enable_tool` is refused with a message that
points the user to Settings > Agent Tools, and the action is dropped from
the function schema, the prompt rule and the tool index so models are not
steered into calling it. The free-form set/delete/reset actions were
already unable to reach `disabled_tools`; a test now pins that.
Fixes#5523
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UAB5sz5srMKR7rbeDFtdMH
* fix(settings): close internal API tool-policy bypass
Refuse global tool-policy replacement from internal-token requests even when an admin owner is impersonated, and block the same write in app_api. Route explicit image re-enable requests to list_tools with a Settings pointer. Keep disabled_tools outside general settings actions even if a default is added later. Remove the retired action from capability metadata. Validation: 573 focused tests passed; the running app rejected handler, generic API and encoded-path internal HTTP attempts while preserving human administrator controls.
* fix(mcp): block agent writes to per-server MCP tool policy
PATCH /api/mcp/servers/{id}/tools only checked require_admin, which accepts
the internal tool token, and app_api still forwarded it, so an agent turn
could re-enable MCP tools an admin had switched off. Refuse the internal
token on that route (same check as POST /api/tools) and add PATCH
/api/mcp/servers to the app_api blocklist, with regression tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BCbhLNC7mJsunxqfwwshqn
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Afonso Junior <chatgpt@yack.com.br>
Use configured quality only for omitted arguments in generation, MCP and edit helpers. Keep explicit medium choices from being overwritten by a low or high default.
* fix(email): guard user-supplied IMAP/SMTP hosts against SSRF
POST /api/email/accounts/test opened raw imaplib/smtplib connections to
whatever host:port the request body named, and POST/PUT /api/email/accounts
and PUT /api/email/config persisted arbitrary hosts for the background
pollers to contact. All of these sit behind require_user/require_owner with
no admin gate, so any authenticated user could make the server connect to
loopback, LAN and cloud-metadata targets (169.254.169.254) and read the
first 200 characters of the remote banner back out of the friendly error.
THREAT_MODEL.md lists email as admin-only; the tool layer honours that but
the HTTP routes did not.
Add check_outbound_host() to src/url_safety.py, a bare-host companion to
check_outbound_url() with the same classification (link-local, multicast,
reserved and unspecified always rejected; private/loopback only with
block_private) plus localhost-name handling and an explicit policy for names
that do not resolve. Wire it into the four email routes via
_check_mail_host(): non-admin owners (and everyone when
EMAIL_BLOCK_PRIVATE_IPS=true) also lose private/loopback targets, admins
keep the local-first LAN/Dovecot setups. Unresolvable names pass through
because the connect that follows fails with the same DNS error.
Regression tests cover the helper's policy table and each route: the test
route rejects before any socket opens, save routes reject before touching
the DB, admins keep loopback unless locked down, and an update that fails
validation leaves the stored row untouched.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TwcwfY2uQMxNRsK3y6xBAp
* fix(email): harden IMAP/SMTP SSRF protection against rebinding, TOCTOU, and unvetted hosts
- P0: Implement safe_create_connection() in src/url_safety.py to resolve hostnames once, validate every returned address against SSRF policy, and connect directly to the pinned IP. This eliminates DNS rebinding and TOCTOU while preserving TLS SNI and certificate hostname verification across IMAP, IMAPS, STARTTLS, SMTP, and SMTP_SSL.
- P1: Ensure unresolvable or deferred hostnames cannot become an unrestricted future connection capability by enforcing runtime SSRF validation fail-closed on every connection, while maintaining compatibility with offline configurations.
- P2: Extend SSRF enforcement to the core connection layer (_open_imap_connection, _send_smtp_message, _imap_connect) using ContextVar owner scoping, protecting pre-existing DB rows, restored configs, background pollers, and scheduled operations.
- P3: Allow admin and single-user access to IPv6 loopback (::1, [::1]) for local mail servers (e.g. Dovecot) by checking loopback before is_reserved, while keeping link-local, cloud metadata (169.254.169.254, 100.100.100.200), and non-admin restrictions strictly enforced.
- Add comprehensive regression and adversarial test coverage in tests/test_email_account_host_ssrf.py.
* fix(email): reject unresolved saved hosts without stdlib monkeypatches
Use the existing explicit policy IMAP/SMTP classes and canonical connect_outbound_tcp helper. Remove process-wide stdlib assignments, context-variable plumbing, redundant aliases and resolver-signature retries. Reject unsafe or unresolved hosts before account/config persistence, keep DNS work off the event loop, return fixed client errors, identify Alibaba metadata accurately and reject invalid TCP ports. Validation: 654 relevant email, URL-safety and CodeQL regression tests passed, 3 skipped. Three unrelated browser checks were excluded after environment failures (missing Playwright browser and Chromium timeout). Running-app checks verified rejected create/update/config requests leave stored data unchanged, administrator LAN configuration still works, and metadata/non-admin loopback are blocked.
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: PR #6029 Reviewer <reviewer@odysseus-review.local>
Co-authored-by: Afonso Junior <chatgpt@yack.com.br>
* fix(security): block read-only cookbook tools for non-admin owners
list_cached_models, list_served_models, tail_serve_output, list_downloads,
search_hf_models, list_serve_presets and list_cookbook_servers were in the
plan-mode read-only allowlist but in neither NON_ADMIN_BLOCKED_TOOLS nor the
dispatcher's _ADMIN_TOOLS. Each implementation calls a require_admin route
over the internal loopback with the process-wide INTERNAL_TOOL_TOKEN and
never checks the owner, so a non-admin user's agent inherited admin reach
through them. list_cached_models is the worst case: a caller-chosen `host`
is forwarded to /api/model/cached, which SSHes to that host with the app
user's keys and walks caller-chosen model_dirs. THREAT_MODEL.md lists model
serving as admin-only, so gate the read side the same way as serve/download.
Adding the names to NON_ADMIN_BLOCKED_TOOLS also hides them from non-admin
tool schemas and caps bearer-token agent runs; _ADMIN_TOOLS blocks them at
dispatch. Regression test covers both lists, schema hiding, and that the
loopback is never reached for a non-admin owner.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TwcwfY2uQMxNRsK3y6xBAp
* fix(cookbook): bind request authority in admin gate regressions
Keep all seven read-only Cookbook tools in the admin prompt inventory. Grant request authority in dispatch tests so the admin gate itself is exercised, and verify administrators still reach each handler. Validation: 359 focused authorization, tool-policy and schema tests passed. Against the running app with real auth accounts, all seven tools reject a non-admin and an admin reaches the authenticated Cookbook HTTP API.
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Afonso Junior <chatgpt@yack.com.br>
* fix(agent): give file tools the agent workspace as root when no workspace is set (#6679)
With no workspace bound, bash/python run in data/agent_workspace, but
RequestAuthority sealed no filesystem root, so every native file tool
failed with "requires a sealed resource root", and an approved
write_file failed with "no sealed resource identity". Seal the agent
workspace instead, and resolve relative paths in _resolve_tool_path
against it rather than the server's cwd.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test: re-seal resource roots when a test swaps the authority owner (#6679)
A no-workspace authority now seals the agent workspace for its owner, so
replace(authority, owner="bob") copied alice's root and was rejected
before the approval-binding check the tests are about.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(agent): only seal a vetted agent workspace as the file-tool root (#6679)
Apply the agent_cwd() check before sealing data/agent_workspace: it must
be a real directory in the readable data subdirs, otherwise no root is
sealed. A symlinked workspace no longer redirects the file tools or the
bash/python launch scope. Tighten the dispatch test assertions to the
exact 'outside the sealed roots' message.
* refactor(agent): vet the no-workspace file root with default_agent_workspace (#6679)
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
#6683 made an exact approval seal a launch boundary (and a filesystem root)
in the request's workspace, but only when a workspace is bound. Agent mode
with no folder picked is the default, and there the approval still sealed
nothing, so "Allow" always ended in "Approved action has no sealed
process/job identity" (bash/python) or "Approved filesystem action has no
sealed resource identity" (write_file). In Docker every agent turn carries
the backend runtime context, so bash and write_file always ask for approval
and never ran.
With no workspace, seal the root in the agent workspace, which is where
agent_cwd() already runs bash/python in that case. Inherited (child)
authorities still cannot widen.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The manage_endpoints agent tool accepts arbitrary user-supplied base_urls
and writes them to the database without calling check_outbound_url(),
while the HTTP route (POST /api/model-endpoints) enforces full SSRF
hardening. A prompt-injection payload smuggled into a skill, note, fetched
page, or email could instruct the model to register a cloud-metadata or
internal-network endpoint via this tool path.
Add the same check_outbound_url() guard used by the HTTP route to the
agent tool's 'add' action. Also sanitize the error message in the except
block to avoid leaking raw exception text.
Add 8 regression tests covering: cloud metadata, non-HTTP schemes,
loopback (local-first), strict mode, safe URLs, empty URLs, gopher
scheme, and error message sanitization.
Co-authored-by: isharak7m <192635824+isharak7m@users.noreply.github.com>
Catch OverflowError in addition to TypeError and ValueError when converting
research_max_tokens, research_extraction_timeout_seconds, and
research_extraction_concurrency to int. Non-finite numbers parsed from JSON
(such as 1e309 or -1e309 representing float infinity) raise OverflowError on
int() conversion, which previously escaped the handler and crashed scheduled
research execution. Also add unit tests covering exponent overflow values.
POST /api/tasks and the scheduler both refuse action tasks whose action
is in ADMIN_ONLY_TASK_ACTIONS (run_local, run_script, ssh_command,
cookbook_serve) for owners without admin task privileges: those actions
reach subprocess.run(shell=True) or SSH with no sandbox. The model-facing
manage_tasks tool applied neither check, so an agent could create or edit
such a task, resume it, or trigger it, and only the scheduler's
execution-time guard stood in the way (pausing the task after the fact).
Route the tool through the same src.task_action_policy helpers the HTTP
route uses: create checks the requested type/action, edit checks the task
as it would be after the edit (so an llm task cannot be flipped into a
shell action), and resume/run check the stored task. Admins, non-
privileged actions and llm/research tasks are unaffected; pause still
works for everyone since it only narrows.
Fixes#6021
Claude-Session: https://claude.ai/code/session_01UAB5sz5srMKR7rbeDFtdMH
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Alexandre Teixeira <111787685+alteixeira20@users.noreply.github.com>
* fix(security): deny generated-image access when ownership verification fails
The except Exception block in serve_generated_image catches all
exceptions during ownership verification and falls through to serve
the file anyway. If the database is down or corrupted, the ownership
check is bypassed entirely — any authenticated user can download any
generated image by guessing the content-hash filename.
Raise HTTP 503 (Service Unavailable) instead of falling through,
so the endpoint fails closed when verification is impossible. This
matches the defense-in-depth principle: deny access on verification
errors rather than granting it.
* fix(security): enforce fail-closed ownership and secure caching for generated images
- Fail closed on missing gallery records, null/empty owners, and DB errors
- Support both cookie session and bearer API token owners via effective_user
- Add in-memory registry for freshly generated images to preserve access
- Eliminate file existence oracle by verifying ownership before disk probing
- Update Cache-Control to private, no-cache, must-revalidate
- Pass owner attribution through tool execution and image MCP server
- Add comprehensive hermetic regression tests for ownership gates and caching
* fix(security): sanitize tool owner arguments and drop unneeded MCP env fallback
- Remove redundant and unsafe ODYSSEUS_MCP_IMAGE_OWNER fallback in image_gen_server.py to fix 4 tests in test_env_reference.py.
- Sanitize caller-supplied tool arguments across all MCP and email tool dispatch paths in src/tool_execution.py: strip untrusted _odysseus_* keys from arguments.
- Only inject _odysseus_owner when verified trusted server context is present, preventing untrusted/anonymous users from spoofing ownership claims.
- Sanitize empty/whitespace owner strings in image_gen_server.py to None.
- Add comprehensive regression tests in test_generated_image_ownership.py for argument sanitization, owner injection, fallback propagation, and server sanitization.
---------
Co-authored-by: isharak7m <192635824+isharak7m@users.noreply.github.com>
Co-authored-by: PR #6166 Reviewer <reviewer@odysseus-review.local>
Skills tagged remote-only in frontmatter are excluded from the injected
skill index when the active model's endpoint is local (loopback/private/
tailscale per is_local_endpoint). Big on-demand reference lookups
overwhelm small local contexts; local models never see the skill, while
remote models and API/UI listings are unaffected.
- SkillsManager.index_for gains optional keyword-only local_model flag
(None = don't gate, same contract as active_toolsets=None)
- _build_system_prompt/_build_base_prompt thread endpoint_url from
_build_route_request_state; unknown endpoint fails open
- tests: cover hidden-on-local, visible-on-remote/unknown, untagged
skills never gated
Refs: discussed for the local-model overwhelm problem
Co-authored-by: Alexandre Teixeira <111787685+alteixeira20@users.noreply.github.com>
