* fix(milvus): support multilingual BM25 retrieval
* fix(milvus): preserve source metric and tighten collection matching
Read the source embedding index metric during BM25 migration, match
collections by {prefix}_{dimension}, and route Japanese/Korean text to
the ICU fallback.
* fix(milvus): satisfy lint on multilingual BM25 changes
---------
Co-authored-by: Jonesxq <239089032+Jonesxq@users.noreply.github.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
Long-term memory chose its candidates before it looked at the question.
Recall listed 400 items by `importance DESC, valid_from DESC`, loaded the
vectors belonging to that list, and scored those; extraction did the same
with a 200-item pool. Importance is uncorrelated with relevance, so the
vector stage could only re-order what importance had already picked, and a
memory below the cut was unreachable no matter how well it matched. The
capacity cap goes up to 2000, so everything past the 400th most important
memory was permanently unrecallable, and `memory_item_embeddings.vector`
being a BYTEA blob meant pgvector was never involved.
The store now answers "which of this person's memories is this question
about":
- Migration 000095 (and the SQLite mirror) add `embedding halfvec` plus a
`(tenant_id, subject_id, model_id, dims)` index. The column is
conditional on the vector extension, so a deployment whose retrieval
driver excludes postgres keeps working. No HNSW: an approximate index is
filtered after the fact, and one subject holds at most a couple of
thousand rows, so exact ordering inside the scope is both correct and a
few milliseconds.
- `SearchItemsByVector` ranks in SQL over every vector the subject has,
with the status, expiry and kind filters joined in so reaching past the
candidate list cannot reach past the filters. SQLite and pgvector-less
PostgreSQL fall back to scoring in process, still over the whole scope.
- Recall fuses the lexical ranking with those hits and grows the candidate
pool with matches the pool did not hold, which is what lets the search
escape it. The lexical pool is now the capacity cap, so it can no longer
be a second, invisible limit under the one that decides what stays active.
- Extraction picks the memories it shows the model the same way. This
matters more than recall: the memory a new statement contradicts is not
reliably an important one, so an importance-ordered window hid exactly
the collision that stops a duplicate being written.
- Vectors are written to the searchable column in the upsert's transaction;
rows predating the column are moved by `SyncVectorColumn` during
maintenance, which costs no model calls. The embedding backfill goes from
50 to 200 per run — at 50, a subject at the capacity cap took over a
month to become fully searchable.
- Traces and logs report `outside_pool`, the number of semantic matches the
lexical pool did not contain. It was structurally zero before, so it is
the number to watch when asking whether semantic recall does anything.
Tests cover whole-scope ranking, the lifecycle and kind filters, pool
growth, the pool-size invariant, and extraction seeing a low-importance
memory it would otherwise duplicate. A pgvector integration test
(WEKNORA_MEMORY_TEST_POSTGRES_DSN) exercises the migration, the halfvec
write, the SQL ranking and the column sync.
StartChildSpan records sandbox housekeeping only when a parent trace already exists, and skill install/remove/maintenance open that boundary themselves.
* feat(browser): enhance BrowserIcon component and update localization strings
- Updated BrowserIcon component to accept dynamic width and height props.
- Revised localization strings in English, Japanese, Korean, Russian, and Chinese for improved clarity and user guidance regarding browser connection and usage instructions.
- Enhanced BrowserConnectionSettings view with better structure and user feedback elements.
* Default BrowserSkill tasks to a separate window
* refactor(prompts): streamline agent system prompts and context templates
- Updated agent system prompts to enhance clarity and specificity in user interactions.
- Revised context templates to better separate metadata and user requests, improving overall structure.
- Adjusted fallback and intent prompts to ensure consistent language handling and response requirements.
- Introduced new localization strings for browser search instructions across multiple languages.
- Added tests for agent prompt serialization and hydration to ensure template integrity and functionality.
* chore: wrap prompt strings to satisfy golangci-lint
Keep the settings, separate-window, and prompt commits pushable together
without rewriting their history.
* fix(frontend): match agent editor save test to serialized payload
The prompt-inheritance save path still closes after update; the test
should assert that payload, not the unsanitized form object.
* feat(browser): implement browser connection management and related tests
- Introduced a new `browserConnection` store to manage the state of browser connections, including online status and device information.
- Updated the `Input-field` component to handle browser connection states, enabling/disabling features based on the connection status.
- Added tests for the `browserConnection` store to ensure correct behavior under various scenarios.
- Enhanced the `Input-field` and `BrowserConnectionSettings` components to provide user feedback based on the browser connection state.
- Adjusted related components and views to integrate the new browser connection logic, improving overall user experience.
* fix(frontend): update model-save assertion for serialized agent payload
The agent editor still persists the selected model ID via a spread of
formData; the source test should match that payload, not the raw form.
* feat(sandbox): add safe live file management to session sandboxes
* fix(sandbox): keep live files from waking paused sandboxes
Serialize directory entries as "directory", refuse Files-tab Connect on paused runtimes, and cap list/delete trees so oversized workspaces fail closed.
---------
Co-authored-by: wizardchen <wizardchen@tencent.com>
* feat(sandbox): use a classic colored user@host prompt in the PTY
Cube, Docker, and E2B now share user@host:last-dir with WeKnora brand green, overriding E2B's default `\w $ ` prompt.
* fix(sandbox): redraw the PTY prompt after connect instead of waiting for Enter
Reattach does not replay bash's already-printed prompt, and FitAddon can clear the first paint. Fit before flushing output, then Ctrl-L if the screen is still empty.
* feat(sandbox): add Debian-style ll/la/l aliases in the interactive PTY
Interactive bash now gets colored ls/grep shortcuts; non-interactive skill scripts still see the real commands.
* fix(sandbox): restore conventional ls colors so directories stay blue
Map ANSI blue/cyan back to Tango colors so directories and executables are distinguishable; keep brand green only for user@host.
* fix(sandbox): correct runtime env-var hint to this config and create time
The copy claimed workspace-wide injection. Scope it to sandboxes created from this config, point per-run secrets to the Sandbox secrets page, and keep the hint off the add-button row so Chinese no longer wraps mid-word.
* feat(sandbox): attach a running session sandbox when the terminal opens
Opening the panel used to always show the start overlay. Lookup first so a running sandbox reconnects, while paused or missing sandboxes still wait for an explicit click and Connect cannot silently resume.
* fix(sandbox): do not inject PTY input or auto-connect a non-running sandbox
Empty-screen reconnect now SIGWINCHes so a reattached busy PTY is not fed
Ctrl-L. Lookup-only opens require a click unless List confirms running, and
closing the panel drops the terminal mount so a Files-tab reopen cannot attach.
---------
Co-authored-by: hailongzhao <hailongzhao@tencent.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
* feat(stream): persist live-run markers and per-run steer queues
Mid-run sends need a cross-replica pointer to the generating
assistant and a control-plane list that never hits the SSE stream.
* feat(agent): inject queued user messages at ReAct round boundaries
The engine drains the steer sink after each round, persists accepted
rows, and delays Done until after a loop-end inject so the turn can continue.
* feat(session): expose /steer and hand leftover messages to a follow-up turn
Owner-scoped steer routes queue or inject into the live run, single-flight
AgentQA via the live-run marker, and history replay keeps multiple user rows
inside one turn.
* feat(frontend): queue mid-run messages while an agent turn streams
Stop and Send stay visible together, the overlay follows server delivery,
and history splits the transcript around injected user bubbles.
* fix(frontend): type pending and client_id on SteerQueueItem
vue-tsc failed because the overlay sets those fields while POST /steer is in flight.
* fix(session): harden steer persist, handoff, and new_run
Keep SetLiveRun ahead of SSE and roll back orphan turns, make inject
persist idempotent, append-then-delete on rebind, cap overrun rounds,
and stop aborting a live stream when /steer returns new_run.
* fix(session): satisfy pre-push lint on steer paths
Wrap long lines, check discarded Error/Emit results, and keep
context as the first argument so the fork push hook can pass.
---------
Co-authored-by: hailongzhao <hailongzhao@tencent.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
* feat(mcp)!: persist tool catalogs and add guided configuration
* fix(mcp): write caller-owned catalogs and require a real tool sync
OAuth viewers can now persist their own directory, and the setup wizard no longer treats an unsynced service as complete. Missing catalogs auto-fetch on the tools step; already-used MCP functions are restored across engine turns.
* fix(mcp): skip redundant list_servers and tighten catalog UI
When the advertised server directory already fits in discover_mcp_tools, models should list_tools or describe directly. Chat and settings now show server names, collapsed docs, and initialize debug metadata for empty optional fields.
* fix(mcp): surface catalog summaries and polish tool detail UI
List responses now include a saved catalog count without live MCP calls, so settings cards and @ mentions can show sync state. Tool details open in a popup instead of expanding the row.
* fix(mcp): drop incomplete Anthropic calls and gate shared catalog refresh
Unclosed streamed tool_use was emitted as {} and executed. list_tools refresh=true now re-lists upstream instead of returning the saved snapshot, and only Admin (or MCP-capable API keys) can overwrite tenant-wide static-auth catalogs.
* fix(mcp): wrap client refresh path to satisfy lll
The metadata refresh URL made doRequest exceed the 120-character limit in the client module CI lint.
* fix(graph): feed OCR chunks to graph extraction instead of image-link placeholders
KnowledgePostProcessService.Handle gathered "text-like" chunks through
chunkService.ListChunksByKnowledgeID, which is text-only by design, so the
image_ocr / image_caption chunks written by multimodal parsing were never
returned. For a scanned PDF the text chunks contain nothing but markdown
image links, so graph extraction ran on empty input: most calls failed
with "unexpected end of JSON input" and the rest echoed the few-shot
example from the extract config into the graph.
- Fetch the exact chunk types the post-process enriches via
chunkRepo.ListChunksByKnowledgeIDAndTypes (text, image_ocr,
image_caption); the service now receives interfaces.ChunkRepository
through dig like knowledgeService already does
- Build a dedicated graphChunks list for graph extraction: skip
image_caption chunks (they duplicate the OCR text) and chunks that
contain only image links (chunkHasExtractableText)
- graphChunkCount, the enqueue loop and the log line use graphChunks;
summary, wiki, auto-tag and question generation keep using textChunks
- Tests: fake ChunkRepository in the wiki-enqueue test, table test for
chunkHasExtractableText
* fix(graph): skip figure OCR and harden empty-chunk detection
Reuse the CommonMark image scanner plus extractRealText so titles with
parentheses and <image> wrappers are not treated as prose. Only enqueue
OCR for graph extraction when the parent text chunk has no extractable
text, and skip link-only chunks for question generation as well.
* style(graph): wrap long post-process extract log line
---------
Co-authored-by: wizardchen <wizardchen@tencent.com>
* feat(sandbox): stream interactive PTYs from Cube and E2B
Give session sandboxes a provider-neutral terminal capability so a later
WebSocket layer can attach without knowing Cube vs E2B. Long-lived
streams use a timeout-free client; idle terminals keep the sandbox alive
by refreshing TTL, and a workspace can set how long to wait before
disconnect.
* feat(session): bridge browser terminals to the session sandbox
Mint a short-lived ticket over a normal authenticated POST so the browser
WebSocket never puts the access JWT in the query string. After upgrade,
recheck the bound token, membership, and session ownership so logout or
revocation tears the PTY down. Provisioning stays opt-in: a GET
handshake only creates a sandbox when the client sends provision=1.
* feat(chat): add a sandbox terminal panel beside the conversation
Open a side panel with an xterm attached to the session sandbox. Starting
the terminal is an explicit click so merely opening the panel does not
create or resume a microVM; reconnects stay lookup-only until the user
confirms create-and-start. Persist the remote shell PID in
sessionStorage so a refresh reattaches instead of spawning an orphan.
* feat(sandbox): let workspaces set terminal idle disconnect
Expose the idle-disconnect duration in the sandbox config editor so an
operator can close unused terminals and let the provider TTL pause the
sandbox, instead of leaving PTYs open for the built-in 15 minutes.
* feat(chat): show generated files in the sandbox side panel
Move the per-message artifact list into the first tab of the right-hand panel so generated files stay in view with the terminal. Keep the old drawer only for embedded chat.
* fix(sandbox-terminal): keep tickets out of logs and live PTYs across refresh
Gin was copying the handshake JWT into response_body logs, silent access-token
rotation dropped open terminals as AUTH_REVOKED, and first-use provisioning
missed shared agents. Local echo now skips password prompts.
---------
Co-authored-by: hailongzhao <hailongzhao@tencent.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
* refactor(resource-access): unify scoped authorization and transfer recovery
* fix(resource-access): close reparse and transfer review gaps
* fix: enforce explicit file bindings and recover move cleanup
* fix: align mutation admission and cleanup scope checks
* fix(agent): improve web search and persistent page reading
* fix(agent): make web fetch snapshots waitable and retryable
Keep in-flight page fetches independent of a caller's short deadline so
content=true and continuations cannot cancel a shared download or miss a
snapshot that is still loading. Mark snapshot_expired as retryable, fetch
only the first three search hits in parallel, and stop forcing Brave's
country to US.
* feat(web-search): add Bocha AI Search provider
Add a Bocha (bochaai.com) web search provider following the Metaso
adapter pattern: hardcoded official endpoint (SSRF-safe), Bearer auth,
4MB response cap, error-detail extraction, and RFC3339 date parsing
(datePublished preferred, dateLastCrawled fallback).
Tenant-facing config fields:
- freshness: noLimit/oneDay/oneWeek/oneMonth/oneYear (default noLimit)
- summary: request long text summaries, preferred as snippets (default on)
Wiring: registry factory in container, provider type enum + /types
metadata (dynamic form fields), frontend provider union type, and
swagger docs enum entries.
* fix(web-search): match Bocha's real wire format for codes and error messages
Live endpoint probing showed errors return {"message": .., "code": "401"}
(string code, message field) instead of the documented {msg, code:int}.
Parse the code tolerantly (string or number) and extract the detail from
message/msg so auth and rate-limit failures surface their reason.
* fix(web-search): allow Bocha in provider CRUD validation
In-service integration testing caught a fourth wiring point the adapter,
registry, /types metadata and frontend covered but CRUD did not: the
service-layer whitelist (isValidProviderType) and parameter validation
switches rejected provider type 'bocha', so providers could be tested
via POST /web-search-providers/test but never created or updated.
Add Bocha to both switches (delegating to ValidateBochaParameters like
Metaso/Zhipu) with service-level test coverage.
* fix(web-search): correct Bocha legacy timestamp timezone
---------
Co-authored-by: vicfei <vicfei@users.noreply.github.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
StringArray is stored as a BLOB on SQLite, so the TEXT equality used by the directory listing never matched and every wiki folder listed empty on the SQLite build. Compare the text form there.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DFSMrLZZ3oq1dKmJFAofz6
* feat(datasource): persist source document creation time and use content edit time for Feishu
Follow-up to 160c806b (Refs #3016), which started persisting the source
last-modified time as metadata.source_updated_at. Two gaps remained:
1. Creation time was never carried. FetchedItem gains CreatedAt and
ingestItem persists it as metadata.source_created_at (RFC3339, UTC)
with the same absent-means-unknown rule as source_updated_at.
2. The Feishu wiki connector filled FetchedItem.UpdatedAt from
NodeEditTime, which moves when a node is renamed or dragged in the
tree without any content change. The connector's own cursor already
prefers ObjEditTime for exactly this reason; use the same rule for the
persisted value (contentEditTime / contentCreateTime, node_* as
fallback), so source_updated_at reflects document edits.
Feishu Drive populates CreatedAt from created_time. GitLab stops
fabricating UpdatedAt = time.Now(): a fetch timestamp is not a document
timestamp, and an absent value is the documented "unknown" (the commit
time is not fetched today; leaving it unset is honest until it is).
Tests: ingestItem covers both keys present/absent; the wiki fetch test
asserts obj_* times win and node_* are the fallback.
* style(gitlab): break up the FetchedItem literal (lll)
---------
Co-authored-by: xiaoyaner0201 <>
A folder named 概念/实体/Concepts was stripped from category_path as model noise, so a page moved into it kept depth 0 and a root-level wiki_path and rendered at the root of the directory tree. Folder paths are authoritative and now map to category_path verbatim; the model-label cleaner still applies to pages without a folder.
Fixes#3053
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DFSMrLZZ3oq1dKmJFAofz6
Wiki ingest makes several independent LLM calls per document and already
retries transient failures (408/429/5xx, timeouts, connection resets) with
exponential backoff, but some gateways report QPM/QPS throttling as HTTP
403 instead of 429 (e.g. DeepSeek MaaS: code 0x04030020, message
"调用频率(qpm)超限"). Those 403s were classified as permanent caller
errors and the wiki span failed immediately with EXTRACT_FAILED /
SUMMARY_FAILED, so a QPM burst aborted generation for the whole batch
(issue #2688).
isTransientLLMError now also matches "status 403" errors whose embedded
response body carries rate-limit wording (qpm/qps, rate limit, too many
requests, throttled, 调用频率/频率超限/请求过于频繁/繁忙, try again
later, slow down). Authorization-shaped 403s (invalid key, forbidden,
permission denied) and unrelated bodies stay permanent, so the change
cannot turn authz failures into retry loops.
Tests: 14-case table covering rate-limit 403 acceptance and
authorization-shaped 403 rejection, all existing isTransientLLMError
cases unchanged and passing.
* feat(sandbox): run sandboxed execs as root by default
Every chat session owns its own sandbox, so the in-container account is
not a tenant boundary and root is the least surprising default for an
agent that installs packages and writes wherever it needs. Host and
cross-tenant isolation stay at the container boundary.
DefaultSandboxExecUser becomes "root" (remote_client.go) and the sandbox
image ends with USER root. The three backends keep naming the account
explicitly on every exec so the account a call lands on does not depend
on which backend a workspace selected; the adapter default, the file
operations and the artifact-directory bootstrap all follow the constant.
The template still provisions a uid 1000 "user" account for interop with
E2B/Cube tooling that addresses it by name.
Root also makes the skill-install flow's account hand-over pointless, so
it goes: verification no longer runs after a chown/chmod lock-down and a
repair round no longer reopens the tree (normalizeSkillPermissions and
reopenSkillDirForRepair are gone), and the prepare and scratch commands
stop chowning. The scratch wipe now also clears /tmp, and the cache
budget guard sizes only root's caches, since no second account can grow
one. Verification is asserted to stay on the ordinary executor (no
AsRoot, no AllowSkillsRoot) rather than on which account it ran as.
* refactor(skills): install chat-time extras into the skill's own env
PYTHONPATH / NODE_PATH pointed at a per-session overlay
(/workspace/.skill-packages/<skill>) because the skill tree was frozen
after install: root-owned, mode 555, and often built by `uv venv` with no
pip, so `pip install --target` somewhere else was the only way a chat
session could add a package.
None of that holds now. The lock-down is gone and execs run as root, so a
missing package goes straight into the skill's own environment, and the
write is already session-local: a chat sandbox is built from the config's
image and its filesystem dies with the session, so nothing reaches the
image other sessions start from. The overlay bought nothing over that
while splitting one skill's packages across two locations.
SessionSkillPackageDir goes, and applySessionPackagePath becomes
applySkillNodePath: it appends the skill's own node_modules after
whatever the caller supplied, replacing a `+=` that prepended a bare
separator once the overlay stopped seeding the variable. No PYTHONPATH is
injected at all. Python dependencies are reached through the skill's venv
interpreter, which the shell wrapper already puts first on PATH; the
overlay-era value named the venv root, which is not an importable path.
PYTHONPATH stays on the injected-name blacklist so a stored value cannot
shadow the venv.
The runtime hints and every prompt that described the overlay now name
`<skillDir>/.venv/bin/python -m pip install` and `npm install` in the
skill directory. frozenSkillTreeGuidance / isFrozenSkillVenvFailure are
renamed for what they detect now that nothing is frozen.
* docs(prompts): reserve /workspace/output for finished deliverables
/workspace/output is the one directory an artifact collector reads, so
everything written there is offered to the user as a deliverable. The
prompts described it as the place for "deliverables" or "downloadable
artifacts" without saying it is exclusively that, which reads as "the
directory you write results to" — and a model doing multi-step work has
results at every step. Drafts, intermediate renders and debug dumps land
next to the finished file and the user is handed all of them.
The four places that describe the workspace now say the same thing: the
collector reads only /workspace/output, so finished deliverables go there
and everything in progress goes anywhere else under /workspace. The
attachment block gains the matching half for the other direction —
/workspace/input holds the user's own files and is not ours to write to,
which it previously implied by calling the paths "read-only inputs" while
the tools that enforce it say so plainly.
Only the /workspace/input half is enforced (write_sandbox_file refuses
that tree). Whether a file is finished is a judgement no tool can make,
so the output half stays guidance and is stated wherever the model is
already reading about the workspace rather than in one place it might
not see.
* fix(skills): restore executable bundles and package recovery
* style(sandbox): wrap prompt text for repository lint checks
* fix(sandbox): guard recovery paths and document root execution
---------
Co-authored-by: hailongzhao <hailongzhao@tencent.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
Connectors already fill FetchedItem.UpdatedAt (Feishu from the real edit
time), but ingestItem dropped it: the persisted metadata only carried
external_id, source_resource_id and datasource_id, and the knowledge
row's own UpdatedAt moves on every re-parse, so nothing downstream could
tell how old a document is.
Write it into metadata as source_updated_at (RFC3339, UTC) when the
connector supplied one, and leave the key out otherwise so an absent
time stays distinguishable from a zero one.
Refs #3016
* fix(chat): keep OCR/caption text of scanned-page image hits in merge
When pdf_force_scanned is enabled, the recognized page text lives only
in image_ocr/image_caption child chunks, while the parent markdown is
cut from image placeholders alone. The merge stage's image_parent_resolve
overwrote the hit content with the pruned parent markdown, so the OCR
text never reached the LLM context and answers degraded to placeholders.
Save the hit chunk's own body before the parent expansion and re-attach
it via JoinChunkContent, which collapses duplicates, so normal multimodal
hits (whose text is already covered by the surrounding content) are not
duplicated.
Fixes#3052
* fix(chat): avoid duplicating scanned OCR after parent merge
Re-attaching the image hit body into Content would be injected again
from ImageInfo during chat enrichment. Clear the matching OCR/caption
field so the model sees one copy, and cover caption plus pruned-placeholder paths.
* style(chat): wrap long merge parent-resolve test fixture
---------
Co-authored-by: LuckVd <LuckVd@users.noreply.github.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
Local storage prefixes now join under the base directory instead of
filepath.Join, FAQ import rejects client task IDs that can leak into
object names, and wiki list ORDER BY uses quoted identifiers instead of
sprintf.
* feat(mcp): add per-tool enable toggle
* style: format migration test
* fix(mcp): keep enable and approval updates from overwriting each other
Partial policy writes now touch only the supplied columns, and a missing
tenant fails closed so a disabled tool cannot slip through.
---------
Co-authored-by: myysophia <myysophia@users.noreply.github.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
IndexMessageToKB returned silently when getChatHistoryConfig yielded nil,
so an operator who enabled message indexing in the UI could see
"indexed messages: 0" with nothing in the logs to explain it.
The two paths disagree on what "enabled" means:
GetChatHistoryKBStats -> cfg != nil && cfg.Enabled
IndexMessageToKB -> cfg.IsConfigured(), which also requires
EmbeddingModelID != "" && KnowledgeBaseID != ""
So Enabled=true with either field empty reports the feature as on while
every message is skipped, and the skip emitted no log line at all.
Log the specific missing precondition instead. Behaviour is unchanged -
this only makes an existing silent path attributable.
* feat(agent): unify sandbox tools and drop host preloaded skills
Read workspace and skill files through read_file, and run installed or staged skills through shell_exec(skill_name). Remove the unused skills/preloaded tree and keep generated-file previews stable across re-renders.
* fix(frontend): do not assume classList in markdown sanitize hooks
Node unit tests stub anchors without a DOM classList. Detect download cards from class/download attributes instead.
* feat(sandbox): add the tenant network policy with its secrets and validation
Administrators need control over sandbox network access, so a named sandbox
config gains a network policy block. It lives inside the existing config JSONB,
so there is no schema change and old rows read back as the zero value.
Both booleans are phrased so that the zero value IS the intended default --
outbound open, inbound closed. That makes "no network block" and "explicitly
saved defaults" provably the same thing, and avoids a tri-state pointer where
nil would have to mean "allowed" for egress and "closed" for ingress.
Injected header values are credentials, so they ride the encrypt-at-rest and
mask-on-read path EnvVars already uses instead of a parallel mechanism. One
CloneWithSecrets traversal serves encrypt, decrypt and mask so the three
directions cannot disagree about which fields are secret.
Validation runs at save time because the alternative is failing at
sandbox-create time, where the admin no longer has the form in front of them.
It rejects both what a provider would refuse and what a provider would accept
without doing what the admin meant -- a domain allowlist with no deny-all
fallback still leaves every unresolved destination reachable.
* feat(sandbox): apply the network policy on both providers, persist the inbound token
Resolve the stored policy into one fully specified provider-neutral policy so
that nothing downstream re-derives a default. Both booleans come out non-nil on
every path, including the nil-tenant one, because a partially specified policy
would let the adapters fall back to their providers' more permissive defaults.
The three inversions (deny-egress to allow-internet, public-inbound, rule-deny
to rule-allow) happen in exactly one place.
Every sandbox-creating path carries it, so sessions, skill installs and the
deep connectivity check all run under the same config's policy rather than
each wiring its own.
Inbound access defaults to CLOSED, deliberately stricter than either
provider's own default: a sandbox URL reachable by anyone who learns the ID is
not a reasonable default for a tenant's workspace. Closing it makes the
provider issue a per-sandbox traffic token, and only at create time -- so the
token is stored with the session binding and replayed through Connect on every
reconnect. Without that, a sandbox that was paused on TTL or rebound after a
restart would answer 403 on every exec and file read, and 403 classifies as an
authentication error rather than NotFound, so the binding would never be
replaced and the session would stay wedged.
Connect therefore takes a request struct instead of a bare ID, which gives the
token a single documented way in rather than implicit state inside an adapter.
Metadata-based recovery is the one reconnect path with no binding to read the
token from. Rather than adopt a sandbox that would be permanently 403, it is
deleted and replaced -- but only for providers that have the credential at
all, since Docker has no inbound token and its containers are still adoptable.
go-e2b keeps the token and never sends it, so for E2B the data-plane
RoundTripper attaches the header from a process-local registry. Cube's SDK
already does this itself.
* fix(sandbox): classify data-plane traffic by authority shape, not the configured domain
isDataPlane suffix-matched the configured sandbox domain, which was wrong in
both directions. Both required a configured proxy_url, so E2B Cloud's default
usage was unaffected; the affected deployments are exactly the self-hosted and
E2B-compatible ones the gateway exists for.
A configured gateway could be silently ignored. sandbox_domain is optional for
E2B -- MissingRequiredFields asks only for api_key and template_id -- and with
it empty the check returned false for everything. The general form is worse
than the empty case: go-e2b's envdBaseURL prefers the domain the control plane
reported over the client-wide one, so the authority actually dialled need not
appear in the config at all. The failure is quiet whenever public DNS happens
to resolve the sandbox domain, and the operator believes traffic is passing
through a gateway that never sees it.
Control-plane calls could be sent to the gateway. E2B's own defaults put the
API host under the sandbox domain (api.e2b.app beneath e2b.app), so the suffix
match classified every Create, Connect and List as data-plane traffic. That
is loud where the gateway does not serve the API, but where it proxies both
planes it succeeds while downgrading an X-API-Key request to plaintext for an
http:// gateway. Cube was immune to the first problem, since it requires both
fields, but not to this one.
Classify on the "<port>-<sandboxID>.<domain>" shape both SDKs generate, using
the SandboxIDFromDataPlaneHost helper that token injection on the very same
request already relies on -- so routing and injection now agree by
construction instead of applying two different rules. That helper exists
precisely because the configured domain is unreliable here; it was simply
never used for the routing half.
The control host stays as an explicit exclusion. The shape test already
rejects api.<domain>, so this only covers an API host that itself wears the
shape, such as 8080-api.example.com.
One existing assertion was replaced rather than kept: the E2B pool test proved
that the Cube fields were ignored by asserting a cube.app authority was not
data-plane traffic, which tested the domain matching rather than the field
selection. It now asserts the transport dials the E2B gateway address and not
the Cube one.
* feat(sandbox): add the network policy editor to the sandbox config drawer
The policy sits in the runtime step, directly under the settings it
constrains, rather than in a step of its own. The existing execution timeouts
were compressed into a two-column grid to make room, keeping each field's own
label and its own tip -- the reason they were one per row was that three
unlabelled numbers sharing a footnote gave no clue which one you were editing,
and two columns with per-field tips does not regress that.
Provider-specific sections are shown only for the backend that has them: the
Cube L7 rule cards, E2B's per-host header transforms, and for Docker just the
bridge/none selector, since Docker can only express the overall switch.
Injected header values are write-only. A stored one shows as configured and is
preserved when the input is left blank, which is how env-var secrets already
behave. Renaming a rule or a header is the case worth care: the merge indexes
credentials on the rule-and-header pair, so a rename means the old value can no
longer be recovered. One helper decides both the placeholder text and whether
the payload re-attaches the value, so the form cannot claim a credential is
kept while the save blanks it.
The add buttons use keys of their own instead of reusing the env-var table's
"add variable", which read wrongly under a list of network destinations and
which the i18n key check cannot catch, since the key exists in every bundle.
* docs(sandbox): document the network policy and its inbound default change
The existing network bullet described the pre-feature state -- both providers
open by default, tunable only through the neutral policy at create time --
which is no longer true of the defaults or of the admin-facing surface.
Three things needed writing down rather than restating.
The inbound default flips from publicly reachable to credential-required. A
deployment that relies on a browser or an external service reaching a sandbox
port directly, rather than going through WeKnora, has to opt back in on that
config.
The binding store now holds a bearer credential in plaintext. Leaving it
unencrypted was a deliberate trade -- per-reconnect crypto on the hot path
against Redis's own access control -- and the trade is only sound if the
requirement is actually stated, along with the code-side constraint it depends
on: a binding or a handle is never logged whole, only the sandbox ID and
whether a credential is present.
On binding loss, a closed-inbound Cube or E2B config no longer adopts its old
sandbox by metadata. The token went with the binding and the provider will not
reissue it, so adoption would produce a session that answers 403 on every
data-plane call with no way back. The sandbox is deleted and rebuilt, which
costs the /workspace scratch. Docker has no such credential and still adopts.
* fix(sandbox): fail closed when a closed-inbound create has no traffic token
A Cube/E2B sandbox created without the inbound token would be bound anyway,
then 403 on every data-plane call as an unreplaceable authentication error.
Refuse the create, destroy the sandbox, and stop Docker from persisting
inbound/egress radios it cannot honour. Canonicalize IPv4 /0 deny-all onto
0.0.0.0/0 and reject CR/LF in injected header names and values.
* fix(sandbox): satisfy pre-push lint on the network policy files
Unexport the data-plane host parser, drop file-level package docs that
revive rejected, and lowercase error-string prefixes that staticcheck
ST1005 treats as capitalized English.
* fix(sandbox): persist rotated inbound tokens and stop misreporting restricted egress
Connect can reissue the traffic token after pause/resume; write it back onto
the binding so a restart does not restore a stale credential that 403s and
cannot be replaced. Restore the volume-support i18n key, map Docker
network_mode=none onto deny-egress, inject the E2B inbound header without a
gateway pool, and fold header names when merging redacted secrets.
---------
Co-authored-by: hailongzhao <hailongzhao@tencent.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
Provider prefix caches only hit when system/tools stay byte-stable and
cache routing is actually on the request. Stop rewriting those prefixes
mid-turn, attach prompt_cache_key / cache_control, and log hit rate
instead of showing it in the conversation tree.
Large write/edit calls used to sit on a spinner until the full JSON
arrived. Emit path and running line counts from partial arguments so
the UI can update in place, without putting file bodies on the SSE wire.
Refuse truncated tool calls, compact-and-retry once on overflow, and let
sandbox files be appended, paged, and batch-edited instead of rewritten.
Models can set context_window (default 200K) so compaction sizes itself
from the real window; agent max_iterations of -1 means unlimited.
* fix(models): show usage details when deletion is blocked
* fix(models): open the matching agent section and cap usage lists
Agent "open configuration" now lands on the tab that owns the binding
(VLM/ASR vs follow-ups), and delete failures keep the server message when
structured details cannot be parsed. Listed KB/agent rows are capped at 50
while totals still drive the deletion guard and dialog headings.
---------
Co-authored-by: Leon <leon_hfut@foxmail.com>
Co-authored-by: wizardchen <wizardchen@tencent.com>
Those /skills-sh/owner/repo/slug listings are federated, not native
@owner/slug pages, and the GitHub folder is often deeper than the URL
slug. Resolve them through ClawHub so the commit-pinned handoff names
the real skill path.
- Introduced StopSkill method to allow users to abort ongoing skill installations, enabling retry or uninstall options.
- Updated API documentation to include the new stop endpoint for skills.
- Enhanced frontend components to support stopping skills, including UI updates and user feedback messages.
- Added tests to ensure the StopSkill functionality works as intended, including handling various skill states.
- Updated internationalization files to include new messages related to stopping skills.
- Added write_skill_file and edit_skill_file tools to facilitate writing and editing files directly within the skill directory during installation.
- Updated the agent service to register these tools exclusively for the built-in skill installer, ensuring they operate within the correct scope.
- Enhanced error handling and validation to prevent unauthorized file access outside the designated skill directory.
- Improved documentation and prompts to clarify usage and limitations of the new tools, emphasizing the importance of proving import resolutions by executing them.
- Added tests to validate the functionality and security of the new file management tools.
- Updated .env.example to clarify skill zip upload limits and their implications for Nginx configuration.
- Refactored frontend logic to differentiate between skill zip uploads and other uploads, ensuring accurate error messages for oversized requests.
- Improved utility functions for managing upload limits and added tests to validate behavior for oversized JSON bodies.
- Enhanced error handling in various handlers to provide clearer feedback on upload size constraints, particularly for sandbox and catalog installations.