load_skill returns SKILL.md in full by design — it is the model's behavioral
contract, and pagination by default would let the model silently miss later
mandatory sections. read_file / readSkillFile / load_skill are therefore on
the spill-exclusion list so their output is never replaced by a disk pointer.
The exclusion only covered half the path. In spillRawOrTruncate, an excluded
tool's result came back from persistIfOversized unchanged (no spill), failed
the SPILL_MARKER_PREFIX check, and fell through to truncateToolResult(8000) —
so an 8261-char SKILL.md was hard-cut through the middle and stamped with a
'[TRUNCATED: ... middle omitted]' marker. Weaker models ignore the attached
fidelity note and fabricate the removed span, inventing tool calls against
endpoints the skill never described.
- spillRawOrTruncate now returns retrieval-excluded results raw; the per-turn
aggregate budget stays the backstop.
- Outsized SKILL.md degrades to resumable pagination instead of an unbounded
inline dump. Never a lossy middle-cut.
buildWikiContext enumerated an agent's bound knowledge-base pages into
the system prompt capped only by maxContextChars (default 10000, sized
for large cloud models). On a small-context model a large KB therefore
consumed a big fixed slice of the window on every turn — the "tool token
estimate fills the context" report in #521 (the growth lands in the
system-prompt bucket, not the tool-schema bucket; wiki tool schemas are
fixed-size and do not scale with file count).
Add a budgeted buildWikiContext(agentId, budgetTokens) overload mirroring
buildRelevantContext: the page enumeration also stops once the estimated
token total exceeds the budget, appending the existing
'... and more (use wiki_list_pages)' hint. AgentGraphBuilder passes the
same prefix budget it already applies to the memory block; the legacy
Integer.MAX_VALUE path keeps chars-only behavior for large models.
Tests cover null-budget (all pages), token-budget truncation, and
zero-budget skip.
Wire MCP standard notifications/progress into the existing SSE stream so long-running MCP tool calls surface live progress instead of a bare spinner. A per-call progressToken maps back to (conversationId, toolCallId); ProgressAwareMcpToolCallback injects it into tools/call _meta and calls McpSyncClient directly (falling back to the delegate on error, and applying identity forwarding first). Progress events skip the ring buffer and are replayed from a latest-value snapshot on SSE reconnect. Frontend renders a gradient progress bar in ToolCallSegment when a running tool reports progress.
- ProgressLedgerService.upsert: the reserved-prefix guard rejects the write; the
comment said 'strip the prefix and continue', which no longer matches. Rewrite
it to describe the actual reject behavior.
- ActionNode: translate the class javadoc (including the B2 pinned-constraints and
B5 auto-backfill notes) to English per code style.
The identity forwarded to opt-in MCP servers was a one-dimensional string
(ChatOrigin.requesterId): a MateClaw username for web logins, but a webchat
visitorId for visitors and an IM sender id for IM — indistinguishable to the
REST backend. The signed-token mode (d204b702) made this worse: an RS256
signature over an unauthenticated visitorId reads as "MateClaw authenticated
this user" to any backend that trusts the signature.
Introduce an identity-typing dimension at McpIdentityForwardService:
- classify() branches on ChatOrigin: authenticated (web login, sub=immutable
userId), anonymous (webchat visitor, trust=anonymous), external (IM sender,
trust=external), or none (cron/system → nothing injected, fail-closed).
- mint() adds `trust` and `channel_type` claims; plaintext value is prefixed
`trust:subject` so backends can tell the kinds apart without a JWT.
The immutable userId reaches resolve() without coupling it to the user store:
JwtAuthFilter stamps user.id into auth.setDetails() (both JWT and PAT paths),
and ChatController.memoryOrigin carries it on a new ChatOrigin.requesterUserId
field (only-add, per the record's evolution rule).
Resolves the webchat semantic mismatch raised in #459 and the "sub should be
an immutable user id" follow-up. 82 tests green (4 identity classes covered
with claim assertions + full ChatOrigin/MCP regression).
(cherry picked from commit b5d2cfbf98b39848d7139c743a0b81fea71e8ffe)
Lets an employee pin an ordered fallback chain of (provider, model) entries; the same provider may appear multiple times with different models. Build-time dedup keys on exact (provider, model).
- resolveAgentBasePath: the relative-override branch now normalizes the
resolved path and rejects values that escape the workspace root via "../"
(the absolute branch already did this), keeping attachment/media/tool I/O
contained when an agent's workspaceBasePath is a relative override.
- cleanAttachmentFiles: return early on a null/blank conversationId so a bare
upload root can never be walked and deleted wholesale.
- Translate the chat-upload Javadoc/comments to English (cleanAttachmentFiles,
BaseAgent image-path resolver) per code style.
- Add a resolver test for the relative-override escape fallback.
Strict OpenAI-compatible providers reject the /chat/completions request
with HTTP 400 when an assistant message in history carries a tool call
whose function.arguments is not parseable JSON. Normalize blank or
non-JSON arguments to "{}" at the send chokepoint so streaming,
history-replay, and older-persisted tool calls all stay well-formed.
execute_code (bash/sh/shell) bypassed the workspace boundary guard, so shell
code run through it could read/write/delete paths outside the workspace sandbox
(e.g. cat /etc/passwd) while the same paths were blocked for read_file and the
shell tools. Bring execute_code under the guard (scan only shell-language code,
report the code param), and trust the tool-result spill roots so a legitimate
spilled result stays readable. Adds regression tests.
- Remove mateclaw-ui/package-lock.json and yarn.lock. This is a pnpm
monorepo where pnpm-lock.yaml is the only lockfile; the npm/yarn locks
were stray duplicates. Add a .gitignore rule so they are not committed
again by mistake.
- SourceEvidenceLedger: reference Pattern/Matcher by their imported
simple names instead of inline fully-qualified names.
Backend (SourceEvidenceLedger):
- appendWikiSourceTable now normalizes existing source lines in-place to
canonical "[N] Title - section - page N" format instead of skipping them
- Added replaceSourceLine helper that matches a full source line by regex
and replaces it with the canonical form
- When source lines exist without a "来源：" header, automatically insert
one so the frontend preprocessor can locate the source table
Frontend (useMarkdownRenderer):
- Added data-citation-index / data-citation-title to DOMPurify whitelist
- Added preprocessWikiCitations preprocessor: parses the canonical source
table to build an index-to-title map, replaces [n] markers in the answer
body with clickable <a> links, and wraps entire source-table rows so the
full line is clickable
- Integrated into the render pipeline after wikilink substitution and
before Marked parsing
Frontend (useGlobalWikilinkClick):
- Extended the click delegation selector to match both .wiki-link and
.wiki-citation elements
- Title extraction falls back: data-citation-title || data-wiki-title
Tests: added three test cases for source-line normalization, idempotency,
and automatic header insertion
Live lifecycle board (grid<->board toggle) plus an assignee-swimlane plan
board that groups follow-up re-runs of one goal into a single xN card.
Custom right-side detail/goal panels with markdown output. Fixes plans
being persisted under the per-run trace id so the board actually populates.
Closes#385
Issue #304. Operators who want an agent with NO knowledge base had no
way to express it: leaving the KB picker empty fell through to "inherit
workspace-wide" (every KB visible), so the agent ended up ingesting
every KB's context. This adds the same opt-out toggle that
skills_disabled (V126) / tools_disabled already provide.
Backend:
- V154 migration (h2 + mysql + kingbase): mate_agent.wiki_disabled
BOOLEAN/TINYINT/SMALLINT NOT NULL DEFAULT FALSE. Legacy agents stay
bit-identical.
- AgentEntity.wikiDisabled: Boolean field, @TableField("wiki_disabled").
- AgentBindingService.getBoundKbIds: short-circuit at the top —
wiki_disabled=true returns Set.of() regardless of binding rows. Mirrors
the precedence contract of getBoundSkillIds vs skills_disabled.
- AgentBindingService.setKbBindings: a non-empty save auto-clears a
stale wiki_disabled flag (same contract as setSkillBindings /
setToolBindings on their respective flags). Empty saves leave the flag
untouched — the UI toggle owns the bit, not the binding writer.
- AgentBindingServiceWikiDisabledTest: 5 cases covering all three
return states + the stale-flag auto-clear + empty-save no-op.
Frontend:
- Agents.vue KB picker: add the "此智能体不使用任何知识库" /
"This agent uses no knowledge bases" toggle, mirroring the skills /
tools picker layout. Tab badge shows "Off" when the toggle is on.
- types/index.ts: add Agent.wikiDisabled?: boolean.
- Save logic: when wikiDisabled is on, send an empty KB list (the
setKbs contract then leaves the flag alone server-side, exactly as
setSkills / setTools behave for their opt-out flags).
- i18n (zh + en): new strings for toggle label, hint, badge, and the
scope description shown when the toggle is on.
Stacked on top of #382 (which introduced AgentBindingResolver
.getBoundKbIds). No agent-runtime changes — wiki tools already degrade
cleanly when getBoundKbIds returns Set.of().
Add GET /api/v1/channels/webchat/wiki/pages mirroring /skills, so
downstream integrators can build a [[slug]] picker UI that points the
LLM at specific wiki pages. The picker token format is the universal
Obsidian/Wikipedia wikilink convention; the LLM consumes [[slug]] via
the existing wiki_read_page(slug=...) tool, so no agent-runtime changes
are needed.
- AgentBindingResolver.getBoundKbIds(agentId): three-state mirror of
getBoundSkillIds. null = no rows (fall through to workspace-wide KBs),
Set.of() = explicitly scoped to zero KBs, non-empty = explicit scope.
- WebChatController.listWikiPages: API Key + visitorToken auth chain,
agentId workspace anti-escalation, visibility excludes pageType=
synthesis (LLM intermediate artifacts), 100-page cap forces keyword
filter, response carries only display-level metadata.
- WebChatWikiPageView DTO: kbId/kbName/slug/title/summary/pageType;
content/embedding/sourceRawIds deliberately stay admin-console-only.
- WikiTool.wiki_read_page @Tool description: document the [[slug]]
convention so the LLM treats each token as a wiki-page reference.
- WebChatWikiPageListTest: 8 cases covering happy path, keyword filter,
synthesis exclusion, anti-escalation, auth failures, cap behavior,
and the no-binding → workspace-wide fallback.
Closes#381.
Turn a single natural-language requirement into a ready-to-review
employee: the model proposes name, persona, runtime type and a
validated set of skills/tools/knowledge base, which the user confirms
or tweaks before the agent is created.
- backend: POST /api/v1/agents/generate builds a draft from the
workspace's real capability catalog; every suggested tool/skill/KB is
re-validated against the catalog so nothing hallucinated is offered
- frontend: 3-step wizard at /agents/create reusing the existing
create + binding endpoints; reusable capability picker shows selected
items as compact chips with an on-demand searchable catalog
- goal: continue (not skip) on max-iterations and evidence-insufficient turns.
A max-iterations turn grants a fresh iteration budget ("hard continuation"),
bounded per run and sized into the graph recursion ceiling, so a task too big
for one budget keeps going instead of stalling until the next user message.
- plan-execute: re-plan the remaining work on a step exception, and on a
signature-based stall (repeated failures / identical results / no usable
result) instead of advancing dependent steps with junk; bounded by a per-run
re-plan cap, with a graduated change-strategy nudge before the hard stop.
- plan-execute: auto-derive a goal from a genuine multi-step plan, seeding the
acceptance criteria from the plan steps, so the goal subsystem engages without
the model calling setGoal; broadcast goal_created so the UI hydrates.
- react: refund the iteration for setup-only rounds (load_skill / enable_tool)
so a tight budget is not eaten by the load-then-use two-step.
- ui: re-fetch the active goal when a turn finishes so a goal created or mutated
mid-conversation surfaces without depending on an SSE event.
- streaming: make retry backoff / total-time budget instance fields with a
test-only seam; clarify that the wall-clock budget (not max-retries) bounds a
sustained SERVER_ERROR loop to ~8 attempts, fixing the slow/flaky retry test.
The knowledge-base trust verification recorded wiki citations from every
non-readFile tool response by sniffing its JSON for a top-level title /
pages / chunks field. Tools like getGoalStatus return a top-level title,
which falsely populated the citation set and then forced [n] citations on
the final answer (otherwise flagged EVIDENCE_INSUFFICIENT). Gate citation
mining on the wiki_* tool name instead.
Likewise, the grounded answer contract (cite-or-refuse) was appended to
every ReasoningNode call unconditionally, degrading general agents that
have no knowledge base. Append it only when the agent has a wiki_* tool
bound, scoping the strict regime to KB-grounded scenarios.
Adds a regression test asserting a non-wiki tool with a top-level title
creates no wiki citations.
Add ErrorClassificationTest regression cases for the AI-gateway retry
hardening: 5xx-before-4xx ordering (a proxy 502 whose body says
"bad request" stays retryable), Chinese / numeric provider billing
patterns, and DNS / TLS infrastructure-fatal detection.
Fix the cert-trust pattern while adding its test: Java's ValidatorException
emits "PKIX path building failed" with an uppercase PKIX and the error
chain is not lower-cased, so the previous lowercase pattern never matched
— an untrusted/expired cert chain fell through to the retryable
SERVER_ERROR bucket and was retried in vain instead of failing over.
