Replace hardcoded 20-entry IANA_TIMEZONES with getAllIanaTimezones()
using Intl.supportedValuesOf (~400 zones). Switch Select dropdowns
to searchable Combobox in cron, heartbeat, and system config.
Add defense-in-depth timezone validation:
- Backend: validate in heartbeat.set handler and SetDefaultTimezone()
- Frontend: isValidIanaTimezone() guard before save in all 3 dialogs
Closes#614
- Add "low confidence" instruction to memory_search tool description
to prevent models from fabricating memories when no results found
- Add dedicated ## Memory Recall section in system prompt (supplements
recency reminder) with clear instructions for memory_search/memory_get
- Update flush prompts: replace YYYY-MM-DD with actual date at runtime,
cleaner append-only wording
- Update AGENTS.md memory privacy section for multi-tenant: remove
implementation details (per-user scoping), keep group chat output
guardrails that work for both shared and isolated memory configs
- RegenerateDialog now handles its own progress via WS events instead of
opening SummoningModal (inline spinner + auto-close on completion)
- Clean up SummoningModal: remove mode/isRegenerate prop, summon-only
- Agent create: default to Predefined, collapse Open type behind toggle
with warning banner explaining per-user context trade-off
- Add 4 new agent presets: Coder, Support, Writer, Translator (en/vi/zh)
- Remove 15 dead summoning.regenerate* i18n keys
- Add providerTypeOf() to extract provider_type via type assertion
(e.g. "chatgpt_oauth") instead of config name (user-set "openai")
- Exclude "compat" providers (openai_compat → OpenRouter/DeepSeek/Groq)
from strict mode and SOUL echo — they proxy to non-OpenAI models
- Fix isOpenAIStrict matching openai_compat incorrectly
GPT models have strong recency bias and lose persona in long prompts.
Extract Style/Vibe sections from SOUL.md and echo them at the end of
the system prompt (~200 chars each) so GPT sees personality traits right
before generating. Only applies to OpenAI/Codex — Claude respects early
system prompt instructions well and doesn't need this.
Single transient errors (e.g. 504 from upstream proxy) no longer
instantly disconnect MCP servers. Requires healthFailThreshold (3)
consecutive failures before setting connected=false and triggering
reconnect. Applies to both Manager.healthLoop and poolHealthLoop.
Prevent invalid UTF-8 from being persisted when auto-setting followups by sanitizing message content and truncating by rune count. Add regression tests for emoji truncation and malformed byte sequences.
Instead of all-or-nothing when MCP tool count exceeds threshold,
keep first 40 tools registered inline and only defer the excess
to BM25 search via mcp_tool_search. System prompt now shows both
inline descriptions and search guidance in hybrid mode.
Also raises skill inline count from 40 to 60 (token limit is the
real bottleneck for skills).
- Add mode prop to SummoningModal (summon vs regenerate) so Edit with AI
shows appropriate text instead of summoning language
- Fix memory document and KG entity detail dialogs using sm:max-w-* to
properly override base sm:max-w-lg from DialogContent
- Expose isFetching from useMCP hook so refresh button animation works
Combine forward+reverse traversal into single recursive branch using
CASE to fix SQLSTATE 42P19 (PostgreSQL parses triple UNION ALL as
left-associative, putting recursive ref in non-recursive term).
Also widen entity detail dialog from max-w-5xl to max-w-7xl.
Replace replaceLastMediaTag with replaceFirstMediaTag across all 5 media
enrichment functions. Forward iteration + first-match produces natural
positional pairing, fixing reversed tag alignment when multiple media
refs exist in one message.
Also fixes same latent bug in enrichDocumentPaths, enrichAudioIDs, and
enrichVideoIDs. Supersedes #608.
Zero temperature was too rigid, causing LLM to miss implied entities
and relations. 0.2 allows picking up contextual connections while
staying deterministic for structured JSON output.
Add 3 new entity types: technology, product, document — reducing
concept catch-all bucket. Add 4 new relation types: authored,
references, provides, requires. Improve prompt with disambiguation
guide between similar types, stricter related_to usage, and varied
confidence examples. Update graph view colors and mass for new types.
Add reverse-edge UNION ALL to recursive CTE so traversal follows both
source→target and target→source edges. Reverse edges prefixed with ~
in via field (e.g. ~manages). Tool output shows directional arrows.
* fix(discord): preserve attachment source URL in media tags (#602)
- Add SourceURL field to MediaInfo struct
- Populate SourceURL from Discord attachment URL in resolveMedia()
- Emit <media:image url="..."> in BuildMediaTags() when SourceURL is set
- Refactor enrichImageIDs/enrichImagePaths with helper functions
- Add comprehensive unit tests for media URL handling
* fix: update system prompt for url attribute + add enrichment regression tests
- System prompt now documents the url attribute in <media:image> tags
- Add TestEnrichImageIDs_BareTag for non-Discord channel enrichment
- Add TestEnrichImageIDs_SkipsAlreadyEnriched for double-enrichment safety
- Add TestEnrichImagePaths_NoDoubleEnrich for historical message safety
- Add TestEnrichImagePaths_AttributeOrderIndependence for url-before-id tags
---------
Co-authored-by: viettranx <viettranx@gmail.com>
grok-imagine-video and grok-2-image were not in the isNonChatModel()
allowlist, causing verify to call Chat API which fails with malformed
error. Also fix friendlyVerifyError() fallback splitting inside JSON
values via LastIndex.
Three layered bugs caused OpenAI-compatible providers to silently
produce empty tool call arguments when max_tokens was hit mid-JSON:
1. FinishReason override: all providers unconditionally overwrote
"length" → "tool_calls" when tool calls existed, preventing the
agent loop's truncation guard from firing.
2. Silent parse failure: JSON unmarshal errors were logged but args
stayed as empty map with no signal to the caller.
3. No fallback for unreliable providers: some proxies don't emit
finish_reason:"length" at all, leaving no detection path.
Fixes:
- Add ParseError field to ToolCall struct for explicit error signal
- Guard FinishReason override with `!= "length"` in OpenAI, Codex
- Set ParseError in all provider parsers (OpenAI, Anthropic, Codex)
- Add hasParseErrors() fallback guard in agent loop for EC-5 scenario
- Cap consecutive truncation retries (maxTruncationRetries=3) to
prevent burning all iterations when max_tokens is persistently low
Closes#605
- Use senderID (id|username) instead of userID in group no-mention path,
preventing duplicate contacts for the same Telegram user
- Use full name (FirstName + LastName) in both contact insert paths
- Show contact_type (User/Group) instead of peer_kind (Direct/Group) in
contacts table TYPE column and filter dropdown
- Add contact_type filter support in HTTP handler, PG and SQLite stores
- Revert leader workspace override: leader keeps personal workspace as
default, team workspace accessible via ToolTeamWorkspaceFromCtx
- Auto-copy: when leader creates team_tasks, scan subject+description
for file paths → copy from personal to team workspace so members can
access them
- Safety: Lstat (reject symlinks), 10MB size cap, .env excluded from
allowed extensions, path traversal blocked
- Prompt hint: clarify members can only access team workspace files,
referenced files are auto-copied
- New /subagents and /subagent <id> commands for viewing subagent tasks
from the persistent DB table
- Inline keyboard with sa: callback prefix for detail view
- Refactor telegram.New() to functional options pattern (WithAgentStore,
WithTeamStore, WithSubagentTaskStore, WithPendingMessageStore)
- Wire SubagentTaskStore via WithSubagentTaskStore option
- Token cost tracking: accumulate input/output tokens per subagent,
include in announce messages and persist to DB
- Per-edition rate limits: MaxSubagentConcurrent/Depth on Edition struct,
tenant-scoped concurrency enforcement in Spawn/RunSync
- WaitAll action: spawn(action=wait, timeout=N) blocks until all
children complete, returns merged summary
- Auto-retry: configurable MaxRetries (default 2) with linear backoff
for transient LLM failures
- Producer-consumer announce queue: merges staggered subagent results
into single LLM run (same pattern as team task announces)
- Raw metadata in bus messages to prevent double-formatting
- Fire-and-forget DB persistence with detached context + tenant scope
- Split oversized files for <200 line compliance
- Migration 000034: subagent_tasks table with tenant scope, JSONB
metadata + GIN index, partial index for archival candidates
- SubagentTaskStore interface with Create/Get/UpdateStatus/List/Archive
- PG implementation with parameterized queries and tenant isolation
- SQLite schema v3→4 migration + no-op stub for Lite edition
- Wire into store.Stores and factories
Compaction summaries were too generic ("provide a concise summary"),
causing loss of task progress, decisions, and identifier corruption
after summarization.
Port from OpenClaw TS (compaction.ts):
- Structured MUST PRESERVE sections: active tasks, progress, last
request, decisions, TODOs, commitments
- Identifier preservation: preserve UUIDs, hashes, URLs, file names
exactly as written (no shortening/reconstruction)
- Prioritize recent context over older history
- Shared prompt constant used by both mid-loop and background
compaction paths
Problem: Agent sessions accumulated 71K+ input tokens (83% history)
because read_file and exec had no output limits. SOUL personality
drowned by massive context.
Changes:
- read_file: add offset/limit params + 50K char output cap with
pagination hints (model can re-read with offset)
- exec/shell: cap output at 30K chars with smart head+tail truncation
(preserves errors/summaries at tail)
- pruning: add per-result 30% context guard, tune softTrimRatio
0.3→0.25 and softTrimMaxChars 4K→3K, add tail-aware soft trim
- mid-loop: allow pruning to re-trigger each iteration (was one-shot)
Design: cap at source, preserve full data in session, prune at
consumption time. read_file offset/limit enables recovery of
truncated content.
Closes#532
- Replace prefix truncation with SHA-256 hash-based shortening for oversized tool call IDs (40-char OpenAI/Azure limit)
- Normalize provider-prefixed model IDs (e.g. openai/o3-mini) before capability checks for temperature and max_completion_tokens
- Add regression tests for ID collision, correlation, and prefixed model routing
* feat(reasoning): add capability-aware effort resolution
- resolve requested reasoning levels against exact model capabilities
- persist requested effort on agents and expose effective effort in traces
- add backend tests for provider models, agent store, and resolution logic
Refs #591
* feat(ui): gate reasoning controls by model capabilities
- only show supported reasoning levels when provider model metadata is available
- preserve expert reasoning selections during async model loading
- surface effective reasoning details in trace dialogs and localized copy
Refs #591
* docs(api): document capability-aware reasoning controls
- describe exact-match capability lookup and downgrade behavior
- update provider model metadata and trace response documentation
- refresh the generated OpenAPI spec for the new reasoning fields
Refs #591
* feat: add provider-first reasoning controls
* docs: refresh PR 593 UI evidence callouts
* refactor: deduplicate reasoning normalize functions and remove PR evidence
- Export NormalizeReasoningEffort/NormalizeReasoningFallback from providers
package; store package now delegates instead of duplicating
- Store reasoning fallback constants alias providers canonical definitions
- Export deriveLegacyThinkingLevel from types/provider.ts; remove local
copies from agent-advanced-dialog and provider-overview
- Remove unused _providerType param from useProviderModels hook
- Fix reasoning debug log to fire for all cases with a reason (not just
non-off efforts)
- Remove docs/pr-593-evidence/ binary screenshots from repo
---------
Co-authored-by: viettranx <viettranx@gmail.com>
Port missing group chat style guidance from OpenClaw TS:
- "Write like a human" — prevents robotic/formal GPT responses
- "Avoid Markdown tables" — GPT tends to spam tables in groups
- "Use real line breaks sparingly"
These instructions exist in OpenClaw TS (groups.ts buildGroupIntro)
but were missing in GoClaw's group prompt.
GPT-4o+ models prioritize "developer" messages over "system" for
instruction adherence. GoClaw was sending "system" for all providers,
causing GPT models to poorly follow SOUL/system prompts.
Map "system" → "developer" only for native OpenAI endpoints
(api.openai.com). Non-OpenAI backends keep "system" role unchanged.
Ported from OpenClaw TS: model-compat.ts → isOpenAINativeEndpoint()
Session tools (sessions_list, session_status, sessions_history, sessions_send)
were using resolveAgentIDString(ctx) which returns the agent UUID, but session
keys are built using agent_key. This caused all session tool operations to fail
silently or return "access denied" for every channel.
Replace resolveAgentIDString() with ToolAgentKeyFromCtx() in all four session
tools. Add fail-closed guard for empty agent key.
* fix(ui): restore provider-owned codex pool inherit state
* docs(pr): add before-after UI evidence
* refactor: remove dead hasProviderDefaults param and harden pool rendering
- Remove unused _hasProviderDefaults parameter from buildDraftRouting
and routingDraftSignature, clean up all call sites and useMemo deps
- Filter deleted providers from selectedPoolProviderNames to prevent
ghost entries when a saved pool member no longer exists
- Add symmetric backend test for inherit + non-nil provider defaults
---------
Co-authored-by: viettranx <viettranx@gmail.com>
Upload handler previously archived skills immediately when deps were
missing. Now calls InstallDeps() first (owner/master tenant only) and
falls back to archive on failure.
Changes:
- Auto-install missing deps during skill upload (same flow as seeder)
- Atomic DB persist: deps state written with CreateSkillManaged in one call
- Per-slug upload mutex prevents concurrent race conditions
- Frontend: warning state (amber triangle) instead of throwing error
- Non-cancellable context for DB write after dep install
- SQLite StoreMissingDeps now works for custom skills (not just system)
- Comprehensive unit + integration tests
Closes#468
- Reject incompatible explicit provider embedding dimensions on create/update
- Ignore previously saved incompatible dimensions at runtime, fall back to 1536
- Share RequiredMemoryEmbeddingDimensions constant from store package
- Remove dimensions input from provider UI (always 1536 per pgvector schema)
- Add HTTP, runtime, and validation unit tests
Closes#548
Supersedes #410
Telegram: senderLabel now includes full name alongside username, e.g.
[From: @nguyennlt (Nguyễn Trần)] instead of just [From: @nguyennlt].
Also uses FirstName + LastName when no username is set.
WhatsApp: add missing [From:] annotation — was the only channel not
annotating sender identity for the LLM.
## Summary
- Agent rename now updates Name field in IDENTITY.md (agent-level + per-user copies for open agents) via both HTTP and WS paths
- Fixes lossy identity rebuild: previous logic reconstructed IDENTITY.md from only Name/Emoji/Avatar, silently dropping LLM-written fields (Creature, Purpose, Vibe, etc.)
- Uses targeted field replacement (bootstrap.UpdateIdentityField) preserving original formatting
- Adds ListUserContextFilesByName to AgentContextStore (PG + SQLite)
- Fixes LastIndex bug in UpdateIdentityField where values containing colons (e.g. Avatar URLs) would break field replacement
- Removes ~335 lines of dead config.json fallback code (agentStore is always set)
- Adds 9 unit tests for UpdateIdentityField covering plain/markdown formats, URLs, edge cases
* fix(cron): Run Now executes correctly and reset stale running status (#568)
- PG store RunJob now sets next_run_at = NULL before executeOneJob so
loadClaimedJob (which requires next_run_at IS NULL) succeeds instead
of silently skipping the forced run
- recomputeStaleJobs on startup resets last_status 'running' → 'interrupted'
for jobs that were mid-execution when the server crashed
- UI: add typed CronJobPatch interface matching backend store.CronJobPatch
(deliverChannel/deliverTo/stateless) — replaces Record<string, unknown>
on updateJob and onUpdate props across cron detail components
- UI: fix enabled/disabled label in overview tab to reflect local state
instead of stale job prop
* fix(cron): add error handling for ExecContext calls per code review
- RunJob: return error if claiming job (next_run_at = NULL update) fails
instead of silently proceeding to executeOneJob which would then skip
- recomputeStaleJobs: log warning on failure, log count of reset jobs
- Add log/slog import to cron_exec.go
* fix(providers): claude-cli ChatSync image format mismatch (#577)
When images are present, Chat() now uses stream-json output format to
match the stream-json input format required by Claude CLI >= v2.1.87.
Also adds concurrent execution guard to RunJob() preventing
double-execution when Run Now is clicked rapidly.
---------
Co-authored-by: viettranx <viettranx@gmail.com>
- Fix hardcoded dark edge color on deselect — use neutral #64748b
- Reduce graph fetch limit 200→50 to match component render cap
- Fix O(n*m) node lookup on theme change — use Map for O(1)
- Remove double fitView (prop + manual call)
- Add KGEdgeData interface replacing `as any` casts for type safety
- Use entityMap (useMemo Map) for O(1) entity lookup instead of O(n) find
- Pre-compute EDGE_STYLE_DEFAULT/FADED constants to reduce GC pressure
- Skip edge object recreation when style unchanged during selection
- Wrap EntityNode in memo() with memoized container style
- Cap force simulation ticks at 200 for graphs with 100+ nodes
Duc Nguyen