* fix(sandbox): avoid shell in FsBridge writes
Replace sh -c with interpolated path by shell-free 'tee -- <path>' argv form,
piping content via stdin. Prevents command injection through filenames
containing shell metacharacters inside the sandbox container.
Co-authored-by: evgyur <evgyur@gmail.com>
* fix(security): fail-closed on pairing DB errors across channels
On IsPaired lookup error, deny instead of granting access. Covers the shared
CheckDMPolicy/CheckGroupPolicy helpers (Slack/Discord/Feishu/WhatsApp/Zalo) and
the four inline Telegram pairing checks.
Co-authored-by: Srini <srinis.k@gmail.com>
* fix(security): harden provider URL validation against SSRF
Enforce scheme check for all provider types; restrict local types (ollama,
claude_cli, acp) to an explicit localhost allowlist instead of skipping checks;
resolve remote hostnames and reject any IP in a private/reserved range via the
shared security.IsBlocked CIDR list (covers loopback, link-local, metadata,
multicast, and unspecified 0.0.0.0/::). Closes the wildcard-DNS bypass and the
local-type escape hatch. Operator opt-in via GOCLAW_ALLOW_PRIVATE_PROVIDER_URLS.
Exports security.IsBlocked as the single source of truth for blocked ranges.
Co-authored-by: Linh Vo Van <linh.vo@e-cq.net>
* feat(pipeline): add fail-closed tool call authorization gate
Gate tool execution against the server-side AllowedTools allowlist built from the
RBAC/tenant-aware filtered tool set. Resolve the tool-call prefix before the
allowlist lookup so prefixed agents are not wrongly blocked, re-check deny on lazy
MCP activation, and expand IsDenied to cover aliased tool names.
Co-authored-by: Huy Doan <tui@pm.me>
* fix(security): expand file-serve deny-list defense-in-depth
Add absolute-path deny prefixes (/home, /Users, /srv, /var/lib, /var/www, /opt)
and an explicit fail-closed log when no file-serving boundary is configured.
Co-authored-by: Linh Vo Van <linh.vo@e-cq.net>
* fix(providers): allow claude cli executable paths
Refs: #1185
---------
Co-authored-by: evgyur <evgyur@gmail.com>
Co-authored-by: Srini <srinis.k@gmail.com>
Co-authored-by: Linh Vo Van <linh.vo@e-cq.net>
Co-authored-by: Huy Doan <tui@pm.me>
Squash merge PR #115 after resolving changelog and SQLite migration-map conflicts with current dev. Renumbered channel-context PostgreSQL migration to 000075 and bumped PG required schema to 75 plus SQLite schema to 44 so it follows the run timeline migration. Local checks passed: go test ./..., go build ./..., go build -tags sqliteonly ./..., go vet ./..., and pnpm -C ui/web build. PR CI run 26705617311 passed release-versioning, go, and web.
Squash merge PR #114 after resolving changelog and pipeline input conflicts with current dev. Local checks passed: go test ./internal/agent ./cmd ./internal/channels/..., go build ./..., go build -tags sqliteonly ./..., and go vet ./.... PR CI run 26705332779 passed release-versioning, go, and web.
Squash merge PR #113 after resolving the project changelog conflict with current dev. Local checks passed: Go store/http/gateway/agent/pipeline tests, SQLite-tagged tests, both Go builds, web Vitest, and web build. PR CI run 26705098712 passed release-versioning, go, and web.
* feat(mcp): filter tools at registration + detect FastMCP session reset with force-reconnect
Two foundational MCP reliability improvements:
(1) Tool allow/deny filtering at BridgeTool registration: Previously the runtime grant-check at execute time surfaced "grant revoked" errors when the LLM called a tool it wasn't allowed to call. Filter upfront in both the per-agent registration path (manager_connect.go) and per-user registration path (loop_mcp_user.go). Adds tool_filter.go with IsToolAllowed + tests. This eliminates the "registered then runtime-denied" loop.
(2) FastMCP session reset detection + force-reconnect: FastMCP/Python mcp servers reject tools/call as "invalid during session initialization" when the upstream session lifecycle resets but our pool still holds the old Mcp-Session-Id. Detector matches three known phrasings (FastMCP, mcp-go, mcp-go transport ErrSessionTerminated) in session_reset.go. On detection, BridgeTool.Execute requests a force-reconnect via atomic CAS dedup so N concurrent failing calls collapse to one reconnect. Health loops skip ping while pending so a server answering ping in "initializing" state cannot clobber connected=true before the fresh Initialize completes. Includes 30s timeout, structured slog telemetry, concurrent CAS dedup test.
Files: tool_filter.go + tool_filter_test.go (new), session_reset.go + session_reset_test.go (new), manager_connect.go (connectServer/connectViaPool signatures + registerBridgeTools/registerPoolBridgeTools filter logic + reconnPending skip), manager.go (connectAndFilter + connectServer call signature changes), loop_mcp_user.go (filter-at-register block + WithForceReconnect wiring), pool.go (reconnPending skip), bridge_tool.go (session reset detection + WithForceReconnect callback).
* fix(mcp): self-heal grant cache + bypass per-user grant for system/empty userID
Two production fixes for "MCP tool: grant revoked" recurring on song-nhi-v2.
(1) System-user bypass in ListAccessible: Registration uses LoadForAgent(ctx, agentID, "") while execute uses IsAllowed(ctx, agentID, "system", ...). The LEFT JOIN on mcp_user_grants could match a stale disabled row keyed user_id='system' and silently filter the server out only at execute. Skip the join entirely for synthetic owner identities (userID="" or "system") so registration and execute see the same set. Applied to both PostgreSQL (mcp_servers_access.go) and SQLite (mcp_servers_access.go).
(2) Grant-checker no-cache on empty allowByServer: grant_checker.loadEntry now skips the cache write when allowByServer is empty. Without this, a single transient empty result pinned permanent denial until a bus invalidate fired. Re-queries until the empty condition clears, then caches normally. Includes TestStoreGrantChecker_EmptyEntryNotCached.
* fix(exec): surface chdir error instead of misleading fork/exec
Linux Go forkAndExecInChild conflates chdir + execve failures into one
PathError naming only argv0. When agent workspace points to a stale
directory, exec reports "fork/exec /usr/bin/gh: no such file" even though
the binary exists — the real failure is chdir on cmd.Dir.
Two fixes:
- loop_context.go: fall back to l.workspace when per-user MkdirAll fails
(3 sites: user workspace, dispatched team, auto-resolved team)
- shell.go + credentialed_exec.go: preflight cwd via validateExecCwd
before exec.Command, so the error names the directory, not the binary
Regression test TestValidateExecCwd covers empty / existing / missing /
file-instead-of-dir.
* feat(upgrade): auto-normalize stale agent workspace paths
Detects non-portable workspace values left by deployment migrations
(Docker → bare-metal, host path drift) and rewrites them to the current
configured base. Registered as data hook 072_normalize_agent_workspaces;
runs automatically after `migrate up`.
Stale patterns:
- /app/workspace/* (Docker container path persisted on bare-metal)
- ~ prefix (Go doesn't expand tildes; MkdirAll creates literal ~ dir)
Base resolved with gateway precedence: GOCLAW_WORKSPACE env >
config.Agents.Defaults.Workspace > default ~/.goclaw/workspace.
Idempotent (skips rows already at proposed value). Conservative
(leaves absolute non-stale paths untouched as intentional custom config).
Resolve the PR #8/#9/#10 stack on current dev, including Bitrix24 install callback hardening, migration renumbering, duplicate-domain fail-closed routing, UI textarea/mobile cleanup, and review hardening.
- Use maps.Copy in hooks dispatcher instead of manual map loop
- Remove implicit loop variable capture in router_abort_test (Go 1.26 semantics)
- Use range without index in http_test where index unused
- Use min() builtin in script_test instead of manual min computation
- Add vault_documents.chat_id + composite index (migration 000056)
- Filter vault_search by chat_id when team.workspace_scope=isolated
- Stamp chat_id on AfterWrite/AfterWriteMedia for isolated teams
- Deny cross-chat vault_read in isolated teams (M2 fix)
- RunContext.TeamIsolated flag resolved once per run
- Fallback WorkspaceChatID → ChatID in loop_context for entry points
that don't set WorkspaceChatID explicitly (WS direct, HTTP, cron)
Fixes cross-chat doc leak where agent in chat A could see vault docs
from chat B within the same isolated team.
* refactor(providers): migrate ToolDefinition.Function to pointer + add image response fields
ToolDefinition.Function becomes *ToolFunctionSchema with omitempty so native tool types (image_generation, web_search, etc.) can be declared without a function body. All 9 internal construction sites updated. CleanToolSchemas refactored — function-shape cleaning extracted into cleanFunctionSchema helper, outer pass-through handles native tools.
Added image response fields needed by the next commits: ChatResponse.Images, StreamChunk.Images, ImageContent.Partial (distinguishes partial frames from final images).
* feat(providers): native image_generation for Codex + OpenAI-compat tracks
Codex native (POST /codex/responses): emit image_generation tool object in request tools[] (type, action, model, output_format, partial_images). Handle SSE events response.image_generation_call.partial_image + response.output_item.done (type image_generation_call) + response.completed output[] walk for non-stream. Dedup per item_id. Extend codexSSEEvent/codexItem with output_format, result, partial_image_b64, partial_image_index.
OpenAI-compat (/v1/chat/completions): serialize ToolDefinition{Type:'image_generation'} as {type:'image_generation'} pass-through. Parse choices[0].message.images[] + delta.images[] (data URLs) via new parseDataURL helper; append to ChatResponse.Images.
ProviderCapabilities.ImageGeneration flag; Codex provider + adapter set true. Other providers default false.
* feat(agent,http,store): persist assistant images + tri-level image_generation gate
Agent loop tri-level gate: (provider capability) AND (AgentConfig.AllowImageGeneration, default true, stored in other_config.allow_image_generation) AND (request lacks x-goclaw-no-image-gen header). Gate in loop_tool_filter.go appends ToolDefinition{Type:'image_generation'} only when all three pass. Per-request opt-out parsed in chat_completions.go and propagated via RunRequest.NoImageGen.
Media persistence: persistAssistantImages writes final images (Partial:false) to {workspace}/media/{sha256}.{ext}, returns MediaRef entries, clears inline Images[] from the message. Idempotent on hash, traversal-safe, symlink-guarded. Invoked from pipeline.FinalizeStage via new Deps.PersistAssistantImages callback — covers both stream-final and non-stream paths.
Agent store reads AllowImageGeneration from other_config JSONB with absent/nil = true default (matches V3Flags pattern). No DB migration — code-only default.
* feat(ui/web): image_generation toggle + streaming placeholder + download filename
Composer chip 'Images' visible only when active agent's provider has ImageGeneration capability. Per-agent localStorage persistence via useImageGenToggle hook. When off, sends noImageGen:true to WS chat method (maps to x-goclaw-no-image-gen on upstream HTTP call path).
ActiveRunZone renders a skeleton placeholder while streaming partial_image frames arrive. MediaGallery assigns generated-YYYYMMDD-HHmmss.png as the download filename for hex/UUID PNGs.
i18n keys added to en/vi/zh chat.json: imageGenToggle, imageGenGenerating, imageGenDownloadName. 8 vitest tests for the toggle hook.
* docs: add Image Generation section to codebase-summary + changelog entry
Documents the new native image_generation pipeline across providers layer (Codex + OpenAI-compat), agent gate, media persistence, and web UI surface.
* fix(ui/web): match Codex-routed providers for image_generation toggle
Image-gen toggle visibility was hard-coded to provider id 'chatgpt_oauth' but real Codex-routed agents in production use provider ids like 'cliproxy-codex'. The toggle never rendered.
Replace the Set-has check with a small helper that accepts the literal ids plus any provider string containing 'codex' (case-insensitive). Same logic applied in both chat-input.tsx (composer chip) and chat-page.tsx (streaming placeholder gate).
Verified against a live Codex-routed agent: toggle now renders, noImageGen:true propagates on toggle-off.
* docs(pr-1002): targeted-mode UX evidence report
Captures the UI integration surface for native image_generation against a live Codex-routed agent on the remote dev backend.
Includes: composer toggle chip (rendered), streaming skeleton placeholder, and honest failure-path capture showing the legacy create_image builtin fallback. Self-contained HTML report in .github/pr-assets/1002/index.html.
* fix(permissions): classify sessions.compact as write method
CI RBAC-drift test (TestMethodRole_DriftCoverage_AllProtocolMethodsClassified) was failing because the new sessions.compact method added upstream was not classified in any of isPublicMethod / isAdminMethod / isWriteMethod / isReadMethod.
Sessions compaction mutates session history (compacts messages into summaries), so it belongs with the other sessions.* write methods.
* fix(tests): remove duplicate contains() in integration package
Both tts_gemini_live_test.go and mcp_grant_revoke_test.go declared a file-local func contains(s, substr string) bool with identical bodies, causing 'contains redeclared in this block' at compile time in the integration job.
Replace all call sites with strings.Contains (same semantics, stdlib) and drop the duplicates. No behavior change.
* feat(providers): NativeImageProvider interface + Codex implementation
Defines a provider-level contract (NativeImageProvider.GenerateImage) that OAuth-backed providers can implement to serve image generation without exposing static API credentials. Re-uses the PR's Track A native wire format (POST /codex/responses with image_generation tool, item.result decoding, SSE fallback).
CodexProvider + CodexAdapter implement it. Also adds MediaRef.Prompt field so downstream layers can propagate the generating prompt alongside the asset.
* feat(tools): route create_image through NativeImageProvider for OAuth providers
create_image.callProvider now checks for a NativeImageProvider implementation before the credentialProvider interface. When the provider chain points at a Codex-family provider (no static API key), the tool delegates to the provider's GenerateImage which executes the native ChatGPT Responses API image_generation flow.
Resolves 'provider X does not expose API credentials required for image generation' errors for openai-codex / cliproxy-codex chains. On success the tool embeds the user's prompt as a PNG tEXt 'Description' chunk (file-local pngEmbedPrompt helper to avoid tools→agent import cycle), writes the image to /tmp, and threads the prompt through result.MediaPrompts for downstream MediaRef propagation.
* feat(agent,pipeline): propagate image prompt through MediaRef + PNG tEXt embed helper
Adds EmbedPNGPrompt public helper in internal/agent/png_metadata.go that inserts a tEXt 'Description' chunk (plus 'Software: goclaw') into PNG byte streams before the IEND chunk. Non-PNG inputs are passed through unchanged — the helper never errors on unknown formats.
FinalizeStage wires MediaResult.Prompt (from create_image tool output) onto MediaRef.Prompt so the UI can render the generating prompt alongside the image. Per-image prompt list threaded via pipeline RunState.
* feat(ui/web): show generating prompt as caption under each image in MediaGallery
When a MediaRef carries a prompt, MediaGallery renders it as a muted italic caption (line-clamp-2) beneath the image with the full text in the title tooltip. Caption is hidden when the prompt is absent so non-assistant images (user uploads, legacy data) look unchanged.
MediaItem + session media_refs types extended with an optional prompt field; the chat-message adapter threads ref.prompt through when converting WS payloads to UI state.
* fix(providers/codex): stream:true + instructions for native image_generation
The ChatGPT Responses API on /codex/responses rejects two things hard:
- stream:false → HTTP 400 "Stream must be set to true"
- missing instructions → HTTP 400 "Instructions are required"
buildNativeImageRequestBody now sets stream:true and a purpose-specific instructions string ("Generate an image matching the user's description using the image_generation tool. Return only the image; do not describe it in text."). The existing parseNativeImageSSE path was already in place for stream parsing; routing changed from the non-stream branch to the SSE branch.
Regression assertions added to TestCodexGenerateImage_BuildsNativeRequest so these two fields can't silently regress.
* feat(providers,tools,ui): image_model whitelist (gpt-image-2 default, gpt-image-1.5 legacy)
Replaces the hardcoded "gpt-image-2" literal in buildNativeImageRequestBody with a user-configurable field threaded through NativeImageRequest.ImageModel. The whitelist is enforced by ValidateImageModel which rejects anything outside {gpt-image-2, gpt-image-1.5} with a clear error — prevents silent upstream 400s from model names the Responses API would reject.
create_image.callProvider reads params.image_model from the chain entry and threads it through. Empty / absent falls back to DefaultImageModel (gpt-image-2).
UI: added an 'Image model' select inside the existing openai-codex Settings panel on the Create Image Provider Chain dialog. Options: Default · gpt-image-2 (recommended) and Legacy · gpt-image-1.5. i18n keys in en/vi/zh tools.json under builtin.mediaChain.
Tests: TestCodexGenerateImage covers default/legacy/rejected model cases; TestCreateImageTool_ThreadsImageModel covers params→request threading with empty/legacy/explicit sub-cases.
* fix(tools): raise media chain default timeout to 600s/1 retry for image gen
gpt-image-2 on complex prompts (dense Vietnamese text, infographic layouts) legitimately takes 4–8 minutes to complete. The previous default of 120s × 2 retries routinely died mid-generation with 'context deadline exceeded' — the upstream run was still producing bytes when our ctx cancelled.
Default is now Timeout: 600 / MaxRetries: 1. Retries dropped to 1 because image generation is stateful per upstream run: a mid-flight timeout leaves orphan server work, and retrying a fresh generation doubles cost for no gain. Surface the failure fast so operators can widen the timeout instead.
Operators can still set a tighter value explicitly via the Chain dialog.
* refactor: remove user-facing Images toggle, keep admin-level AllowImageGeneration
The per-request opt-out toggle (composer chip + streaming placeholder + noImageGen header plumbing) was a support footgun — users toggle OFF, forget, then can't generate images and think it's broken. Removed in full.
Kept: AgentConfig.AllowImageGeneration (admin kill-switch, stored in other_config.allow_image_generation, default true). Tri-level gate simplifies to two tiers: provider capability AND agent config allows.
Removed: useImageGenToggle hook, IMAGE_GEN_PROVIDER_IDS set in chat-input, supportsImageGenProvider helper, agentProvider/agentKey props on ChatInput, showImageGenPlaceholder prop on MessageBubble/ActiveRunZone/ChatThread, noImageGen param on use-chat-send, parseNoImageGen in chat_completions.go, NoImageGen on RunRequest, no_image_gen_header_test.go, imageGenToggle/imageGenGenerating i18n keys. Kept imageGenDownloadName — used by MediaGallery for generated-\*.png filename resolution.
* docs(pr-1002): refreshed UX trace + updated codebase notes
Replaces the earlier stealth-state evidence with a clean three-capture trace from a real successful run: inline image + prompt caption, MediaGallery lightbox expansion, Chain dialog with the new Image model dropdown open. Skill-routing rows scrubbed from the capture — they reflect per-agent skill setup, not anything this PR introduces.
codebase-summary + changelog: reflect final state (toggle removed, image_model selector, 600s default chain timeout, gpt-image-2 as quality baseline).
* fix(pipeline): preserve mid-loop image_generation output across iterations
FinalizeStage previously read state.Think.LastResponse.Images, which holds
only the last iteration's response. If the LLM emitted image_generation_call
in iteration N alongside a function_call, then responded text-only in N+1,
the image from N was silently dropped on finalize.
Accumulate final (non-partial) images into state.Observe.AssistantImages
across every iteration via ObserveStage, and source FinalizeStage from the
accumulator instead of LastResponse. Partial streaming frames are filtered
defensively; response.Images is cleared on drain to prevent double-counting
on re-exec.
* test(pipeline): regression coverage for mid-loop image accumulation
Six ObserveStage cases covering image accumulation semantics:
- single-iter image-only, image+tool_call same iter, mid-loop image
surviving text-only final iter, multiple images across iters, partial
frame filtering, nil-response safety.
Two FinalizeStage cases verifying accumulator is the source of truth:
- PersistsFromObserveAccumulator: image in Observe + empty LastResponse
must still be persisted via PersistAssistantImages.
- NoPersistWhenAccumulatorEmpty: no call when no images were emitted.
---------
Co-authored-by: viettranx <viettranx@gmail.com>
- Change default mode from "off" to "cache-ttl" (enabled by default)
- Increase softTrimRatio from 0.25 to 0.3 to match TS defaults
- Update tests to reflect new default-enabled behavior
* fix(vault): prevent vault_read id-namespace collision
vault_search was leaking KG/episodic entity ids into result sets even when
narrow `types` were requested, and callers then passed those ids to
vault_read which returned a generic "document not found". The cause was
threefold:
1. `types` filter was only applied to the vault fan-out; KG and episodic
ran unconditionally. Now gated by shouldFanout(types, key).
2. vault_search output lacked a per-source tool hint. Each result now ends
with " → use <tool>" naming the correct follow-up (vault_read,
knowledge_graph_search, or memory_search).
3. vault_read miss returned "document not found" without checking whether
the id belonged to a foreign namespace. It now probes KG then episodic
and returns a namespace-specific redirect error. Stores are injected
via SetKGStore/SetEpisodicStore, nil-safe, tenant-scoped.
Adds red→green characterization tests plus an end-to-end integration
scenario seeding a vault doc + KG entity with identical basenames.
* test(agent): bump none-mode prompt size budget to 3100
vault_read wiring (#948) added ~95 chars to read_file tool summary,
pushing none-mode prompt from <3000 to 3075 chars. Bump budget to
3100 (~775 tokens) to match the intentional addition.
* test(integration): ensure data_migrations table exists in reset helper
The reset helper runs before RunPendingHooks, but RunPendingHooks is
what normally creates data_migrations. On a fresh CI database the
DELETE fails with 'relation does not exist'. Create the table
defensively so reset works regardless of execution order.
* refactor(vault): per-source id fields + wire episodic into search
Align vault_search output fields with downstream tool input params:
doc_id (vault_read), entity_id (knowledge_graph_search), episodic_id
(memory_expand). Prevents LLMs from pattern-matching a generic `id:`
and misrouting a foreign-namespace uuid into vault_read. Fallback
redirect in vault_read now quotes id + names the correct param so
the LLM can self-correct in one turn.
Also wire stores.Episodic into VaultSearchService (stale comment
claimed pending-impl; PGEpisodicStore has existed and been in use
since v3). Unifies search fan-out with vault_read namespace probe.
---------
Co-authored-by: viettranx <viettranx@gmail.com>
Models occasionally emit NO_REPLY with trailing underscores, quotes,
markdown emphasis, or short explanations (e.g. `NO_REPLY_`,
`**NO_REPLY**`, `NO_REPLY: offline`). Previous word-boundary check
treated `_` as a word char so `NO_REPLY_` slipped through and was
delivered to end users.
Rewrite IsSilentReply to strip decorative wrappers, then prefix-match
case-insensitively. Token is only rejected when glued to another
word character (e.g. `NO_REPLYING`). Intentional divergence from
upstream #19537 — documented in the doc comment.
Guard unit coverage:
- MessageTargetEnforced helper table (10 session shapes)
- TestMessageToolCrossTargetGuard — 9 scenarios: DM/group same-target
pass, cross-target block, forward opt-in pass with notice; cron,
heartbeat, subagent, team FREE
- TestMessageToolCrossTargetGuard_TraceReplay — anchors the 019d9fcf
production trace against regression
- NoticeFallbackSender — msgBus-less deployment still gets the audit
breadcrumb via t.sender
- NoNoticeOnSendFailure — failed forward does not post a phantom notice
System prompt coverage:
- <current_reply_target> emitted for direct + group, omitted when
ChatID is empty
Pre-existing TestSelfSendGuard/text_to_different_chat_allowed updated
to pass forward=true — the test now correctly isolates the self-send
guard from the new cross-target guard.
Goon