progress.md

• Completed externalagent foundation phase 10 (shared resolve orchestrator extraction):

  • extracted shared externalagent resolve/approval policy handling into pkg/externalagent/orchestrator.go.
  • WebUI now uses the shared orchestrator for launch-policy preview and approval gating instead of carrying its own private duplicate logic.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'TestHandleResolveExternalAgentSession(CreatesSession|IncludesMatchedPermissionRulePreview|ReturnsPendingApprovalForAskRule|ReturnsPendingApprovalForManualMode|RejectsDeniedByPermissionRule|StartsProcessWhenApproved)|TestApproveExternalAgentPendingRequest(StartsProcessImmediately|AllowsSubsequentResolveForSameSession)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 13 (gateway shared starter adoption):

  • gateway externalagent resolve now also uses the shared externalagent process starter path, so approved gateway launches can immediately create a runtime process instead of stopping at session resolution.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestResolveExternalAgentSessionEndpoint(CreatesSession|ResolvesRelativeWorkspace|RejectsWorkspaceOutsideConfiguredRoot|IncludesMatchedPermissionRulePreview|StartsProcessWhenApproved)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 12 (shared process starter extraction):

  • extracted the common externalagent process-start path into pkg/externalagent/starter.go.
  • WebUI now uses the shared starter instead of keeping a private process-start implementation, making the spawn/continue layer align with the shared resolve orchestrator direction.

• Verification run:

  • go test -count=1 ./pkg/externalagent ./pkg/webui -run 'TestHandleResolveExternalAgentSession(CreatesSession|IncludesMatchedPermissionRulePreview|ReturnsPendingApprovalForAskRule|ReturnsPendingApprovalForManualMode|RejectsDeniedByPermissionRule|StartsProcessWhenApproved)|TestApproveExternalAgentPendingRequest(StartsProcessImmediately|AllowsSubsequentResolveForSameSession)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 11 (gateway shared orchestrator adoption):

  • gateway externalagent resolve now uses the shared pkg/externalagent/orchestrator.go preview path instead of keeping a private duplicate of launch-policy evaluation logic.
  • this is the first cross-consumer adoption of the shared orchestrator layer.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestResolveExternalAgentSessionEndpoint(CreatesSession|ResolvesRelativeWorkspace|RejectsWorkspaceOutsideConfiguredRoot|IncludesMatchedPermissionRulePreview)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 9 (approval handler direct continue):

  • approving an externalagent-origin pending request now immediately calls the same externalagent process-start seam, so approval itself continues launch without waiting for another resolve request.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'Test(ApproveExternalAgentPendingRequestStartsProcessImmediately|ApproveExternalAgentPendingRequestAllowsSubsequentResolveForSameSession|HandleResolveExternalAgentSessionStartsProcessWhenApproved)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 8 (immediate continue-on-approve):

  • approving an externalagent-origin pending request now immediately continues the launch by starting the session process, instead of waiting for a second resolve call.
  • this removes the last obvious two-step approval awkwardness from the WebUI externalagent flow.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'Test(ApproveExternalAgentPendingRequestStartsProcessImmediately|ApproveExternalAgentPendingRequestAllowsSubsequentResolveForSameSession|HandleResolveExternalAgentSessionStartsProcessWhenApproved)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 7 (process spawn/continue seam):

  • WebUI externalagent resolve now attempts to start the session process once the launch is approved or otherwise not blocked by approval.
  • subsequent resolve on an approved session now reuses the same session and ensures the runtime process is started if it is not already running.
  • kept the slice intentionally narrow: this is a shared start-attempt seam with launch metadata/event persistence, not a full cross-consumer externalagent orchestrator.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'Test(HandleResolveExternalAgentSessionStartsProcessWhenApproved|ApproveExternalAgentPendingRequestStartsProcessOnSubsequentResolve|HandleResolveExternalAgentSessionReturnsPendingApprovalForAskRule|HandleResolveExternalAgentSessionRejectsDeniedByPermissionRule)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent consumer phase 3 (gateway consumer wiring):

  • gateway now exposes /api/v1/external-agents/resolve-session as the next real externalagent consumer.
  • this path reuses the same externalagent launch normalization, so canonical launcher allowlist and workspace policy now apply on gateway-owned session resolution too.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestResolveExternalAgentSessionEndpoint(CreatesSession|ResolvesRelativeWorkspace|RejectsWorkspaceOutsideConfiguredRoot|RejectsCommandMismatch)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 6 (approve-after-pending resume seam):

  • approving an externalagent-origin pending request now upgrades that session into approval.ModeAuto, so the same session can be resolved again without immediately re-entering the same pending gate.
  • kept the seam intentionally small: this is a session-level approval override for the externalagent resolve path, not a full process spawn/resume orchestrator.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'TestHandleResolveExternalAgentSession(ApproveExternalAgentPendingRequestAllowsSubsequentResolveForSameSession|ReturnsPendingApprovalForAskRule|ReturnsPendingApprovalForManualMode|RejectsDeniedByPermissionRule)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent consumer phase 2 (WeChat codex consumer wiring):

  • WeChat codex runtime creation now reuses externalagent launch normalization, so workspace policy and launcher allowlist apply on the first non-WebUI consumer path.
  • this is the first real channel consumer beyond the WebUI resolve endpoint.

• Verification run:

  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent foundation phase 5 (real pending approval flow):

  • upgraded WebUI externalagent resolve from a read-only policy preview to a minimal real approval bridge.
  • permission rule deny now rejects the launch with 403; permission rule ask or fallback manual approval mode now return 202 with a pending approval request ID and sync session pending-state into taskStore.
  • kept the slice intentionally narrow: session creation still happens first, and this flow does not yet execute a downstream external-agent process or add approve/deny resume behavior for consumers.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'TestHandleResolveExternalAgentSession(ReturnsPendingApprovalForAskRule|ReturnsPendingApprovalForManualMode|RejectsDeniedByPermissionRule|CreatesSession|IncludesMatchedPermissionRulePreview)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui passed.

• Completed gateway control-plane hardening phase 21 (member remote_addr redaction):

  • member-scoped control-plane reads now redact remote_addr from connection list/detail payloads, while admin/owner visibility stays unchanged.
  • kept this as response-level redaction only; no connection metadata model changes.

• Verification run:

  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui passed.

• Completed externalagent foundation phase 4 (read-only permission/elicitation bridge):

  • extended /api/external-agents/resolve-session with a read-only launch_policy preview so callers can see the current externalagent launcher's approval_mode and permission-rule match result without starting a real approval flow.
  • kept the slice intentionally narrow: no real approval request is enqueued yet, and no channel/gateway runtime consumer behavior changed.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'TestHandleResolveExternalAgentSession(CreatesSession|IncludesMatchedPermissionRulePreview|RejectsUnknownAgentKind|RejectsToolMismatchForKnownAgentKind|RejectsCommandThatDoesNotMatchTool)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui passed.

• Completed gateway control-plane hardening phase 20 (bulk-delete remaining visibility):

  • refined DELETE /api/v1/connections so remaining now reports the caller-visible remaining connection count instead of the global live connection count.
  • this closes a small information leak where member users could infer other users' live gateway connections from the bulk-delete response body.

• Verification run:

  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui passed.

• Completed externalagent foundation phase 3 (launcher allowlist):

  • locked /api/external-agents/resolve-session and externalagent.Manager to a fixed launcher identity allowlist instead of treating the entrypoint as a generic shell-command resolver.
  • supported agent_kind launchers are now codex, claude, opencode, and aider; blank tool/command still default from the canonical launcher, while explicit mismatches are rejected with 400.
  • kept the slice intentionally narrow: no args policy, no aliases, no channel/gateway consumer wiring yet.

• Verification run:

  • go test -count=1 ./pkg/externalagent ./pkg/webui -run 'Test(ResolveSessionRejectsCommandThatDoesNotMatchTool|ResolveSessionAllowsCommandThatMatchesTool|ResolveSessionRejectsUnknownAgentKind|ResolveSessionRejectsToolMismatchForKnownAgentKind|HandleResolveExternalAgentSessionRejectsCommandThatDoesNotMatchTool|HandleResolveExternalAgentSessionRejectsUnknownAgentKind|HandleResolveExternalAgentSessionRejectsToolMismatchForKnownAgentKind)$' passed.
  • go test -count=1 ./pkg/externalagent ./pkg/webui passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui passed.

• Completed externalagent foundation phase 2 (workspace policy gate):

  • made externalagent.Manager config-aware so external-agent session resolution can enforce workspace policy instead of trusting arbitrary absolute paths.
  • blank workspace now defaults to cfg.WorkspacePath(), and relative workspace values now resolve under that configured root before session identity/reuse checks.
  • when agents.defaults.restrict_to_workspace=true, resolved workspaces must stay inside cfg.WorkspacePath(); disabling that flag keeps the existing absolute-path behavior.
  • kept the slice intentionally narrow: no channel/gateway wiring yet, only manager/WebUI policy enforcement and regression coverage.

• Verification run:

  • go test -count=1 ./pkg/externalagent ./pkg/webui -run 'Test(ResolveSessionRejectsWorkspaceOutsideConfiguredRootWhenRestricted|ResolveSessionAllowsWorkspaceOutsideConfiguredRootWhenRestrictionDisabled|HandleResolveExternalAgentSessionRejectsWorkspaceOutsideConfiguredRoot|HandleResolveExternalAgentSessionCreatesSession|HandleResolveExternalAgentSessionReusesExistingSession)$' passed.
  • go test -count=1 ./pkg/externalagent ./pkg/webui passed.

• Completed gateway control-plane hardening phase 19 (member single-connection delete parity):

  • aligned DELETE /api/v1/connections/{id} with the already-landed bulk-delete member semantics.
  • member can now delete only their own live gateway connection; attempts against other users or missing targets return 404 to avoid leaking existence.
  • kept admin/owner semantics unchanged and reused the existing gateway auth/IP/rate-limit boundary instead of introducing new pairing or ownership persistence.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(DeleteConnectionEndpoint(RemovesClient|ReturnsNotFoundForUnknownClient|ReturnsNotFoundForMemberWithoutOwnedTarget|RejectsMemberRoleForOtherUsersConnection|RequiresAuth)|DeleteConnectionsEndpoint(RemovesAllClientsForAdmin|AllowsMemberRoleForOwnedClientsOnly|DeletesOwnedClientsWhenMemberHasOnlyOwnedConnections|RequiresAuth))$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed gateway control-plane hardening phase 18 (pairing source breakdown in status):

  • extended /api/v1/status with paired_generated_connections, paired_requested_connections, and paired_legacy_connections.
  • kept the slice read-only and derived from existing live connection/session-source classification without introducing new pairing persistence.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(StatusEndpointReportsPairingSourceBreakdown|StatusEndpointReportsPairedConnections|StatusEndpointCountsConnectionsDeterministically)$' passed.
  • go test -count=1 ./pkg/gateway passed.

Progress Log

2026-03-26

2026-04-03

2026-04-04

• Continued WeChat presenter/interaction protocol follow-up:

  • aligned the skill-install confirmation prompt copy with the newly supported /select 1 and /select 2 aliases.
  • kept the slice intentionally narrow: copy-only alignment on top of the already-landed /select alias behavior.

• Verification run:

  • go test -count=1 ./pkg/channels/wechat -run 'Test(ParseWeChatInteractionAction|ResolvePendingInteraction(Confirm|Deny|SelectConfirmAlias|SelectDenyAlias)|FormatWeChatPrompt(IncludesSelectAliasesForSkillInstall)?)' passed.

• Continued WeChat presenter/interaction protocol follow-up:

  • extended the existing skill-install confirmation flow so /select 1 now aliases confirm and /select 2 aliases deny, instead of rejecting all select actions as unsupported.
  • kept the slice intentionally narrow: only the existing skill-install confirmation path changed, without broadening into generic multi-option interaction state.

• Verification run:

  • go test -count=1 ./pkg/channels/wechat -run 'TestResolvePendingInteractionSelect(ConfirmAlias|DenyAlias)$' passed.
  • go test -count=1 ./pkg/channels/wechat -run 'Test(ParseWeChatInteractionAction|ResolvePendingInteraction(Confirm|Deny|SelectConfirmAlias|SelectDenyAlias|DelegatesToRuntimeApprovals)|FormatWeChatPrompt|ControlService(BindRuntimeRejectsEmptyChatIDBeforeBinding|SendToRuntimeRejectsEmptyChatIDBeforeExplicitRouting|CreateRuntimeRejectsEmptyChatIDBeforeCreatingSession))' passed.

• Completed Slack interactive flow phase 8:

  • added a seventh real shortcut/modal business closure: start shortcut now opens a modal and submission reuses the existing /start command semantics.
  • kept the slice intentionally narrow by treating the modal as a thin shell over the existing command path instead of inventing Slack-only onboarding behavior.

• Verification run:

  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpensStartModal|ViewSubmissionExecutesStartCommand)$' passed.
  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpens(FindSkills|Settings|Model|Help|Status|Agent|Start)Modal|ViewSubmissionExecutes(FindSkills|Settings|Model|Help|Status|Agent|Start)Command)$' passed.

• Completed Slack interactive flow phase 7:

  • added a sixth real shortcut/modal business closure: agent shortcut now opens a modal and submission reuses the existing /agent command semantics.
  • kept the slice intentionally narrow by treating the modal as a thin shell over the existing command path instead of inventing Slack-only agent behavior.

• Verification run:

  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpensAgentModal|ViewSubmissionExecutesAgentCommand)$' passed.
  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpens(FindSkills|Settings|Model|Help|Status|Agent)Modal|ViewSubmissionExecutes(FindSkills|Settings|Model|Help|Status|Agent)Command)$' passed.

• Completed Slack interactive flow phase 6:

  • added a fifth real shortcut/modal business closure: status shortcut now opens a modal and submission reuses the existing /status command semantics.
  • kept the slice intentionally narrow by treating the modal as a thin shell over the existing command path instead of inventing Slack-only status behavior.

• Verification run:

  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpensStatusModal|ViewSubmissionExecutesStatusCommand|ShortcutOpens(FindSkills|Settings|Model|Help|Status)Modal|ViewSubmissionExecutes(FindSkills|Settings|Model|Help|Status)Command)$' passed.

• Continued WeChat consumer-validation slice:

  • absorbed the runtime-create input-boundary fix so empty chat ids are rejected before any session/process side effects occur.
  • kept the slice consumer-local inside pkg/channels/wechat/control.go instead of reopening shared binding contracts.

• Verification run:

  • go test -count=1 ./pkg/channels/wechat -run 'TestControlServiceCreateRuntimeRejectsEmptyChatIDBeforeCreatingSession|TestRuntimeBindingService|TestControlService(DescribeBindings|CreateRuntime|DeleteRuntime|StopRuntime|RouteMessageToBoundRuntime)' passed.

• Continued browser session dual-mode / reliability follow-up:

  • absorbed the browser mode-preservation slice so internal navigate reuse in screenshot/execute_script now keeps the requested mode.
  • absorbed the browser reliability slice so relative URLs are rejected before any browser startup/navigation attempt.
  • kept the existing select-action correctness helper path (buildSelectScript) in the same working state so the browser package remains self-consistent.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'TestBrowserToolExecuteRejectsRelativeNavigateURL|TestBrowserToolGetTextRejectsRelativeURLBeforeNavigation|TestBrowserToolNavigationParamsPreserveMode|TestBrowserToolStartModeFromParams|TestBrowserToolExecuteRejectsInvalidMode|TestBrowserToolParametersIncludeRelayMode' passed.
  • go test -count=1 ./pkg/tools passed earlier in the same worktree state.

• Continued conversation/thread binding slice 4 (target primary preservation on rebind):

  • absorbed the contract-first shared-layer slice so rebinding a conversation onto a target session no longer steals that target session's existing primary conversation key.
  • preserved the earlier deterministic source-session promotion behavior by updating the test ordering to reflect the new two-sided contract instead of the old single-sided assumption.

• Verification run:

  • go test -count=1 ./pkg/conversationbindings -run 'TestService(RebindingPromotesDeterministicPrimaryConversation|RebindPreservesExistingPrimaryConversationOnTargetSession)$' passed.
  • go test -count=1 ./pkg/conversationbindings ./pkg/channels/wechat passed.

• Continued channel capability matrix phase 5 (Telegram streaming scope):

  • wired pkg/channels/telegram/telegram.go to respect the declared streaming capability before sending the thinking/streaming placeholder message.
  • kept the scope intentionally narrow: Telegram now preserves thinking messages in private chats and skips them in group/supergroup chats, matching the default streaming=dm matrix.
  • added regression coverage so group chats no longer get the streaming placeholder while private chats still do.

• Verification run:

  • go test -count=1 ./pkg/channels/telegram -run 'TestSendThinkingMessageSkipsGroupsWhenStreamingUnsupported|TestSupportsInlineButtonsRespectsDefaultCapabilityScope|TestScopedInlineKeyboardDropsButtonsOutsideSupportedScope|TestSkillInstallPromptFallsBackToTextConfirmationWithoutInlineButtons' passed.
  • go test -count=1 ./pkg/channels/telegram passed.

• Completed Slack interactive flow phase 5:

  • added a fourth real shortcut/modal business closure: help shortcut now opens a modal and submission reuses the existing /help command semantics.
  • kept the slice intentionally narrow by treating the modal as a thin UI shell over the existing command path instead of inventing Slack-only help behavior.

• Verification run:

  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpensHelpModal|ViewSubmissionExecutesHelpCommand)$' passed.
  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpens(FindSkills|Settings|Model|Help)Modal|ViewSubmissionExecutes(FindSkills|Settings|Model|Help)Command)$' passed.

• Completed gateway control-plane hardening phase 17 (paired connection status count):

  • extended /api/v1/status with paired_connections so the control plane can observe pairing volume without fetching the full connection list.
  • kept the slice read-only and derived from existing live connection state only.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(StatusEndpoint|StatusEndpointReportsPairedConnections|StatusEndpointCountsConnectionsDeterministically)$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed gateway control-plane hardening phase 16 (pairing observability):

  • extended gateway connection list/detail payloads with explicit paired and paired_session_id fields.
  • kept the slice intentionally read-only: no changes to pairing protocol, ownership, or auth behavior.
  • made unpaired live connections report paired=false instead of relying on callers to infer it from missing session metadata.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(ConnectionsEndpoint|GetConnectionEndpointReturnsConnectionDetails)$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed gateway control-plane hardening phase 15 (member self-managed delete):

  • refined DELETE /api/v1/connections/{id} so it no longer treats every member as globally forbidden.
  • kept admin / owner semantics unchanged while allowing member to delete only their own live gateway connection.
  • returned 404 for member attempts against other users' connections or nonexistent targets, so the endpoint does not leak connection existence.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestDeleteConnectionEndpoint(AllowsMemberRoleForOwnedConnection|RejectsMemberRoleForOtherUsersConnection|ReturnsNotFoundForMemberWithoutOwnedTarget|RemovesClient|ReturnsNotFoundForUnknownClient|RequiresAuth)$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed channel capability matrix phase 2 (first runtime consumer):

  • extracted the shared capability evaluation/model code into pkg/channelcapabilities so channel runtime packages can consume the matrix without importing pkg/channels and creating a cycle.
  • kept pkg/channels/capabilities.go as a thin compatibility wrapper so existing callers and tests continue to use the same API surface.
  • wired pkg/channels/whatsapp/whatsapp.go to respect the declared native_commands capability before routing slash commands, making WhatsApp's default native_commands=off matrix real at runtime.
  • added a regression test proving WhatsApp now forwards /help ... as plain inbound text instead of executing the command handler when native commands are disabled.

• Verification run:

  • go test -count=1 ./pkg/channels/whatsapp -run TestHandleInboundTreatsSlashCommandAsPlainTextWhenNativeCommandsDisabled failed first, then passed after the fix.
  • go test -count=1 ./pkg/channels/whatsapp ./pkg/channels -run 'Test(HandleInboundTreatsSlashCommandAsPlainTextWhenNativeCommandsDisabled|GetDefaultCapabilitiesForChannel|IsCapabilityEnabled|MergeCapabilities)$' passed.
  • go test -count=1 ./pkg/channels/... passed.

• Continued channel capability matrix phase 3 (Telegram inline button scope):

  • wired pkg/channels/telegram/telegram.go to evaluate the declared inline_buttons capability before attaching inline keyboards to the settings menu and skill-install confirmation flow.
  • kept the scope mapping intentionally narrow: private chats map to dm; group and supergroup map to group.
  • made Telegram's default inline_buttons=dm capability real at runtime, so group/supergroup chats no longer receive those inline keyboards while private chats keep them.
  • added a text fallback for skill-install confirmation prompts when inline buttons are suppressed, so group chats still get a usable /yes /no confirmation path.
  • added regression tests for capability scope evaluation, keyboard suppression outside the supported scope, and the text fallback prompt.

• Verification run:

  • go test -count=1 ./pkg/channels/telegram -run 'Test(SupportsInlineButtonsRespectsDefaultCapabilityScope|ScopedInlineKeyboardDropsButtonsOutsideSupportedScope|SkillInstallPromptFallsBackToTextConfirmationWithoutInlineButtons)$' failed first, then passed after the fix.
  • go test -count=1 ./pkg/channels/telegram passed.

• Continued channel capability matrix phase 4 (WeWork native command scope):

  • added an explicit default capability profile for wework, with native_commands=off instead of falling back to the generic all baseline.
  • wired pkg/channels/wework/wework.go to respect the declared native-command capability before routing slash commands.
  • added runtime and matrix regression coverage so WeWork slash text now stays on the normal inbound bus path when native commands are disabled.

• Verification run:

  • go test -count=1 ./pkg/channels/wework ./pkg/channels -run 'Test(ProcessMessageTreatsSlashCommandAsPlainTextWhenNativeCommandsDisabled|GetDefaultCapabilitiesForChannel)$' failed first, then passed after the fix.
  • go test -count=1 ./pkg/channels/wework ./pkg/channels passed.

• Completed browser session migration phase 3 (relay attach-only mode):

  • added the first narrow relay browser connection mode instead of continuing to reject it at the tool boundary.
  • kept the semantics intentionally strict: relay mode only attaches to an existing browser instance and never launches a new one, so the first slice is real but bounded.
  • added regression coverage for relay mode parsing, relay start behavior with and without an existing instance, and invalid-mode rejection using a truly unsupported value.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'Test(ResolveBrowserMode|BrowserSessionStartWithModeRelay|BrowserToolStartModeFromParams|BrowserToolExecuteRejectsInvalidMode)$' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed conversation/thread binding slice 3 (deterministic rebinding promotion):

  • added a rebinding regression test to prove that when one conversation moves to a different session, the old session promotes its next primary conversation deterministically instead of depending on metadata write order.
  • updated the rebinding cleanup and persistence paths in pkg/conversationbindings/service.go to reuse the same stable state ordering already used by binding queries.
  • verified the existing WeChat runtime consumer still works without adaptation after the rebinding contract tightened.

• Verification run:

  • go test -count=1 ./pkg/conversationbindings -run 'TestServiceRebindingPromotesDeterministicPrimaryConversation$' passed.
  • go test -count=1 ./pkg/conversationbindings passed.
  • go test -count=1 ./pkg/toolsessions ./pkg/conversationbindings ./pkg/channels/wechat passed.

• Completed gateway control-plane hardening phase 14 (pairing consistency hardening):

  • allowed legacy gateway sessions with empty source metadata to be reused during websocket pairing, preserving old gateway session compatibility instead of rejecting them as non-gateway sessions.
  • rejected duplicate live attaches for the same paired session_id and serialized the attach window so concurrent websocket upgrades cannot race into the same paired session.
  • enforced websocket inbound session_id consistency: when clients send a non-empty session id, it must match the active paired gateway session or the message is rejected before routing.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(WSChat(RejectsUnknownRequestedSessionBeforeUpgrade|UsesRequestedExistingGatewaySession|AllowsRequestedLegacyGatewaySessionWithEmptySource|RejectsSecondLiveConnectionForRequestedSession)|ProcessMessage(RejectsMismatchedInboundSessionID|AllowsMatchingInboundSessionID)|ProcessMessageUsesPairedSessionIDForRouterAndResponse)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed Slack interactive flow phase 4:

  • added a third concrete shortcut/modal business closure: model shortcut now opens a modal and submission reuses the existing /model command semantics.
  • kept the slice intentionally narrow by using the existing command path and returning the result as an ephemeral Slack response, instead of inventing Slack-only model inspection behavior.
  • added regression coverage for model shortcut modal open and modal submission command execution, plus a combined regression pack covering all three real modal flows.

• Verification run:

  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpensModelModal|ViewSubmissionExecutesModelCommand)$' passed.
  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpens(FindSkills|Settings|Model)Modal|ViewSubmissionExecutes(FindSkills|Settings|Model)Command)$' passed.
  • go test -count=1 ./pkg/channels/slack ./pkg/commands passed.

• Completed gateway control-plane hardening phase 13 (pairing handshake validation order):

  • fixed the first pairing slice so invalid websocket session_id values are rejected before websocket upgrade instead of after the connection has already been hijacked.
  • added a real websocket-dial regression test to prove clients now receive an actual HTTP 400 for unknown requested sessions.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(WSChatRejectsUnknownRequestedSessionBeforeUpgrade|ResolveGatewaySessionIDRejectsUnknownRequestedSession)$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed gateway control-plane hardening phase 11 (member read-scope correction):

  • corrected the previous endpoint-scope auth slice so member no longer reads global gateway status.
  • narrowed member access on /api/v1/connections and /api/v1/connections/{id} to only the caller's own uid-owned connections, while keeping delete operations restricted to admin / owner.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(Gateway(StatusEndpointRejectsMemberRole|ConnectionsEndpointAllowsMemberRoleForOwnedConnectionsOnly)|AuthenticateRequestAllowsMemberRoleForWebsocketPath|GetConnectionEndpoint(AllowsMemberRoleForOwnedConnection|RejectsMemberRoleForOtherUsersConnection))$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed gateway control-plane hardening phase 12 (websocket session pairing reuse):

  • added a thin pairing path so /ws/chat?session_id=<gateway-session> can reuse an existing gateway session instead of always minting a transient connection-scoped one.
  • rejected unknown session_id values and non-gateway sessions with 400, keeping the first pairing slice tightly bounded inside gateway-owned sessions only.
  • aligned router input and websocket reply session_id values to the paired gateway session, so reconnect/reuse flows keep a stable session identity.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(ResolveGatewaySessionID(UsesRequestedExistingGatewaySession|RejectsUnknownRequestedSession|RejectsNonGatewaySession)|ProcessMessageUsesPairedSessionIDForRouterAndResponse)$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed gateway control-plane hardening phase 10 (endpoint-scoped REST auth):

  • refined the previous role gate into explicit read/manage scopes in pkg/gateway/server.go instead of treating every REST control-plane endpoint as equally privileged.
  • allowed member tokens to read /api/v1/status, /api/v1/connections, and /api/v1/connections/{id} while keeping DELETE /api/v1/connections/{id} restricted to admin / owner.
  • preserved websocket chat compatibility by continuing to allow any valid authenticated token there, and preserved legacy tokens without a role claim as admin-compatible.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(Gateway(StatusEndpointAllowsMemberRole|ConnectionsEndpointAllowsMemberRole)|DeleteConnectionEndpointRejectsMemberRole|GetConnectionEndpointAllowsMemberRole|AuthenticateRequestAllowsMemberRoleForWebsocketPath)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Continued backlog execution from task_plan.md:

  • re-read the persistent planning artifacts and GoalX run state.
  • confirmed the previous GoalX run is complete, while the next approved mainline remains gateway control-plane hardening.
  • narrowed the next minimal slice to gateway.allowed_ips so gateway gets a shared IP allowlist across websocket handshake and REST control-plane entrypoints.

• Completed gateway control-plane hardening phase 7 (IP allowlist):

  • added gateway.allowed_ips to pkg/config.GatewayConfig and kept it runtime-reloadable.
  • validated allowed_ips entries so blanks and non-IP literals fail fast in config validation instead of deferring to runtime surprises.
  • added a shared checkClientIP() gate in pkg/gateway/server.go and applied it to both websocket handshake and REST control-plane entrypoints.
  • kept the first version intentionally narrow: exact literal IP matching only, no CIDR, no forwarded-header trust chain, no pairing/scope semantics.

• Verification run:

  • go test -count=1 ./pkg/gateway ./pkg/config -run 'Test(Gateway(CheckClientIP(AllowsRequestsWhenListUnset|AllowsConfiguredIP|RejectsUnconfiguredIP)|StatusEndpoint(RejectsDisallowedIP|AllowsConfiguredIP)|WSChatRejectsDisallowedIP)|ValidatorRejects(BlankGatewayAllowedIPs|InvalidGatewayAllowedIPs))$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed gateway control-plane hardening phase 8 (per-IP rate limit):

  • added gateway.rate_limit_per_minute to pkg/config.GatewayConfig and kept it runtime-reloadable.
  • validated negative rate_limit_per_minute values as config errors.
  • added shared per-IP limiter state in pkg/gateway/server.go and enforced it at both REST control-plane and websocket handshake entrypoints.
  • kept the first version intentionally narrow: remote-IP buckets only, no session/user scope, no pairing integration, no advanced eviction policy.

• Verification run:

  • go test -count=1 ./pkg/gateway ./pkg/config -run 'Test(Gateway(RateLimit(AllowsRequestsWhenUnset|RejectsSecondRequestFromSameIP|UsesPerIPBuckets)|StatusEndpointRejectsRateLimitedRequest|WSChatRejectsRateLimitedRequest)|ValidatorRejectsNegativeGatewayRateLimitPerMinute)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed gateway control-plane hardening phase 9 (control-plane auth scope):

  • added a shared gateway JWT parser in pkg/gateway/server.go so gateway can consistently read sub / uid / role from existing WebUI-issued tokens.
  • tightened REST control-plane authorization so /api/v1/status and connection-management endpoints now require admin / owner, instead of accepting any valid JWT.
  • preserved websocket chat compatibility by continuing to allow any valid authenticated token there, and preserved legacy control-plane tokens by treating missing role claims as admin-compatible.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestGateway(StatusEndpointRejectsMemberRole|ConnectionsEndpointRejectsMemberRole|AuthenticateRequestAllowsMemberRoleForWebsocketPath)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed Slack interactive flow phase 3:

  • added a second real shortcut/modal business closure on top of the existing Slack interactive routing skeleton: settings shortcut now opens a modal and submission reuses the existing /settings command semantics.
  • kept the slice intentionally narrow by treating the modal as a thin UI shell over the already-existing settings command, instead of inventing a parallel Slack-only settings protocol.
  • added regression coverage for settings shortcut modal open and modal submission command execution.

• Verification run:

  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpens(Settings|FindSkills)Modal|ViewSubmissionExecutes(Settings|FindSkills)Command)$' passed.
  • go test -count=1 ./pkg/channels/slack ./pkg/commands passed.

• Completed context economy preflight execution telemetry slice:

  • added preflight.applied to the shared preflight decision shape so route metadata can distinguish "recommended" from "actually executed".
  • updated both legacy and blades orchestrators to mark applied=true only when compact_before_run really executes, while warning-only paths stay false.
  • extended websocket route payloads and Chat UI rendering so the frontend can show when preflight compaction was actually applied.

• Verification run:

  • go test -count=1 ./pkg/agent -run 'TestChatWithPromptContextDetailed_(IncludesContextPressurePreview|DoesNotAutoCompressWarningPreflightBeforeModelCall|AutoCompressesCriticalPreflightBeforeBlades)$' passed.
  • go test -count=1 ./pkg/webui -run '^TestChatRouteStateJSONIncludesContextPressureFields$' passed.
  • go test -count=1 ./pkg/agent ./pkg/webui passed.

• Completed gateway control-plane hardening phase 2 (REST auth gate):

  • added a shared requireAuthenticatedAPI() gate in pkg/gateway/server.go.
  • changed GET /api/v1/status and GET /api/v1/connections to require the same JWT authentication path already used by websocket clients instead of exposing control-plane data anonymously.
  • added regression coverage in pkg/gateway/server_test.go for both unauthorized and authorized REST access paths.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(Gateway(StatusEndpointRequiresAuth|ConnectionsEndpointRequiresAuth)|StatusEndpoint|ConnectionsEndpoint)$' failed first, then passed after the fix.
  • go test -count=1 ./pkg/gateway passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed gateway control-plane hardening phase 3 (connection delete endpoint):

  • added DELETE /api/v1/connections/{id} in pkg/gateway/server.go so the control plane can actively terminate a single websocket client connection.
  • kept the endpoint behind the same JWT auth gate used by the other control-plane REST endpoints.
  • added regression coverage in pkg/gateway/server_test.go for successful removal, unknown-client 404, and unauthenticated 401.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestDeleteConnectionEndpoint(RemovesClient|ReturnsNotFoundForUnknownClient|RequiresAuth)$' failed first, then passed after the fix.
  • go test -count=1 ./pkg/gateway passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed gateway control-plane hardening phase 4 (stable connection list metadata):

  • extended gateway clients with connection metadata capture for connected_at and remote_addr.
  • changed GET /api/v1/connections to return a stable, sorted connection list with session_id, connected_at, and remote_addr fields instead of depending on Go map iteration order.
  • added regression coverage in pkg/gateway/server_test.go for sorted connection output and status reporting with active connections.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(ConnectionsEndpoint|StatusEndpointCountsConnectionsDeterministically)$' failed first, then passed after the fix.
  • go test -count=1 ./pkg/gateway passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed gateway control-plane hardening phase 5 (connection detail endpoint):

  • added GET /api/v1/connections/{id} in pkg/gateway/server.go so the control plane can inspect a single websocket connection without fetching the full list.
  • reused the same JWT auth gate and shared connection serialization shape as the list endpoint.
  • added regression coverage in pkg/gateway/server_test.go for success, unknown-client 404, and unauthenticated 401.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestGetConnectionEndpoint(ReturnsConnectionDetails|ReturnsNotFoundForUnknownClient|RequiresAuth)$' failed first, then passed after the fix.
  • go test -count=1 ./pkg/gateway passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed gateway control-plane hardening phase 6 (max connection limit):

  • added gateway.max_connections to pkg/config.GatewayConfig with 0 meaning unlimited.
  • validated negative max_connections values as config errors.
  • added a server-side checkConnectionLimit() guard so websocket upgrades are rejected once the configured limit is reached.
  • added regression coverage in pkg/gateway/server_test.go and pkg/config/path_test.go for limit enforcement and config validation.

• Verification run:

  • go test -count=1 ./pkg/gateway ./pkg/config -run 'Test(Gateway(RejectsConnectionsAboveConfiguredLimit|AllowsConnectionsWhenLimitUnset)|ValidatorRejectsNegativeGatewayMaxConnections)$' failed first, then passed after the fix.
  • go test -count=1 ./pkg/gateway passed.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed conversation/thread binding slice 2 (deterministic query order):

  • added TestServiceBindingQueriesReturnDeterministicConversationOrder in pkg/conversationbindings/service_test.go to lock a real missing contract with TDD.
  • confirmed RED first: ListBindings() returned [chat-b chat-a] when bindings were created in non-sorted order.
  • updated pkg/conversationbindings/service.go so sessionToBindingRecords, ListBindings, and GetBindingsBySession now return a stable order by conversation_id instead of leaking metadata write order.
  • verified the existing WeChat runtime consumer remained compatible without code changes.

• Verification run:

  • go test -count=1 ./pkg/conversationbindings -run 'TestServiceBindingQueriesReturnDeterministicConversationOrder' failed first, then passed after the fix.
  • go test -count=1 ./pkg/conversationbindings passed.
  • go test -count=1 ./pkg/toolsessions ./pkg/conversationbindings ./pkg/channels/wechat passed.

• Completed Slack interactive flow phase 2:

  • extended the Slack API abstraction with OpenView(...) so shortcut handlers can open modals through the existing channel abstraction.
  • implemented the first real shortcut/modal business flow: find_skills shortcut now opens a modal, and find_skills_modal submission re-runs the existing find-skills command with the submitted query.
  • kept the flow narrow by reusing the existing skill-install confirmation message path instead of inventing a second install-confirm mechanism for modals.
  • added regression coverage in pkg/channels/slack/slack_test.go for shortcut modal open and view-submission command execution.

• Verification run:

  • go test -count=1 ./pkg/channels/slack -run 'TestHandle(ShortcutOpensFindSkillsModal|ViewSubmissionExecutesFindSkillsCommand)' failed first, then passed after the fix.
  • go test -count=1 ./pkg/channels/slack passed.
  • go test -count=1 ./pkg/channels/slack ./pkg/commands passed.

• Completed gateway control-plane hardening phase 1 (origin allowlist):

  • added gateway.allowed_origins to pkg/config.GatewayConfig.
  • replaced the fully-open websocket CheckOrigin behavior with a server-level allowlist check.
  • kept requests without an Origin header allowed so existing non-browser clients still work.
  • added regression coverage in pkg/gateway/server_test.go for allowed and blocked origins, and updated pkg/config/path_test.go for the new config field.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestGatewayCheckOrigin(AllowsConfiguredOrigins|AllowsRequestsWithoutOrigin)' failed first, then passed after the fix.
  • go test -count=1 ./pkg/gateway ./pkg/config passed.

• Completed channel capability matrix phase 1:

  • added pkg/channels/capabilities.go to import the core goclaw capability model: capability types, scope parsing, enablement checks, merge behavior, and per-channel default matrices.
  • kept this slice intentionally low risk by landing the shared declaration/evaluation layer first without forcing all channel runtimes to consume it in the same commit.
  • added regression coverage in pkg/channels/capabilities_test.go for default matrices, scope-based enablement, and override merge behavior.

• Verification run:

  • go test -count=1 ./pkg/channels -run 'GetDefaultCapabilitiesForChannel|IsCapabilityEnabled|MergeCapabilities' passed.
  • go test -count=1 ./pkg/channels passed.

• Completed browser advanced extraction phase 5 (get_text):

  • extended pkg/tools/browser.go so the browser tool now exposes a get_text action for plain-text extraction on top of the existing get_html path.
  • imported the lightweight goclaw HTML-to-text slice as a local helper instead of adding a new browser tool type or extra dependency surface.
  • added regression coverage in pkg/tools/browser_test.go for action exposure and HTML tag stripping.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'BrowserToolParametersIncludeGetText|HTMLToTextStripsTags' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed browser advanced extraction phase 4 (extract_structured_data):

  • extended pkg/tools/browser.go so the browser tool now exposes an extract_structured_data action with extract_type modes for all, schema_org, json_ld, and meta.
  • imported the goclaw structured-data extraction slice by adding reusable extraction-script builders and CDP result formatting directly into the existing BrowserTool.
  • added regression coverage in pkg/tools/browser_test.go for action exposure and extractor script composition.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'BrowserTool(ParametersIncludeExtractStructuredData|BuildExtractionScript)' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed browser advanced extraction phase 3 (print_pdf):

  • extended pkg/tools/browser.go so the browser tool now exposes a print_pdf action plus PDF-specific parameters for orientation, header/footer, background rendering, and margins.
  • imported the goclaw BrowserPrintToPDF slice in a nekobot-native form by reusing the existing session manager and output directory instead of adding a separate CDP tool type.
  • added regression coverage in pkg/tools/browser_test.go for print_pdf action exposure and PDF argument construction defaults/overrides.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'BrowserTool(ParametersIncludePrintPDF|BuildPrintToPDFArgs)' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed MaixCAM outbound response routing:

  • re-audited pkg/channels/maixcam/maixcam.go and found slash-command responses were already written back to the active device connection, so the real remaining gap was outbound bus replies being broadcast to every connected device.
  • updated SendMessage to parse maixcam:<device> session IDs and route outbound replies only to the targeted device connection while preserving broadcast behavior when no device target is present.
  • added regression coverage in pkg/channels/maixcam/maixcam_test.go for device-targeted send, broadcast fallback, and session parsing.

• Verification run:

  • go test -count=1 ./pkg/channels/maixcam -run 'SendMessage(TargetsSessionDevice|BroadcastsWithoutTargetSession)|MaixcamDeviceIDFromSession' failed first, then passed after the fix.
  • go test -count=1 ./pkg/channels/maixcam passed.

• Completed memory quality pack phase 4 (embedding cache):

  • added pkg/memory/embedding_cache.go to import the useful goclaw LRU cache idea in a nekobot-appropriate form: caching embedding vectors by input text instead of redundantly caching objects already held by the in-memory store.
  • updated pkg/memory/manager.go so both Add and Search reuse cached embedding vectors for repeated text, reducing duplicate provider calls while keeping the store and search interface unchanged.
  • added regression coverage in pkg/memory/manager_cache_test.go to prove repeated Add and repeated Search on the same text only invoke the embedding provider once.

• Verification run:

  • go test -count=1 ./pkg/memory -run 'Manager(SearchCachesQueryEmbeddings|AddCachesEmbeddingsForRepeatedText)' failed first, then passed after the fix.
  • go test -count=1 ./pkg/memory passed.

• Completed browser session migration phase 2:

  • extended pkg/tools/browser.go so the browser tool schema now exposes a mode parameter with auto/direct options instead of hiding session startup strategy inside tool internals.
  • wired navigate to pass the resolved startup mode into BrowserSession.StartWithMode, so callers can explicitly request direct attach semantics while keeping auto-mode reuse as the default.
  • added regression coverage in pkg/tools/browser_test.go for default/direct mode parsing and explicit rejection of unsupported modes like relay before any browser startup happens.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'BrowserToolStartMode|BrowserToolExecuteRejectsInvalidMode|BrowserSession|ResolveBrowserMode' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed browser session migration phase 1:

  • extended pkg/tools/browser_session.go with explicit auto/direct connection modes instead of only a fixed single-path startup flow.
  • added a reuse-first strategy so browser sessions now try to attach to existing Chrome debug ports before falling back to launching a new headless instance.
  • added regression coverage in pkg/tools/browser_session_test.go for mode parsing, auto-mode fallback-to-launch, and direct-mode reuse of an existing browser instance.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'BrowserSession|ResolveBrowserMode' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed memory quality pack phase 3 (temporal decay):

  • added pkg/memory/temporal_decay.go to import the core goclaw time-aware ranking slice for builtin memory search.
  • extended pkg/memory/types.go with TemporalDecayConfig and SearchOptions.TemporalDecay, then applied temporal decay inside pkg/memory/manager.go before MMR so age-adjusted scores feed later diversity re-ranking.
  • added regression coverage in pkg/memory/search_manager_test.go for pure decay ordering and manager-level search behavior with temporal decay enabled.

• Verification run:

  • go test -count=1 ./pkg/memory passed.

• Completed memory quality pack phase 2 (MMR):

  • added pkg/memory/mmr.go to import the core goclaw MMR re-ranking slice for builtin memory search.
  • extended pkg/memory/types.go with MMRConfig and SearchOptions.MMR, then applied MMR inside pkg/memory/manager.go after raw store search so diversity re-ranking stays isolated from storage code.
  • added regression coverage in pkg/memory/search_manager_test.go for direct MMR ordering and manager-level search behavior with MMR enabled.

• Verification run:

  • go test -count=1 ./pkg/memory passed.

• Completed memory quality pack phase 1 (citations):

  • added pkg/memory/citations.go to import the useful citation-formatting slice from goclaw in a way that fits nekobot's existing memory types.
  • extended pkg/memory/types.go with EndLineNumber, Timestamp, and result-level Citation / AgeInDays fields so later memory-quality slices have a compatible shape.
  • updated pkg/memory/manager.go and pkg/tools/memory.go so both direct memory context rendering and the memory tool render unified citation strings like path#Lx-Ly instead of bare file paths.
  • added regression coverage in pkg/memory/search_manager_test.go and pkg/tools/memory_test.go for citation decoration and display formatting.

• Verification run:

  • go test -count=1 ./pkg/memory ./pkg/tools passed.

• Completed conversation/thread binding migration phase 1:

  • extended pkg/conversationbindings/service.go from a thin bind/resolve wrapper into a reusable binding layer with BindWithOptions, rich BindingRecord views, GetBinding, ListBindings, GetBindingsBySession, and CleanupExpired.
  • kept persistence on top of existing tool-session records to avoid schema churn while still importing the useful goclaw ideas: binding metadata, target kind/placement, conversation view, and expiry cleanup.
  • tightened List behavior so the service only returns sessions that actually match the configured channel + prefix instead of every session from the same source.
  • added regression coverage for filtered listing, metadata-bearing binds, session-based lookup, and expired-binding cleanup; verified WeChat runtime binding tests remain green.

• Verification run:

  • go test -count=1 ./pkg/conversationbindings passed.
  • go test -count=1 ./pkg/channels/wechat passed.
  • go test -count=1 ./pkg/toolsessions passed.

• Completed Tool Sessions / QMD runtime admin smoke pack:

  • added pkg/webui/server_toolsessions_test.go as a dedicated WebUI regression pack for tool sessions.
  • covered owner isolation, attach-token create/consume flow, OTP generation + access login, one-time password consumption, process status/input/output/kill flow, terminated-session archival, and tool-event cleanup.
  • re-used existing QMD handler coverage in pkg/webui/server_status_test.go as the backend smoke baseline for status/update/install/cleanup behavior, so Batch C now has both prompts and runtime-admin smoke coverage recorded.

• Verification run:

  • go test -count=1 ./pkg/webui -run 'ToolSession|QMD|Status|Session' passed.
  • go test -count=1 ./... passed.
  • npm --prefix pkg/webui/frontend run build could not run in the current shell because npm is missing; pnpm is present on disk but also fails because node is not available on PATH.

• Completed Runtime Prompts regression pack and checklist:

  • added pkg/prompts/manager_test.go to cover scope override semantics, disabled prompt/binding filtering, session binding replacement, and render-context separation between system_text and user_text.
  • found and fixed a real bug in pkg/prompts/manager.go: when the same prompt was bound in multiple scopes, Resolve previously let earlier query order win, so global could incorrectly override channel or session; resolution now explicitly prefers narrower scope, then lower priority.
  • added WebUI regression coverage in pkg/webui/server_prompts_test.go for scope override plus render-context fields (channel, route, workspace, custom).
  • added docs/RUNTIME_PROMPTS.md with behavior notes and a reusable smoke checklist, and linked it from README.md.

• Verification run:

  • go test -count=1 ./pkg/prompts passed.
  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/prompts ./pkg/webui passed.

2026-03-25

• Completed Slack interactive flow phase 1:

  • added Slack-side pending interaction state for skill install confirmations, aligned with Telegram/Discord semantics.
  • changed Slack slash-command skill install confirmation from “send a pseudo inbound message” to “store pending interaction, require same-user confirm/cancel, expire after 15 minutes, re-run the original command with confirmation metadata, and update the original Slack message with the result”.
  • introduced a narrow Slack API interface to make the channel logic testable without live Slack I/O.
  • added placeholder shortcut / view-submission routing hooks so later modal/shortcut flows have a stable entry point instead of being hard-coded into the event switch.
  • added regression tests for pending-state storage, confirm execution path, cancel path, and expiry cleanup in pkg/channels/slack/slack_test.go.

• Verification run:

  • go test -count=1 ./pkg/channels/slack passed.
  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/channels/slack ./pkg/commands passed.

• Re-baselined the post-WeChat migration plan:

  • confirmed the latest WeChat migration commits are already pushed to origin/main.
  • updated task_plan.md to mark the WeChat workstream as stage-complete for now and narrowed the next execution target to Slack interactive flow completion.
  • rewrote the Slack backlog item to reflect the real gap: missing pending state, expiry cleanup, message update path, and extensible shortcut/modal routing rather than only “callback exists / not exists”.

• Added WeChat presenter-style output guidance for agent turns:

  • prepended WeChat-specific output rules before user messages so the agent is explicitly told to avoid Markdown and prefer local absolute file paths for rich content.
  • included workspace-root hints in the injected WeChat instructions so generated attachment files have a stable preferred location.
  • added regression tests for presenter prompt assembly.

• Verification run:

  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/channels/wechat ./pkg/wechat/... ./pkg/webui passed.

• Added the first WeChat weak-interaction slice:

  • wired command responses with commands.InteractionTypeSkillInstallConfirm into the WeChat channel.
  • added pending interaction state per WeChat user and command-style confirmation handling for /yes, /no, and /cancel.
  • aligned the confirmation execution path with Telegram/Discord by re-running the command with __confirm_install__ <repo> and skill_install_confirmed_repo metadata.
  • added regression tests for action parsing, pending interaction expiry, confirm execution, deny handling, and prompt rendering.

• Verification run:

  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/channels/wechat ./pkg/wechat/... ./pkg/webui passed.

• Removed the obsolete channel-local WeChat protocol layer:

  • deleted pkg/channels/wechat/protocol.go after confirming channel runtime, send path, and WebUI QR bind flow all use shared pkg/wechat packages.
  • simplified pkg/channels/wechat/channel.go to keep only bot-backed channel glue instead of duplicated client/credential protocol state.

• Verification run:

  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/channels/wechat ./pkg/wechat/... ./pkg/webui passed.

• Completed WeChat channel shared-SDK migration and attachment send pipeline:

  • switched pkg/channels/wechat runtime monitor, typing keepalive, outbound text/image/file/video sending, and QR binding helpers to shared pkg/wechat SDK primitives.
  • replaced rendered markdown image sending from channel-local inline payloads with shared uploader-based image sending.
  • added outbound file-path extraction/cleanup so reply text can promote local absolute paths into WeChat image/video/file attachments while removing those paths from the final text body.
  • aligned credential storage with shared pkg/wechat/types.Credentials.
  • added regression tests for file-path extraction and attachment classification in pkg/channels/wechat/attachments_test.go.

• Verification run:

  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/channels/wechat ./pkg/wechat/... ./pkg/webui passed.

• Completed gua/libc/wechat SDK baseline migration into nekobot/pkg/wechat:

  • added shared types / client / auth / cdn / messaging / monitor / parse / typing / voice / bot packages under pkg/wechat.
  • kept existing pkg/channels/wechat working while introducing the new shared SDK layer, so follow-up channel enhancements can build on stable primitives instead of channel-local protocol code.

• Verification run:

  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/wechat/... ./pkg/channels/wechat passed.

• Re-ordered the WeChat workstream per latest requirement:

  • promoted gua/libc/wechat SDK full migration into nekobot/pkg/wechat as the current feature slice.
  • moved WeChat attachment/file-path send-path enhancement behind the shared SDK migration.

• Re-scoped the next channel migration slice to WeChat SDK/send-path improvements:

  • switched reference source from goclaw to gua for WeChat-specific presenter / formatter / upload behavior.
  • identified the highest-value low-risk gap in nekobot: outbound replies cannot yet turn local file paths into WeChat image/video/file attachments.
  • updated task_plan.md to prioritize a WeChat attachment send pipeline before broader Slack interaction work.

• Implemented subagent completion notification flow and spawn context propagation:

  • Added pkg/subagent notification payload + outbound sender abstraction so finished tasks can render origin-channel notifications without coupling the package to the bus implementation.
  • Wired agent startup to enable subagents and bridge notifications into the message bus outbound path.
  • Registered the spawn tool in agent runtime and propagated channel/session route context into spawn tool execution.
  • Updated direct channel agent call sites (Telegram / ServerChan / WeChat) to use ChatWithPromptContext so tool execution has origin channel/session metadata.

• Added regression tests for:

  • subagent notification rendering/sending (pkg/subagent/notify_test.go)
  • spawn context route propagation (pkg/tools/spawn_test.go)
  • agent spawn tool registration (pkg/agent/agent_test.go)

• Verification run:

  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./pkg/subagent ./pkg/tools ./pkg/agent passed.
  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./... passed.

• Updated planning artifacts after completing this feature:

  • marked Subagent 完成通知真正回推 origin channel complete in task_plan.md
  • marked Batch B Subagent origin notify 接线 complete

• Re-audited nekobot against current code, ~/code/goclaw, and ~/code/gua, then rewrote the task backlog to distinguish completed baseline vs actual remaining gaps.

• Cleared stale backlog items that are already implemented:

  • /gateway restart and /gateway reload are implemented in pkg/gateway/controller.go.
  • memory hybrid text similarity already exists in pkg/memory/store.go.
  • skills version/tool comparison already exists in pkg/skills/eligibility.go.
  • cron at / every / delete_after_run / run-now already exist in pkg/cron/*, WebUI, and CLI.

• Confirmed current stable baseline now includes:

  • Web-first runtime admin for prompts / tool sessions / QMD / skills runtime status
  • provider fallback + cooldown + route override
  • session history sanitize / safe history / context compression
  • multi-path skills with snapshots/versioning

• Added new migration backlog sourced from goclaw:

  • general thread/conversation binding layer
  • memory quality pack (MMR / temporal decay / citations / cache)
  • gateway control-plane hardening
  • browser dual-mode session and advanced extraction
  • OAuth credential manager

• Added new migration backlog sourced from gua:

  • user-scoped external agent runtime foundation
  • permission / elicitation bridge for external agents
  • WeChat presenter and attachment-output pipeline
  • runtime prompt detection / tmux-style interactive control
  • channel interaction model for weak-interaction platforms

• Updated task_plan.md to reflect:

  • completed capabilities
  • real unfinished gaps
  • new Batch A-E execution order
  • rule that each completed feature must be committed and pushed individually

• Added runtime-backed prompt management with Ent schemas for prompt and prompt_binding, including CRUD, binding resolution, and render helpers in pkg/prompts.

• Wired prompt manager into FX/runtime startup and exposed WebUI prompt APIs with server-side tests in pkg/webui/server_prompts_test.go.

• Added frontend Prompts page and usePrompts hook, plus supporting textarea component and i18n entries.

• Expanded runtime admin flows around tool sessions, config, providers, marketplace, QMD inspection, and status endpoints to support the web-first dashboard model.

• Added workspace-aware QMD path resolution and improved session export defaults/visibility, including resolved export directory and cleanup support.

• Improved skills runtime metadata handling with snapshot/version coverage and added regression tests for snapshot/version behavior.

• Added provider cooldown tests and related runtime integration updates.

• Updated README and QMD docs to reflect the current Web-first setup and Docker/QMD behavior.

• Created and pushed commit 58877a5 (feat(runtime): add web-managed prompts and tool session controls).

• Follow-up needed on next device/session:

  • Run go test -count=1 ./...
  • Run npm --prefix pkg/webui/frontend run build
  • Manually smoke test Prompts page, tool session controls, and QMD admin flow in WebUI

2026-02-15

• Initialized planning artifacts for provider DB migration task.
• Inspected current provider backend/frontend implementation paths.
• Confirmed provider CRUD currently depends on config file persistence and draft-based frontend flow.
• Next: implement DB-backed provider store and wire into WebUI handlers.
• Implemented pkg/providerstore with SQLite provider CRUD and runtime config sync.
• Wired provider store into WebUI APIs and gateway/CLI startup module graphs.
• Updated dashboard provider dialog: clicking dialog Apply now directly persists provider changes.
• Adjusted storage to reuse the existing single DB file tool_sessions.db per user request.
• Verification: go test ./... passed.
• Refactored Ent location to a single shared path: moved generated code from pkg/toolsessions/ent to pkg/storage/ent and updated all imports.
• Re-ran verification after Ent refactor: go test ./... passed.
• Implemented runtime config DB store (config_sections) on shared tool_sessions.db and startup overlay logic.
• Migrated WebUI save paths for init password, channel updates, global config save, and chat routing persistence from file writes to DB writes.
• Added config DB store tests and verified with go test ./....
• Updated config/config.example.json, pkg/config/config.example.json, and docs/CONFIG.md to match latest minimal-file + DB-runtime behavior.
• Replanned config UX: switched dashboard Config from whole-document JSON editing to section-scoped editing with section selector/reset/save.
• Removed outdated provider-in-config snippets from README/docs and aligned examples to bootstrap-only config.
• Added WebUI config section into /api/config read/write path and persisted via DB section storage.
• Validation run: go test ./... passed.

2026-02-28

• Completed memory storage abstraction hardening with MemoryBackend implementations for file/db/kv/noop.
• Fixed file backend I/O in pkg/agent/memory_backend.go to use os.MkdirAll + os.ReadFile + os.IsNotExist while keeping atomic writes.
• Fixed NewMemoryStore fallback typing in pkg/agent/memory.go to safely degrade to noop backend when file backend init fails.
• Added pkg/agent/memory_backend_test.go to verify KV backend selection, DB backend selection, and KV-unavailable fallback to file backend.
• Verification run: go test ./pkg/agent ./pkg/config ./cmd/nekobot passed.
• Added ACP stdio entrypoint command nekobot acp with FX wiring and lifecycle management in cmd/nekobot/acp.go.
• Extended ACP session state and adapter mapping so session/new mcpServers are converted into config.MCPServerConfig and stored per ACP session.
• Updated blades orchestrator tool resolver path to honor ACP session-level MCP overrides, while keeping existing provider fallback and tool execution flow unchanged.
• Added MCP transport compatibility for ACP sse by mapping to blades HTTP transport and expanded config validation to accept sse transport values.
• Added ACP adapter tests for session creation/mode/cancel/prompt validation and MCP mapping coverage in pkg/agent/acp_adapter_test.go.
• Added ACP session/update bridge in adapter Prompt flow to emit agent text chunks via ACP connection while preserving existing provider fallback and tool execution semantics.
• Wired ACP adapter to AgentSideConnection in CLI startup so session update notifications are available in real ACP runtime.
• Expanded ACP adapter tests to cover session update emission, session update failure/cancel handling, and connection-detach behavior.
• Verification run: go test ./pkg/agent ./pkg/config ./cmd/nekobot passed.
• Added ACP session/load support in adapter with absolute CWD validation, in-memory session bootstrap, and per-session MCP override mapping.
• Updated ACP initialize capability to advertise loadSession=true so ACP clients can restore existing session IDs.
• Added ACP adapter tests for Initialize loadSession capability plus LoadSession success and validation failure paths.
• Added ACP session model state exposure in session/new and session/load responses to reflect the session’s active model.
• Implemented ACP experimental session/set_model handling with per-session model override updates and validation.
• Expanded ACP adapter tests to cover loaded/new session model state plus session/set_model success and invalid-param cases.
• Verification run: go test -count=1 ./pkg/agent ./pkg/config ./cmd/nekobot passed.
• Added ACP current_mode_update session notifications in session/set_mode so clients receive mode-change updates immediately.
• Added ACP adapter tests for session/set_mode notification emission plus update failure/cancel handling.
• Verification run: go test -count=1 ./pkg/agent ./pkg/config ./cmd/nekobot passed.
• Added trailing user-message dedup normalization in pkg/agent/context.go (BuildMessages) to avoid double-injecting the current prompt when callers pre-append user turns.
• Applied the same trailing user-message normalization in pkg/agent/blades_runtime.go before hydrating blades session history, keeping blades runtime behavior aligned with legacy prompt construction semantics.
• Added regression tests in pkg/agent/agent_test.go for trailing-current-user dedup and non-matching history preservation.
• Verification run: go test -count=1 ./pkg/agent ./pkg/gateway ./pkg/webui ./cmd/nekobot ./pkg/config passed.
• Note: full go test -count=1 ./... still fails in pkg/cron with known upstream fatal error: concurrent map writes in Ent atlas migration path (unchanged by this batch).
• Fixed runtime Ent schema migration race by serializing EnsureRuntimeEntSchema calls with a package-level mutex in pkg/config/runtime_client.go.
• Added regression test TestEnsureRuntimeEntSchemaConcurrentCalls in pkg/config/db_store_test.go to verify concurrent schema init no longer fails.
• Verification run: go test -count=1 ./pkg/config ./pkg/cron passed.
• Verification run: go test -count=1 ./... passed.
• Added deterministic tool description ordering in pkg/agent/context.go by sorting tool descriptions before assembling the tools section, improving prompt cache stability.
• Added regression test TestBuildToolsSection_SortsToolDescriptionsDeterministically in pkg/agent/agent_test.go.
• Verification run: go test -count=1 ./pkg/agent passed.
• Verification run: go test -count=1 ./... passed.
• Wired Agent.callLLMWithFallback to provider failover semantics (providers.ClassifyError + shared CooldownTracker) so retriable failures continue fallback and non-retriable format errors fail fast.
• Added provider cooldown skip behavior in agent fallback path, including contextual logging for skip reason and remaining cooldown window.
• Added agent failover regression tests for retriable fallback continuation, non-retriable short-circuit behavior, and cooldown-based skip on subsequent attempts (pkg/agent/agent_test.go).
• Added provider failover/cooldown regression tests covering cooldown skipping, non-retriable stop, reason tracking, all-cooldown exhaustion, and 24h failure-window reset (pkg/providers/loadbalancer_test.go).
• Added always frontmatter support in skills metadata (pkg/skills/manager.go) with eligibility-aware always-skill selection.
• Updated skill prompt assembly to emit a dedicated Always Skills XML section and keep regular skills in deterministic name order.
• Added compatibility parsing for metadata.openclaw.always in pkg/skills/loader.go.
• Extended validation with ValidateAlways to warn when always-on skills are disabled.
• Added regression tests for always-skill loading, prompt rendering, and validation (pkg/skills/manager_test.go, pkg/skills/loader_test.go).
• Updated docs with always field and Always Skills behavior (docs/CONFIG.md).
• Verification run: go test -count=1 ./pkg/skills ./pkg/agent passed.
• Continued Skills follow-up: switched # Available Skills from full markdown bodies to compact XML summary (<skills><skill ... /></skills>) in pkg/skills/manager.go, keeping Always Skills full XML instructions unchanged.
• Added instructions_length summary field using rune count for deterministic lightweight metadata and better token budgeting hints.
• Added/updated regression tests for compact summary output and non-ASCII length handling in pkg/skills/manager_test.go.
• Verification run: go test -count=1 ./pkg/skills ./pkg/agent passed.
• Verification run: go test -count=1 ./... passed.
• Added WebUI Cron API routes and handlers in pkg/webui/server.go for list/create/delete/enable/disable/run-now operations.
• Added structured error logging for Cron mutation failures in WebUI handlers (delete/enable/disable/run).
• Added WebUI Cron handler tests in pkg/webui/server_cron_test.go covering unavailable manager, CRUD flow, invalid RFC3339 at_time, not-found run-now, and disabled-job run-now behavior.
• Added frontend Cron integration with new hooks and page (pkg/webui/frontend/src/hooks/useCron.ts, pkg/webui/frontend/src/pages/CronPage.tsx), plus routing/sidebar wiring in pkg/webui/frontend/src/App.tsx and pkg/webui/frontend/src/components/layout/Sidebar.tsx.
• Added Cron i18n strings in pkg/webui/frontend/public/i18n/en.json, pkg/webui/frontend/public/i18n/zh-CN.json, and pkg/webui/frontend/public/i18n/ja.json.
• Verification run: go test ./pkg/webui ./pkg/cron ./pkg/agent ./pkg/config ./cmd/nekobot passed.
• Verification run: go test ./... passed.
• Verification run: npm --prefix pkg/webui/frontend run build passed (after installing frontend deps with npm --prefix pkg/webui/frontend ci).
• Added CLI command nekobot cron run <job-id> to trigger immediate execution for existing jobs.
• Aligned blades runtime tool error semantics with legacy orchestrator: tool execution failures now return Error: ... tool results instead of aborting the whole run.
• Added blades runtime regression tests for tool-error result fallback and role/parts mapping (pkg/agent/agent_test.go).
• Updated architecture docs for Cron capabilities to reflect DB-backed persistence and run-now support.
• Verification run: go test ./pkg/agent ./pkg/cron ./cmd/nekobot passed.
• Verification run: go test ./... passed.
• Added CLI regression tests for nekobot cron run command wiring and arg validation in cmd/nekobot/cron_test.go.
• Verification run: go test ./cmd/nekobot passed.
• Verification run: go test ./... passed.
• Fixed blades tool-result history conversion in pkg/agent/blades_runtime.go: when hydrating prior RoleTool messages, each blades.ToolPart now maps to its own providers.UnifiedMessage so multiple tool results in one blades message are preserved for provider context reconstruction.
• Added regression tests for blades tool history conversion in pkg/agent/agent_test.go:
  • TestBladesModelProvider_ConvertMessagesPreservesMultipleToolResults
  • TestBladesModelProvider_ConvertMessagesToolFallbackToRequest
• Verification run: go test ./pkg/agent passed.
• Verification run: go test ./... passed.
• Feature Batch #2 收口：chatWithBladesOrchestrator 会话历史注入现在保留 assistant 的 tool-calls turns（即使 text 为空），避免在重建 blades history 时丢失工具调用上下文，保证与 legacy 的 tool 执行链路语义一致。
• Added hasBladesHistoryContent + enhanced toBladesMessage in pkg/agent/blades_runtime.go to preserve assistant tool call metadata when hydrating history into blades session.
• Added regression tests in pkg/agent/agent_test.go:
  • TestToBladesMessage_AssistantToolCallsPreserved
  • TestHasBladesHistoryContent
• Verification run: go test -count=1 ./pkg/agent passed.
• Implemented static prompt caching in pkg/agent/context.go with file-state/tool-signature invalidation and dynamic current-time substitution to reduce repeated full prompt rebuilds while keeping fresh time output.
• Added context prompt regression tests in pkg/agent/agent_test.go for current-time placeholder replacement plus cache invalidation on bootstrap file and tool-description changes.
• Verification run: go test -count=1 ./pkg/agent passed.
• Verification run: go test -count=1 ./... passed.

2026-03-29

• Completed harness feature review and fixes:

  • Reviewed 5 recent commits: 1b7c3d0, 580741d, 46026ac, 583245d, c409cf1
  • Fixed snapshot incremental delta calculation to avoid message duplication
  • Added snapshot JSONL rewrite after Undo() for persistence
  • Added session ID passthrough for streaming updates
  • Added explicit fallback notice for streaming without handler
  • Changed watcher path matching to use RLock for better concurrency
  • Updated ConfigPage section metadata with proper i18n labels
  • Committed and pushed all fixes as 7412fdd

• Verification run:

  • GOPROXY=https://goproxy.cn,direct go test -count=1 ./... passed
  • npm --prefix pkg/webui/frontend run build passed

• Updated task_plan.md:

  • Marked Batch A verification complete
  • Marked Batch C frontend build and backend test complete
  • Marked 2026-03-29 review batch Phase 4 complete

• Added missing i18n labels for harness config sections:

  • Added audit, undo, preprocess, learnings, watch section labels
  • Added translations for en, zh-CN, ja
  • Committed as 28a8093

• Added line range support for @file mentions:

  • Support @file.txt:50-100 syntax for extracting specific line ranges
  • Added extractLineRange method for line-based content extraction
  • Updated formatFileReference to show line range in output
  • Added test cases for line range functionality
  • Committed as 6707d22

• Verification run:

  • go test -count=1 ./pkg/preprocess passed
  • npm --prefix pkg/webui/frontend run build passed

• Completed next browser advanced CDP action slice from task_plan.md:

  • added browser actions get_metrics, emulate_device, set_viewport, list_pages, new_page, activate_page, and close_page in pkg/tools/browser.go.
  • extended browser tool parameter schema with the new actions plus device profile selection and target_id for page control.
  • taught BrowserSession to expose a reusable DevTools management seam from the active attach endpoint, so higher-level CDP control actions do not need to reconstruct attach logic.
  • added deterministic helpers for device-profile / viewport validation and page-target control, with regression tests in pkg/tools/browser_test.go and pkg/tools/browser_session_test.go.

• Verification run:

  • go test -count=1 ./pkg/tools passed.

• Completed follow-up browser cookie/session-control slice from task_plan.md:

  • added browser actions set_cookie and clear_cookies in pkg/tools/browser.go.
  • added cookie-parameter schema support for name, value, secure, http_only, and same_site.
  • added validation helpers for cookie-setting arguments and regression tests covering required fields, absolute URL enforcement, and SameSite normalization.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'TestBrowserToolParametersIncludeCookieControlActions|TestBrowserToolBuildSetCookieArgs|TestBrowserToolBuildSetCookieArgsRejectsMissingName|TestBrowserToolBuildSetCookieArgsRejectsRelativeURL|TestBrowserToolBuildSetCookieArgsRejectsInvalidSameSite' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed follow-up browser storage control slice from task_plan.md:

  • added browser actions get_storage, set_storage, remove_storage, and clear_storage.
  • kept the slice page-scoped and relay-friendly by implementing it as generated JavaScript over the already attached browser session, without reopening attach/session orchestration.
  • added validation and script-shape regression tests for storage scope, key/value requirements, and generated JS behavior.

• Completed browser console capture closure that was already partially in flight:

  • finished the missing get_console helper types/options/collection utilities so the browser control surface compiles and returns bounded console/log entries.
  • kept the implementation minimal: level filtering + bounded collection over existing Runtime/Log subscriptions, without introducing persistent console session state.

• Verification run:

  • go test -count=1 ./pkg/tools passed.

• Completed browser storage/console control follow-up slice:

  • wired browser storage actions (get_storage, set_storage, remove_storage, clear_storage) into the runtime execute switch and restored their JS script builder implementation.
  • extended get_storage so it can return either the full storage map or a single key lookup.
  • implemented minimal get_console log capture with level filtering and bounded result size so the already-exposed browser action is now functional.

• Verification run:

  • go test -count=1 ./pkg/tools passed.

• Completed follow-up browser cookie/session-control slice from task_plan.md:

  • added browser actions set_cookie and clear_cookies in pkg/tools/browser.go.
  • added cookie-parameter schema support for name, text, secure, http_only, and same_site.
  • added validation helpers for cookie-setting arguments and regression tests covering required fields, absolute URL enforcement, and SameSite normalization.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'TestBrowserToolParametersIncludeCookieControlActions|TestBrowserToolBuildSetCookieArgs|TestBrowserToolBuildSetCookieArgsRejectsMissingName|TestBrowserToolBuildSetCookieArgsRejectsRelativeURL|TestBrowserToolBuildSetCookieArgsRejectsInvalidSameSite' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed follow-up browser relay/CDP slice from task_plan.md:

  • added browser action get_console in pkg/tools/browser.go.
  • added console filter options errors_only, warnings_only, info_only, and max_entries.
  • added deterministic console entry normalization helpers for log/runtime events with regression tests in pkg/tools/browser_test.go.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'TestBrowserToolParametersIncludeGetConsole|TestBrowserToolBuildConsoleOptionsDefaultsAndFilters|TestBrowserConsoleEntryMatchesRespectsPriority|TestBrowserConsoleEntryFrom(Log|Runtime)' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed browser minimal network-observation slice from task_plan.md:

  • added browser action get_network in pkg/tools/browser.go.
  • implemented a bounded request/response/finished/failed event collector using the existing CDP session without expanding into HAR/interception.
  • added regression tests for network action exposure and event-summary shaping.

• Verification run:

  • go test -count=1 ./pkg/tools -run 'TestBrowserToolParametersIncludeGetNetwork|TestBrowserNetworkEntryFromRequest|TestBrowserNetworkEntryFromResponse|TestBrowserNetworkEntryFromFinishedAndFailed' passed.
  • go test -count=1 ./pkg/tools passed.

• Completed gateway control-plane hardening phase 18 (metrics auth gate):

  • moved GET /metrics behind the same authenticated control-plane gate as the rest of the gateway admin surface.
  • kept the scope intentionally narrow by requiring manage-level access only, without changing metric shape or pairing semantics.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'TestMetricsEndpoint(RequiresAuth|RejectsMemberRole)?$|TestMetricsEndpoint$' passed.
  • go test -count=1 ./pkg/gateway passed.

• Completed externalagent consumer phase 6 (gateway approval UX closure):

  • gateway now exposes approval endpoints for the externalagent flow (GET /api/v1/approvals, POST /api/v1/approvals/:id/approve, POST /api/v1/approvals/:id/deny).
  • pending gateway externalagent launches can now be approved and immediately continued to process start without leaving the gateway control plane.

• Verification run:

  • go test -count=1 ./pkg/gateway -run 'Test(GatewayApproveExternalAgentPendingRequestStartsProcessImmediately|GatewayListApprovalsReturnsPendingRequests|ResolveExternalAgentSessionEndpoint(ReturnsPendingApprovalForAskRule|ReturnsPendingApprovalForManualMode|RejectsDeniedByApprovalMode|StartsProcessWhenApproved))$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent consumer phase 5 (WeChat shared starter adoption):

  • WeChat codex runtime creation now uses the shared externalagent process starter instead of its own private process-start path.
  • this is the first channel-side adoption of the shared starter layer.

• Verification run:

  • go test -count=1 ./pkg/channels/wechat -run 'Test(BuildRuntimePresetCodex.*|ControlServiceCreateCodexRuntime.*)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat passed.

• Completed externalagent approval UX phase 2 (shared HTTP response body helper):

  • added shared ResolveFlowResult.HTTPStatus() and ResolveFlowResult.ResponseBody() helpers so WebUI and gateway now emit their externalagent resolve responses from the same contract helper instead of hand-assembling maps separately.

• Verification run:

  • go test -count=1 ./pkg/webui ./pkg/gateway -run 'Test(HandleResolveExternalAgentSession(CreatesSession|IncludesMatchedPermissionRulePreview|ReturnsPendingApprovalForAskRule|ReturnsPendingApprovalForManualMode|RejectsDeniedByPermissionRule|StartsProcessWhenApproved)|ApproveExternalAgentPendingRequest(StartsProcessImmediately|AllowsSubsequentResolveForSameSession)|ResolveExternalAgentSessionEndpoint(CreatesSession|ResolvesRelativeWorkspace|RejectsWorkspaceOutsideConfiguredRoot|IncludesMatchedPermissionRulePreview|ReturnsPendingApprovalForAskRule|ReturnsPendingApprovalForManualMode|RejectsDeniedByApprovalMode|StartsProcessWhenApproved)|GatewayApproveExternalAgentPendingRequestStartsProcessImmediately|GatewayListApprovalsReturnsPendingRequests)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat ./pkg/agent passed.

• Completed externalagent approval UX phase 1 (consumer contract convergence):

  • aligned the three main consumers around the same externalagent approval concepts: status, request_id, reason, and launch_policy.
  • WeChat now exposes the same pending/denied semantics through a standardized text reply helper instead of ad-hoc phrasing.

• Verification run:

  • go test -count=1 ./pkg/channels/wechat -run 'Test(ControlServiceResolveCodexRuntimeReturnsPendingApproval|ControlServiceResolveCodexRuntimeReturnsDeniedApproval|FormatRuntimeApprovalReply(Pending|Denied)|BuildRuntimePresetCodex.*|ControlServiceCreateCodexRuntime.*)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat ./pkg/agent passed.

• Completed externalagent consumer phase 7 (WeChat shared resolve orchestrator adoption):

  • WeChat codex runtime creation now uses the shared externalagent resolve orchestrator for approval/policy gating before process start.
  • this brings the third consumer into the shared resolve layer instead of only reusing normalization/starter helpers.

• Verification run:

  • go test -count=1 ./pkg/channels/wechat ./pkg/agent ./pkg/externalagent passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat ./pkg/agent passed.

• Evaluated /home/czyt/code/fastclaw for importable capabilities and added prioritized adoption candidates to task_plan.md.

• Recommended follow-up parallel plan:

  • Hook Pipeline
  • Webhook Trigger Server
  • Policy Engine
  • Setup Wizard / First-Run
  • OpenAI-Compatible API Hardening

• User clarified an additional requirement for all fastclaw-derived work: pair backend implementation with corresponding nekobot frontend completion (management, observability, or interaction UI), rather than landing backend-only features.

• Current fastclaw-derived execution status:

  • Hook Pipeline landed
  • Webhook Trigger Server landed
  • Policy Engine landed
  • Setup Wizard / First-Run landed
  • OpenAI-compatible hardening started (first header-hardening slice landed)

• Completed WeChat approval UX phase 1 (managed externalagent /yes /no closure):

  • WeChat now resolves pending managed externalagent launches through the existing /yes, /no, and /select interaction surface instead of only showing a pending message.
  • approving a pending managed launch now immediately continues startup; denying keeps the process stopped.

• Verification run:

  • go test -count=1 ./pkg/channels/wechat -run 'Test(ControlServiceResolvePendingInteraction(ContinuesManagedRuntime|DeniesManagedRuntime)|ResolvePendingInteractionDelegatesToRuntimeApprovals|ResolvePendingInteractionDelegatesNumericSelectToRuntimeApprovals)$' passed.
  • go test -count=1 ./pkg/gateway ./pkg/externalagent ./pkg/webui ./pkg/channels/wechat ./pkg/agent ./pkg/tools ./pkg/approval ./pkg/tasks ./pkg/policy passed.

• Completed channel trace follow-up phase 2 (shared tool_call trace helper + ServerChan adoption):

  • extracted the WeChat-only tool trace formatter into shared pkg/channeltrace so direct session-based channels can reuse the same tool_call/tool_result rendering contract.
  • ServerChan now prepends the same compact tool trace before the final reply, and WeChat now consumes the shared helper instead of keeping a private implementation.
  • kept the slice intentionally narrow: only direct ChatWithPromptContext channels were touched; bus-driven channels like Telegram still need a separate event-path design to surface tool traces.

• Verification run:

  • go test -count=1 ./pkg/channeltrace ./pkg/channels/wechat ./pkg/channels/serverchan passed.