Chat Agent Feature Checklist

Chat Agent Feature Checklist

Frozen acceptance boundary for chatagent product orchestration and Web UI remediation (C+T3).
Regression and PRs are judged against this list plus the linked tests. If behavior exists in code but is missing here, update this document before changing code.

Related engine docs: README, Architecture. Maintainer guides: internal/server/chatagent/AGENTS.md, pkg/agent/AGENTS.md.

Dual-channel SSE protocol

Both channels use chatagent.StreamEvent JSON (type discriminator) from internal/server/chatagent/protocol.go.

Channel Message send (primary turn stream) Live event subscribe Subscribe filter
REST POST /chatagent/sessions/:id/messages — full SSE (delta / thinking / tool / turn_trace / … / done) GET /chatagent/sessions/:id/events confirm, confirm_resolved, canceled, mode_change, run_complete
Web POST /service/web/agents/:id/messages — turn execution; UI may use fetch/SSE depending on handler GET /service/web/agents/:id/events Same subset (chatagent_web_stream.go)

Event shape (StreamEvent) is shared. Both /events subscribers use the same type filter (not full deltas). Primary turn tokens come from the messages/send path (REST StreamAPIRun), not from /events.

1. Platform DM chat

ID Behavior Entry Edge cases Tests
P-01 Bind/create chat session for platform DM and run agent turn internal/server/chatagent_handler.go → Service.Run Closed session rejected; disabled agent chatagent service/handler unit tests
P-02 Slack: stream placeholder + SlackStreamSink updates Slack platform only Placeholder send failure → non-streaming fallback slack_sink_test.go, handler stream tests
P-03 Non-Slack: single reply after run completes Other platforms Persist assistant message failure surfaces user-visible error handler tests
P-04 Persist assistant message to store After successful Run DB error handler tests

2. REST API (/chatagent/*)

Auth: ScopeChatAgentChat. Owner checks on session-scoped routes.

ID Behavior Entry Edge cases Tests
R-01 Agent info GET /chatagent/info Disabled → 503 chatagent_http_test.go, BDD chat HTTP
R-02 List / create / close sessions GET|POST /chatagent/sessions, DELETE …/:id Closed → not found for owner ops; create body workspace (empty = config root, one first-level subdir); nested / hidden / missing dir → 400 and no leftover active session unit + chat_agent_chat_spec_test.go
R-03 List messages / plans / todos GET …/messages, GET …/plans, GET …/todos Empty history; user rows may include attachments[]; missing session workspace dir still lists history unit HTTP tests
R-03a Session trajectory GET …/trajectory Old sessions omit SYSTEM/CONTEXT when no turn_trace; owner 404/403 same as messages trajectory_test.go, chatagent_http_test.go
R-04 Send message (full SSE stream) POST …/messages JSON {text, attachments?} Empty text and no attachments → 400; run in flight 409; max 8 attachments; MIME allowlist; model missing modality → 400; turn_trace is primary-channel only (not /events) chat_agent_chat_spec_test.go, multimodal unit tests, protocol_test.go
R-04a Upload session media POST …/media multipart Strict MIME allowlist; binds (session_id, owner); requires media public base when FS signing needed media upload tests
R-04b Modality gate Before run Audio/video Reject until catalog marks ModalityAudioIn/ModalityVideoIn; unknown models allow image only modality reject tests
R-05 Subscribe session events GET …/events Client disconnect; buffered hub session_events_test.go
R-06 Confirm / cancel run POST …/confirm, POST …/cancel Unknown confirm 404; already resolved 409 confirm unit + BDD plan mode
R-07 Session mode get/set GET|PUT …/mode Plan vs normal chat_agent_chat_spec_test.go, plan mode tests
R-08 Clear session permission grants DELETE …/permission-grants permission session tests
R-09 Permissions get/put/delete GET|PUT|DELETE /chatagent/permissions Invalid config HTTP + permission tests
R-09a Approval mode get/put/delete GET|PUT|DELETE /chatagent/approval Invalid mode approval store/API + hook mode matrix
R-10 Context usage / compact GET …/context, POST …/compact Compaction no-op; missing session workspace dir → 400 (no fallback to config root) context_usage tests, agents page context
R-11 Export session GET …/export export tests
R-12 Resolve resource URI GET /chatagent/resources plan://, file:// chat_agent_spec_test.go resources It
R-13 Scheduled tasks CRUD + runs /chatagent/scheduled-tasks… One-shot complete; cancel scheduled_api_test.go, chat_agent_scheduled_task_spec_test.go
R-14 Session model / thinking settings GET|PUT …/settings; create body model / thinking_level / workspace; AgentInfo.selectable_models Unknown model / invalid level → 400; empty fields → runtime yaml/default; PUT workspace → 400 (immutable after create) settings_test.go, chatagent_http_test.go

3. Web chat UI (/service/web/agents/*)

ID Behavior Entry Edge cases Tests
W-01 Agents home: composer + session list GET /service/web/agents Unauthenticated → login; composer workspace options SSR from config root + first-level non-hidden subdirs agents_page_spec_test.go
W-01a Session list: auto title, last-message preview, day groups, pin/archive, status filters GET /service/web/agents/list?filter=; POST|DELETE …/:id/pin|archive Default hides archived; running / needs_approval from runtime gates agents_webservice_test.go, helpers/day-group unit tests
W-02 Create session (+ optional pending prompt) POST /service/web/agents ?prompt= / sessionStorage pending key; create body workspace (empty = config root); nested / hidden / missing dir → 400 and no leftover active session agents page + JS pending prompt; agents_webservice_test.go
W-03 Chat page hydrate history GET /service/web/agents/:id Closed session; missing session workspace dir still renders history agents page
W-04 Send message / cancel / confirm Web posts under /agents/:id/… Approval once/always/reject; reopen page replays pending confirm; run_complete reloads history; empty text OK when attachments present chat BDD helpers; unit confirm; pending SSR
W-04a Composer media (upload + paste image) Two-step: upload media then send with file_ids MIME allowlist; max 8; modality Reject / missing public_base_url surfaced in UI chatagent-chat.js + media unit/HTTP tests
W-05 Context ring + popover GET …/context + JS Token window zero; Skills list collapsed until opened agents_page_spec_test.go context It
W-06 Streaming markdown + tool cards + thinking + todo panel public/js/chatagent-*.js Open code fence delay; tool upsert; tool/thinking collapse; codeblock chrome; jump-to-bottom chat BDD stream done; chatagent_message_test.go
W-11 HTML artifact preview (present_html) tools/htmlpreview, chatagent-html.js, tool card iframe sandbox="allow-scripts" (no allow-same-origin) + injected CSP meta 256KB cap; same id latest iframe only; superseded cards drop Preview/Source; no new SSE type (html/artifact_id/call_id on existing tool events); plan mode allowed; not on subagent/pipeline htmlpreview tests, transform redact test, event_stream_test.go, memory_registry_test.go, pipeline_run_test.go, chatagent_message_test.go
W-10 Trajectory view Chat | Trajectory on the thread; GET …/trajectory; Duration gantt + Preview/Raw inspector ?view=trajectory; composer stays; no SYSTEM/CONTEXT without turn_trace; thread.js forwards turn_trace SSE agents_page_spec_test.go, chatagent_message_test.go, agents_webservice_test.go
W-07 Close session DELETE /service/web/agents/:id agents page
W-08 Model + thinking controls Composer + thread settings bar; GET|PUT …/settings localStorage defaults; empty DB falls back to yaml chat_model; workspace picker is create-time only (thread shows read-only label; PUT workspace → 400) agents page + chatagent-chat.js
W-09 Approval flow UX Nav/Home pending-approval badge; docked session approval bar; Allow once / Always allow matching copy Badge count from runtime gates; always button only when pattern suggested; auto-mode escalate hides Always session_activity_test.go, helpers/home/message unit tests; GET /service/web/approval-badge
W-09d Auto approval mode Permissions page mode selector; flagged tools reviewed by aux LLM; deny circuit breaker Autonomous ignores auto; reviewer fail → escalate pkg/agent/approval, approval_hook_test.go
W-12 Outbound MCP tools chat_agent.mcp_servers; tools mcp_<server>_<tool>; permission mcp.<name> No self /mcp; stdio YAML only; DCG skipped tools/mcp, pkg/agent/permission MCP tests
W-09e Tool loop detection chat_agent.loop_detection; generic/no-progress/ping-pong warn→block via doom_loop; global + post-compaction Terminate Blocked calls not recorded; overflow-retry compaction arms post-compaction only pkg/agent/loopdetect, loopdetect_hook_test.go
W-09a Background attention for approval Tab title ● Needs approval while confirm pending; desktop Notification when tab is hidden (permission via user gesture) Permission denied/default is silent; focus tab on notification click; clear title on resolve public/js/app.js, public/js/chatagent-approval.js
W-09b Approval inbox + deferred external Framework agent.approval via GatewaySend (inapp ± default); open session / resolve / cancel marks read and cancels deferred Plan mode still blocks send_notification; ConfirmGate not persisted approval_notify.go, confirm_test.go; Inbox /service/web/inbox*
W-09c Inbox UI Global bell + /service/web/inbox; unread default; deep-link open marks read System inapp channel immutable; open redirect only /service/web/* inbox_webservice.go / inbox_webservice_test.go

4. Permissions UI

ID Behavior Entry Edge cases Tests
Q-01 Permissions page render GET /service/web/chatagent-settings chatagent_permissions_webservice_test.go
Q-02 Save form / JSON / reset POST …, POST …/reset Field validation errors same
Q-03 REST permissions parity /chatagent/permissions Defaults when empty BDD permissions It

5. Skills / Memory / Subagents / Scheduled (Web admin)

ID Behavior Entry Edge cases Tests
A-01 Skills CRUD + files + zip import + enable/disable /service/web/agent-skills… Enable/disable; Import zip agent skills webservice tests / pages
A-02 Memory facts list/upsert/delete /service/web/agent-memory… Scope isolation; pin memory webservice tests
A-02b Session summaries browse/search/retry /service/web/agent-session-summaries… Archive-only generation session summary tests
A-03 Subagents CRUD + tasks /service/web/agent-subagents… Seed defaults on enable subagent specs/tests
A-04 Scheduled tasks list/detail/state /service/web/agent-scheduled-tasks… Pause/resume agent_scheduled_tasks_page_spec_test.go
A-05 Session inspect UI /service/web/agent-sessions… Entry payload, events, confirm, todos agent_sessions_page_spec_test.go

6. Orchestration capabilities

ID Behavior Entry Edge cases Tests
O-01 Interactive Service.Run / RunAPI service.go Session lock; harness pool reuse service_test.go, BDD
O-02 Plan mode blocks writes until confirm/normal mode + permission hooks Return to normal allows write chat_agent_spec_test.go plan It
O-03 Confirm gate (once / always / reject) confirm.go Pattern suggest confirm tests + BDD
O-04 Harness pool TTL / config hash refresh harness_pool.go Evict on close/abort; missing session workspace dir hard-fails harness create (no fallback to config root) harness-related tests
O-05 Session title generation title.go LLM disabled for tests; media-only turns use [image]/[audio]/[video] placeholders title_test.go, BDD title wait helpers
O-05a Multimodal user turns RunRequest.Attachments → MediaPart (file_id); provider-aware URL or Binary hydrate Session-scoped ownership; re-sign/hydrate per run; tool-model strip media parts transform + service multimodal tests
O-06 Manual + automatic compaction CompactSession, ctxmgr compaction / context tests
O-08 Pipeline agent step (ephemeral) pipeline_run.go / RunPipelineAgent Tools/skills allowlist; memory default off pipeline_run_test.go, ephemeral_run_test.go
O-09 Scheduled autonomous run + delivery scheduled_run.go, scheduler Isolated session; permission policy chat_agent_scheduled_task_spec_test.go, scheduled_* tests
O-10 Skills tool / memory tool / delegate_subagent tool / todo tools / present_html registry + tools Allowlists for subagents; present_html not on subagent/pipeline skills/memory/subagent/todo/htmlpreview tests
O-11 Sensors / progress / usage recording sensors, progress, usage_record unit tests
O-12 Prompt cache prompt_cache.go Invalidation on config change prompt_cache_test.go
O-13 Persist turn_trace executeRun after prompt assembly, h.Session().Append Skip RunKindPipeline; one node per Run; publish SSE when API publisher is set service.go / trajectory persist tests

7. Out of scope / non-goals for this remediation

  • Rewriting pkg/agent loop/harness core
  • Merging /chatagent and /service/web/agents route prefixes
  • Desktop agent protocol (pkg/types/agent.go / internal/server/agent.go)
  • Public HTML artifact share URLs, workspace HTML export, get_html, right-hand canvas

QA self-test (manual)

Use after each vertical slice and before freeze sign-off.

  1. Platform: send one DM (Slack if available) — streaming or final reply, message persisted.
  2. Web: create session from /service/web/agents, send a turn, see SSE done and history hydrate.
  3. Approval: trigger a gated tool — Allow once; repeat with Always allow matching; Deny path. Confirm Nav/Home badge count and docked approval bar. With the tab in background, confirm tab title shows ● Needs approval and (if Notification permission granted) a desktop notification; resolve clears the title.
  4. Plan mode: set plan → blocked write → switch normal → write works.
  5. Compact: call compact on a long session; context ring updates.
  6. Scheduled: create one-shot task, wait for completed + run row.
  7. Skills: toggle/enable a skill and confirm agent can read_skill.
  8. Knowledge: create a markdown doc in Agent Knowledge; confirm agent can search_knowledge then get_knowledge by path.
  9. Memory: fact CRUD via UI or memory_* tools; session summaries via archive + search tool.
  10. Subagent: run a delegate_subagent tool delegation; progress/tool events appear.
  11. Permissions: save form, reset to defaults; session grants clear; Knowledge key defaults to allow.
  12. Export: download/export session JSONL or documented format.
  13. Title: first user message eventually updates session title (or test-disabled path).
  14. REST smoke: POST /chatagent/sessions/:id/messages SSE reaches done; GET …/events receives hub events when a run publishes.
  15. Multimodal: upload an image via POST …/media, send with attachments (optional empty text) on a vision-capable model; history shows attachment; audio/video on unmarked models returns 400.
  16. Workspace: create a session from a first-level subdirectory on the composer picker; thread shows a read-only workspace label; changing it via settings is rejected. Delete that subdirectory and confirm history still loads while a new run fails.

Appendix A — Schedule / stream duplication audit (WS-B)

See chatagent-remediation-audit.md.

Appendix B — chatagent ↔ pkg/agent boundary audit (WS-C)

See chatagent-remediation-audit.md.

Maintainer guides: internal/server/chatagent/AGENTS.md, internal/modules/web/AGENTS.md (chatagent JS guardrails).

Warranty note

Regressions introduced directly by this remediation are in scope for fix within two weeks of checklist freeze. Checklist gaps discovered later require updating this document before code changes.