cb39a16c · samitouri/QOSami_Agents_Zero-1

extensions/python/_functions/AGENTS.md

+1

@@ -16,6 +16,7 @@

16 - Extension functions must match the implicit hook's supplied arguments.

17 - Preserve ordering prefixes where exception handling, watchdog registration, or cleanup depends on them.

18 - Hooks that mirror persisted AI responses into UI logs must reuse existing stream log items and avoid duplicating live response-tool logs.

19 +- The `AgentContext.run_task/end` hook attaches integration callbacks to the returned `DeferredTask`; keep terminal side effects out of `agent.py`.

20 - Recovery-loop circuit breakers must stop at the General Settings limit and render their user-visible cost warning from a core framework prompt.

21 - Prompt settings snapshots must be task-local, accessed through `get_settings_for_prompt()`, and end with the matching `Agent.prepare_prompt` call, including exceptional exits.

22

extensions/python/_functions/agent/AgentContext/run_task/end/_10_mark_state_dirty.py new

+14

@@ -0,0 +1,14 @@

1 +from helpers.defer import DeferredTask

2 +from helpers.extension import Extension

3 +from helpers.state_monitor_integration import mark_dirty_all

4 +

5 +

6 +class MarkStateDirty(Extension):

7 + def execute(self, data: dict | None = None, **kwargs) -> None:

8 + task = data.get("result") if isinstance(data, dict) else None

9 + if isinstance(task, DeferredTask):

10 + task.add_done_callback(

11 + lambda _future: mark_dirty_all(

12 + reason="agent.AgentContext.run_task_done",

13 + )

14 + )

extensions/python/response_stream/AGENTS.md

+2

@@ -13,6 +13,8 @@

13 - Keep streaming output synchronized with UI log items.

14 - Treat parsed stream snapshots as partial data; nested tool fields may be `None`

15 until their values arrive.

16 +- Live root responses carry `finished: false` until the response tool completes

17 + them so the WebUI can defer expensive Markdown rendering while content grows.

18 - Preserve include-alias replacement semantics where prompts/tools rely on them.

19 - Do not expose unmasked secrets in live responses.

20

extensions/python/response_stream/_20_live_response.py

+1

@@ -41,6 +41,7 @@ class LiveResponse(Extension):

41 type="response",

42 heading=f"icon://chat {self.agent.agent_name}: Responding",

43 id=shared_id,

44 + finished=False,

45 )

46 )

47

helpers/defer.py

+5

@@ -101,6 +101,11 @@ class DeferredTask:

101 self._start_task()

102 return self

103

104 + def add_done_callback(self, callback: Callable[[Future], Any]) -> None:

105 + if not self._future:

106 + raise RuntimeError("Task hasn't been started")

107 + self._future.add_done_callback(callback)

108 +

109 def __del__(self):

110 self.kill()

111

helpers/defer.py.dox.md

+2

@@ -17,6 +17,7 @@

17 - `ChildTask` (no explicit base class)

18 - `DeferredTask` (no explicit base class)

19 - `start_task(self, func: Callable[..., Coroutine[Any, Any, Any]], *args, **kwargs)`

20 + - `add_done_callback(self, callback: Callable[[Future], Any]) -> None`

21 - `is_ready(self) -> bool`

22 - `result_sync(self, timeout: Optional[float]=...) -> Any`

23 - `async result(self, timeout: Optional[float]=...) -> Any`

@@ -30,6 +31,7 @@

31

32 - Helper modules own reusable framework APIs and must preserve public callers unless all callers, tests, and docs are updated together.

33 - `DeferredTask` retains its callable and arguments only while an invocation is active; completion and `kill()` clear those references after the running coroutine has taken its own snapshot.

34 +- `add_done_callback()` forwards to the current invocation's concurrent future and rejects calls before `start_task()`; callbacks observe `is_alive() == False` and must remain lightweight.

35 - Task results remain available after completion. `restart()` can restart an active invocation, but a completed invocation has no retained call recipe and must be started again explicitly.

36 - Update this file whenever public functions, classes, persistence behavior, path/security assumptions, side effects, or cross-module contracts change.

37 - Observed side-effect areas: scheduler state.

helpers/log.py

+2 -1

@@ -36,7 +36,7 @@ def _lazy_mark_dirty_for_context(context_id: str, *, reason: str | None = None)

36 from helpers.state_monitor_integration import mark_dirty_for_context

37

38 _MARK_DIRTY_FOR_CONTEXT = mark_dirty_for_context

39 - _MARK_DIRTY_FOR_CONTEXT(context_id, reason=reason)

39 + _MARK_DIRTY_FOR_CONTEXT(context_id, reason=reason, include_collections=False)

40

41

42 T = TypeVar("T")

@@ -309,6 +309,7 @@ class Log:

309 if kwargs:

310 kwargs_out = copy.deepcopy(kwargs)

311 kwargs_out = self._mask_recursive(kwargs_out)

312 + kwargs_out = _truncate_value(kwargs_out)

313

314 with self._lock:

315 item = self.logs[no]

helpers/log.py.dox.md

+2

@@ -43,6 +43,8 @@

43

44 - Important called helpers/classes observed in the source: `TypeVar`, `dataclass`, `_MARK_DIRTY_ALL`, `_MARK_DIRTY_FOR_CONTEXT`, `truncate_text_by_ratio`, `cast`, `threading.RLock`, `self.set_initial_progress`, `self._update_item`, `self._notify_state_monitor`, `_lazy_mark_dirty_all`, `_lazy_mark_dirty_for_context`, `self._mask_recursive`, `_truncate_progress`, `self.set_progress`, `LogOutput`, `_truncate_value`, `json.dumps`, `time.time`, `self.log._update_item`.

45 - Keep request/response, tool, or helper semantics documented here at the same time as source changes.

46 +- Keyword KVP updates use the same recursive key/value truncation as an explicit `kvps` mapping, preventing streamed reasoning and other live fields from bypassing payload limits.

47 +- Existing log-item and progress updates mark only the selected context stream dirty and explicitly omit unchanged context/task collections; creating a new log item still broadcasts a full collection refresh.

48

49 ## Work Guidance

50

helpers/state_monitor.py

+43 -6

@@ -32,6 +32,7 @@ class ConnectionProjection:

32 # pushes indefinitely during continuous activity (throttled coalescing).

33 dirty_version: int = 0

34 pushed_version: int = 0

35 + collections_dirty_version: int = 0

36 # Development-only diagnostics - last known cause of the most recent dirty wave.

37 dirty_reason: str | None = None

38 dirty_wave_id: str | None = None

@@ -41,7 +42,7 @@ class ConnectionProjection:

42 class StateMonitor:

43 """Per-sid dirty tracking with debounced snapshot push scheduling."""

44

44 - def __init__(self, debounce_seconds: float = 0.025) -> None:

45 + def __init__(self, debounce_seconds: float = 0.1) -> None:

46 self.debounce_seconds = float(debounce_seconds)

47 self._lock = threading.RLock()

48 self._projections: dict[ConnectionIdentity, ConnectionProjection] = {}

@@ -95,7 +96,13 @@ class StateMonitor:

96 for namespace, sid in identities:

97 self.mark_dirty(namespace, sid, reason=reason, wave_id=wave_id)

98

98 - def mark_dirty_for_context(self, context_id: str, *, reason: str | None = None) -> None:

99 + def mark_dirty_for_context(

100 + self,

101 + context_id: str,

102 + *,

103 + reason: str | None = None,

104 + include_collections: bool = True,

105 + ) -> None:

106 if not isinstance(context_id, str) or not context_id.strip():

107 return

108 target = context_id.strip()

@@ -111,7 +118,13 @@ class StateMonitor:

118 if projection.request is not None and projection.request.context == target

119 ]

120 for namespace, sid in identities:

114 - self.mark_dirty(namespace, sid, reason=reason, wave_id=wave_id)

121 + self.mark_dirty(

122 + namespace,

123 + sid,

124 + reason=reason,

125 + wave_id=wave_id,

126 + include_collections=include_collections,

127 + )

128

129 def update_projection(

130 self,

@@ -142,6 +155,7 @@ class StateMonitor:

155 *,

156 reason: str | None = None,

157 wave_id: str | None = None,

158 + include_collections: bool = True,

159 ) -> None:

160 identity: ConnectionIdentity = (namespace, sid)

161 loop = self._dispatcher_loop

@@ -157,22 +171,36 @@ class StateMonitor:

171 running_loop = None

172

173 if running_loop is loop:

160 - self._mark_dirty_on_loop(identity, reason=reason, wave_id=wave_id)

174 + self._mark_dirty_on_loop(

175 + identity,

176 + reason=reason,

177 + wave_id=wave_id,

178 + include_collections=include_collections,

179 + )

180 return

181

163 - loop.call_soon_threadsafe(self._mark_dirty_on_loop, identity, reason, wave_id)

182 + loop.call_soon_threadsafe(

183 + self._mark_dirty_on_loop,

184 + identity,

185 + reason,

186 + wave_id,

187 + include_collections,

188 + )

189

190 def _mark_dirty_on_loop(

191 self,

192 identity: ConnectionIdentity,

193 reason: str | None = None,

194 wave_id: str | None = None,

195 + include_collections: bool = True,

196 ) -> None:

197 with self._lock:

198 projection = self._projections.get(identity)

199 if projection is None:

200 return

201 projection.dirty_version += 1

202 + if include_collections:

203 + projection.collections_dirty_version = projection.dirty_version

204 if runtime.is_development():

205 projection.dirty_reason = (

206 reason.strip()

@@ -228,6 +256,7 @@ class StateMonitor:

256 namespace, sid = identity

257 task = asyncio.current_task()

258 base_version = 0

259 + include_collections = True

260 dirty_reason: str | None = None

261 dirty_wave_id: str | None = None

262 try:

@@ -250,10 +279,17 @@ class StateMonitor:

279 if request is None:

280 return

281 base_version = projection.dirty_version

282 + include_collections = (

283 + not request.collections_delta

284 + or projection.collections_dirty_version > projection.pushed_version

285 + )

286 dirty_reason = projection.dirty_reason

287 dirty_wave_id = projection.dirty_wave_id

288

256 - snapshot = await build_snapshot_from_request(request=request)

289 + snapshot = await build_snapshot_from_request(

290 + request=request,

291 + include_collections=include_collections,

292 + )

293

294 with self._lock:

295 projection = self._projections.get(identity)

@@ -288,6 +324,7 @@ class StateMonitor:

324 ws_debug(

325 f"[StateMonitor] emit state_push namespace={namespace} sid={sid} seq={seq} "

326 f"context={request.context!r} logs_len={logs_len} "

327 + f"include_collections={include_collections} "

328 f"reason={dirty_reason!r} wave={dirty_wave_id!r}"

329 )

330 await manager.emit_to(

helpers/state_monitor.py.dox.md

+4 -2

@@ -17,9 +17,9 @@

17 - `register_sid(self, namespace: str, sid: str) -> None`

18 - `unregister_sid(self, namespace: str, sid: str) -> None`

19 - `mark_dirty_all(self, reason: str | None=...) -> None`

20 - - `mark_dirty_for_context(self, context_id: str, reason: str | None=...) -> None`

20 + - `mark_dirty_for_context(self, context_id: str, reason: str | None=..., include_collections: bool=...) -> None`

21 - `update_projection(self, namespace: str, sid: str, request: StateRequestV1, seq_base: int) -> None`

22 - - `mark_dirty(self, namespace: str, sid: str, reason: str | None=..., wave_id: str | None=...) -> None`

22 + - `mark_dirty(self, namespace: str, sid: str, reason: str | None=..., wave_id: str | None=..., include_collections: bool=...) -> None`

23 - Top-level functions:

24 - `get_state_monitor() -> StateMonitor`

25 - `_reset_state_monitor_for_testing() -> None`

@@ -35,6 +35,8 @@

35 ## Key Concepts

36

37 - Important called helpers/classes observed in the source: `threading.RLock`, `field`, `ws_debug`, `_ws_debug_enabled`, `context_id.strip`, `loop.call_soon_threadsafe`, `self._schedule_debounce_on_loop`, `asyncio.get_running_loop`, `asyncio.current_task`, `loop.is_closed`, `self._debounce_handles.pop`, `self._push_tasks.pop`, `self._projections.pop`, `self.mark_dirty`, `self._mark_dirty_on_loop`, `runtime.is_development`, `loop.call_later`, `StateMonitor`, `ConnectionProjection`, `handle.cancel`.

38 +- The default monitor coalesces dirty waves into at most one state push per 100 ms per connection without postponing an already scheduled push.

39 +- Collection-dirty versions are coalesced independently from ordinary stream dirties. Only clients that explicitly negotiate `collections_delta` may receive `contexts: null` and `tasks: null`; legacy clients remain on full snapshots.

40 - Keep request/response, tool, or helper semantics documented here at the same time as source changes.

41

42 ## Work Guidance

helpers/state_monitor_integration.py

+11 -2

@@ -7,7 +7,16 @@ def mark_dirty_all(*, reason: str | None = None) -> None:

7 get_state_monitor().mark_dirty_all(reason=reason)

8

9

10 -def mark_dirty_for_context(context_id: str, *, reason: str | None = None) -> None:

10 +def mark_dirty_for_context(

11 + context_id: str,

12 + *,

13 + reason: str | None = None,

14 + include_collections: bool = True,

15 +) -> None:

16 from helpers.state_monitor import get_state_monitor

17

13 - get_state_monitor().mark_dirty_for_context(context_id, reason=reason)

18 + get_state_monitor().mark_dirty_for_context(

19 + context_id,

20 + reason=reason,

21 + include_collections=include_collections,

22 + )

helpers/state_monitor_integration.py.dox.md

+2 -1

@@ -12,7 +12,7 @@

12 - `state_monitor_integration.py.dox.md` owns durable notes about responsibilities, contracts, side effects, and verification for that implementation.

13 - Top-level functions:

14 - `mark_dirty_all(reason: str | None=...) -> None`

15 -- `mark_dirty_for_context(context_id: str, reason: str | None=...) -> None`

15 +- `mark_dirty_for_context(context_id: str, reason: str | None=..., include_collections: bool=...) -> None`

16

17 ## Runtime Contracts

18

@@ -24,6 +24,7 @@

24 ## Key Concepts

25

26 - Important called helpers/classes observed in the source: `get_state_monitor.mark_dirty_all`, `get_state_monitor.mark_dirty_for_context`, `get_state_monitor`.

27 +- Global dirty waves always include context/task collections; high-frequency context-local owners may explicitly omit them when their mutation cannot change collection metadata.

28 - Keep request/response, tool, or helper semantics documented here at the same time as source changes.

29

30 ## Work Guidance

helpers/state_snapshot.py

+23 -12

@@ -3,7 +3,7 @@ from __future__ import annotations

3 import types

4 from typing import Any, Mapping, TypedDict, Union, get_args, get_origin, get_type_hints

5

6 -from dataclasses import dataclass

6 +from dataclasses import dataclass, replace

7

8 import pytz # type: ignore[import-untyped]

9

@@ -17,8 +17,8 @@ from helpers.task_scheduler import TaskScheduler

17 class SnapshotV1(TypedDict):

18 deselect_chat: bool

19 context: str

20 - contexts: list[dict[str, Any]]

21 - tasks: list[dict[str, Any]]

20 + contexts: list[dict[str, Any]] | None

21 + tasks: list[dict[str, Any]] | None

22 logs: list[dict[str, Any]]

23 log_guid: str

24 log_version: int

@@ -37,6 +37,7 @@ class StateRequestV1:

37 log_from: int

38 notifications_from: int

39 timezone: str

40 + collections_delta: bool = False

41

42

43 class StateRequestValidationError(ValueError):

@@ -159,6 +160,7 @@ def parse_state_request_payload(payload: Mapping[str, Any]) -> StateRequestV1:

160 log_from = payload.get("log_from")

161 notifications_from = payload.get("notifications_from")

162 timezone = payload.get("timezone")

163 + collections_delta = payload.get("collections_delta", False)

164

165 if context is not None and not isinstance(context, str):

166 raise StateRequestValidationError(

@@ -184,6 +186,12 @@ def parse_state_request_payload(payload: Mapping[str, Any]) -> StateRequestV1:

186 message="timezone must be a non-empty string",

187 details={"timezone": timezone},

188 )

189 + if not isinstance(collections_delta, bool):

190 + raise StateRequestValidationError(

191 + reason="collections_delta_type",

192 + message="collections_delta must be a boolean",

193 + details={"collections_delta_type": type(collections_delta).__name__},

194 + )

195

196 tz = timezone.strip()

197 try:

@@ -203,6 +211,7 @@ def parse_state_request_payload(payload: Mapping[str, Any]) -> StateRequestV1:

211 log_from=log_from,

212 notifications_from=notifications_from,

213 timezone=tz,

214 + collections_delta=collections_delta,

215 )

216

217

@@ -245,15 +254,16 @@ def advance_state_request_after_snapshot(

254 except (TypeError, ValueError):

255 pass

256

248 - return StateRequestV1(

249 - context=request.context,

257 + return replace(

258 + request,

259 log_from=log_from,

260 notifications_from=notifications_from,

252 - timezone=request.timezone,

261 )

262

263

256 -async def build_snapshot_from_request(*, request: StateRequestV1) -> SnapshotV1:

264 +async def build_snapshot_from_request(

265 + *, request: StateRequestV1, include_collections: bool = True

266 +) -> SnapshotV1:

267 """Build a poll-shaped snapshot for both /poll and state_push."""

268

269 localization = Localization.get()

@@ -269,7 +279,8 @@ async def build_snapshot_from_request(*, request: StateRequestV1) -> SnapshotV1:

279 from_no = _coerce_non_negative_int(request.log_from, default=0)

280 notifications_from_no = _coerce_non_negative_int(request.notifications_from, default=0)

281

272 - _prune_missing_saved_contexts()

282 + if include_collections:

283 + _prune_missing_saved_contexts()

284

285 active_context = AgentContext.get(ctxid) if ctxid else None

286

@@ -291,9 +302,9 @@ async def build_snapshot_from_request(*, request: StateRequestV1) -> SnapshotV1:

302 ctxs: list[dict[str, Any]] = []

303 tasks: list[dict[str, Any]] = []

304 processed_contexts: set[str] = set()

294 - agent_profile_labels = _get_agent_profile_labels()

305 + agent_profile_labels = _get_agent_profile_labels() if include_collections else {}

306

296 - all_ctxs = AgentContext.all()

307 + all_ctxs = AgentContext.all() if include_collections else []

308 for ctx in all_ctxs:

309 if ctx.id in processed_contexts:

310 continue

@@ -345,8 +356,8 @@ async def build_snapshot_from_request(*, request: StateRequestV1) -> SnapshotV1:

356 snapshot: SnapshotV1 = {

357 "deselect_chat": bool(ctxid) and active_context is None,

358 "context": active_context.id if active_context else "",

348 - "contexts": ctxs,

349 - "tasks": tasks,

359 + "contexts": ctxs if include_collections else None,

360 + "tasks": tasks if include_collections else None,

361 "logs": logs,

362 "log_guid": active_context.log.guid if active_context else "",

363 "log_version": log_end,

helpers/state_snapshot.py.dox.md

+2 -1

@@ -25,7 +25,7 @@

25 - `parse_state_request_payload(payload: Mapping[str, Any]) -> StateRequestV1`

26 - `_coerce_state_request_inputs(context: Any, log_from: Any, notifications_from: Any, timezone: Any) -> StateRequestV1`

27 - `advance_state_request_after_snapshot(request: StateRequestV1, snapshot: Mapping[str, Any]) -> StateRequestV1`

28 -- `async build_snapshot_from_request(request: StateRequestV1) -> SnapshotV1`: Build a poll-shaped snapshot for both /poll and state_push.

28 +- `async build_snapshot_from_request(request: StateRequestV1, include_collections: bool=...) -> SnapshotV1`: Build a poll-shaped snapshot for both /poll and state_push.

29 - `_notify_timezone_changed(previous_timezone: str, current_timezone: str) -> None`

30 - `async build_snapshot(context: str | None, log_from: int, notifications_from: int, timezone: str | None) -> SnapshotV1`

31 - Notable constants/configuration names: `_SNAPSHOT_V1_SCHEMA`, `SNAPSHOT_SCHEMA_V1_KEYS`.

@@ -42,6 +42,7 @@

42 - Important called helpers/classes observed in the source: `dataclass`, `_build_schema_from_typeddict`, `get_origin`, `timezone.strip`, `StateRequestV1`, `localization.get_timezone`, `localization.set_timezone`, `ctxid.strip`, `_coerce_non_negative_int`, `AgentContext.get_notification_manager`, `notification_manager.output`, `_get_agent_profile_labels`, `ctxs.sort`, `tasks.sort`, `validate_snapshot_schema_v1`, `_coerce_state_request_inputs`, `super.__init__`, `get_args`, `_annotation_to_isinstance_types`, `TypeError`.

43 - Snapshot building prunes non-running in-memory contexts that were previously saved but no longer have a `chat.json`, preventing stale sidebar rows after chat files are deleted outside `/chat_remove`.

44 - Notification payloads use the manager's matching GUID and cursor from the same atomic read, preventing a concurrent notification from being skipped by the WebUI.

45 +- `StateRequestV1.collections_delta` is an optional, false-by-default capability. Negotiated state pushes may use `null` for both `contexts` and `tasks` when those collections are unchanged; HTTP polling and legacy WebSocket clients always receive full lists.

46 - Keep request/response, tool, or helper semantics documented here at the same time as source changes.

47

48 ## Work Guidance

tests/test_defer_lifecycle.py

+36

@@ -48,6 +48,42 @@ def test_completed_task_releases_call_references_and_children():

48 task.kill(terminate_thread=True)

49

50

51 +def test_run_task_end_extension_marks_state_dirty_after_completion(monkeypatch):

52 + from extensions.python._functions.agent.AgentContext.run_task.end import (

53 + _10_mark_state_dirty as task_done_extension,

54 + )

55 +

56 + task = make_task()

57 + callback_called = threading.Event()

58 + observations: list[tuple[str | None, bool]] = []

59 +

60 + def mark_dirty(*, reason=None):

61 + observations.append((reason, bool(task.is_alive())))

62 + callback_called.set()

63 +

64 + monkeypatch.setattr(

65 + task_done_extension,

66 + "mark_dirty_all",

67 + mark_dirty,

68 + )

69 +

70 + async def run():

71 + return "done"

72 +

73 + try:

74 + with pytest.raises(RuntimeError, match="Task hasn't been started"):

75 + task.add_done_callback(lambda _future: None)

76 + task.start_task(run)

77 + task_done_extension.MarkStateDirty(agent=None).execute(

78 + data={"result": task}

79 + )

80 + assert task.result_sync(timeout=2) == "done"

81 + assert callback_called.wait(2)

82 + assert observations == [("agent.AgentContext.run_task_done", False)]

83 + finally:

84 + task.kill(terminate_thread=True)

85 +

86 +

87 def test_kill_clears_stored_call_without_clearing_running_arguments():

88 task = make_task()

89 owner = Owner()

tests/test_multi_tab_isolation.py

+1 -1

@@ -20,7 +20,7 @@ async def test_state_monitor_per_sid_isolation_independent_snapshots_seq_and_cur

20

21 namespace = "/ws"

22

23 - async def fake_build_snapshot_from_request(*, request):

23 + async def fake_build_snapshot_from_request(*, request, include_collections=True):

24 context = request.context

25 log_from = request.log_from

26 notifications_from = request.notifications_from

tests/test_parallel_tool.py

+3 -3

@@ -1154,9 +1154,9 @@ def test_chats_sidebar_projects_parallel_children_as_indented_accordion() -> Non

1154 )

1155

1156 assert "parent_context_id" in store

1157 - assert "const nextExpandedParents = { ...this.expandedParents };" in store

1158 - assert "nextExpandedParents[selectedId] === undefined" in store

1159 - assert "nextExpandedParents[selectedId] = true;" in store

1157 + assert "this.expandedParents[selectedId] === undefined" in store

1158 + assert "...this.expandedParents," in store

1159 + assert "[selectedId]: true," in store

1160 assert "topLevelContexts()" in html

1161 assert "childContexts(context.id)" in html

1162 assert "chat-child-container" in html

tests/test_plain_response_logging.py

+43 -1

@@ -23,6 +23,47 @@ def _agent_with_generating_log():

23 return agent, item

24

25

26 +def test_log_keyword_updates_share_kvp_value_limit():

27 + item = Log().log(type="agent", heading="A0: Reasoning")

28 +

29 + item.update(reasoning="r" * 6_000, finished=True)

30 +

31 + assert len(item.kvps["reasoning"]) <= 5_000

32 + assert "Characters hidden" in item.kvps["reasoning"]

33 + assert item.kvps["finished"] is True

34 +

35 +

36 +def test_live_log_updates_omit_unchanged_collections(monkeypatch):

37 + import helpers.log as log_module

38 +

39 + full_dirty: list[str | None] = []

40 + context_dirty: list[tuple[str, str | None, bool]] = []

41 + monkeypatch.setattr(

42 + log_module,

43 + "_MARK_DIRTY_ALL",

44 + lambda *, reason=None: full_dirty.append(reason),

45 + )

46 + monkeypatch.setattr(

47 + log_module,

48 + "_MARK_DIRTY_FOR_CONTEXT",

49 + lambda context_id, *, reason=None, include_collections=True: context_dirty.append(

50 + (context_id, reason, include_collections)

51 + ),

52 + )

53 +

54 + log = Log()

55 + log.context = SimpleNamespace(id="ctx", streaming_agent=None)

56 + item = log.log(type="agent", heading="Calling LLM")

57 + item.update(content="stream update")

58 + log.set_progress("Receiving")

59 +

60 + assert full_dirty == ["log.Log._notify_state_monitor"]

61 + assert context_dirty == [

62 + ("ctx", "log.Log._update_item", False),

63 + ("ctx", "log.Log.set_progress", False),

64 + ]

65 +

66 +

67 def test_responses_plain_text_completion_finishes_generating_log_as_response():

68 agent, item = _agent_with_generating_log()

69 data = {

@@ -89,7 +130,7 @@ def test_responses_plain_text_completion_does_not_replace_live_response_log():

130

131

132 @pytest.mark.asyncio

92 -async def test_live_response_renders_single_action_wrapper():

133 +async def test_live_response_renders_single_unfinished_action_wrapper():

134 log = Log()

135 generating = log.log(type="agent", id="msg-1")

136 loop_data = SimpleNamespace(params_temporary={"log_item_generating": generating})

@@ -111,6 +152,7 @@ async def test_live_response_renders_single_action_wrapper():

152 assert response.type == "response"

153 assert response.content == "wrapper works"

154 assert response.id == "msg-1"

155 + assert response.kvps["finished"] is False

156

157

158 @pytest.mark.asyncio

tests/test_snapshot_schema_v1.py

+46

@@ -86,6 +86,52 @@ async def test_snapshot_builder_produces_contract_schema_key_set_and_defaults():

86 assert payload["notifications_version"] >= 0

87

88

89 +@pytest.mark.asyncio

90 +async def test_negotiated_incremental_snapshot_uses_null_collection_sentinel():

91 + from helpers import state_snapshot as snapshot

92 +

93 + request = snapshot.StateRequestV1(

94 + context=None,

95 + log_from=0,

96 + notifications_from=0,

97 + timezone="UTC",

98 + collections_delta=True,

99 + )

100 + payload = await snapshot.build_snapshot_from_request(

101 + request=request,

102 + include_collections=False,

103 + )

104 +

105 + snapshot.validate_snapshot_schema_v1(payload)

106 + assert set(payload) == EXPECTED_SNAPSHOT_KEYS

107 + assert payload["contexts"] is None

108 + assert payload["tasks"] is None

109 +

110 +

111 +def test_state_request_collection_delta_is_optional_and_type_checked():

112 + from helpers import state_snapshot as snapshot

113 +

114 + base = {

115 + "context": None,

116 + "log_from": 0,

117 + "notifications_from": 0,

118 + "timezone": "UTC",

119 + }

120 +

121 + assert snapshot.parse_state_request_payload(base).collections_delta is False

122 + assert (

123 + snapshot.parse_state_request_payload(

124 + {**base, "collections_delta": True}

125 + ).collections_delta

126 + is True

127 + )

128 + with pytest.raises(snapshot.StateRequestValidationError) as error:

129 + snapshot.parse_state_request_payload(

130 + {**base, "collections_delta": "yes"}

131 + )

132 + assert error.value.reason == "collections_delta_type"

133 +

134 +

135 def test_snapshot_schema_rejects_unexpected_top_level_keys():

136 from helpers import state_snapshot as snapshot

137

tests/test_state_monitor.py

+99

@@ -8,6 +8,12 @@ if str(PROJECT_ROOT) not in sys.path:

8 sys.path.insert(0, str(PROJECT_ROOT))

9

10

11 +def test_state_monitor_defaults_to_ten_pushes_per_second() -> None:

12 + from helpers.state_monitor import StateMonitor

13 +

14 + assert StateMonitor().debounce_seconds == 0.1

15 +

16 +

17 @pytest.mark.asyncio

18 async def test_state_monitor_debounce_coalesces_without_postponing_and_cleanup_cancels_pending():

19 from helpers.state_monitor import StateMonitor

@@ -101,3 +107,96 @@ async def test_state_monitor_namespace_identity_prevents_cross_namespace_state_p

107

108 assert captured

109 assert all(ns == ns_a for ns, _ in captured)

110 +

111 +

112 +@pytest.mark.asyncio

113 +async def test_collection_delta_tracks_full_and_stream_dirty_waves(monkeypatch) -> None:

114 + import asyncio

115 +

116 + import helpers.state_monitor as state_monitor_module

117 + from helpers.state_monitor import StateMonitor

118 + from helpers.state_snapshot import StateRequestV1

119 +

120 + namespace = "/ws"

121 + sid = "sid-delta"

122 + identity = (namespace, sid)

123 + include_calls: list[bool] = []

124 + emitted: list[dict] = []

125 +

126 + async def fake_snapshot(*, request, include_collections=True):

127 + include_calls.append(include_collections)

128 + return {

129 + "deselect_chat": False,

130 + "context": request.context or "",

131 + "contexts": [] if include_collections else None,

132 + "tasks": [] if include_collections else None,

133 + "logs": [],

134 + "log_guid": "guid",

135 + "log_version": request.log_from,

136 + "log_progress": "",

137 + "log_progress_active": False,

138 + "paused": False,

139 + "notifications": [],

140 + "notifications_guid": "notifications",

141 + "notifications_version": request.notifications_from,

142 + }

143 +

144 + class FakeManager:

145 + def __init__(self, loop):

146 + self._dispatcher_loop = loop

147 +

148 + async def emit_to(self, _namespace, _sid, _event_type, payload, **_kwargs):

149 + emitted.append(payload["snapshot"])

150 +

151 + monitor = StateMonitor(debounce_seconds=60.0)

152 + monitor.bind_manager(FakeManager(asyncio.get_running_loop()))

153 + monitor.register_sid(namespace, sid)

154 + monitor.update_projection(

155 + namespace,

156 + sid,

157 + request=StateRequestV1(

158 + context="ctx",

159 + log_from=0,

160 + notifications_from=0,

161 + timezone="UTC",

162 + collections_delta=True,

163 + ),

164 + seq_base=1,

165 + )

166 + monkeypatch.setattr(

167 + state_monitor_module,

168 + "build_snapshot_from_request",

169 + fake_snapshot,

170 + )

171 +

172 + async def flush() -> None:

173 + handle = monitor._debounce_handles.pop(identity)

174 + handle.cancel()

175 + await monitor._flush_push(identity)

176 +

177 + monitor.mark_dirty(namespace, sid, include_collections=False)

178 + await flush()

179 +

180 + monitor.mark_dirty(namespace, sid, include_collections=False)

181 + monitor.mark_dirty(namespace, sid, include_collections=True)

182 + await flush()

183 +

184 + monitor.update_projection(

185 + namespace,

186 + sid,

187 + request=StateRequestV1(

188 + context="ctx",

189 + log_from=0,

190 + notifications_from=0,

191 + timezone="UTC",

192 + ),

193 + seq_base=1,

194 + )

195 + monitor.mark_dirty(namespace, sid, include_collections=False)

196 + await flush()

197 +

198 + assert include_calls == [False, True, True]

199 + assert emitted[0]["contexts"] is None

200 + assert emitted[0]["tasks"] is None

201 + assert emitted[1]["contexts"] == []

202 + assert emitted[2]["contexts"] == []

tests/test_state_sync_handler.py

+6

@@ -55,6 +55,8 @@ async def _create_manager_with_socketio() -> tuple[WsManager, "WsWebui", FakeSoc

55

56 @pytest.mark.asyncio

57 async def test_state_request_success_returns_wire_level_shape_and_contract_payload():

58 + from helpers.state_monitor import get_state_monitor

59 +

60 _manager, handler = await _create_manager()

61

62 result = await handler.process(

@@ -65,6 +67,7 @@ async def test_state_request_success_returns_wire_level_shape_and_contract_paylo

67 "log_from": 0,

68 "notifications_from": 0,

69 "timezone": "UTC",

70 + "collections_delta": True,

71 },

72 "sid-1",

73 )

@@ -73,6 +76,9 @@ async def test_state_request_success_returns_wire_level_shape_and_contract_paylo

76 assert set(result.keys()) >= {"runtime_epoch", "seq_base"}

77 assert isinstance(result["runtime_epoch"], str) and result["runtime_epoch"]

78 assert isinstance(result["seq_base"], int)

79 + projection = get_state_monitor()._projections[(NAMESPACE, "sid-1")]

80 + assert projection.request is not None

81 + assert projection.request.collections_delta is True

82

83

84 @pytest.mark.asyncio

tests/test_webui_chat_deletion.py

+10 -2

@@ -78,10 +78,18 @@ assert(

78 model.contexts === unchangedContexts,

79 "unchanged snapshots must preserve the Alpine contexts array",

80 );

81 +const unchangedFirstRow = model.contexts[0];

82 model.applyContexts([{{ ...chats[0], name: "Renamed" }}, ...chats.slice(1)]);

83 assert(

83 - model.contexts !== unchangedContexts && model.contexts[0].name === "Renamed",

84 - "changed context metadata must replace the contexts array",

84 + model.contexts === unchangedContexts &&

85 + model.contexts[0] === unchangedFirstRow &&

86 + model.contexts[0].name === "Renamed",

87 + "changed context metadata must update its existing row in place",

88 +);

89 +model.applyContexts([...chats, {{ id: "d", created_at: 40 }}]);

90 +assert(

91 + model.contexts !== unchangedContexts && model.contexts[0].id === "d",

92 + "structural context changes must replace and reorder the contexts array",

93 );

94

95 const tree = [

tests/test_webui_message_ordering_static.py

+8

@@ -19,6 +19,14 @@ def test_full_log_replays_replace_existing_message_dom():

19 assert "normalized.sort(" in messages_js

20

21

22 +def test_unfinished_root_responses_defer_markdown_rendering():

23 + messages_js = read("webui", "js", "messages.js")

24 +

25 + assert "const renderMarkdown = kvps?.finished !== false;" in messages_js

26 + assert "markdown: renderMarkdown," in messages_js

27 + assert "latex: renderMarkdown," in messages_js

28 +

29 +

30 def test_message_ordering_uses_a_bounded_tail_first_renderer_cache():

31 messages_js = read("webui", "js", "messages.js")

32 message_window_js = read("webui", "js", "message-window.js")

tests/test_webui_message_window.py

+14 -1

@@ -77,10 +77,21 @@ assert(

77 );

78

79 const sharedIdGroup = new MessageWindow({{ initialLimit: 60 }});

80 -sharedIdGroup.reset([

80 +const sharedIdKeys = sharedIdGroup.merge([

81 {{ no: 1, id: "shared-run-id", type: "agent", content: "final generation" }},

82 {{ no: 2, id: "shared-run-id", type: "response", content: "final response" }},

83 ]);

84 +assert(

85 + sharedIdKeys.has("id:shared-run-id:type:response"),

86 + "merge must report a newly added step",

87 +);

88 +const updatedSharedIdKeys = sharedIdGroup.merge([

89 + {{ no: 2, id: "shared-run-id", type: "response", content: "updated response" }},

90 +]);

91 +assert(

92 + updatedSharedIdKeys.size === 0,

93 + "merge must not report an existing record update as newly added",

94 +);

95 assert(sharedIdGroup.size === 2, "a shared id must not merge GEN and response records");

96 assert(

97 sharedIdGroup.visibleMessages().map((entry) => entry.type).join(",") ===

@@ -329,6 +340,8 @@ def test_process_groups_are_atomic_and_page_steps_in_fifties():

340 assert '"code_exe",' in MESSAGE_WINDOW_JS.read_text(encoding="utf-8")

341 assert "getUnitKeys: getMessageRenderUnitKeys" in messages

342 assert "getProcessGroupRenderMessages(windowMessages)" in messages

343 + assert "const addedMessageKeys = _messageWindow.merge" in messages

344 + assert "addedMessageKeys.has(getMessageCacheKey(message))" in messages

345 assert 'button.className = "process-group-show-more"' in messages

346 assert "current + PROCESS_GROUP_STEP_PAGE_SIZE" in messages

347 assert "group.dataset.fullStartTimestamp" in messages

tests/test_ws_client_api_surface.py

+27

@@ -39,3 +39,30 @@ def test_websocket_js_exports_minimal_namespaced_api_surface() -> None:

39

40 assert "broadcast" not in exports

41 assert "requestAll" not in exports

42 +

43 +

44 +def test_completed_state_push_cannot_overwrite_disconnected_mode() -> None:

45 + source = (

46 + PROJECT_ROOT / "webui" / "components" / "sync" / "sync-store.js"

47 + ).read_text(encoding="utf-8")

48 +

49 + apply_end = source.split("await applySnapshot(data.snapshot", 1)[1].split(

50 + 'this._setMode(SYNC_MODES.HEALTHY, "push applied");', 1

51 + )[0]

52 + assert "if (!stateSocket.isConnected()) return;" in apply_end

53 +

54 +

55 +def test_partial_snapshot_retains_sidebar_collections_and_extension_shape() -> None:

56 + source = (PROJECT_ROOT / "webui" / "index.js").read_text(encoding="utf-8")

57 + request_builder = source.split(

58 + "export function buildStateRequestPayload", 1

59 + )[1].split("export async function applySnapshot", 1)[0]

60 +

61 + assert "collections_delta: true" in request_builder

62 + assert "const hasCollections =" in source

63 + assert "Array.isArray(snapshot.contexts) && Array.isArray(snapshot.tasks)" in source

64 + assert "snapshot: extensionSnapshot" in source

65 + assert "contexts: chatsStore.contexts" in source

66 + assert "tasks: tasksStore.tasks" in source

67 + assert "if (hasCollections)" in source

68 + assert "snapshot.contexts || []" not in source

webui/components/sidebar/AGENTS.md

+1 -1

@@ -32,7 +32,7 @@

32 - The utility-message preference controls both individual utility steps and utility-only process-group chrome so hidden utility runs cannot leave empty headers in the transcript.

33 - Chat deletion removes the sidebar row optimistically in the same render batch as fallback selection. Keep successful local deletion tombstones for the page session so out-of-order poll or push snapshots cannot reinsert rows; restore the row and clear its tombstone if the delete request fails.

34 - Chat selection must synchronize the sidebar store even when the low-level context has already switched to the requested ID.

35 -- Unchanged context snapshots preserve the Alpine contexts-array identity to avoid chat-list reconciliation, while selection and parent-expansion synchronization still run; changed metadata and deletion tombstones must still replace the visible list.

35 +- Context snapshots preserve the Alpine contexts-array and row identities while their order is stable, updating changed row metadata in place so streaming log counters do not reconcile the whole chat list. Additions, removals, reordering, and deletion tombstones must still replace the visible list; selection and parent-expansion synchronization must not publish unchanged state.

36

37 ## Work Guidance

38

webui/components/sidebar/chats/chats-store.js

+28 -7

@@ -65,7 +65,22 @@ const model = {

65 const contextsJson = JSON.stringify(nextContexts);

66 if (contextsJson !== this.contextsJson) {

67 this.contextsJson = contextsJson;

68 - this.contexts = nextContexts;

68 + const sameRows =

69 + nextContexts.length === this.contexts.length &&

70 + nextContexts.every((context, index) => context?.id === this.contexts[index]?.id);

71 +

72 + if (sameRows) {

73 + nextContexts.forEach((context, index) => {

74 + const current = this.contexts[index];

75 + if (JSON.stringify(current) === JSON.stringify(context)) return;

76 + Object.keys(current).forEach((key) => {

77 + if (!(key in context)) delete current[key];

78 + });

79 + Object.assign(current, context);

80 + });

81 + } else {

82 + this.contexts = nextContexts;

83 + }

84 }

85

86 // Keep selectedContext in sync when the currently selected context's

@@ -74,17 +89,23 @@ const model = {

89 const selectedId = this.selected;

90 const updated = this.contexts.find((ctx) => ctx.id === selectedId);

91 if (updated) {

77 - this.selectedContext = updated;

78 - const nextExpandedParents = { ...this.expandedParents };

92 + if (this.selectedContext !== updated) this.selectedContext = updated;

93 if (updated.parent_context_id) {

80 - nextExpandedParents[updated.parent_context_id] = true;

94 + if (!this.expandedParents[updated.parent_context_id]) {

95 + this.expandedParents = {

96 + ...this.expandedParents,

97 + [updated.parent_context_id]: true,

98 + };

99 + }

100 } else if (

101 this.hasChildren(selectedId) &&

83 - nextExpandedParents[selectedId] === undefined

102 + this.expandedParents[selectedId] === undefined

103 ) {

85 - nextExpandedParents[selectedId] = true;

104 + this.expandedParents = {

105 + ...this.expandedParents,

106 + [selectedId]: true,

107 + };

108 }

87 - this.expandedParents = nextExpandedParents;

109 }

110 }

111 },

webui/components/sync/AGENTS.md

+2

@@ -13,6 +13,8 @@

13 ## Local Contracts

14

15 - Keep sync state compatible with WebSocket state-sync events.

16 +- A queued state push that finishes after transport loss must not overwrite the

17 + `DISCONNECTED` mode or flush reconnect notifications.

18 - Avoid noisy user-facing alerts for transient sync state unless existing UX expects them.

19 - Keep the compact status cluster free of native title tooltips; interactive

20 extensions must provide accessible names directly.

webui/components/sync/sync-store.js

+1

@@ -499,6 +499,7 @@ const model = {

499 await this.sendStateRequest({ forceFull: true });

500 },

501 });

502 + if (!stateSocket.isConnected()) return;

503 this._setMode(SYNC_MODES.HEALTHY, "push applied");

504 await this._flushPendingReconnectToast();

505 }

webui/index.js

+34 -25

@@ -355,6 +355,7 @@ export function buildStateRequestPayload(options = {}) {

355 log_from: forceFull ? 0 : lastLogVersion,

356 notifications_from: forceFull ? 0 : notificationStore.lastNotificationVersion || 0,

357 timezone,

358 + collections_delta: true,

359 };

360 }

361

@@ -382,8 +383,17 @@ export async function applySnapshot(snapshot, options = {}) {

383 return { updated: false };

384 }

385

386 + const hasCollections =

387 + Array.isArray(snapshot.contexts) && Array.isArray(snapshot.tasks);

388 + const extensionSnapshot = hasCollections

389 + ? snapshot

390 + : {

391 + ...snapshot,

392 + contexts: chatsStore.contexts,

393 + tasks: tasksStore.tasks,

394 + };

395 const snapCtx = {

386 - snapshot,

396 + snapshot: extensionSnapshot,

397 willUpdateMessages: lastLogVersion != snapshot.log_version,

398 skip: false,

399 };

@@ -433,25 +443,25 @@ export async function applySnapshot(snapshot, options = {}) {

443 setConnectionStatus(true);

444 }

445

436 - // Update chats list using store

437 - let contexts = snapshot.contexts || [];

438 - chatsStore.applyContexts(contexts);

446 + if (hasCollections) {

447 + // Update chats list using store

448 + chatsStore.applyContexts(snapshot.contexts);

449

440 - // Update tasks list using store

441 - let tasks = snapshot.tasks || [];

442 - tasksStore.applyTasks(tasks);

450 + // Update tasks list using store

451 + tasksStore.applyTasks(snapshot.tasks);

452

444 - // Make sure the active context is properly selected in both lists

445 - if (context) {

446 - // Update selection in both stores

447 - chatsStore.setSelected(context);

453 + // Make sure the active context is properly selected in both lists

454 + // Leave an empty selection unchanged so the welcome screen stays visible.

455 + if (context) {

456 + // Update selection in both stores

457 + chatsStore.setSelected(context);

458

449 - const contextInChats = chatsStore.contains(context);

450 - const contextInTasks = tasksStore.contains(context);

459 + const contextInChats = chatsStore.contains(context);

460 + const contextInTasks = tasksStore.contains(context);

461

452 - if (contextInTasks) {

453 - tasksStore.setSelected(context);

454 - }

462 + if (contextInTasks) {

463 + tasksStore.setSelected(context);

464 + }

465

466 if (!contextInChats && !contextInTasks) {

467 if (chatsStore.contexts.length > 0) {

@@ -466,19 +476,18 @@ export async function applySnapshot(snapshot, options = {}) {

476 deselectChat();

477 }

478 }

469 - } else {

470 - // No context selected: keep it that way so the welcome screen stays visible.

479 }

480 + }

481

473 - // update message queue

474 - messageQueueStore.updateFromPoll();

482 + // update message queue

483 + messageQueueStore.updateFromPoll();

484

476 - // A context switch is visually complete only after its matching snapshot

477 - // has rendered and the surrounding chat state has been synchronized.

478 - finishChatLoading(snapshot.context);

485 + // A context switch is visually complete only after its matching snapshot

486 + // has rendered and the surrounding chat state has been synchronized.

487 + finishChatLoading(snapshot.context);

488

480 - return { updated };

481 - }

489 + return { updated };

490 +}

491

492 export async function poll() {

493 try {

webui/js/AGENTS.md

+3 -1

@@ -53,7 +53,9 @@

53 - Convert standard TeX delimiters before Markdown parsing without touching inline or fenced code. Keep thought-card math rendering local to the agent-message handler rather than adding math flags to generic process-step or key/value rendering.

54 - Do not expose secrets in localStorage, console logs, URLs, or WebSocket payloads.

55 - Full message snapshots that start at backend log `no` 0 must replace the current message DOM before rendering; incremental snapshots should keep patching existing messages.

56 -- Long histories stay cached as raw log data but render a contiguous tail-first DOM window. The initial base view contains one 60-entry page; after paging, the base window contains two aligned pages, retaining the adjacent page and discarding only the far page in either direction. Visible boundaries expand to whole logical process groups so a page never reconstructs a partial group; the unit classifier must include plugin-backed process steps such as `code_exe`, and oversized groups use their own 50-step incremental window. Paging must preserve a visible anchor and occur at the scroll boundary after user intent, using passive loading indicators rather than count-bearing controls. Live entries and late content growth follow the tail until the reader deliberately moves away; historical window rebuilds must cancel pending auto-scroll effects, render in an off-screen staging history, and atomically swap fully laid-out content into the live scroller before restoring its anchor.

56 +- The state request builder advertises collection-delta support. An incremental snapshot may carry `contexts: null` and `tasks: null`; retain the current stores and skip sidebar selection/fallback reconciliation in that case. Extension hooks still receive the cached full collections so existing plugin contracts remain list-shaped.

57 +- Root responses with explicit `finished: false` render as escaped plain text while streaming, then switch to Markdown and LaTeX when finished; legacy responses without the flag remain formatted.

58 +- Long histories stay cached as raw log data but render a contiguous tail-first DOM window. The initial base view contains one 60-entry page; after paging, the base window contains two aligned pages, retaining the adjacent page and discarding only the far page in either direction. Visible boundaries expand to whole logical process groups so a page never reconstructs a partial group; the unit classifier must include plugin-backed process steps such as `code_exe`, and oversized groups use their own 50-step incremental window. A capped process group rebuilds only when a newly added step advances that window; updates to an existing step patch it in place. Paging must preserve a visible anchor and occur at the scroll boundary after user intent, using passive loading indicators rather than count-bearing controls. Live entries and late content growth follow the tail until the reader deliberately moves away; historical window rebuilds must cancel pending auto-scroll effects, render in an off-screen staging history, and atomically swap fully laid-out content into the live scroller before restoring its anchor.

59 - Message-window cache identity must keep different log types distinct even when they share a backend ID; root-agent GEN and response records intentionally use the same ID and must both survive replay, while same-ID/same-type updates still replace their earlier cached version.

60 - Utility records join a process render unit only when a substantive process step follows before the next standalone boundary. Group visibility must use that full-log classification rather than infer utility-only state from partially mounted DOM children. Standalone utility-only runs must not wrap root responses or reopen completed groups, and their group chrome stays hidden unless utility messages are enabled.

61 - Warnings classified into a full-log process render unit must remain process steps during replay even when the current DOM tail is already complete; use the live DOM tail only when render metadata is unavailable.

webui/js/message-window.js

+3

@@ -178,6 +178,7 @@ export class MessageWindow {

178 const previousStartKey = this._records[this.start]?.key || null;

179 const previousEndKey = this._records[this.end - 1]?.key || null;

180 const wasAtTail = followTail && this.end >= this._records.length;

181 + const addedKeys = new Set();

182 let requiresSort = false;

183

184 for (const message of Array.isArray(messages) ? messages : []) {

@@ -202,6 +203,7 @@ export class MessageWindow {

203 this._recordsByKey.set(key, record);

204 this._indexByKey.set(key, this._records.length);

205 this._records.push(record);

206 + addedKeys.add(key);

207 }

208 }

209

@@ -226,6 +228,7 @@ export class MessageWindow {

228 this.end = previousEnd >= 0 ? previousEnd + 1 : this.end;

229 this._clampBounds();

230 }

231 + return addedKeys;

232 }

233

234 showTail() {

webui/js/messages.js

+7 -5

@@ -205,7 +205,7 @@ async function setMessagesNow(messages, generation) {

205 const history = getChatHistoryEl();

206 const followTail = shouldFollowMessageTail();

207

208 - _messageWindow.merge(messages, { followTail });

208 + const addedMessageKeys = _messageWindow.merge(messages, { followTail });

209 bindMessageWindow(history);

210 if (_messageWindowRenderPromise) await _messageWindowRenderPromise;

211

@@ -217,6 +217,7 @@ async function setMessagesNow(messages, generation) {

217 const cappedProcessGroupUpdate = hasCappedProcessGroupUpdate(

218 messages,

219 windowMessages,

220 + addedMessageKeys,

221 );

222 if (initialWindow || compactedTail || cappedProcessGroupUpdate) {

223 return await renderMessageWindow({

@@ -661,7 +662,7 @@ function getProcessGroupRenderMessages(messages) {

662 });

663 }

664

664 -function hasCappedProcessGroupUpdate(messages, windowMessages) {

665 +function hasCappedProcessGroupUpdate(messages, windowMessages, addedMessageKeys) {

666 if (!messages.length) return false;

667 const groupStates = getProcessGroupPageState(windowMessages);

668 return messages.some((message) => {

@@ -670,7 +671,7 @@ function hasCappedProcessGroupUpdate(messages, windowMessages) {

671 const total = groupStates.get(group.key)?.steps.length || 0;

672 const limit = _processGroupStepLimits.get(group.key) ||

673 PROCESS_GROUP_STEP_PAGE_SIZE;

673 - return total > limit;

674 + return total > limit && addedMessageKeys.has(getMessageCacheKey(message));

675 });

676 }

677

@@ -1995,6 +1996,7 @@ export function drawMessageResponse({

1996 // no container or valid process group, create new container

1997 if (!container) container = getOrCreateMessageContainer(id, "left");

1998

1999 + const renderMarkdown = kvps?.finished !== false;

2000 const messageDiv = _drawMessage({

2001 messageContainer: container,

2002 heading: undefined,

@@ -2002,8 +2004,8 @@ export function drawMessageResponse({

2004 kvps: undefined,

2005 messageClasses: [],

2006 contentClasses: [],

2005 - markdown: true,

2006 - latex: true,

2007 + markdown: renderMarkdown,

2008 + latex: renderMarkdown,

2009 mainClass: "message-agent-response",

2010 smoothStream: false, // smooth render disabled, not reliable yet !isMassRender(), // stream smoothly if not in mass render mode

2011 });

Optimize realtime WebUI synchronization