Files
lobe-chat/.agents/skills/local-testing/scripts/agent-gateway/probe-dump.ts
T
Arvin Xu 837a3daa58 feat(chat): consume gateway uiMessages snapshot as SoT at step boundaries (#15153)
* ♻️ refactor(chat-store): useFetchMessages accepts options object

LOBE-9501

Replace the positional `skipFetch?: boolean` second argument with an
`options?: { skipFetch?, revalidateOnFocus? }` object on both
`useChatStore.useFetchMessages` and `useConversationStore.useFetchMessages`.
Plumb `revalidateOnFocus` through to the underlying SWR config so callers
can suppress focus revalidate per-call (default behaviour unchanged).

Mechanically migrate all 7 call sites to the new shape. No behaviour
change in this commit — the streaming-aware `revalidateOnFocus: false`
follow-up lives in the next commit.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

*  feat(chat): consume gateway uiMessages snapshot as SoT at step boundaries

LOBE-9501

Server attaches the canonical UIChatMessage[] snapshot to step_start and
agent_runtime_end events (#15152). The client now uses that pushed payload
as the source of truth instead of refetching from DB:

- step_start handler calls replaceMessages(uiMessages, { context }) when
  the snapshot is present, so the assistant tab-switch / next-step path
  no longer issues a refetch that returns a stale assistant placeholder.
- agent_runtime_end handler does the same for the terminal step — the
  last step has no later step_start to carry a fresh snapshot, so this
  branch is the only one that reconciles the final commit.
- step_complete on phase=tool_execution stops calling refreshMessages.
  That refetch was the direct cause of the assistantGroup→assistant
  clobber regression captured by the agent-gateway probe scripts.
- ChatList disables SWR revalidateOnFocus while the current topic is
  streaming (via operationSelectors.isAgentRuntimeRunningByContext) and
  automatically restores it after the run ends. Tab-focus during a run
  no longer triggers the stale DB read.

Doesn't touch streamingExecutor.ts (homogeneous runtime — parallel path).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* 🐛 fix(chat-store): wire gateway handler to consume server-pushed uiMessages SoT

LOBE-9501

#15152 (server) attaches the canonical UIChatMessage[] snapshot to both
the Redis SSE channel and the gateway /push-event channel. The earlier
client patch wired the consumer into `runAgent.ts`, but that file only
runs on the Group Chat SSE path. The actual gateway entry point
(`createGatewayEventHandler` in `gatewayEventHandler.ts`, used by single
agent, sub-agent, and hetero-CLI flows) ignored the field entirely and
kept refetching from DB.

Fix the gateway handler:

- step_start: consume `event.data.uiMessages` and replaceMessages with
  the pushed SoT. Skipped when absent — hetero adapters don't emit
  step_start at all (HeterogeneousEventType excludes it), so the new
  branch is invisible to hetero.

- agent_runtime_end: same SoT consumption; the existing
  `fetchAndReplaceMessages` becomes the fallback for events without the
  field. Claude Code adapter emits agent_runtime_end with empty data,
  so hetero terminal behavior is preserved by the fallback.

- stream_start: gate the DB fetch on `!newAssistantMessageId`. Native
  gateway streams carry `assistantMessage.id` (the preceding step_start
  also delivered the SoT), so the await is unnecessary — AND it was
  blocking the enqueue chain. Live chunks queued behind that await
  could not dispatch, which manifested as "streaming content never
  lands in messagesMap" during tab-switch and slow-network repros.
  Hetero CLI streams never set `assistantMessage.id`, so the fetch
  still runs for them on every stream_start.

Verified with the agent-gateway probe (separate commit): chunks now
land in real time (cLen grows 3 → 529 monotonically), and tab-switch
mid-stream no longer rolls the streamed assistantGroup back to the
LOADING placeholder (ROLLBACKS=none in the analyzer output).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🧪 chore(local-testing): rewrite agent-gateway probes in TS + add CLI

LOBE-9501

Convert the local-testing agent-gateway probes from .js/.mjs to TypeScript
and add a unified `run.ts` CLI that bundles via Bun.build (no extra
deps) and persists dumps to a gitignored `.agent-gateway/` directory for
use as streaming-replay test fixtures.

- types.ts: shared dump shape (ProbeStreamEvent / ProbeTimelineSample /
  ProbeDump) and `declare global` for the `window.__PROBE_*` surface
- probe-events.ts: WebSocket + fetch interception (gateway WS captures
  any socket with `operationId=`; fetch captures `/api/agent/stream` for
  direct SSE). Per-key timeline samples every 200ms so we can see
  which messagesMap key streaming chunks actually land in
- probe-dump.ts: stops the timeline timer and stashes JSON dump on
  `window.__PROBE_LAST_DUMP_JSON` (runner returns that global)
- analyze-events.ts: stream events (non-chunk) + chunks summary +
  action-call stacks + correlation + per-key assistant growth +
  rollback detection. Per-key growth was added specifically to
  diagnose "chunks arrive but assistant cLen never moves"
- run.ts: `install` | `dump [name]` | `analyze [path]` CLI. Bundles via
  Bun.build, wraps as IIFE with explicit return, pipes to
  `agent-browser eval --stdin`. Dumps land at
  `.agent-gateway/<name>-<YYYYMMDD-HHmmss>.json`

`.agent-gateway/` is gitignored so dumps accumulate across debugging
sessions without polluting git.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🐛 fix(local-testing): repair run.ts after autofix mangled path imports

LOBE-9501

The eslint --fix run during the previous commit applied the unicorn
`import-style` rule and renamed every `join(` / `dirname(` / `resolve(`
to `path.join(` / `path.dirname(` / `path.resolve(`, but the replacement
was a naive text substitution that:

1. rewrote `array.join('\n')` to `array.path.join('\n')` — broke bundle
   error reporting (would TypeError on the build-failure path)
2. produced `const path = path.join(DUMP_DIR, filename)` inside cmdDump
   — shadowed the `path` module with itself, ReferenceError on every
   dump invocation

Rename the local `path` to `dumpPath` and drop the spurious `.path`
prefix on the array `.join`. Verified round-trip: install + dump now
write a valid capture to `.agent-gateway/`.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🧪 chore(local-testing): capture per-call message snapshot in probe

LOBE-9501

The probe's `replaceMessages` wrapper used to record only `count` and
`params` — enough to see "two messages were written" but not WHICH two.
For post-stream collapse debugging we need to see whether each call
restored streamed content (cLen=N) or wiped to LOADING_FLAT (cLen=3).

Two changes:

- Capture `snapshot` field on every replaceMessages call: last 2
  messages' id / role / cLen / rLen / updatedAt. The analyzer prints
  this inline next to each call so reviewers can see content drift /
  collapse without re-reading the dump.

- Make wrapping idempotent across re-installs. The old guard
  `chat.__probeWrapped = true` froze the first-installed wrapper across
  re-installs, so updates to the probe body had no effect without a
  page reload. Stash the originals on
  `window.__PROBE_ORIG_REFRESH_MESSAGES` /
  `window.__PROBE_ORIG_REPLACE_MESSAGES` and re-wrap from those on
  every install.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🧪 chore(local-testing): add mutation log + dispatchMessage wrap to probe

LOBE-9501

The replaceMessages-only wrap couldn't catch chunk-level writes (those go
through internal_dispatchMessage) or attribute post-stream collapses to a
specific writer. Add:

- `__PROBE_MUTATIONS` — unified ordered log of every dbMessagesMap[key]
  reference change, with `last`/`prevLast` summaries and a `delta` field
  that tags interesting transitions (`cLen↓N→M`, `rLen↓`, `id:A→B`,
  `n↓prev→cur`). Both writers — replaceMessages AND internal_dispatchMessage
  — push to the same buffer so a single timeline shows all stores writes.

- Idempotent action wrapping. Originals are stashed on
  `window.__PROBE_ORIG_*` and re-wrapped from there on every install, so
  probe edits take effect without a page reload (previous
  `chat.__probeWrapped` flag froze the first wrapper).

- Snapshot field on replaceMessages — last 2 messages'
  id/role/cLen/rLen/updatedAt — so reviewers can see WHICH content each
  call is writing instead of just the count.

- Dump file now carries the `mutations` array alongside streamEvents,
  actionCalls, timeline.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🐛 fix(chat-store): gate SWR onData by isStreaming for streaming topic

LOBE-9501

Backstop for the post-stream cLen collapse that survives even with the
gateway SoT consume in place. Reproduction (confirmed):

1. Send a stream that lands lots of WS chunks into ChatStore
2. Immediately reload the page

If the page reload races against server-side chunk fan-out into Postgres,
SWR's fresh fetch returns the assistant row in its LOADING_FLAT placeholder
state (cLen=3) and writes that to ChatStore via the conversation-store
mirror — even though the WS push at agent_runtime_end carried the
correct full content moments earlier.

`mergeFetchedMessagesWithLocalState`'s updatedAt tie-breaker handles
this for in-session repros (local message wins when its updatedAt is
newer), but it degenerates when:

- The SoT consume just wrote server's snapshot updatedAt onto the local
  message, equalising the timestamps so the next stale DB fetch wins
- The user reloads (no local state to merge against — fresh fetch wins
  outright)

Add a gate at the bottom of `ConversationStore.useFetchMessages.onData`:
while `isAgentRuntimeRunningByContext(context)` is true, drop the SWR
write entirely. SWR's own cache still updates, so once streaming ends a
normal revalidate writes through correctly.

This is layered defense — it does NOT fix the underlying server-side
fan-out lag (filed as separate Linear issue). It does prevent the
client-side flash users currently see during the lag window.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* 🧪 test(chat-store): align gateway handler tests with SoT contract

The previous assertions still expected `stream_start` to issue a DB refetch
on every native gateway stream — the very behaviour LOBE-9501 removes
(`acb9523a04`). Update the three failing cases to the new contract:

- `stream_start > should associate new message with operation`:
  assert `messageService.getMessages` is NOT called when
  `assistantMessage.id` is present (the SoT snapshot from the preceding
  `step_start` already pre-populated `dbMessagesMap`).
- `sequential processing`: rewrite around the surviving ordering guarantee
  — `associate` (stream_start) must precede `dispatch` (stream_chunk) so
  the chunk targets the new id. Add a sibling case for hetero CLI streams
  (no `assistantMessage.id` → DB fetch is still mandatory).
- `multi-step integration > full LLM → tools → LLM cycle`: keep the
  post-`tool_end` `replaceMessages` assertion (tool_end still refreshes
  from DB), invert the post-`stream_start` assertion for step 2.

42 tests passing (was 41 + 1 new hetero fallback test).

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-24 20:05:58 +08:00

38 lines
1.1 KiB
TypeScript

// Stops the events-probe timeline timer and stashes the full capture as a
// JSON string on `window.__PROBE_LAST_DUMP_JSON`. `run.ts` wraps the bundle
// in an IIFE that returns that global, which `agent-browser eval` prints to
// stdout — the runner then persists it under `.agent-gateway/`.
import type { ProbeDump } from './types';
declare global {
interface Window {
__PROBE_LAST_DUMP_JSON?: string;
}
}
const w = window;
if (w.__PROBE_TIMELINE_TIMER) {
clearInterval(w.__PROBE_TIMELINE_TIMER);
w.__PROBE_TIMELINE_TIMER = null;
}
const mutations = w.__PROBE_MUTATIONS ?? [];
const dump: ProbeDump & { mutations: typeof mutations } = {
meta: {
t0: w.__PROBE_T0 ?? 0,
collectedAt: Date.now(),
sampleCount: (w.__PROBE_MSG_TIMELINE ?? []).length,
eventCount: (w.__PROBE_STREAM_EVENTS ?? []).length,
callCount: (w.__PROBE_ACTION_CALLS ?? []).length,
},
streamEvents: w.__PROBE_STREAM_EVENTS ?? [],
actionCalls: w.__PROBE_ACTION_CALLS ?? [],
timeline: w.__PROBE_MSG_TIMELINE ?? [],
mutations,
};
w.__PROBE_LAST_DUMP_JSON = JSON.stringify(dump);