Skip to content

feat: [ENG-2488] Wire resolveActiveProvider in daemon bootstrap#556

Merged
bao-byterover merged 1 commit into
proj/persis-task-historyfrom
feat/ENG-2488
Apr 27, 2026
Merged

feat: [ENG-2488] Wire resolveActiveProvider in daemon bootstrap#556
bao-byterover merged 1 commit into
proj/persis-task-historyfrom
feat/ENG-2488

Conversation

@bao-byterover
Copy link
Copy Markdown
Collaborator

No description provided.

@bao-byterover bao-byterover requested a review from leehpham as a code owner April 27, 2026 04:22
@bao-byterover bao-byterover merged commit 5263baa into proj/persis-task-history Apr 27, 2026
3 checks passed
@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Apr 27, 2026

Claude finished @bao-byterover's task in 3m 43s —— View job


PR Review: feat: [ENG-2488] Wire resolveActiveProvider in daemon bootstrap

  • Gather context and read changed files
  • Review for correctness, security, and code quality
  • Post summary with score

Summary

Overall score: 4/5

This is a focused, well-scoped change that wires resolveActiveProvider into the daemon bootstrap so the Web UI can display which provider/model handled each task. The design is intentionally best-effort — errors are swallowed by safeResolveActiveProvider in TaskRouter, so task dispatch is never blocked. A few minor concerns below.


Observations

suggestion — Per-task I/O overhead

// src/server/infra/daemon/brv-server.ts
async resolveActiveProvider() {
  const config = await resolveProviderConfig({
    authStateStore,
    providerConfigStore,
    providerKeychainStore,
    tokenRefreshManager,
  })
  ...
}

resolveProviderConfig does a providerConfigStore.read() (disk read) and potentially a keychain lookup or OAuth token refresh on every task creation. Since the callback is async and errors are swallowed (safeResolveActiveProvider), this never blocks dispatch — but it does add per-task I/O overhead. Worth considering a short-lived in-memory cache (e.g. keyed by activeProvider, TTL of a few seconds) to coalesce rapid bursts of task creation. Not a blocker given the fire-and-forget semantics, but worth tracking if task throughput grows.


nitpick — Repeated parameter construction

// GET_PROVIDER_CONFIG handler
resolveProviderConfig({authStateStore, providerConfigStore, providerKeychainStore, tokenRefreshManager})

// resolveActiveProvider callback (new)
resolveProviderConfig({authStateStore, providerConfigStore, providerKeychainStore, tokenRefreshManager})

The exact same 4-dependency object is constructed in two places in brv-server.ts. This is acceptable given the inline closure style used throughout, but if the params evolve, both spots need updating. A small shared helper like resolveCurrentProviderConfig() bound once in the outer scope would centralise this. Low priority.

Fix this →


nitpick — Redundant truthy guard on activeProvider

return {
  ...(config.activeModel ? {model: config.activeModel} : {}),
  ...(config.activeProvider ? {provider: config.activeProvider} : {}),
}

activeProvider is typed as string (non-optional in ProviderConfigResponse), so in practice it should never be an empty string unless the config is in a degraded state. The falsy guard is harmless and defensive, but worth noting that it deviates from the type contract. activeModel is string | undefined so its guard is fully correct.


praise — Failure isolation is correct

The comment "Failures are swallowed by TaskRouter's safeResolveActiveProvider — never blocks dispatch" correctly describes the fault-tolerance contract. Provider stamping is enrichment metadata, not critical path, so this is exactly the right design.


observation — No new tests in this PR

Given the repo's strict TDD policy, it's worth confirming that safeResolveActiveProvider in TaskRouter (on the base branch) has test coverage for the error-swallowing path, and that the resolveActiveProvider wiring in brv-server.ts is exercised by an integration test. If not, a follow-up test would be warranted.

@bao-byterover bao-byterover deleted the feat/ENG-2488 branch April 27, 2026 04:23
bao-byterover added a commit that referenced this pull request May 9, 2026
* feat: [ENG-2486] Extend transport schemas with provider/model (#554)

* feat: [ENG-2487] TaskRouter stamps provider + model on task:create (#555)

* feat: [ENG-2488] Wire resolveActiveProvider in daemon bootstrap (#556)

* feat: [ENG-2491] Domain entity + store interface (Level 2 schema) (#559)

* feat: [ENG-2492] FileTaskHistoryStore: Index + Data API (basic save/get/list)

* feat: [ENG-2498] Transport schemas (pagination, get, delete events) (#560)

* feat: [ENG-2495] FileTaskHistoryStore: Delete + Clear API (#561)

* feat: [ENG-2494] FileTaskHistoryStore: Stale recovery on read (#562)

* feat: [ENG-2496] Daemon llmservice accumulator + real-time TaskHistoryHook (#565)

* feat: [ENG-2497] Daemon bootstrap wiring + per-project store factory + startup audit (#567)

* feat: [ENG-2499] Daemon handlers (list paginated, get, delete, bulk, clear) + task:deleted broadcast (#568)

* feat: [ENG-2493] FileTaskHistoryStore: Prune + Compaction (#569)

* Feat/eng 2515 (#575)

* feat: [ENG-2515] Task history persistence — correctness fixes before WebUI ships

* fix: [ENG-2515] retention design + WebUI provider/model payload

Two correctness fixes surfaced while preparing the M2.10–M2.13 WebUI
handover:

1. Disable age-based task history retention by default. Age prune (30
   days) was a design mistake: it silently deletes user tasks even when
   the count cap (1000) has plenty of room left. Task history is a
   business artifact for audit/review, not a log — count-based rotation
   is the correct retention policy. Set the default to 0; deployments
   that want time-based eviction can still opt in via the constructor
   option.

2. Expose provider/model on the WebUI task:created subscriber. Daemon
   stamps both fields on every task:create (M1) but the local
   TaskCreatedPayload omitted them, so the chip M1.04 builds on would
   have no source. Extend the payload type and forward both fields into
   upsertStatus.

* Feat/eng 2515 (#576)

* feat: [ENG-2515] Task history persistence — correctness fixes before WebUI ships

* fix: [ENG-2515] retention design + WebUI provider/model payload

Two correctness fixes surfaced while preparing the M2.10–M2.13 WebUI
handover:

1. Disable age-based task history retention by default. Age prune (30
   days) was a design mistake: it silently deletes user tasks even when
   the count cap (1000) has plenty of room left. Task history is a
   business artifact for audit/review, not a log — count-based rotation
   is the correct retention policy. Set the default to 0; deployments
   that want time-based eviction can still opt in via the constructor
   option.

2. Expose provider/model on the WebUI task:created subscriber. Daemon
   stamps both fields on every task:create (M1) but the local
   TaskCreatedPayload omitted them, so the chip M1.04 builds on would
   have no source. Extend the payload type and forward both fields into
   upsertStatus.

* fix: [ENG-2515] eliminate task-history test flake — close cache-invalidation race and add deterministic flush API

Two pathologies surfaced as ~13% flake on local and similar on CI:

1. Cache-invalidation race in production code. A firePrune pass that
   started before a save() could finish AFTER the save's
   indexCache = undefined, overwriting the invalidation with a snapshot
   missing the just-saved row. Symptom: list(), clear(), deleteMany()
   returning N-1 entries after N awaited saves. Fix: bump indexEpoch on
   every index write (save / tombstone / recovery / compaction);
   doReadIndexDedup samples the epoch at the start and skips its
   setCache step if the epoch advanced. Also clear indexDedupInFlight
   on every write so a list() cannot pick up a pre-write in-flight
   promise.

2. Timing-based polling in tests. waitForPruneToSettle polled at 80 x
   4ms with a 4-stable-check exit, which under CI parallel-worker load
   missed the setTimeout(0) macrotask window. Fix: expose
   flushPendingOperations() — each firePrune wraps its setTimeout in a
   promise assigned to pruneChain; the public method awaits the chain
   (loops because pruneRequested re-extends it) and drains
   operationLock. Test callers replace polling with this single
   await store.flushPendingOperations().

Plus: tempDir entropy upgraded from Math.random() to randomUUID() in
two test files to remove a residual collision risk in CI parallel runs.

Verification: a new race-regression test
(file-task-history-store-cache-race.test.ts) reproduces the race at
~27% rate on production code and drops to 0% post-fix; running each
affected test file 30x consecutively yields 120/120 pass; full suite
npm test reports 7103 passing, 0 failing.

* fix: [ENG-2515] FileTaskHistoryStore — chunk tombstone appends under 4 KB to keep concurrent saves from corrupting JSON lines

Large clear() / deleteMany() batches previously issued a single appendFile
of all tombstones joined. POSIX guarantees nothing about regular-file write
atomicity (PIPE_BUF applies to FIFOs/sockets only). On filesystems that
don't serialize appends per inode, an unlocked concurrent save() can
interleave content into the middle of a multi-line tombstone write,
corrupting a JSON line. The line is then silently skipped by
IndexLineSchema.safeParse, so the tombstone is "missing" from the dedup
map even though the data file was already unlinked → list() returns a
ghost row whose getById() returns undefined.

Fix: introduce chunkLinesByBytes (exported for direct unit-test) and
issue tombstone appends in chunks under MAX_APPEND_CHUNK_BYTES (3.5 KB).
Each individual appendFile call stays under the 4 KB page boundary so
the kernel page-cache write path treats it as a single sector-level
write on common filesystems. Chunks are sequential under the existing
withOperationLock so concurrent saves can only land BETWEEN chunks,
never WITHIN one.

Tests: chunker unit test covers boundary + oversized-line + 200-tombstone
realistic case; integration test covers (a) clear() with 200 entries
removing all tombstones + data files, (b) every appendFile chunk well-formed
JSON post-clear, (c) clear() interleaved with 50 concurrent fresh saves
keeping the index parseable.

* refactor: [ENG-2515] move TaskHistoryEntry into shared/ + runtime layer-isolation test

The persisted-entry shape lived in src/server/core/domain/entities/. The
WebUI Tasks tab needs the same TS type (TaskGetResponse, stored detail),
which forced a webui→server import — inverting the boundary the ESLint
rule on tui/ already enforces.

Move the TS shape (TaskHistoryEntry, TaskHistoryEntryBase, TaskErrorData,
TaskHistoryStatus, TASK_HISTORY_SCHEMA_VERSION) into
src/shared/transport/events/task-events.ts. The Zod schema stays in
src/server/core/domain/entities/task-history-entry.ts and now carries
`satisfies z.ZodType<TaskHistoryEntry>`, so any drift between the runtime
schema and the type is a typecheck error. The entity file re-exports the
type so existing server-side imports continue to compile unchanged.

test/unit/shared/layer-isolation.test.ts: runtime walk over src/shared/
asserting no `from '...server/...'` (or other higher layers) appears in
any .ts/.tsx file. Catches regressions even if a future ESLint rule is
bypassed.

Incidental: src/server/infra/process/task-router.ts picks up formatting
churn from the prettier/eslint-fix pass on the same change set (line
wrapping only, no logic change).

* feat: [ENG-2490] add provider chip column to WebUI task list

* feat: [ENG-2500] wire WebUI delete + bulk delete + clear-completed mutations

* feat: [ENG-2501] render task detail from disk via task:get for cold tasks

* feat: [ENG-2503] dim styling + tooltip for daemon-interrupted task entries

* feat: [ENG-2490] cover (provider, '') edge case in formatProviderModel

* feat: [ENG-2500] keep selection on bulk-delete failure + cover Delete failed fallback

* feat: [ENG-2501] gate task:get on live status + pin enabled + infinite staleTime

* feat: [ENG-2503] tighten isInterrupted status type + cover precedence

* feat: [ENG-2503] use design-system Tooltip for interrupted-row hint

* feat: [ENG-2539] M2.16 task:list server-side filter/search + numbered… (#594)

* feat: [ENG-2539] M2.16 task:list server-side filter/search + numbered pagination

Schema: drop cursor (before/limit/nextCursor), add page/pageSize/total/pageCount
+ searchText, provider[], model[], createdAfter/Before, minDuration/maxDurationMs.
Response adds counts (Model A: matches current filter scope), availableProviders,
availableModels (history-derived, exclude pivots so dropdowns dont shrink).

Handler: 2-pass search. Pass-1 over index summary (content + error.message).
Pass-2 lazy-cracks data file via store.getById for completed tasks where full
result text might match. In-memory tasks read task.result directly (no I/O).
Race-safe: getById errors swallowed with log-once-per-query.

Store: ITaskHistoryStore.list rename after/before -> createdAfter/createdBefore,
add provider/model coarse pre-filter, drop limit (handler paginates).

Tests: 5 new test files updated. New cases cover all 7 filter dims, AND
combination, pass-2 happy path (full result match past byte 1500), pass-2
in-memory path, file-race swallow, counts Model A invariant counts.all === total,
availableProviders/Models exclude pivots + symmetric empty-string guard.

* fix: [ENG-2539] M2.16 PR review fixes

Address all 8 inline comments from PR #594:

- Docs: counts is Model A post-filter, counts.all === total invariant
  (was incorrectly described as pre-status-filter)
- Docs: page is echoed back as-sent without upper-bound clamp against
  pageCount; documented contract on JSDoc
- Interface: document why provider/model/status options remain on
  ITaskHistoryStore.list (direct-caller use; handler bypasses pivots)
- mapBounded signature: explicit fn returns R | undefined
- handleTaskList: refactor 2-pass to single-pass over candidatesNoSearch;
  derive merged inline; pass-2 reuses the map (was 2xN, now 1xN)
- Comment: clarify pass2Filter spread auto-inherits new non-pivot dims
- Schemas: bind via satisfies z.ZodType<X> to task-events.ts interfaces;
  re-export TaskListRequest/Response/Counts/AvailableModel from shared so
  the interface is single source of truth (compile-time drift prevention)
- Tests: add maxDurationMs upper-bound + no-store fallback coverage

* feat: [ENG-2502] paginate task list with useInfiniteQuery + Load more table row

* feat: [ENG-2502] keyboard-accessible Load more row + indent describe blocks

* feat: [ENG-2547] M2.17 WebUI numbered pagination + Filter dropdown + active filter tags

* feat: [ENG-2547] address bot review: memo serverStatus + closeTask, dedup useGetProviders + isDurationPreset, memo tags

* feat: [ENG-2610] M2.18 pagination polish + default page size 20 + friendly provider names

* feat: [ENG-2610] memoize providerNames against providersResponse, not derived array

---------

Co-authored-by: ncnthien <nhatthien185@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant