fix: reduce GitHub cleanup and watch polling load (#3866)
This commit is contained in:
Carlos Florêncio
@@ -0,0 +1,138 @@
|
||||
---
|
||||
created: 2026-09-22
|
||||
status: complete
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-001
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-002
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-POLLING-001
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-POLLING-002
|
||||
- REQ-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/watch-task-cleanup.md
|
||||
- ../../specs/integrations/system-design/github-pr-discovery-health.md
|
||||
- ../../specs/integrations/system-design/github-workflow-attention.md
|
||||
legacy_specs: []
|
||||
---
|
||||
|
||||
# Implementation plan: Watch task cleanup
|
||||
|
||||
## Overview
|
||||
|
||||
Prevent historical tasks from consuming provider quota. Implement GitHub
|
||||
eligibility first, GitHub batch termination second, and sibling filters third.
|
||||
Then reduce cleanup reads, bound fallback, add adaptive discovery, and cache
|
||||
Actions observations. All seven sequential work orders are complete.
|
||||
|
||||
## Source documents
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/watch-task-cleanup.md)
|
||||
- [System design](../../specs/integrations/system-design/watch-task-cleanup.md)
|
||||
- [Discovery requirements](../../specs/integrations/requirements/github-pr-discovery-health.md)
|
||||
- [Discovery design](../../specs/integrations/system-design/github-pr-discovery-health.md)
|
||||
- [Actions requirements](../../specs/integrations/requirements/github-workflow-attention.md)
|
||||
- [Actions design](../../specs/integrations/system-design/github-workflow-attention.md)
|
||||
|
||||
## Confirmed root cause and reproduction
|
||||
|
||||
GitHub cleanup list queries read historical deduplication rows without task
|
||||
eligibility checks. Service gates fetch remote state before engagement checks.
|
||||
Fetch errors become false eligibility results, so the batch continues.
|
||||
|
||||
A temporary in-memory SQLite reproduction executed the actual SQL extracted
|
||||
from both GitHub review list methods. Both returned task IDs `active`,
|
||||
`archived`, `deleted`, and an empty reservation. Only the first and last should
|
||||
be cleanup candidates. No production database or provider API was accessed.
|
||||
|
||||
GitLab review and issue queries also lack task eligibility filters. Azure
|
||||
DevOps review and work-item queries filter watch generation only. Both
|
||||
providers fetch upstream state before checking task engagement.
|
||||
|
||||
The current GitHub feedback helper requests PR data, reviews, comments, checks,
|
||||
and optional workflow attention. The exact request count differs from the
|
||||
reported older implementation, but historical-row amplification remains.
|
||||
|
||||
## Scope and compatibility
|
||||
|
||||
Preserve cleanup policies, lifecycle-prompt protection, empty GitHub
|
||||
reservations, explicit watch deletion, and reset behavior. Ignore historical
|
||||
records without changing deduplication. No UI or schema changes are required.
|
||||
Unarchived Done tasks remain eligible. Archive races after query selection
|
||||
remain outside the atomicity guarantee.
|
||||
|
||||
## Work orders
|
||||
|
||||
- [x] [Task 01: Filter GitHub cleanup candidates](task-01-github-candidates.md)
|
||||
- [x] [Task 02: Stop GitHub cleanup on rate limits](task-02-github-rate-limits.md)
|
||||
- [x] [Task 03: Filter sibling provider cleanup candidates](task-03-sibling-candidates.md)
|
||||
|
||||
- [x] [Task 04: Use minimal PR reads for cleanup](task-04-minimal-cleanup.md)
|
||||
- [x] [Task 05: Bound per-workspace PR fallback](task-05-bounded-fallback.md)
|
||||
- [x] [Task 06: Back off idle searching watches](task-06-adaptive-discovery.md)
|
||||
- [x] [Task 07: Cache Actions observations with bounded expiry](task-07-actions-cache.md)
|
||||
|
||||
## Polling and cache values
|
||||
|
||||
| Function | Shipped behavior | Bound/details |
|
||||
| --- | --- | --- |
|
||||
| PR poller scheduler tick | 1 minute | Fixed scheduler tick |
|
||||
| Searching, running or idle below 2 hours | 1 minute | Fast discovery |
|
||||
| Searching, idle 2 to below 24 hours | 15 minutes | Activity restores next-tick eligibility |
|
||||
| Searching, idle at least 24 hours | 30 minutes | Never permanently stops |
|
||||
| Known PR background checks | 1 minute | Fixed cadence |
|
||||
| Review-watch discovery and cleanup | 5 minutes | Eligible tasks only |
|
||||
| Issue-watch discovery and cleanup | 5 minutes | Eligible tasks only |
|
||||
| Status/search response TTL | 30 seconds | Outer response cache |
|
||||
| Explicit feedback response TTL | 8 seconds | Explicit refresh invalidates |
|
||||
| PR sync freshness window | 30 seconds | Passive status freshness |
|
||||
| Actions changing, empty, or attention-required | 30 seconds | Run and job observations |
|
||||
| Actions all complete, no attention required | 5 minutes | Same-SHA reruns appear after expiry |
|
||||
| Actions runs/jobs capacity | 512 entries each | In-memory bounded caches |
|
||||
| Actions batch concurrency / total budget | 4 / 5 seconds | Preserved |
|
||||
| Fallback target budget | 5 per workspace, 10 globally | Rotating per cycle |
|
||||
|
||||
Activity restores next-tick eligibility. Explicit refresh bypasses idle and
|
||||
Actions cache delays but never quota/auth admission. Poll ticks can add up to
|
||||
1 minute to the interval. Missing activity information selects fast polling.
|
||||
A completed-SHA external rerun may take up to 6 minutes to appear automatically.
|
||||
|
||||
Retain discovery rate-limit retry delays of 1, 2, 4, 8, then 15 minutes when
|
||||
no provider retry time is available. Existing provider retry deadlines are
|
||||
capped at 15 minutes by discovery health. Auth/config circuit backoff starts
|
||||
at 2 minutes, caps at 6 hours, and uses 25 percent jitter. The poller rate-limit
|
||||
sleep cap remains 10 minutes per wait; it does not clear the exhaustion state.
|
||||
|
||||
## Tests
|
||||
|
||||
Each work order lists new regression names, existing test locations, and exact
|
||||
commands. Use store-backed service tests to prove candidate selection reaches
|
||||
provider-call behavior. Browser tests add no evidence for this backend-only
|
||||
change. Run each regression before production edits and record its expected
|
||||
failure. Then run the affected package suites.
|
||||
|
||||
## Verification results
|
||||
|
||||
- `python3 scripts/list-docs.py validate`: passed, 299 decisions and 1110 specifications.
|
||||
- `python3 scripts/lint-spec-files.py --all`: passed.
|
||||
- `python3 scripts/lint-spec-files.test.py`: passed, 36 tests.
|
||||
- `git diff --check`: passed.
|
||||
|
||||
Implementation is complete. Production, permanent test, specification, plan,
|
||||
and public documentation changes are present in the working tree and remain
|
||||
uncommitted.
|
||||
|
||||
## Risks
|
||||
|
||||
- Existing tests often attach fictitious task IDs without task rows. Update
|
||||
active fixtures rather than weakening eligibility.
|
||||
- GitHub by-watch queries also support explicit watch deletion. Preserve its
|
||||
unfiltered inventory through existing task-ID methods.
|
||||
- A batch can stop only after an error becomes visible. Concurrent requests
|
||||
within its current feedback fetch may already be running.
|
||||
- Ignored historical deduplication rows remain stored. This fix bounds polling,
|
||||
not storage retention or quota consumed by other active work.
|
||||
|
||||
## Documentation impact
|
||||
|
||||
Internal requirements, designs, and work orders describe the expanded scope.
|
||||
No controls change. The public integration guide documents delayed idle
|
||||
discovery and bounded Actions freshness as shipped behavior.
|
||||
@@ -0,0 +1,103 @@
|
||||
---
|
||||
id: "01-github-candidates"
|
||||
title: "Filter GitHub cleanup candidates"
|
||||
status: done
|
||||
wave: 1
|
||||
depends_on: []
|
||||
plan: "plan.md"
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-001
|
||||
acceptance_criteria:
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.1
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.2
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.3
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.4
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.5
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/watch-task-cleanup.md
|
||||
---
|
||||
|
||||
# Task 01: Filter GitHub cleanup candidates
|
||||
|
||||
## Summary
|
||||
|
||||
Filter GitHub cleanup candidates. Implement with TDD and record the regression results.
|
||||
|
||||
## In scope
|
||||
|
||||
Add mixed-state regressions in `store_task_cleanup_test.go` and
|
||||
`service_review_cleanup_test.go`. Name the new tests
|
||||
`TestCleanupCandidateListsExcludeHistoricalTasks` and
|
||||
`TestCleanupMergedReviewTasksSkipsHistoricalTasks`.
|
||||
Cover by-watch, global, orphan, and workspace cleanup for reviews and issues.
|
||||
Assert zero upstream calls for archived and missing tasks. Cover terminal and
|
||||
nonterminal empty reservations, all policies, engagement, and lifecycle prompts.
|
||||
|
||||
Apply the candidate predicate to all four GitHub list methods. Preserve
|
||||
explicit watch deletion with unfiltered task-ID queries in `service_reviews.go`
|
||||
and `service_issues.go`. Add archived-child deletion assertions to existing
|
||||
watch deletion tests. Assert deduplication remains present through unfiltered
|
||||
lookups, not filtered cleanup queries.
|
||||
|
||||
Update obsolete missing-task cleanup expectations to zero provider calls and
|
||||
zero deletions. Keep task-not-found race coverage by deleting an initially
|
||||
eligible task after selection.
|
||||
|
||||
## Out of scope
|
||||
|
||||
New policies, UI, schema migrations, and unrelated provider changes.
|
||||
|
||||
## Acceptance
|
||||
|
||||
- Mixed candidate lists and store-backed cleanup satisfy AC 001.1 through 001.4.
|
||||
- Explicit watch deletion, reset, scope, and deduplication satisfy AC 001.5.
|
||||
- Regression tests fail before the query changes and pass afterward.
|
||||
|
||||
## Verification
|
||||
|
||||
Run from the repository root.
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
git diff --check
|
||||
```
|
||||
|
||||
## Files likely touched
|
||||
|
||||
- `apps/backend/internal/github/store.go`
|
||||
- `apps/backend/internal/github/service_reviews.go`
|
||||
- `apps/backend/internal/github/service_issues.go`
|
||||
- `apps/backend/internal/github/store_task_cleanup_test.go`
|
||||
- `apps/backend/internal/github/service_review_cleanup_test.go`
|
||||
- `apps/backend/internal/github/service_cleanup_policy_test.go`
|
||||
- `apps/backend/internal/github/store_watch_reset_test.go`
|
||||
- `apps/backend/internal/github/store_watch_disable_test.go`
|
||||
|
||||
## Dependencies
|
||||
|
||||
None.
|
||||
|
||||
## Risks
|
||||
|
||||
Preserve task inventory outside cleanup. Seed actual active tasks in fixtures.
|
||||
|
||||
## Parallelism
|
||||
|
||||
`sequential`
|
||||
|
||||
## Inputs
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/watch-task-cleanup.md)
|
||||
- [Design](../../specs/integrations/system-design/watch-task-cleanup.md)
|
||||
- Existing source and tests listed above.
|
||||
|
||||
## Results
|
||||
|
||||
Implemented the task eligibility predicate for all four GitHub cleanup list
|
||||
queries. Explicit watch deletion keeps its unfiltered task-ID inventory, so
|
||||
archived and missing task rows are preserved for deduplication while active
|
||||
tasks and empty reservations remain cleanup candidates. Added mixed-state
|
||||
store and service regressions, updated active-task fixtures, and retained a
|
||||
post-selection task-not-found race regression.
|
||||
|
||||
Verification: `go test ./internal/github -count=1` passed.
|
||||
@@ -0,0 +1,111 @@
|
||||
---
|
||||
id: "02-github-rate-limits"
|
||||
title: "Stop GitHub cleanup on rate limits"
|
||||
status: done
|
||||
wave: 2
|
||||
depends_on: ["01-github-candidates"]
|
||||
plan: "plan.md"
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-002
|
||||
acceptance_criteria:
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-002.1
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-002.2
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-002.3
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-002.4
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/watch-task-cleanup.md
|
||||
---
|
||||
|
||||
# Task 02: Stop GitHub cleanup on rate limits
|
||||
|
||||
## Summary
|
||||
|
||||
Stop GitHub cleanup on rate limits. Implement with TDD and record the regression results.
|
||||
|
||||
## In scope
|
||||
|
||||
Add `TestCleanupBatchStopsOnRateLimit` in
|
||||
`service_cleanup_policy_test.go`, with review and issue cases. Include an empty
|
||||
reservation first, partial deletion counts, wrapped 403/429 errors, unrelated
|
||||
403 errors, authenticated-user errors, context cancellation, and later retries.
|
||||
Add `TestCleanupPollCycleSkipsRemainingCleanupAfterRateLimit` in
|
||||
`poller_test.go`. Cover per-watch and final orphan calls for both poll loops.
|
||||
|
||||
Propagate errors through the private deletion gates, batch helpers, and public
|
||||
cleanup methods in `service_cleanup.go`. Use the existing typed classifier,
|
||||
including normalization of GitHub CLI stderr rate-limit failures at the
|
||||
`GHClient` boundary. Thread the resolved automation tracker into batch
|
||||
admission and stop against the resource used by the resolved client (including
|
||||
GraphQL for CLI-backed PR and issue reads). Include a stub that exhausts the
|
||||
tracker while returning successful feedback. Preserve scope: do not consult
|
||||
another workspace's tracker.
|
||||
|
||||
Preserve ordinary row-error continuation and completed counts. In `poller.go`,
|
||||
stop further cleanup calls for the current cycle after a rate-limit error.
|
||||
Keep scheduled recovery and discovery behavior intact.
|
||||
|
||||
## Out of scope
|
||||
|
||||
New policies, UI, schema migrations, and unrelated provider changes.
|
||||
|
||||
## Acceptance
|
||||
|
||||
- Review and issue batches stop on errors and tracker exhaustion.
|
||||
- Public cleanup methods preserve partial counts and expose the stop error.
|
||||
- Poller regressions prove remaining cleanup and orphan calls are suppressed.
|
||||
|
||||
## Verification
|
||||
|
||||
Run from the repository root.
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
git diff --check
|
||||
```
|
||||
|
||||
## Files likely touched
|
||||
|
||||
- `apps/backend/internal/github/service_cleanup.go`
|
||||
- `apps/backend/internal/github/poller.go`
|
||||
- `apps/backend/internal/github/service_cleanup_policy_test.go`
|
||||
- `apps/backend/internal/github/service_review_cleanup_test.go`
|
||||
- `apps/backend/internal/github/poller_test.go`
|
||||
- `apps/backend/internal/github/poller_cleanup_log_test.go`
|
||||
|
||||
## Dependencies
|
||||
|
||||
01-github-candidates.
|
||||
|
||||
## Risks
|
||||
|
||||
Preserve task inventory outside cleanup. Seed actual active tasks in fixtures.
|
||||
|
||||
## Parallelism
|
||||
|
||||
`sequential`
|
||||
|
||||
## Inputs
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/watch-task-cleanup.md)
|
||||
- [Design](../../specs/integrations/system-design/watch-task-cleanup.md)
|
||||
- Existing source and tests listed above.
|
||||
|
||||
## Results
|
||||
|
||||
Implemented rate-limit and cancellation stop propagation through review and
|
||||
issue cleanup batches. Batches preserve completed deletion counts, continue
|
||||
after ordinary row errors, honor the resolved workspace tracker (Core for REST
|
||||
clients and GraphQL for GHClient), and return recognized 403/429 errors.
|
||||
GHClient now normalizes CLI rate-limit
|
||||
stderr into the typed error path, and cleanup admission checks the resolved
|
||||
client's resource bucket so GraphQL-backed CLI cleanup does not continue after
|
||||
its quota is exhausted. Poller cycles stop before later watches and the final
|
||||
orphan sweep after a stop-worthy cleanup error. Added service and poller
|
||||
regressions for wrapped errors, tracker exhaustion, CLI-shaped rate limits,
|
||||
cancellation, partial counts, ordinary 403 continuation, and later retry
|
||||
recovery.
|
||||
Cleanup admission also reopens automatically once the tracked resource reset
|
||||
deadline has passed, even when no intervening provider call refreshed the
|
||||
tracker.
|
||||
|
||||
Verification: `go test ./internal/github -count=1` passed.
|
||||
@@ -0,0 +1,102 @@
|
||||
---
|
||||
id: "03-sibling-candidates"
|
||||
title: "Filter sibling provider cleanup candidates"
|
||||
status: done
|
||||
wave: 3
|
||||
depends_on: ["02-github-rate-limits"]
|
||||
plan: "plan.md"
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-001
|
||||
acceptance_criteria:
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.1
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.2
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.4
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.5
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/watch-task-cleanup.md
|
||||
---
|
||||
|
||||
# Task 03: Filter sibling provider cleanup candidates
|
||||
|
||||
## Summary
|
||||
|
||||
Filter sibling provider cleanup candidates. Implement with TDD and record the regression results.
|
||||
|
||||
## In scope
|
||||
|
||||
Add `TestCleanupCandidateListsExcludeHistoricalTasks` and
|
||||
`TestCleanupSkipsHistoricalTasks` to the GitLab and Azure DevOps packages.
|
||||
Create `gitlab/service_cleanup_test.go` beside the existing
|
||||
`service_cleanup_reason_test.go` and `service_cleanup_mr_automation_test.go`.
|
||||
Use their existing store and cleanup test fixtures. Cover active, archived,
|
||||
missing, and empty-task rows together. Assert no remote status calls for
|
||||
historical tasks and preserve active-task policies.
|
||||
|
||||
Filter GitLab review and issue queries in `store_watches.go`, including global,
|
||||
workspace, and by-watch methods. Filter Azure DevOps work-item and PR list
|
||||
queries in `watch_store.go`. Preserve workspace and generation constraints.
|
||||
GitLab and Azure DevOps continue to skip empty reservations in cleanup.
|
||||
Do not change sibling rate-limit handling or task deletion adapters.
|
||||
|
||||
## Out of scope
|
||||
|
||||
New policies, UI, schema migrations, and unrelated provider changes.
|
||||
|
||||
## Acceptance
|
||||
|
||||
- Both provider packages exclude historical tasks from upstream cleanup calls.
|
||||
- Active policy behavior and workspace/generation boundaries remain intact.
|
||||
- Empty reservations and reset inventory preserve existing semantics.
|
||||
|
||||
## Verification
|
||||
|
||||
Run from the repository root.
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/gitlab ./internal/azuredevops -count=1)
|
||||
git diff --check
|
||||
```
|
||||
|
||||
## Files likely touched
|
||||
|
||||
- `apps/backend/internal/gitlab/store_watches.go`
|
||||
- `apps/backend/internal/gitlab/store_task_cleanup_test.go`
|
||||
- `apps/backend/internal/gitlab/service_cleanup_test.go`
|
||||
- `apps/backend/internal/azuredevops/watch_store.go`
|
||||
- `apps/backend/internal/azuredevops/watch_store_test.go`
|
||||
- `apps/backend/internal/azuredevops/watch_cleanup_test.go`
|
||||
|
||||
## Dependencies
|
||||
|
||||
02-github-rate-limits.
|
||||
|
||||
## Risks
|
||||
|
||||
Preserve task inventory outside cleanup. Seed actual active tasks in fixtures.
|
||||
|
||||
## Parallelism
|
||||
|
||||
`sequential`
|
||||
|
||||
## Inputs
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/watch-task-cleanup.md)
|
||||
- [Design](../../specs/integrations/system-design/watch-task-cleanup.md)
|
||||
- Existing source and tests listed above.
|
||||
|
||||
## Results
|
||||
|
||||
Implemented the shared task eligibility predicate for GitLab review and issue
|
||||
candidate queries, including global, workspace, and by-watch reads. Azure
|
||||
DevOps work-item and pull-request reads now apply the same archived-or-missing
|
||||
task filter while preserving watch generation scoping. Empty reservations stay
|
||||
visible to cleanup, which continues to skip them before provider status calls.
|
||||
|
||||
Added mixed-state store regressions and instrumented cleanup tests for both
|
||||
providers. Active task policies still delete terminal records, while archived
|
||||
and missing task IDs generate zero provider calls. Updated SQLite fixtures to
|
||||
provide the central task table and seed active task rows where cleanup reads
|
||||
task eligibility.
|
||||
|
||||
Verification: `go test ./internal/gitlab ./internal/azuredevops -count=1`
|
||||
passed.
|
||||
@@ -0,0 +1,92 @@
|
||||
---
|
||||
id: "04-minimal-cleanup"
|
||||
title: "Use minimal PR reads for cleanup"
|
||||
status: done
|
||||
wave: 4
|
||||
depends_on: ["03-sibling-candidates"]
|
||||
plan: "plan.md"
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-001
|
||||
acceptance_criteria:
|
||||
- AC-INTEGRATIONS-WATCH-CLEANUP-001.6
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/watch-task-cleanup.md
|
||||
---
|
||||
|
||||
# Task 04: Use minimal PR reads for cleanup
|
||||
|
||||
## Summary
|
||||
|
||||
Use minimal PR reads for cleanup. Implement with TDD and preserve workspace identity.
|
||||
|
||||
## In scope
|
||||
|
||||
Replace cleanup-only GetPRFeedback calls with GetPR and conditional review
|
||||
reads. Fetch viewer identity only when an approval candidate requires it.
|
||||
Preserve all cleanup policies and approval matching. Propagate each fetch error
|
||||
to the batch handling from Task 02. Update existing cleanup stubs and tests.
|
||||
|
||||
Add TestCleanupUsesMinimalPRReads to service_review_cleanup_test.go. Assert
|
||||
zero comments, checks, workflow-run, and job reads. Terminal PRs require only
|
||||
GetPR. Open PRs retain approval-based cleanup. Include rate limits at each
|
||||
required fetch and preserve partial counts.
|
||||
|
||||
## Out of scope
|
||||
|
||||
New UI controls, credential policies, global schedulers, and unrelated providers.
|
||||
|
||||
## Acceptance
|
||||
|
||||
- Meet every linked acceptance criterion with deterministic regressions.
|
||||
- Preserve existing policy, scope, and error-handling contracts.
|
||||
- Record the expected red test result and passing package results.
|
||||
|
||||
## Verification
|
||||
|
||||
Run from the repository root.
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
git diff --check
|
||||
```
|
||||
|
||||
## Files likely touched
|
||||
|
||||
- `apps/backend/internal/github/service_cleanup.go`
|
||||
- `apps/backend/internal/github/service_review_cleanup_test.go`
|
||||
- `apps/backend/internal/github/service_cleanup_policy_test.go`
|
||||
|
||||
## Dependencies
|
||||
|
||||
03-sibling-candidates.
|
||||
|
||||
## Risks
|
||||
|
||||
Mocks must implement the smaller read path. Preserve existing review approval semantics.
|
||||
|
||||
## Parallelism
|
||||
|
||||
`sequential`
|
||||
|
||||
## Inputs
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/watch-task-cleanup.md)
|
||||
- [Design](../../specs/integrations/system-design/watch-task-cleanup.md)
|
||||
- Source files and existing adjacent tests listed above. New helper/test files are created as needed.
|
||||
|
||||
## Results
|
||||
|
||||
Implemented the minimal GitHub cleanup read path. Cleanup now fetches PR state
|
||||
with `GetPR`, fetches reviews only for open PRs, and requests the authenticated
|
||||
user only when at least one approval review exists. Cleanup no longer requests
|
||||
comments, check runs, workflow runs, or workflow jobs through the feedback
|
||||
helper. Fetch errors use the Task 02 batch handling, including rate-limit stop
|
||||
and partial-count behavior.
|
||||
|
||||
Added deterministic coverage for terminal and open PR reads, approval-based
|
||||
cleanup, historical-row cleanup, and rate limits. The focused red tests failed
|
||||
before the production change as expected. Verification passed:
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
```
|
||||
@@ -0,0 +1,109 @@
|
||||
---
|
||||
id: "05-bounded-fallback"
|
||||
title: "Bound per-workspace PR fallback"
|
||||
status: done
|
||||
wave: 5
|
||||
depends_on: ["04-minimal-cleanup"]
|
||||
plan: "plan.md"
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-POLLING-002
|
||||
acceptance_criteria:
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-002.1
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-002.2
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-002.3
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-002.4
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/github-pr-discovery-health.md
|
||||
---
|
||||
|
||||
# Task 05: Bound per-workspace PR fallback
|
||||
|
||||
## Summary
|
||||
|
||||
Bound per-workspace PR fallback. Implement with TDD and preserve workspace identity.
|
||||
|
||||
## In scope
|
||||
|
||||
Replace the all-workspace boolean result from tryBatchedPRWatchCheck with
|
||||
per-workspace outcomes. Retain auth, rate-limit, and invalid-query suppression.
|
||||
Publish successful workspace results even when another workspace fails.
|
||||
|
||||
Permit at most 5 fallback targets per workspace and 10 per cycle globally.
|
||||
Rotate both workspaces and targets across cycles. Unsupported GraphQL clients
|
||||
use the same budget. Stop the affected workspace on auth or rate-limit failure.
|
||||
Do not mark deferred work as checked or consume admission permits indefinitely.
|
||||
|
||||
Add TestPRWatchFallbackBudgetAndFairness to poller_test.go. Cover multiple
|
||||
workspaces, successful earlier batches, unsupported clients, transient errors,
|
||||
rate limits during fallback, cancellation, event publication, and fair rotation.
|
||||
Retain TestCheckPRWatches_BatchedAuthFailureSkipsPerWatchFallback.
|
||||
|
||||
## Out of scope
|
||||
|
||||
New UI controls, credential policies, global schedulers, and unrelated providers.
|
||||
|
||||
## Acceptance
|
||||
|
||||
- Meet every linked acceptance criterion with deterministic regressions.
|
||||
- Preserve existing policy, scope, and error-handling contracts.
|
||||
- Record the expected red test result and passing package results.
|
||||
|
||||
## Verification
|
||||
|
||||
Run from the repository root.
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
git diff --check
|
||||
```
|
||||
|
||||
## Files likely touched
|
||||
|
||||
- `apps/backend/internal/github/poller.go`
|
||||
- `apps/backend/internal/github/poller_test.go`
|
||||
- `apps/backend/internal/github/poller_circuit_test.go`
|
||||
- `apps/backend/internal/github/service_pr_watch_batched.go`
|
||||
|
||||
## Dependencies
|
||||
|
||||
04-minimal-cleanup.
|
||||
|
||||
## Risks
|
||||
|
||||
A target can cost several HTTP requests. The budget limits targets, not pagination.
|
||||
|
||||
## Parallelism
|
||||
|
||||
`sequential`
|
||||
|
||||
## Inputs
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/github-pr-discovery-health.md)
|
||||
- [Design](../../specs/integrations/system-design/github-pr-discovery-health.md)
|
||||
- Source files and existing adjacent tests listed above. New helper/test files are created as needed.
|
||||
|
||||
## Results
|
||||
|
||||
Implemented per-workspace PR batch outcomes and bounded REST fallback. A
|
||||
successful workspace publishes its results immediately and is never replayed
|
||||
through REST when another workspace fails. Authentication, rate-limit, and
|
||||
invalid-query outcomes remain deferred; transient failures and clients without
|
||||
GraphQL support use a rotating round-robin fallback with a five-target
|
||||
workspace budget and ten-target cycle budget. Fallback stops the affected
|
||||
workspace on rate-limit or authentication errors and stops the cycle on
|
||||
cancellation. The passive workspace refresh path applies the same five-target
|
||||
workspace and ten-target rolling one-minute budgets with fair target rotation;
|
||||
it classifies batch failures before fallback and runs fallback checks
|
||||
sequentially so an authentication or rate-limit error leaves no unscheduled
|
||||
tail of calls for that workspace.
|
||||
Equivalent workspace targets are deduplicated before workspace and global
|
||||
budgets are admitted, and one fallback result is applied to every watch in the
|
||||
target group.
|
||||
|
||||
Added deterministic coverage for mixed workspaces, successful event
|
||||
publication, unsupported and transient clients, target rotation, fallback
|
||||
budgets, rate-limit stopping, and cancellation. Verification passed:
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
```
|
||||
@@ -0,0 +1,153 @@
|
||||
---
|
||||
id: "06-adaptive-discovery"
|
||||
title: "Back off idle searching watches"
|
||||
status: done
|
||||
wave: 6
|
||||
depends_on: ["05-bounded-fallback"]
|
||||
plan: "plan.md"
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-POLLING-001
|
||||
acceptance_criteria:
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-001.1
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-001.2
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-001.3
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-001.4
|
||||
- AC-INTEGRATIONS-GITHUB-PR-POLLING-001.5
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/github-pr-discovery-health.md
|
||||
---
|
||||
|
||||
# Task 06: Back off idle searching watches
|
||||
|
||||
## Summary
|
||||
|
||||
Back off idle searching watches. Implement with TDD and preserve workspace identity.
|
||||
|
||||
## In scope
|
||||
|
||||
Add a due-target selector with an injected clock. Keep the scheduler tick at
|
||||
1 minute. Searching intervals are 1 minute below 2 hours idle or while running,
|
||||
15 minutes from 2 to below 24 hours, and 30 minutes thereafter.
|
||||
New watches receive a next-tick attempt. Known PRs remain at 1 minute.
|
||||
|
||||
Load task activity in bulk through a narrow adapter. Reuse task-domain query
|
||||
patterns without treating generic UpdatedAt or provider notifications as activity.
|
||||
Include session state, human conversation, agent execution, and observed branch
|
||||
commit/push activity. Missing evidence retains fast discovery. No provider reads
|
||||
or remote Git scans are allowed for activity classification.
|
||||
|
||||
Use the fastest eligible member for shared discovery targets. Preserve due
|
||||
state across reconciliation and restart using existing persisted activity and
|
||||
LastCheckedAt. Passive sync must honor scheduling; explicit refresh bypasses
|
||||
idleness only. Keep quota/auth admission intact.
|
||||
|
||||
Add TestSearchingWatchAdaptiveSchedule in a new poller_schedule_test.go.
|
||||
Cover exact boundaries, restart, activity after a slow check, unknown activity,
|
||||
all-idle and mixed shared targets, archive/delete, and external PR discovery
|
||||
without a new commit. Add service-level passive/explicit refresh regressions.
|
||||
Inspect LoadTaskLastActivity in task/repository/sqlite/task_status_summary.go
|
||||
as a query pattern, not as an assumed clean activity signal.
|
||||
|
||||
Update `docs/public/integrations.md` during implementation to describe the
|
||||
shipped freshness bounds. Keep the public guide unchanged during planning.
|
||||
Run these additional checks after that documentation edit:
|
||||
|
||||
```bash
|
||||
node --test scripts/validate-public-docs.test.mjs
|
||||
node scripts/validate-public-docs.mjs
|
||||
```
|
||||
|
||||
## Out of scope
|
||||
|
||||
New UI controls, credential policies, global schedulers, and unrelated providers.
|
||||
|
||||
## Acceptance
|
||||
|
||||
- Meet every linked acceptance criterion with deterministic regressions.
|
||||
- Preserve existing policy, scope, and error-handling contracts.
|
||||
- Record the expected red test result and passing package results.
|
||||
|
||||
## Verification
|
||||
|
||||
Run from the repository root.
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
git diff --check
|
||||
```
|
||||
|
||||
## Files likely touched
|
||||
|
||||
- `apps/backend/internal/github/poller.go`
|
||||
- `apps/backend/internal/github/poller_schedule.go`
|
||||
- `apps/backend/internal/github/poller_schedule_test.go`
|
||||
- `apps/backend/internal/github/store.go`
|
||||
- `apps/backend/internal/github/models.go`
|
||||
- `apps/backend/internal/github/service_pr_watch.go`
|
||||
- `apps/backend/internal/github/service_pr_watch_batched.go`
|
||||
- `apps/backend/internal/task/repository/interface.go` (only if the activity adapter needs a new task-domain query)
|
||||
- `apps/backend/internal/task/repository/sqlite/task_status_summary.go` and adjacent tests
|
||||
|
||||
If the task activity query changes, also run:
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/task/repository/sqlite -run "Activity|Watch" -count=1)
|
||||
```
|
||||
|
||||
- `docs/public/integrations.md`
|
||||
|
||||
## Dependencies
|
||||
|
||||
05-bounded-fallback.
|
||||
|
||||
## Risks
|
||||
|
||||
Provider writes can contaminate generic timestamps. Passive UI requests can defeat backoff.
|
||||
|
||||
## Parallelism
|
||||
|
||||
`sequential`
|
||||
|
||||
## Inputs
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/github-pr-discovery-health.md)
|
||||
- [Design](../../specs/integrations/system-design/github-pr-discovery-health.md)
|
||||
- Source files and existing adjacent tests listed above. New helper/test files are created as needed.
|
||||
|
||||
## Results
|
||||
|
||||
Implemented adaptive passive discovery scheduling with a one-minute scheduler
|
||||
tick. Searching watches check every minute while running or recently active,
|
||||
then back off to 15 minutes after two hours of inactivity and 30 minutes after
|
||||
24 hours. A newly created watch uses the newer of its creation time and the
|
||||
authoritative task activity timestamp, so an old task does not make a fresh
|
||||
watch enter a slow tier after its first check. Known pull requests remain on
|
||||
the one-minute cadence, and activity, restart, archive, and deletion behavior
|
||||
retain due state correctly.
|
||||
|
||||
Added a bulk task-activity projection and backend adapter. The projection uses
|
||||
human conversation, agent execution, session state, and observed branch
|
||||
activity, while excluding generic task timestamps, lifecycle messages, and
|
||||
live-monitor snapshots. Shared discovery targets use the fastest eligible
|
||||
watch, while numbered and searching watches sharing a branch keep separate due
|
||||
admission. Passive refresh and its bounded per-watch fallback honor due admission;
|
||||
explicit refresh remains immediate.
|
||||
|
||||
Passive workspace reads also use a one-minute per-workspace admission cooldown
|
||||
when no stale task requires refresh. New or reset watches clear that cooldown,
|
||||
and stale-task refreshes bypass it.
|
||||
|
||||
The task WebSocket sync boundary now carries an explicit refresh bit. Automatic
|
||||
subscription, reconnect, and retry requests use passive due admission, while a
|
||||
user refresh bypasses idle admission and still obeys quota and authentication
|
||||
gates. An explicit refresh supersedes an active passive provider attempt rather
|
||||
than reusing its result; concurrent explicit refreshes still share provider
|
||||
work, and discovery retry deadlines remain authoritative. Added deterministic scheduler, repository-query, production-shaped
|
||||
handler, and passive/explicit refresh regressions. Verification passed:
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
(cd apps/backend && go test ./internal/task/repository/sqlite -run "Activity|Watch" -count=1)
|
||||
node --test scripts/validate-public-docs.test.mjs
|
||||
node scripts/validate-public-docs.mjs
|
||||
```
|
||||
@@ -0,0 +1,137 @@
|
||||
---
|
||||
id: "07-actions-cache"
|
||||
title: "Cache Actions observations with bounded expiry"
|
||||
status: done
|
||||
wave: 7
|
||||
depends_on: ["06-adaptive-discovery"]
|
||||
plan: "plan.md"
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002
|
||||
acceptance_criteria:
|
||||
- AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.1
|
||||
- AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.2
|
||||
- AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.3
|
||||
- AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.4
|
||||
system_design:
|
||||
- ../../specs/integrations/system-design/github-workflow-attention.md
|
||||
---
|
||||
|
||||
# Task 07: Cache Actions observations with bounded expiry
|
||||
|
||||
## Summary
|
||||
|
||||
Cache Actions observations with bounded expiry. Implement with TDD and preserve workspace identity.
|
||||
|
||||
## In scope
|
||||
|
||||
Share workflow-run and job observations across batch enrichment and service
|
||||
status/feedback reads. Scope by credential generation and repository; runs also
|
||||
use head SHA, and jobs use run ID plus attempt. Classify each PR separately.
|
||||
|
||||
Use 30 seconds for empty, pending, running, unknown, or attention-required data.
|
||||
Use 5 minutes only for nonempty completed data without attention requirements.
|
||||
Keep 512 entries each for runs and jobs, and coalesce concurrent misses. Do not
|
||||
cache errors or change merge/check eligibility. Preserve the existing 4-worker,
|
||||
5-second batch enrichment budget.
|
||||
|
||||
Explicit refresh invalidates relevant outer and Actions cache entries. Prevent
|
||||
in-flight old observations from repopulating after invalidation. Head and
|
||||
credential changes must select fresh entries. Future mutation hooks can reuse
|
||||
this invalidation seam; no approval or rerun UI is added.
|
||||
|
||||
Add TestWorkflowAttentionCacheExpiryAndInvalidation to a new
|
||||
workflow_attention_cache_test.go. Cover same-SHA reruns, new run attempts,
|
||||
empty/approval/error results, concurrent callers, credential isolation, different
|
||||
PRs sharing one SHA, explicit refresh, and bounded eviction. Use fake time.
|
||||
Retain the existing workflow attention classification suite.
|
||||
|
||||
Update `docs/public/integrations.md` during implementation to describe the
|
||||
shipped freshness bounds. Keep the public guide unchanged during planning.
|
||||
Run these additional checks after that documentation edit:
|
||||
|
||||
```bash
|
||||
node --test scripts/validate-public-docs.test.mjs
|
||||
node scripts/validate-public-docs.mjs
|
||||
```
|
||||
|
||||
## Out of scope
|
||||
|
||||
New UI controls, credential policies, global schedulers, and unrelated providers.
|
||||
|
||||
## Acceptance
|
||||
|
||||
- Meet every linked acceptance criterion with deterministic regressions.
|
||||
- Preserve existing policy, scope, and error-handling contracts.
|
||||
- Record the expected red test result and passing package results.
|
||||
|
||||
## Verification
|
||||
|
||||
Run from the repository root.
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
git diff --check
|
||||
```
|
||||
|
||||
## Files likely touched
|
||||
|
||||
- `apps/backend/internal/github/service_pr_watch_batched.go`
|
||||
- `apps/backend/internal/github/service_pr.go`
|
||||
- `apps/backend/internal/github/service.go`
|
||||
- `apps/backend/internal/github/ttl_cache.go`
|
||||
- `apps/backend/internal/github/workflow_attention.go`
|
||||
- `apps/backend/internal/github/workflow_attention_cache.go`
|
||||
- `apps/backend/internal/github/workflow_attention_cache_test.go`
|
||||
- `apps/backend/internal/github/workflow_attention_test.go`
|
||||
|
||||
- `docs/public/integrations.md`
|
||||
|
||||
## Dependencies
|
||||
|
||||
06-adaptive-discovery.
|
||||
|
||||
## Risks
|
||||
|
||||
Same-SHA results can change. Never freeze a completed workflow collection indefinitely.
|
||||
|
||||
## Parallelism
|
||||
|
||||
`sequential`
|
||||
|
||||
## Inputs
|
||||
|
||||
- [Requirements](../../specs/integrations/requirements/github-workflow-attention.md)
|
||||
- [Design](../../specs/integrations/system-design/github-workflow-attention.md)
|
||||
- Source files and existing adjacent tests listed above. New helper/test files are created as needed.
|
||||
|
||||
## Results
|
||||
|
||||
Implemented bounded Actions observation caches shared by batched workflow
|
||||
enrichment and service status, feedback, and watch reads. Run entries are
|
||||
scoped by credential, repository, and head SHA. Job entries are scoped by
|
||||
credential, repository, run ID, and attempt. Completed nonempty observations
|
||||
without attention use a five-minute TTL; empty, changing, unknown, and
|
||||
attention-required observations use 30 seconds. Each cache has 512 entries,
|
||||
singleflight coalescing, and provider-error omission.
|
||||
|
||||
Explicit PR refresh invalidates the affected outer status or feedback entry
|
||||
and repository Actions entries. Automatic task subscriptions, reconnects, and
|
||||
retries use passive sync and preserve a completed Actions observation; only a
|
||||
user refresh carries explicit refresh intent and invalidates the relevant
|
||||
entries. Per-key epochs prevent an older in-flight read from repopulating an
|
||||
invalidated cache. Credential changes clear the caches and credential
|
||||
generations remain part of every workspace cache key. Existing workflow
|
||||
classification, fork identity, check state, merge eligibility, and the
|
||||
four-worker, five-second batch budget remain unchanged.
|
||||
|
||||
Added `TestWorkflowAttentionCacheExpiryAndInvalidation` and an in-flight
|
||||
invalidation regression covering same-SHA reruns, attempts, empty and
|
||||
attention data, provider errors, concurrent callers, credential isolation,
|
||||
shared SHAs, explicit refresh, and bounded eviction. The focused test failed
|
||||
before the cache seams existed as expected. Verification passed:
|
||||
|
||||
```bash
|
||||
(cd apps/backend && go test ./internal/github -count=1)
|
||||
node --test scripts/validate-public-docs.test.mjs
|
||||
node scripts/validate-public-docs.mjs
|
||||
```
|
||||
@@ -109,6 +109,19 @@ The **PR discovery failed** warning is separate from the quota values:
|
||||
|
||||
When discovery succeeds, the warning clears and the status revision advances. If the warning remains after a quota refresh, wait for the retry time or verify the selected connection and repository scope.
|
||||
|
||||
#### Pull request watch freshness
|
||||
|
||||
Background checks for a known pull request run at least once per minute. A
|
||||
watch that is still searching for a pull request also checks once per minute
|
||||
while its task is running or has activity within the last two hours. Searching
|
||||
checks use a 15-minute interval after two hours of inactivity and a 30-minute
|
||||
interval after 24 hours. The checks continue at the 30-minute interval, so a
|
||||
pull request opened later on an unchanged branch remains discoverable.
|
||||
|
||||
Passive workspace refreshes follow these intervals. An explicit pull request
|
||||
refresh checks immediately, subject to the workspace's authentication and
|
||||
GitHub rate limits.
|
||||
|
||||
### Automation and personal identity
|
||||
|
||||
PAT and CLI connections are human identities. They provide both workspace automation and the fallback identity for **My GitHub** views and user-triggered actions. Settings show this shared identity inside **Workspace GitHub access** instead of repeating it as a separate **My GitHub identity** section.
|
||||
@@ -341,6 +354,17 @@ For a task with linked GitHub pull requests, open the PR status control above th
|
||||
|
||||
If a current pull-request head has a GitHub Actions workflow that requires maintainer approval, the PR status control shows **Awaiting maintainer approval**, even when GitHub reports no checks. Detailed desktop and mobile views show the workflow name, the reason, and a link to GitHub. Approval-only workflow attention is not a failed check, does not start **Auto-fix CI & address comments**, and does not make the pull request ready for **Auto-merge or requeue when ready**. If GitHub does not provide enough evidence, Kandev keeps the workflow state unavailable or marks the last same-head observation as stale instead of claiming approval.
|
||||
|
||||
Kandev refreshes GitHub Actions observations for empty, running, changing, or
|
||||
attention-required results within 30 seconds. Nonempty completed results with
|
||||
no attention requirement remain cached for up to five minutes. The cache is
|
||||
Run observations are scoped to the selected credential, repository, and head
|
||||
SHA. Job observations are scoped to the selected credential, repository, run
|
||||
ID, and workflow attempt.
|
||||
An explicit pull request refresh clears the relevant observation immediately.
|
||||
Reruns on the same SHA appear after the completed-result cache expires or after
|
||||
an explicit refresh. Provider errors are retried and are not stored as an
|
||||
empty workflow result.
|
||||
|
||||
This is a GitHub-only lifecycle feature. Kandev reuses the existing lightweight task PR poller, which checks watched linked PRs roughly once per minute; it does not add a separate scheduler. Saving enabled options also evaluates the task's current linked PRs without waiting for the next poll.
|
||||
|
||||
Auto-fix waits for the pull request checks to settle, then sends one repair round for newly failed checks, review comments, an ordinary merge conflict, or an actionable merge-queue removal. It snapshots each check, comment, conflict, and removal state, so repeated observations do not create duplicate rounds. An authoritative conflict resolution clears its checkpoint without using a round; an unknown mergeability state preserves the checkpoint. Updating an already queued auto-fix message does not use another round, and auto-fix pauses after 10 rounds unless the user disables and re-enables it.
|
||||
|
||||
@@ -72,6 +72,21 @@ and provider status. Task and UI systems consume its results.
|
||||
immediate fallback using the same invalid query or an automatic retry more
|
||||
often than once per minute for that discovery target.
|
||||
|
||||
### REQ-INTEGRATIONS-GITHUB-PR-POLLING-001: Adaptive background discovery
|
||||
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-001.1:** Searching watches shall use a 1-minute interval while their task runs or has less than 2 hours of inactivity. They shall use 15 minutes from 2 hours to less than 24 hours, then 30 minutes.
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-001.2:** Task execution, task conversation activity, or an observed branch commit or push shall restore fast discovery. A new watch shall receive an initial attempt on the next poll tick.
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-001.3:** Idle discovery shall not stop permanently. A PR opened externally on an unchanged branch shall remain discoverable. Known PRs shall retain their 1-minute background cadence.
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-001.4:** Explicit PR refresh shall bypass the idle schedule but obey existing rate-limit and authentication admission. Passive frontend refreshes shall obey the idle schedule for searching targets. An explicit refresh shall not reuse the result of an active passive provider attempt.
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-001.5:** Poll attempts and provider status writes shall not count as task activity. Shared targets shall use the fastest eligible member's schedule. Archived and deleted members shall not accelerate it.
|
||||
|
||||
### REQ-INTEGRATIONS-GITHUB-PR-POLLING-002: Bounded batch fallback
|
||||
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-002.1:** Rate-limit, authentication, and invalid-query batch failures shall not trigger per-watch fallback for the affected workspace. Successful workspace batches shall not be repeated because another workspace failed.
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-002.2:** Transient failures shall permit at most 5 fallback target checks per workspace and 10 across one poll cycle. Deferred targets shall retain their state and receive a fair opportunity on later cycles.
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-002.3:** A rate-limit or authentication error during fallback shall stop remaining checks for that workspace. Cancellation shall stop the cycle. Existing retry deadlines remain authoritative.
|
||||
- **AC-INTEGRATIONS-GITHUB-PR-POLLING-002.4:** Clients without GraphQL support shall retain REST support under the same bounded schedule. Deferral shall not imply a successful provider observation or discard pending status events.
|
||||
|
||||
## Exclusions
|
||||
|
||||
No new credentials, automatic PR creation, inferred PR links from chat, changed
|
||||
@@ -86,3 +101,7 @@ runtime status, not a permanent incident archive.
|
||||
- [System design](../system-design/github-pr-discovery-health.md)
|
||||
- [Implementation package](../../../plans/github-pr-discovery-health/plan.md)
|
||||
- [Watch reconciliation repair](../../../plans/github-pr-watch-reconciliation/plan.md)
|
||||
|
||||
## Polling efficiency package
|
||||
|
||||
- [Implementation plan](../../../plans/watch-task-cleanup/plan.md)
|
||||
|
||||
@@ -26,11 +26,18 @@ The shared [task summary](../../ui/requirements/pr-task-status-summary.md) owns
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.2:** Approval-only workflows shall not count as failed, passed, or running checks. They shall not trigger CI repair rounds or enable automatic merging.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.3:** Desktop task summaries and the existing phone PR drawer shall show the same localized reason. Detailed surfaces shall link to GitHub.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.4:** When approval and failed checks coexist, Kandev shall show both. Each linked PR shall retain its own status.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.5:** After approval, cancellation, rerun, or a head change, the next successful refresh shall reconcile the reason. Reloads shall retain the latest stored observation.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.5:** After approval, cancellation, rerun, or a head change, the next successful provider refresh shall reconcile the reason. Explicit refresh shall bypass cached Actions observations. Reloads shall retain the latest stored observation.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.6:** Missing checks or unstable mergeability alone shall not imply approval. Unavailable provider evidence shall remain distinct from an observed absence of approval gates.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.7:** When GitHub requires an action but the reason is ambiguous, Kandev shall show "Workflow needs attention" instead of an approval claim.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001.8:** An approval reason shall replace unexplained "unstable" copy in the affected summary. Independent conflicts, reviews, queue state, and actual failures shall remain visible.
|
||||
|
||||
### REQ-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002: Bounded Actions caching
|
||||
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.1:** Equivalent Actions reads shall share cached results and concurrent fetches within one credential scope, repository, and head SHA. Different credentials shall never share results.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.2:** Background observations shall expire after 30 seconds for empty, pending, running, or attention-required results. Nonempty completed results without attention requirements shall expire after 5 minutes, including success and failure.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.3:** A head change, credential change, or explicit refresh shall prevent reuse of the affected stale observation. Same-SHA reruns shall become visible after expiry and the next successful poll.
|
||||
- **AC-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002.4:** Provider errors shall not become cached absence. Cached observations shall not become merge eligibility evidence. Classification shall remain specific to each PR, including fork identity.
|
||||
|
||||
## Out of scope
|
||||
|
||||
- Approving, rerunning, cancelling, or merging GitHub workflows or pull requests.
|
||||
@@ -41,3 +48,5 @@ The shared [task summary](../../ui/requirements/pr-task-status-summary.md) owns
|
||||
## Implementation plans
|
||||
|
||||
- [Workflow attention](../../../plans/github-workflow-attention/plan.md)
|
||||
|
||||
- [Polling efficiency](../../../plans/watch-task-cleanup/plan.md)
|
||||
|
||||
@@ -0,0 +1,41 @@
|
||||
---
|
||||
status: draft
|
||||
system: integrations
|
||||
created: 2026-09-22
|
||||
owners:
|
||||
- kandev
|
||||
---
|
||||
|
||||
# Watch task cleanup
|
||||
|
||||
## Overview
|
||||
|
||||
Integration cleanup must preserve provider capacity for active work. Historical
|
||||
tasks must not generate repeated provider requests. Integrations own this
|
||||
contract because they select cleanup candidates and request provider state.
|
||||
|
||||
## Requirements
|
||||
|
||||
### REQ-INTEGRATIONS-WATCH-CLEANUP-001: Cleanup eligibility
|
||||
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-001.1:** Every archived or deleted task shall generate zero provider state requests during a cleanup sweep. This applies to GitHub reviews and issues, GitLab reviews and issues, and Azure DevOps reviews and work items.
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-001.2:** In a collection containing active, archived, deleted, and unassigned records, each existing unarchived task shall remain eligible independently. Eligibility shall not depend on another record's state.
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-001.3:** GitHub reservations without an assigned task shall remain eligible. Terminal reservations shall be removed under `auto` and `always`, without increasing the deleted-task count. `never` shall preserve them without provider requests.
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-001.4:** Existing policies shall remain effective for active tasks. `never` shall make no cleanup state requests. `always` shall delete terminal tasks. `auto` shall preserve user-engaged tasks and existing lifecycle-prompt protections.
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-001.5:** Background and manual cleanup shall use the same task eligibility rules. Watch reset, explicit watch deletion, workspace authorization, and deduplication shall retain their existing behavior.
|
||||
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-001.6:** Review cleanup shall request only PR state and, for open PRs, review and viewer identity evidence. It shall not request comments, checks, workflow runs, or workflow jobs.
|
||||
|
||||
### REQ-INTEGRATIONS-WATCH-CLEANUP-002: GitHub rate-limit termination
|
||||
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-002.1:** After a GitHub cleanup batch receives a recognized rate-limit error, it shall start no further row checks. Completed deletions shall remain counted. Unprocessed tasks and reservations shall remain intact.
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-002.2:** Rate-limit recognition shall include HTTP 429 and HTTP 403 responses indicating rate limits or abuse detection. An unrelated HTTP 403 shall not be classified as a rate limit.
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-002.3:** A cleanup sweep shall return its rate-limit error to its caller. That poll cycle shall skip its remaining cleanup calls after this error. Ordinary row errors shall preserve the row and retain existing continuation behavior.
|
||||
- **AC-INTEGRATIONS-WATCH-CLEANUP-002.4:** A later scheduled sweep shall remain able to retry. Cancellation shall stop remaining batch work. Requests already running within one feedback fetch may finish or be cancelled.
|
||||
|
||||
## Exclusions
|
||||
|
||||
No cleanup policy, task retention setting, UI control, or provider credential
|
||||
contract is added. Unarchived tasks on a Done step remain active candidates.
|
||||
This contract does not guarantee spare quota against unrelated API consumers.
|
||||
GitLab and Azure DevOps reservation semantics remain unchanged.
|
||||
@@ -2,6 +2,8 @@
|
||||
status: draft
|
||||
system: integrations
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-POLLING-001
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-POLLING-002
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-DISCOVERY-001
|
||||
- REQ-INTEGRATIONS-GITHUB-PR-DISCOVERY-HEALTH-001
|
||||
created: 2026-09-11
|
||||
@@ -187,3 +189,86 @@ historical scope and results.
|
||||
- [Sync coordination](github-task-pr-sync-coordination.md)
|
||||
- [Authentication ownership ADR](../../../decisions/0047-github-authentication-ownership.md)
|
||||
- [Implementation package](../../../plans/github-pr-discovery-health/plan.md)
|
||||
|
||||
## Adaptive discovery schedule
|
||||
|
||||
`REQ-INTEGRATIONS-GITHUB-PR-POLLING-001` maps to this section.
|
||||
Keep `defaultPRPollInterval` at 1 minute. Select due searching targets before
|
||||
batch construction. Known numbered watches keep the current cadence.
|
||||
|
||||
| Searching task state | Minimum interval |
|
||||
| --- | --- |
|
||||
| Running or activity age below 2 hours | 1 minute |
|
||||
| Activity age from 2 hours to below 24 hours | 15 minutes |
|
||||
| Activity age at least 24 hours | 30 minutes |
|
||||
|
||||
Use injected time and `PRWatch.LastCheckedAt` for due checks. Round execution
|
||||
to the next 1-minute tick, so actual delay may include one extra tick.
|
||||
Use watch creation as an activity baseline for the first 2 hours.
|
||||
A never-checked watch is immediately eligible. Missing activity evidence
|
||||
fails open to fast polling. No new random delay is required in this package.
|
||||
|
||||
Add a narrow task-activity projection, loaded in bulk for eligible watch tasks.
|
||||
Reuse the bounded query pattern in
|
||||
`task/repository/sqlite/task_status_summary.go:LoadTaskLastActivity`, but not
|
||||
its generic task-update source. Preserve that existing UI query unchanged.
|
||||
Use persisted session execution state and conversation timestamps, plus observed
|
||||
commit/push activity. Do not use generic task or watch `UpdatedAt`: GitHub sync
|
||||
can update those fields itself. Exclude automatic provider notifications from
|
||||
conversation activity. Include human messages and agent execution activity.
|
||||
Do not scan remote branches or invoke GitHub to calculate idleness.
|
||||
Record observed branch activity from existing local Git event paths; absence of
|
||||
such an event is not proof that no remote push occurred.
|
||||
|
||||
An activity transition restores next-tick eligibility even after a slow poll.
|
||||
Persisted task activity and last-check timestamps survive restart. A runtime
|
||||
branch-activity marker may reset on restart without stopping periodic discovery.
|
||||
Expire runtime markers after 24 hours and remove them with their watch.
|
||||
Reconciliation of an unchanged watch must preserve its age and last check.
|
||||
|
||||
Group equivalent discovery targets before admission. Any active member selects
|
||||
the 1-minute schedule for that target; do not duplicate the provider query.
|
||||
Numbered targets and searching targets remain separate groups even when they
|
||||
share a repository and branch, because their polling cadences differ.
|
||||
Passive task/page refresh must use the same admission rule or it defeats the
|
||||
background savings. Explicit user refresh bypasses only idle admission, never
|
||||
quota or auth gates. Preserve `PRSyncFreshnessWindow` at 30 seconds for existing
|
||||
sync consumers and keep the distinction between automatic and explicit reads.
|
||||
If a passive provider attempt is already active, an explicit refresh starts a
|
||||
new admitted attempt instead of joining the passive result. The superseded
|
||||
attempt cannot publish its result, and concurrent explicit refreshes share the
|
||||
explicit batch singleflight key. A retry deadline from an authentication,
|
||||
rate-limit, or invalid-query failure remains authoritative.
|
||||
When a passive workspace read finds no stale task, admit workspace discovery
|
||||
reads at most once per minute. A newly created or reset watch clears that
|
||||
cooldown, and stale-task refreshes bypass it.
|
||||
|
||||
## Bounded fallback
|
||||
|
||||
`REQ-INTEGRATIONS-GITHUB-PR-POLLING-002` maps to this section.
|
||||
Replace the poller's all-or-nothing boolean batch outcome with per-workspace
|
||||
outcomes: completed, deferred, fallback-eligible. Publish successful results
|
||||
without waiting for every workspace to succeed. Preserve existing shared
|
||||
health admission for rate limits and invalid queries, and auth circuits.
|
||||
|
||||
Allow 5 fallback target checks per workspace, capped at 10 per cycle globally.
|
||||
Apply this limit to unsupported-GraphQL clients too. Rotate workspace and target
|
||||
selection between cycles so fixed ordering cannot starve later targets.
|
||||
These are target-check limits, not HTTP-request limits: a check may paginate
|
||||
or fetch several resources. Stop a workspace on the first authentication,
|
||||
configuration, rate-limit, or invalid-query error and check cancellation before
|
||||
each target. Never retry successful batches via REST.
|
||||
Do not mark deferred targets as successfully checked.
|
||||
Deduplicate equivalent watches before consuming either budget and apply one
|
||||
fallback result to every watch in the selected target group.
|
||||
|
||||
Retain existing failure schedules: discovery retry starts at 1 minute and caps
|
||||
at 15 minutes; auth/config circuits start at 2 minutes and cap at 6 hours with
|
||||
25 percent jitter. Provider reset deadlines and credential changes retain
|
||||
existing handling. This package adds no independent retry loop.
|
||||
|
||||
## Polling efficiency delivery
|
||||
|
||||
See [the implementation plan](../../../plans/watch-task-cleanup/plan.md).
|
||||
Use deterministic clock tests for exact 2-hour and 24-hour boundaries,
|
||||
restart, mixed groups, passive reads, explicit refresh, and fallback fairness.
|
||||
|
||||
@@ -2,6 +2,7 @@
|
||||
status: draft
|
||||
system: integrations
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002
|
||||
- REQ-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-001
|
||||
---
|
||||
|
||||
@@ -144,3 +145,42 @@ Desktop and mobile Playwright tests verify the existing entry points and the pro
|
||||
- [Requirements](../requirements/github-workflow-attention.md)
|
||||
- [Shared task summary design](../../ui/system-design/pr-task-status-summary.md)
|
||||
- [Implementation plan](../../../plans/github-workflow-attention/plan.md)
|
||||
|
||||
## Expiring Actions observations
|
||||
|
||||
`REQ-INTEGRATIONS-GITHUB-WORKFLOW-ATTENTION-002` maps to this section.
|
||||
Share raw workflow-run observations across batched enrichment and service
|
||||
feedback/status reads. Keep per-PR classification outside the shared cache.
|
||||
Key entries by credential scope and generation, normalized repository identity,
|
||||
and head SHA. Reuse the bounded `ttlCache` and its singleflight/invalidation
|
||||
patterns; add per-entry expiry support if needed. Keep at most 512 run entries
|
||||
and 512 job entries. Errors are not cached.
|
||||
|
||||
| Observation | TTL |
|
||||
| --- | --- |
|
||||
| Empty list or any pending/running/unknown status | 30 seconds |
|
||||
| Any approval/action-required or ambiguous attention result | 30 seconds |
|
||||
| Nonempty, all completed, no attention requirement | 5 minutes |
|
||||
|
||||
A SHA does not freeze the workflow run collection. Reruns and new runs can
|
||||
arrive for the same SHA. Never use an infinite TTL. Cache job observations by
|
||||
credential scope, repository, run ID, and attempt, with the same expiry classes.
|
||||
A run-attempt change must not reuse jobs from an earlier attempt.
|
||||
|
||||
Keep existing batched enrichment limits: concurrency 4 and total budget
|
||||
5 seconds. Preserve the outer status/search cache TTL of 30 seconds and the
|
||||
feedback TTL of 8 seconds. Long Actions entries only affect workflow attention,
|
||||
not checks, review status, or auto-merge decisions.
|
||||
|
||||
Explicit refresh invalidates the relevant Actions and outer response entries,
|
||||
then coalesces concurrent refreshes. Generation checks prevent an older
|
||||
in-flight request from repopulating invalidated data. Passive page polling
|
||||
uses the cache. Route future approval/rerun mutations through this invalidation
|
||||
seam; adding those mutations is outside this package.
|
||||
|
||||
External reruns on a completed SHA can take up to 5 minutes plus the next
|
||||
1-minute poll tick to appear. Attention and running results use the short TTL.
|
||||
Provider errors retain unknown/stale-positive semantics and existing retry
|
||||
admission. Terminal PRs continue to skip Actions reads altogether.
|
||||
|
||||
See [the implementation plan](../../../plans/watch-task-cleanup/plan.md).
|
||||
|
||||
@@ -0,0 +1,126 @@
|
||||
---
|
||||
status: draft
|
||||
system: integrations
|
||||
requirements:
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-001
|
||||
- REQ-INTEGRATIONS-WATCH-CLEANUP-002
|
||||
created: 2026-09-22
|
||||
owners:
|
||||
- kandev
|
||||
---
|
||||
|
||||
# Watch task cleanup design
|
||||
|
||||
## Ownership and mapping
|
||||
|
||||
Integration stores select candidates. Integration services preserve cleanup
|
||||
policies and call the existing task deletion adapters.
|
||||
|
||||
| Requirement | Design section |
|
||||
| --- | --- |
|
||||
| REQ-INTEGRATIONS-WATCH-CLEANUP-001 | Candidate selection and compatibility |
|
||||
| REQ-INTEGRATIONS-WATCH-CLEANUP-002 | GitHub batch termination |
|
||||
|
||||
## Candidate selection and compatibility
|
||||
|
||||
Cleanup queries use this predicate for each deduplication row `d`:
|
||||
|
||||
```sql
|
||||
d.task_id = '' OR EXISTS (
|
||||
SELECT 1 FROM tasks t
|
||||
WHERE t.id = d.task_id AND t.archived_at IS NULL
|
||||
)
|
||||
```
|
||||
|
||||
Keep watch, workspace, and generation predicates outside this parenthesized
|
||||
expression. A plain inner join loses reservations. A left join with only an
|
||||
archive-null check incorrectly includes deleted tasks.
|
||||
|
||||
Apply this filter to GitHub `ListReviewPRTasksByWatch`, `ListAllReviewPRTasks`,
|
||||
`ListIssueWatchTasksByWatch`, and `ListAllIssueWatchTasks` in `store.go`.
|
||||
Workspace cleanup already uses the all-record queries and watch ownership.
|
||||
There are no GitHub workspace-specific list methods in this checkout.
|
||||
|
||||
`DeleteReviewWatch` and `DeleteIssueWatch` also consume the by-watch methods.
|
||||
Move their explicit deletion inventory to the existing unfiltered
|
||||
`ListReviewPRTaskIDsByWatch` and `ListIssueWatchTaskIDsByWatch` methods.
|
||||
Preserve their authorization, empty-ID handling, and best-effort deletion.
|
||||
Reset and deduplication lookups remain unfiltered.
|
||||
|
||||
GitLab applies the predicate to all, workspace, and by-watch review and issue
|
||||
queries in `store_watches.go`. Preserve workspace joins, generation values,
|
||||
and ordering. Its cleanup currently skips empty reservations and retains that
|
||||
behavior.
|
||||
|
||||
Azure DevOps applies the predicate to `ListWorkItemWatchTasks` and
|
||||
`ListPullRequestWatchTasks` in `watch_store.go`. Preserve watch generation
|
||||
scoping. Its cleanup also continues to skip empty reservations.
|
||||
|
||||
Archived and missing-task records are ignored, not purged by this change.
|
||||
Retaining them avoids changing deduplication and allows restored tasks to
|
||||
become eligible again. Existing task-deletion handlers may still remove them.
|
||||
No schema migration or index is needed: the existence query uses the task key.
|
||||
|
||||
Eligibility reflects committed task state at query time. This change does not
|
||||
make remote requests atomic with a concurrent archive or deletion.
|
||||
|
||||
## GitHub batch termination
|
||||
|
||||
Extend the private deletion gates to return errors alongside eligibility and
|
||||
reason. Extend both batch helpers to return a partial count and error.
|
||||
Use `errors.As` with `GitHubAPIError` and the existing
|
||||
`isGitHubRateLimitAPIError` classifier in `errors.go`.
|
||||
|
||||
Check cancellation before each row. On a rate-limit error, return immediately,
|
||||
including on the empty-reservation path. Ordinary fetch errors retain existing
|
||||
per-row failure tracking and continuation. Do not classify every 403 as a
|
||||
rate limit. Propagate rate-limit errors from `GetAuthenticatedUser` as well.
|
||||
|
||||
Per-watch, global, orphan, and workspace cleanup entry points propagate the
|
||||
partial count and error. Global sweeps stop before the next workspace batch.
|
||||
In each review or issue poll cycle, a rate-limit error suppresses subsequent
|
||||
cleanup calls, including its final orphan sweep. Discovery remains subject to
|
||||
its existing search-bucket rules. This is a cycle-local stop, not a new global
|
||||
credential circuit or scheduler.
|
||||
|
||||
Keep existing client rate tracking and reset behavior. A later poll retries
|
||||
normally. Emit one batch-level error through existing caller logging instead
|
||||
of one new failure for each unvisited row. Unvisited rows do not increment
|
||||
failure counters.
|
||||
Cleanup admission evaluates the current reset deadline, so an exhausted
|
||||
resource becomes eligible again after reset even if no provider call refreshes
|
||||
the tracker in between sweeps.
|
||||
|
||||
## Minimal review cleanup lookup
|
||||
|
||||
Replace the cleanup gate's `GetPRFeedback` call with `GetPR`. For merged or
|
||||
closed PRs, apply the existing policy without additional provider reads.
|
||||
For open PRs, fetch `ListPRReviews` and the authenticated user only when needed
|
||||
for the existing approval rule. Preserve approval matching and policy behavior.
|
||||
Never fetch comments, check runs, workflow runs, or workflow jobs for cleanup.
|
||||
Reuse existing authenticated-user caching. Do not introduce another cleanup TTL.
|
||||
Propagate rate-limit errors from each required read to the batch stop path.
|
||||
Thread the resolved automation `RateTracker` into batch admission and honor
|
||||
core exhaustion before each row. Never substitute another workspace's tracker.
|
||||
|
||||
## Test boundaries
|
||||
|
||||
Use real store fixtures and instrumented clients. Cover mixed collections,
|
||||
all cleanup entry points, policy matrices, empty reservations, preserved
|
||||
deduplication, and explicit watch deletion. Fixtures representing active tasks
|
||||
must insert actual task rows. Missing-task fixtures must assert zero provider
|
||||
calls rather than treating an absent task as active.
|
||||
|
||||
Rate tests cover wrapped typed errors, ordinary 403 responses, cancellation,
|
||||
partial counts, an error on the first reservation, tracker exhaustion, and
|
||||
later-cycle recovery. Poller tests prove the final orphan sweep does not restart
|
||||
cleanup after a rate-limit stop.
|
||||
|
||||
## Related contracts
|
||||
|
||||
- [GitHub identity ownership](../../../decisions/0047-github-authentication-ownership.md)
|
||||
- [PR lifecycle prompt protection](../../../decisions/0051-pr-agent-notifications-extend-task-pr-automation.md)
|
||||
- [Requirements](../requirements/watch-task-cleanup.md)
|
||||
|
||||
Existing authorization and lifecycle protections remain authoritative. This
|
||||
local candidate filter needs no new architecture decision.
|
||||
Reference in New Issue
Block a user