[feat] discover_triggers: 30-60s to <1s via a cached trigger catalog - #5189
Conversation
|
@coderabbitai review |
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
📝 WalkthroughWalkthroughThe PR implements cached full-catalog trigger discovery: it fetches Composio events once, caches the snapshot in Redis, scores events in memory, adds integration filtering and ranking changes, and preserves existing discovery contracts and browse routes. It also adds configuration, tests, evaluation tooling, agent guidance, and design documentation. ChangesTrigger Catalog Discovery
Estimated code review effort: 4 (Complex) | ~60 minutes Sequence Diagram(s)sequenceDiagram
participant Client
participant TriggersService
participant Redis
participant ComposioCatalog
Client->>TriggersService: discover_triggers
TriggersService->>Redis: read provider catalog
Redis-->>TriggersService: cached snapshot or miss
TriggersService->>ComposioCatalog: list_all_events on miss
ComposioCatalog-->>TriggersService: paginated catalog snapshot
TriggersService->>Redis: store snapshot with TTL
TriggersService->>TriggersService: score and rank events in memory
TriggersService-->>Client: discovered events and alternatives
Possibly related PRs
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
✅ Action performedReview finished.
|
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 5
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: a846a837-a911-4e62-81f9-262a2b2d9c8e
📒 Files selected for processing (5)
docs/design/agent-workflows/projects/trigger-discovery-catalog/README.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/context.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/plan.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/research.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/status.md
mmabrouk
changed the title
dosubot
Bot
added
Backend
enhancement
| ) | ||
|
|
||
| async def _discover_events_for_use_case( | ||
| async def _cached_event_catalog( |
There was a problem hiding this comment.
Review guide 1/5 — the cache seam. This mirrors ToolsService._cached_search one-to-one: provider-only key (project-agnostic on purpose, connection state is joined fresh per project downstream), namespace triggers:catalog, TTL from COMPOSIO_CATALOG_CACHE_TTL_SECONDS (24h default). The asyncio.wait_for below is the whole-fetch deadline from the design review: one stalled page can no longer stack the 8 per-page 30s timeouts; it converts to the domain AdapterError, same as any provider failure on the old path. The shared cache helper's built-in SET-NX lock single-flights concurrent cold callers for free.
| # "slack message" would be a spurious adjacency. Score-only; matched_terms and | ||
| # exact_phrase stay untouched. | ||
| integration_name = str(integration.name or "").lower() | ||
| content_terms = [ |
There was a problem hiding this comment.
Review guide 2/5 — the only scoring change. Base term weights are untouched; this adds +4 per consecutive pair of use-case terms found adjacent in the normalized event key/name. It replaces the relevance ordering the paged provider search used to give us on ties. Two subtleties both came from live verification: (1) without it, "when a new GitHub issue is created" surfaced ISSUE_COMMENT_CREATED over ISSUE_CREATED (both base-score 26, catalog order broke the tie); (2) integration-name terms are excluded from pairs because they prefix every event key of that integration — a first version rewarded SLACK_MESSAGE_REACTION_ADDED for "slack message". Both cases are pinned in tests with the wrong event listed first.
| os.getenv("COMPOSIO_WEBHOOK_REPLAY_WINDOW_SECONDS") or 300 | ||
| ) | ||
| # Full trigger-types catalog: project-agnostic cache TTL + whole-fetch deadline. | ||
| catalog_cache_ttl_seconds: int = int( |
There was a problem hiding this comment.
Review guide 4/5 — the two knobs. TTL bounds staleness (a trigger removed by Composio inside the window fails loudly at create_subscription, which validates against the live provider), and the deadline bounds a cold call when Composio is degraded. Both follow the existing ComposioConfig field style; nothing reads os.getenv outside env.py.
| adapter.list_all_events.assert_awaited_once() | ||
|
|
||
|
|
||
| async def test_discover_triggers_adjacency_bonus_breaks_catalog_order_tie(): |
There was a problem hiding this comment.
Review guide 5/5 — what the new tests pin. These two ranking tests place the WRONG event first in the snapshot, so they fail if the adjacency bonus (or its integration-term exclusion) regresses — stable sort alone cannot make them pass. Below them: cache-hit skips the adapter, cache-miss writes with the configured TTL, deadline converts to AdapterError, and one-fetch-per-request memoization (assert_awaited_once). Pagination + cursor-guard live in the new test_triggers_catalog_client.py.
|
@coderabbitai review |
✅ Action performedReview finished.
|
|
CI note: the unit-test jobs skipped on the PR because the code push landed while this was still a draft (the workflow gates on draft, and ready_for_review does not retrigger). I dispatched the same workflow on the branch: run 29059521501. Result: 1257 passed, and the only failures are pre-existing on big-agents, byte-identical files this PR does not touch (6 secrets SSRF-rejection tests in oss/tests/pytest/unit/secrets/test_dtos.py and 2 SDK agents tests: test_unknown_harness_is_permissive, test_run_selection_unknown_permission_falls_back_to_allow_reads). The same 8 failures appear on feat/pi-approval-parking's run 29058074329. All triggers and tools tests, including the 194 touched by this PR, pass. The PR checks will run normally on the next push or can be judged by the dispatched run. |
There was a problem hiding this comment.
Actionable comments posted: 3
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 1b10855c-a000-4187-9b92-cfc630381a7c
📒 Files selected for processing (11)
api/oss/src/core/triggers/dtos.pyapi/oss/src/core/triggers/providers/composio/catalog.pyapi/oss/src/core/triggers/service.pyapi/oss/src/utils/env.pyapi/oss/tests/pytest/unit/triggers/test_triggers_catalog_client.pyapi/oss/tests/pytest/unit/triggers/test_triggers_discovery.pydocs/design/agent-workflows/projects/trigger-discovery-catalog/README.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/context.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/plan.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/research.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/status.md
✅ Files skipped from review due to trivial changes (4)
- docs/design/agent-workflows/projects/trigger-discovery-catalog/context.md
- docs/design/agent-workflows/projects/trigger-discovery-catalog/research.md
- docs/design/agent-workflows/projects/trigger-discovery-catalog/README.md
- docs/design/agent-workflows/projects/trigger-discovery-catalog/plan.md
|
|
||
| # Adjacency bonus: consecutive CONTENT terms appearing side by side in the event | ||
| # key/name break score ties the paged provider search used to resolve via its own | ||
| # relevance ordering (e.g. "issue created" prefers ISSUE_CREATED over | ||
| # ISSUE_COMMENT_CREATED). Integration-name tokens are excluded before pairing: | ||
| # they prefix every event key of that integration, so a bigram like | ||
| # "slack message" would be a spurious adjacency. Score-only; matched_terms and | ||
| # exact_phrase stay untouched. | ||
| integration_name = str(integration.name or "").lower() | ||
| content_terms = [ | ||
| term | ||
| for term in terms | ||
| if term not in integration_key and term not in integration_name | ||
| ] | ||
| normalized_key = _normalize_words(event.key) | ||
| normalized_name = _normalize_words(event.name) | ||
| for first, second in zip(content_terms, content_terms[1:]): | ||
| bigram = f"{first} {second}" | ||
| if bigram in normalized_key or bigram in normalized_name: | ||
| score += 4 | ||
|
|
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
Adjacency bonus can pair terms that were never adjacent in the use case.
content_terms removes integration key/name tokens from terms before pairing. That shift means two terms separated only by the integration name become "adjacent" after filtering — e.g. use case "new GitHub issue created" → terms ["new","github","issue","created"] → after removing "github", content_terms = ["new","issue","created"], so the bigram "new issue" gets checked even though "new" and "issue" were never adjacent in the original text (both pinned regression tests below happen to pass only because the spurious bigram doesn't coincidentally match any tested event name).
Also, term not in integration_key/integration_name uses substring containment rather than exact match, so a legitimate content term that happens to be a substring of the integration name/key is wrongly excluded from pairing too.
Compute bigrams from the original ordered terms, and only skip a pair when one endpoint IS (not "contains") the integration key/name:
🐛 Proposed fix
integration_name = str(integration.name or "").lower()
- content_terms = [
- term
- for term in terms
- if term not in integration_key and term not in integration_name
- ]
normalized_key = _normalize_words(event.key)
normalized_name = _normalize_words(event.name)
- for first, second in zip(content_terms, content_terms[1:]):
+
+ def _is_integration_term(term: str) -> bool:
+ return term == integration_key or term == integration_name
+
+ for first, second in zip(terms, terms[1:]):
+ if _is_integration_term(first) or _is_integration_term(second):
+ continue
bigram = f"{first} {second}"
if bigram in normalized_key or bigram in normalized_name:
score += 4📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| # Adjacency bonus: consecutive CONTENT terms appearing side by side in the event | |
| # key/name break score ties the paged provider search used to resolve via its own | |
| # relevance ordering (e.g. "issue created" prefers ISSUE_CREATED over | |
| # ISSUE_COMMENT_CREATED). Integration-name tokens are excluded before pairing: | |
| # they prefix every event key of that integration, so a bigram like | |
| # "slack message" would be a spurious adjacency. Score-only; matched_terms and | |
| # exact_phrase stay untouched. | |
| integration_name = str(integration.name or "").lower() | |
| content_terms = [ | |
| term | |
| for term in terms | |
| if term not in integration_key and term not in integration_name | |
| ] | |
| normalized_key = _normalize_words(event.key) | |
| normalized_name = _normalize_words(event.name) | |
| for first, second in zip(content_terms, content_terms[1:]): | |
| bigram = f"{first} {second}" | |
| if bigram in normalized_key or bigram in normalized_name: | |
| score += 4 | |
| # Adjacency bonus: consecutive CONTENT terms appearing side by side in the event | |
| # key/name break score ties the paged provider search used to resolve via its own | |
| # relevance ordering (e.g. "issue created" prefers ISSUE_CREATED over | |
| # ISSUE_COMMENT_CREATED). Integration-name tokens are excluded before pairing: | |
| # they prefix every event key of that integration, so a bigram like | |
| # "slack message" would be a spurious adjacency. Score-only; matched_terms and | |
| # exact_phrase stay untouched. | |
| integration_name = str(integration.name or "").lower() | |
| normalized_key = _normalize_words(event.key) | |
| normalized_name = _normalize_words(event.name) | |
| def _is_integration_term(term: str) -> bool: | |
| return term == integration_key or term == integration_name | |
| for first, second in zip(terms, terms[1:]): | |
| if _is_integration_term(first) or _is_integration_term(second): | |
| continue | |
| bigram = f"{first} {second}" | |
| if bigram in normalized_key or bigram in normalized_name: | |
| score += 4 |
…paging Composio per use case discover_triggers emulated search with up to 60 sequential uncached Composio calls per use case (30-60s wall time). The full trigger catalog is only 351 events across 41 toolkits, so the service now fetches it whole through the adapter (list_all_events, 8 pages, page-cap + repeated-cursor guards), caches it in Redis for 24h under a project-agnostic key (triggers:catalog), and scores every use case in memory. One asyncio.wait_for deadline bounds the whole fetch. Cold call measured 5.06s, warm 0.7-1.1s, wire contract unchanged. Scoring keeps the existing term weights and adds a content-term adjacency bonus to replace the provider's relevance ordering on ties; terms matching the integration's own key/name are excluded from pairs (else 'slack message' rewards the reaction event). Both live-found ranking cases are pinned by unit tests. New env knobs: COMPOSIO_CATALOG_CACHE_TTL_SECONDS (86400), COMPOSIO_CATALOG_FETCH_DEADLINE_SECONDS (30). Claude-Session: https://claude.ai/code/session_01LJaJ4hpJfTPULrnqKqtTHg
…form restriction, deprecated demotion Validated by an 891-query offline eval (committed as scripts/discovery_eval.py in the design workspace): browse-shaped queries went from 25% empty / 14% wrong toolkit to 100% right toolkit; templated ground truth holds at 98% top-1 / 100% top-4; deprecated events no longer surface. Meta/question words become stopwords; a use case naming a platform is restricted to it; DEPRECATED descriptions get -25; the no-match path returns the closest events with an honest capped-list note. Tool description and build-an-agent skill text updated to match. Claude-Session: https://claude.ai/code/session_01LJaJ4hpJfTPULrnqKqtTHg
7773fc2 to
0b79a52
Compare
dosubot
Bot
added
size:XXL
Follow-up commit 0b79a52: discovery quality pass, validated by an 891-query evalA live check after the cache landed showed a second problem, separate from latency: what agents actually get back. "slack triggers" returned nothing at all. "what events can gmail trigger on" returned a WhatsApp event as the best match, because "what" is a substring of WHATSAPP. "discord message received" returned Slack events even though a Discord trigger exists. Instead of patching cases one by one, this commit adds an offline eval (
Measured on the eval corpus: browse-shaped queries went from 25% fully empty and 14% wrong toolkit to 100% right toolkit and 0% empty. The 686 templated ground-truth queries hold at 98% top-1 and 100% top-4, so nothing regressed. Both previously pinned ranking cases still pass. Variants that measured worse (plural matching, treating "new" as content, down-weighting descriptions) are documented as rejected in research.md §7 with the case each one broke. The same pass fixes the other layers an agent touches: the Verified: 106 triggers unit tests green (4 new pins), the committed eval reproduces the numbers against the final code, and all the cases above were re-run live on the dev stack (warm 0.7s). Reviewed by a staff-level pass; its one required fix (the browse wording overclaimed "full list" while results are capped at |
|
@coderabbitai review |
✅ Action performedReview finished.
|
|
CI note for commit 0b79a52: the three red unit jobs are identical to the base big-agents run at the merge-base SHA 41eacd0 (run 29080056606 fails the same api/sdk/web jobs). api: 6 pre-existing secrets SSRF failures, 1264 passed including all triggers tests. sdk: 2 pre-existing permission-default failures, 1758 passed. web: 22 pre-existing agenta-entities failures where unit tests receive a live 404 HTML page (no web files in this PR). Python/TS lint and format, services unit tests, and runner suites are green. |
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
docs/design/agent-workflows/projects/trigger-discovery-catalog/scripts/discovery_eval.py (1)
178-183: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winUse the extracted
_has_primary_evidenceinstead of inlining its logic.
_has_primary_evidenceis inWANTEDand extracted tons, yetdiscovermanually replicates its gate using_match_signaland_DISCOVERY_MIN_PRIMARY_TERMS. If_has_primary_evidencegains additional checks, the eval script silently drifts and reports inaccurate metrics — exactly the copy-drift the AST extraction approach was designed to prevent.♻️ Proposed fix
surfaced = matches[: 1 + max(limit_alternatives, 0)] - if surfaced: - _score, matched_terms, exact = ns["_match_signal"]( - use_case=use_case, event=surfaced[0][1], integration=surfaced[0][2] - ) - if not (exact or matched_terms >= ns["_DISCOVERY_MIN_PRIMARY_TERMS"]): - surfaced = [] + if surfaced and not ns["_has_primary_evidence"]( + use_case=use_case, event=surfaced[0][1], integration=surfaced[0][2] + ): + surfaced = []
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 169d1362-6545-450a-9af5-1377418beb57
📒 Files selected for processing (14)
api/oss/src/core/triggers/dtos.pyapi/oss/src/core/triggers/providers/composio/catalog.pyapi/oss/src/core/triggers/service.pyapi/oss/src/utils/env.pyapi/oss/tests/pytest/unit/triggers/test_triggers_catalog_client.pyapi/oss/tests/pytest/unit/triggers/test_triggers_discovery.pydocs/design/agent-workflows/projects/trigger-discovery-catalog/README.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/context.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/plan.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/research.mddocs/design/agent-workflows/projects/trigger-discovery-catalog/scripts/discovery_eval.pydocs/design/agent-workflows/projects/trigger-discovery-catalog/status.mdsdks/python/agenta/sdk/agents/adapters/agenta_builtins.pysdks/python/agenta/sdk/agents/platform/op_catalog.py
✅ Files skipped from review due to trivial changes (3)
- docs/design/agent-workflows/projects/trigger-discovery-catalog/README.md
- docs/design/agent-workflows/projects/trigger-discovery-catalog/context.md
- docs/design/agent-workflows/projects/trigger-discovery-catalog/plan.md
🚧 Files skipped from review as they are similar to previous changes (5)
- api/oss/tests/pytest/unit/triggers/test_triggers_catalog_client.py
- api/oss/src/utils/env.py
- api/oss/src/core/triggers/dtos.py
- api/oss/src/core/triggers/providers/composio/catalog.py
- api/oss/src/core/triggers/service.py
| async def test_cached_event_catalog_deadline_raises_adapter_error(monkeypatch): | ||
| service = _service() | ||
|
|
||
| async def _hang(): | ||
| await asyncio.sleep(60) | ||
|
|
||
| service.adapter_registry.get = MagicMock( | ||
| return_value=SimpleNamespace(list_all_events=_hang) | ||
| ) | ||
| monkeypatch.setattr(env.composio, "catalog_fetch_deadline_seconds", 0.01) | ||
|
|
||
| with pytest.raises(AdapterError): | ||
| await service._cached_event_catalog(provider_key="composio") |
There was a problem hiding this comment.
🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win
Add get_cache monkeypatch to the deadline test for isolation.
Unlike test_cached_event_catalog_hit_skips_adapter and test_cached_event_catalog_miss_fetches_once_and_caches, this test does not monkeypatch get_cache. If the real get_cache returns a cached value (e.g., from a prior test), the adapter is never called, AdapterError is never raised, and the test fails for the wrong reason. If get_cache raises or hangs, same outcome. Add a _miss returning None to guarantee the adapter path is reached.
🛡️ Proposed fix
async def test_cached_event_catalog_deadline_raises_adapter_error(monkeypatch):
service = _service()
+
+ async def _miss(**_kwargs):
+ return None
+
+ monkeypatch.setattr("oss.src.core.triggers.service.get_cache", _miss)
+
async def _hang():
await asyncio.sleep(60)📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| async def test_cached_event_catalog_deadline_raises_adapter_error(monkeypatch): | |
| service = _service() | |
| async def _hang(): | |
| await asyncio.sleep(60) | |
| service.adapter_registry.get = MagicMock( | |
| return_value=SimpleNamespace(list_all_events=_hang) | |
| ) | |
| monkeypatch.setattr(env.composio, "catalog_fetch_deadline_seconds", 0.01) | |
| with pytest.raises(AdapterError): | |
| await service._cached_event_catalog(provider_key="composio") | |
| async def test_cached_event_catalog_deadline_raises_adapter_error(monkeypatch): | |
| service = _service() | |
| async def _miss(**_kwargs): | |
| return None | |
| monkeypatch.setattr("oss.src.core.triggers.service.get_cache", _miss) | |
| async def _hang(): | |
| await asyncio.sleep(60) | |
| service.adapter_registry.get = MagicMock( | |
| return_value=SimpleNamespace(list_all_events=_hang) | |
| ) | |
| monkeypatch.setattr(env.composio, "catalog_fetch_deadline_seconds", 0.01) | |
| with pytest.raises(AdapterError): | |
| await service._cached_event_catalog(provider_key="composio") |
1 participant
Context
discover_triggerstook 30 to 60 seconds whilediscover_toolsanswered in under a second. The trigger side had no semantic search to lean on, so it emulated one by paging the Composio catalog with keywords: up to 60 sequential, uncached HTTP calls per use case (integration searches, per-integration event searches, a detail fetch). Three use cases meant up to 180 round trips at ~0.55 s each.The whole Composio trigger catalog is only 351 events across 41 toolkits, and every list item already carries its config schema and sample payload. So this PR fetches the catalog once, caches it, and scores in memory. The design was reviewed and approved on this PR first (the docs under
docs/design/agent-workflows/projects/trigger-discovery-catalog/); the implementation now sits on top, with the five CodeRabbit review comments folded in.Measured on the live EE dev stack: cold call 5.06 s (one 8-page catalog fetch), warm calls 0.66 to 1.11 s with zero Composio catalog calls. Before: 30 to 60 s on every call.
Changes
Catalog fetch (
providers/composio/catalog.py): newlist_all_events()on the catalog client pagesGET /triggers_types(50 per page, Composio's cap) to exhaustion into aTriggerCatalogEventsSnapshot(all events plus a toolkit slug-to-display-name map). Guards: 50-page cap with a warning log when it truncates, and a repeated-cursor break.Cache (
service.py):_cached_event_catalog()mirrors the tools-side_cached_search: Redis namespacetriggers:catalog, provider-only key (project-agnostic on purpose; connection state stays per-project and fresh), TTLCOMPOSIO_CATALOG_CACHE_TTL_SECONDS(default 24 h). The whole fetch runs under oneasyncio.wait_fordeadline,COMPOSIO_CATALOG_FETCH_DEADLINE_SECONDS(default 30 s), converted to the domainAdapterErrorso one stalled page cannot stack per-page timeouts.Discovery (
service.py):discover_triggersloads the snapshot once per request;_discover_events_for_use_casebecame a pure static function scoring all 351 events in memory with the existing scorer. The keyword paginators (_candidate_integrations,_candidate_events) and the per-primaryget_eventcall are deleted;trigger_configandpayloadcome straight from the snapshot.Per-request Composio catalog calls, before and after:
Scorer tie-break (
service.py, found in live verification): with the provider's relevance ordering gone, equal-score ties broke by catalog order, and "when a new GitHub issue is created" surfaced the ISSUE_COMMENT event over ISSUE_CREATED (both scored 26)._match_signalnow adds +4 per consecutive pair of use-case terms found adjacent in the normalized event key or name ("issue created" hitsISSUE_CREATED, notISSUE_COMMENT_CREATED). Terms matching the integration's own key or name are excluded from pairs; a first version without that exclusion made "slack message" rewardSLACK_MESSAGE_REACTION_ADDED, since the integration name prefixes every key. Both cases are pinned by unit tests with the wrong event listed first.The wire contract is unchanged: same request, same
TriggerCapabilitiesResult, router untouched.Scope / risk
list_events,get_event) stay on the live adapter, verified regression-free. Serving them from the snapshot is a follow-up.get_integrationper surfaced, not-ready integration; bounded and needed for freshness).AGENTA_API_CACHING_ENABLED=falseevery request pays one full crawl (~4-5 s) instead of a cache read. Still bounded by the deadline; no crash.create_subscriptiontime, which validates against the live provider.research.md§6). Runtime fetch with our own key is ordinary API usage.Tests
oss/tests/pytest/unit/triggers/+unit/tools/, including new ones: pagination and cursor-guard forlist_all_events, cache hit skips the adapter, cache miss writes with the configured TTL, deadline converts toAdapterError, one-fetch-per-request memoization, and the two ranking pins (GitHub, Slack).secrets/test_dtos.py(SSRF URL rejection), unrelated to this change.triggers:catalog:provider:composioobserved at TTL ~86400, byte-identical responses across warm calls, and the browse route returning 200 with 20 events.How to QA
Prerequisites: the EE dev stack (
load-env hosting/docker-compose/ee/.env.ee.dev.local+run.sh --ee --dev) withCOMPOSIO_API_KEYset. Mint credentials viaPOST /api/admin/simple/accounts/(theaccounts.pyfixture pattern).Steps:
redis-cli --scan --pattern '*triggers:catalog*'andDELthe service key.POST /api/triggers/discoverwith{"use_cases": ["when a new GitHub issue is created", "new Slack message in a channel"]}and yourApiKeyheader. Time it.Expected result: call 1 completes in single-digit seconds and logs one
[composio] list_all_events items=351 toolkits=41; calls 2-3 return in under ~1 s with no Composio fetch lines. The GitHub capability's primary isGITHUB_ISSUE_CREATED_TRIGGER, the Slack oneSLACKBOT_CHANNEL_MESSAGE_RECEIVED, both with non-nulltrigger_config.Automated tests:
Edge cases: the discover response for a nonsense use case must still return the no-match note rather than an arbitrary event (covered by tests). One environment gotcha found during QA: uvicorn watchfiles on the dev stack does not react to host-side writes under
api/oss/src/core/triggers/, so live-testing an edit there needs a forced reload (a byte-identical rewrite ofoss/src/utils/env.pyworks); details instatus.md.https://claude.ai/code/session_01LJaJ4hpJfTPULrnqKqtTHg