chore(http): bump default CLIENT_TIMEOUT from 120s to 300s

@zen-agentchecks n/achecks…zen/bump-client-timeout-300s-dk2b3q → master2 files · +2 −2updated 1mo ago

▸Description· 2 comments

Description

Raises Mercury's default per-TSP-HTTP-call timeout from 120s → 300s in mercury/utils/http.py.

Why

The 120s default was a direct carryover from the original Mercury port (commit cf8f743c3b, 2025-04-08) with no comment, no ADR, no PR rationale, and no upstream constraint forcing it. Since then multiple tools have had to work around it with explicit per-call timeout= overrides:

Override	Tool
180s	`composio_search/actions/vercel_ai_chat.py`
240s	`bubble/actions/data_bulk_create_things.py`
300s	`browserless/actions/scrape_content.py`
300s	`composio/actions/download_s3_file.py`
300s	`googledrive/actions/upload_from_url.py`
600s	`googledrive/actions/download_file.py`
120-600s (user-controlled)	`gemini/actions/generate_image.py`

300s aligns with the median override that tool authors are already choosing, and matches the "5 minutes" intuition several of them wrote in inline comments.

Headroom check

The chain is still constrained by the same upstream limits:

Thermos → Mercury Lambda invoke timeout: 840s (LAMBDA_INVOKE_TIMEOUT_SECONDS, apps/thermos/lib/awsutils/lambda.go)
AWS Lambda hard execution limit: 900s

300s leaves ~540s of headroom under Thermos's ceiling. Tools that legitimately need >300s can still pass an explicit timeout= (the override mechanism is unchanged).

Backwards compatibility

Every call site that already passes timeout= is unaffected.
Tools that previously succeeded under 120s continue to succeed.
The only behavior change is for tools whose TSP call took 120-300s: those now succeed instead of hanging at the previous default.

How did I test this PR

pytest tests/test_utils/test_http.py → 130 passed
ruff check and ruff format --check on changed files → clean
Updated test_client_timeout_value assertion (was the only test pinning the literal 120.0)
Did NOT change any of the tools that already override the timeout — they keep their existing values.

@zen-agent1mo ago

Closing — switching to per-tool timeout override instead of a global default bump (per maintainer direction). Will follow the existing in-repo pattern: pass an explicit timeout= to self.http_request(...) in the action that needs it, or expose a Pydantic timeout field on the request schema with a clamped range (Gemini's generate_image pattern).

@zen-agent1mo ago

Reopening per maintainer direction — proceeding with the global default bump 120s → 300s. 300s matches what other tools already use as their explicit override (browserless, googledrive uploads, composio download_s3_file all hardcode 300s).

loading diff…