feat: make LLM call timeout configurable via env var#58
Conversation
…nv var Add AbortSignal.timeout to both OpenAI-compatible and Anthropic fetch calls in provider.ts, controlled by the MEMORIX_LLM_TIMEOUT_MS environment variable (default: 30000ms). This fixes timeouts when using LLM providers behind proxies (e.g. CLIProxyAPI routing to Chutes/Fireworks) where routing + model cold start regularly exceeds the previous hardcoded 12s timeout. Closes AVIDS2#57
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e2d3fb8ecf
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
src/llm/provider.ts
Outdated
| * Configurable via MEMORIX_LLM_TIMEOUT_MS environment variable. | ||
| * Default: 30000ms (30s) — allows for proxy routing and cold starts. | ||
| */ | ||
| const LLM_CALL_TIMEOUT_MS = parseInt(process.env.MEMORIX_LLM_TIMEOUT_MS || '30000', 10); |
There was a problem hiding this comment.
LLM_CALL_TIMEOUT_MS is parsed once with parseInt(...) and used directly in AbortSignal.timeout(...) without validation. If MEMORIX_LLM_TIMEOUT_MS is set to a non-integer or out-of-range value (for example 60s, empty string, -1, or a very large number), Node throws a RangeError, causing every LLM request to fail before the HTTP call is made. This turns a configuration typo into a full LLM-path outage instead of safely falling back to the default timeout.
Useful? React with 👍 / 👎.
AVIDS2
left a comment
AVIDS2
left a comment
There was a problem hiding this comment.
Thanks for working on this. The direction is useful, but this PR still needs one important fix before we can take it: MEMORIX_LLM_TIMEOUT_MS is parsed and passed directly into AbortSignal.timeout(...) without validation, so values like 60s, -1, or other invalid input can turn a config typo into a hard LLM-path failure. Please validate and clamp/fallback before use. Also, this PR overlaps heavily with #59, so once the validation issue is addressed we’ll likely consolidate toward a single timeout PR instead of merging both.
RaviTharuma
left a comment
RaviTharuma
left a comment
There was a problem hiding this comment.
Thanks for the review! Added input validation for MEMORIX_LLM_TIMEOUT_MS:
- Extracted
parseLLMTimeoutMs()helper that validates the env var is a valid integer - Returns default (30s) for undefined, empty, non-numeric, float, or NaN values
- Clamps valid integers to [1000, 300000] ms range (1s–5min)
- Added 9 unit tests covering all edge cases
Regarding overlap with #59 — happy to consolidate. Our implementation now has the validation that was requested, so whichever PR you prefer to merge is fine.
MEMORIX_LLM_TIMEOUT_MS is now validated: must be a plain positive integer in the range 1000–300000ms. Values like "60s", "-1", or NaN fall back to the default 30000ms with a console.warn explaining the issue.
|
Fixed — added validation: timeout must be a positive integer between 1000-300000ms, falls back to default on invalid input. |
TYPE_WEIGHTS map requires all ObservationType keys — add 'reasoning' entry. archiveExpired document builder requires 'source' field — default to 'agent'. Fixes remaining TS2741 errors surfaced by tsc --noEmit.
- Add 'reasoning' to TYPE_WEIGHTS in evaluate.ts (ObservationType exhaustiveness) - Add source field to MemorixDocument object literal in retention.ts - Change parseLLMTimeoutMs to clamp out-of-range values to min/max instead of falling back to default, matching test expectations
|
@AVIDS2 please check now :) |
AVIDS2
left a comment
AVIDS2
left a comment
There was a problem hiding this comment.
This looks good now. The timeout env handling no longer turns invalid input into an AbortSignal.timeout(...) failure path, the validation/clamping behavior is explicit, and the added tests make the edge cases clear. Between the two overlapping timeout PRs, this is the stronger one to carry forward.
2 participants
Summary
MEMORIX_LLM_TIMEOUT_MSenvironment variable to control LLM call timeouts (default: 30000ms / 30s)AbortSignal.timeout()to both OpenAI-compatible and Anthropicfetchcalls insrc/llm/provider.tsProblem
When using LLM providers behind a proxy (e.g., CLIProxyAPI/Quotio routing to Chutes/Fireworks), the routing + model cold start regularly exceeds 12 seconds. This causes:
LLM compact failed, falling back to heuristiclog spamSee #57 for full details.
Usage
Closes #57