Skip to content

Refactor chaterror.Classify to normalize evidence gathering #1638

Description

@johnstcn

Note: this issue was filed by an over-eager agent with too many MCP tools.

Context

PR coder/coder#27913 fixed a bug where a Bedrock credential resolution failure was incorrectly retried as a generic 500. The fix widened several signal checks in chaterror.Classify to read combinedText (the merged transport wrapper + structured response body) instead of lower (just the wrapper).

This exposed a structural issue: the classifier has two parallel text sources (lower from err.Error() and structured.detail from ProviderError.ResponseBody) and signal checks inconsistently pick which to match against. PR coder/coder#27913 widened the body-only signals, but the underlying design is fragile.

Problems

  1. Two-source inconsistency: Some signals check lower, some check combinedText, and it is unclear which uses which without reading each line. New signal checks can silently regress to lower-only.

  2. Pattern sprawl: Too many string patterns across too many signal lists (overloadedPatterns, authStrongPatterns, configPatterns, timeoutPatterns, etc.) without documentation of which provider incident motivated each pattern.

  3. Signal/display coupling: structured.detail (used for signal matching) is the first line of the response body only (providerErrorResponseMessage truncates at the first newline). If the useful signal text appears on line 2+, it is missed. The signal source and the user-facing detail display are the same string; they should be decoupled.

Proposed approach

Data-flow normalization, not a rule engine:

  1. One gatherEvidence step at the top of Classify that produces a single struct: lowercased combined text (transport error + full structured body, not just first line), status code, and any structured provider fields.
  2. Every signal check reads only from the evidence struct. No signal touches raw inputs.
  3. Keep the prioritized rule ordering as plain Go code. Do not build a pattern-matching DSL or data-driven rule table.
  4. Decouple signal matching text from user-facing detail display text.

Non-goals

  • No new error taxonomy.
  • No change to Classify's public signature.
  • No pattern DSL or data-driven rule engine.

Test strategy

Convert signal tests to a matrix that programmatically asserts each source-agnostic signal classifies correctly when the pattern arrives via body-only and via transport-only. This makes "forgot to check the other source" impossible to reintroduce silently.

Related

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions