Severity Classification

Triage is the second layer of kedge's pipeline. It uses a lightweight AI call to classify each drifted anchor into one of three severity levels, which then drive remediation decisions.

Severity levels

no_update

The code change is purely cosmetic and doesn't affect the documentation's accuracy.

Examples: - Whitespace or formatting changes that got past AST fingerprinting (e.g., in content-hash fallback files) - Internal variable renaming that doesn't affect the public API - Refactoring that preserves behavior

Remediation: No agent invocation. kedge advances the provenance to the current fingerprint.

minor

The code change is small and the documentation needs a minor update.

Examples: - A new optional parameter was added to a method - A variable was renamed in a way that's mentioned in the docs - A default value changed - A new enum variant was added

Remediation: The agent is invoked to update the docs. Whether auto-merge is enabled depends on the auto_merge_severities config.

major

The code change is significant and the documentation needs a substantial rewrite.

Examples: - A required parameter was added or removed - A method's return type changed - A function was removed or renamed - The semantics of an operation changed - A breaking API change

Remediation: The agent is invoked to update the docs. Requires human review before merging in most cases.

How classification works

For each drifted doc, kedge:

1. Builds a prompt containing the documentation content and the code diffs for each drifted anchor
2. Sends the prompt to the configured triage provider (Anthropic, OpenAI, or custom command)
3. Parses the response as a JSON array of { path, symbol, severity } classifications
4. Maps classifications back to anchors

The AI sees both the current documentation and the code changes, so it can judge whether the change affects what the docs describe.

Triage prompt structure

The prompt includes:

• The full documentation content (the markdown body of the steering file)
• For each drifted anchor: the file path, symbol, diff summary, and full diff
• Instructions to classify as no_update, minor, or major with clear criteria
• Expected JSON response format

Doc-level severity

Each doc gets an overall severity equal to the maximum of its anchor severities. If a doc has one minor and one major anchor, the doc severity is major.

Cost control

Triage only runs on drifted docs. Clean docs (where no anchors have drifted) incur zero AI cost. Detection is deterministic and free, and kedge invokes AI only when there is drift to classify.

How severity drives remediation

Severity	Agent invoked?	Provenance advanced?	Typical auto-merge?
no_update	No	Yes (automatic)	N/A
minor	Yes	After agent completes	Often yes
major	Yes	After agent completes	Usually no (human review)

Auto-merge configuration

The auto_merge_severities config controls which severity levels set auto_merge: true in the agent payload:

[remediation]
auto_merge_severities = ["no_update", "minor"]

The agent decides what to do with the auto_merge flag. kedge passes it through without acting on it. In batch mode, auto_merge is true only if every target qualifies individually.

Triage providers

kedge supports three triage backends. See Configuration for setup details.

• anthropic: direct Anthropic API (recommended: claude-haiku-4-5-20251001 for cost efficiency)
• openai: any OpenAI-compatible endpoint (Azure OpenAI, vLLM, local models)
• command: pipe the prompt to an external command via stdin