fix(knowledge): give the reach count the probe's share of the deadline, not the leg's

[waleedlatif1]

Summary

• Counting a caller's reach (how much of the bases they can read, which picks between one graph walk and a per-source search) reads as many index entries as they reach. On a large index a broad member's count consumed the whole vector budget, the leg had no time left to walk, and the search reported the vector leg timed out; the count's stage was also measured twice (an outer wrapper around an inner measured statement), so diagnostics showed it at double its real cost
• The count now runs under the probe's share of the deadline (the same cap the permitted-documents probe already uses), never the whole leg's. A count that runs out of that share answers nothing for this search: the leg keeps its remaining time to walk, is not marked timed out, and nothing is remembered, so the next search counts again
• The reach memo now holds for an hour instead of five minutes. Reach moves slowly and only chooses a ranking strategy; a stale answer costs speed, never access. On a multi-instance deployment this turns a per-instance, per-five-minutes cold count into a rare one
• The duplicated stage wrapper is removed, so permitted_documents reports the statements' real time

Type of Change

• Bug fix

Testing

• The existing timed-out-count test now also asserts the leg's own deadline is not marked spent; it fails when the count's timeout is judged by the leg's budget instead of its own share
• The small-index count test asserts the count's statement timeout is at most the probe budget; it fails with the cap removed
• vitest over lib/knowledge and app/api/knowledge (191 files, 3304 tests), bun run lint, bun run check:audits and native tsc --noEmit pass

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
docs	Skipped		Sep 21, 2026 6:27pm UTC

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; the previous stale-strategy concern was resolved by restoring the five-minute cache TTL, and no new actionable defects remain.

Summary

This PR bounds reach counting to the probe’s share of the search deadline and preserves the remaining vector-leg budget when only that count times out. Follow-up changes restore the reach-cache TTL to five minutes, distinguish an expired leg from an expired count share, and add deadline-accounting tests.

• Caps reach-count database work using a child search budget.
• Avoids memoizing an inconclusive timed-out count.
• Preserves partial-retrieval reporting when the parent vector deadline expires.
• Removes duplicate permitted_documents stage measurement.
• Keeps the reach-strategy cache at five minutes to limit stale planning decisions.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Resolve caller reach] --> B[Create probe-capped count budget]
  B --> C{Reach count finishes?}
  C -->|Yes| D[Choose broad or narrow strategy]
  D --> E[Memoize strategy for five minutes]
  C -->|Count share expires| F[Use broad fallback for this search]
  F --> G[Do not memoize]
  C -->|Parent leg expires| H[Mark vector leg timed out]
  G --> I[Continue retrieval with remaining leg budget]
  H --> J[Return partial retrieval status]

Loading

_{Reviews (2) · Last reviewed commit: "fix(knowledge): keep the reach memo at f..."}

[cubic-dev-ai]

All reported issues were addressed across 2 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[waleedlatif1]

@greptile

[waleedlatif1]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Re-trigger cubic}