Kihagyás

Subagent collateral-bug discovery pattern#

When a subagent's task report includes run-time errors from a side-quest — code paths it tried but failed on while pursuing its primary objective — those errors are a high-value bug-finding signal, not noise to ignore.

The pattern#

  1. Main thread spawns subagent X for task A.
  2. Subagent X, while completing task A, exercises tangentially-related code path Y (e.g. read a dependent module, call a sibling CLI).
  3. Code path Y throws an error.
  4. Subagent X reports task A complete, mentioning the Y-error as an aside.
  5. Main thread treats Y-error as a separate, high-priority bug — extracts the stack trace and roots-causes.

Why this is signal, not noise#

  • Subagent X had no incentive to fabricate the Y-error. It happened to it organically.
  • Main thread might never exercise Y directly. Subagent X's tangential exposure caught a bug that direct unit-tests missed.
  • The Y-bug was likely silent — no user-visible symptom yet.

When it bit us (2026-05-19)#

Three P0/P1 bugs caught this way in a single PM session:

  1. SQLite executescript() implicit-COMMIT — Wave-1 subagent designed the hash-refactor migration; while writing the script, a smoke-test surfaced cannot commit - no transaction is active. Reported as "decision-point #1 for main-thread review". → sqlite-executescript-implicit-commit

  2. vault-gh-bridge idempotency-FAIL — Wave-4 subagent (#14 APPLY first-run) ran second-time-idempotency-test as part of the verify. Reported files_unchanged: 0 as anomaly. → idempotency-frontmatter-date-anchor

  3. vault-ko-ingest.upsert_fact post-#34 P0 schema-bug — Wave-4 Memgraph Phase-3 subagent tried to re-extract → OperationalError: no column named provenance. Reported as blocker. Every new ingest silently broken across the vault until caught. → schema-migration-downstream-grep-checklist

Without the subagent reports, all three bugs would have been latent for days-to-weeks. The Memgraph Phase-3 subagent specifically saved us from a vault-wide production-broken state.

Anti-pattern: dismissing tangential errors#

The temptation: "subagent says task A complete, that's the deliverable — Y-error is its problem, move on."

The cost: production-broken state that surfaces in 3 days when an unrelated workflow trips it. Now the debugging context (the subagent's intermediate state, the exact stack trace, the path that led there) is gone.

How to extract the signal#

When a subagent report includes: - "Run-time error during smoke-test" - "Decision-point N for main-thread review" - "Anomaly N: ..." - "Did not verify because threw"

Always: 1. Read the full error message in the report 2. Trace to the offending file:line 3. Diagnose root-cause yourself (don't ask a subagent — context-loss between agents wastes tokens) 4. Patch + smoke-test 5. Add to the propagation log as a real bug-fix, not as session-noise