When I used sub-agents to write documentation, they didn’t include the verification tokens at all. As a result, those sections effectively bypassed verification and slipped through review without any citation checks.
The root cause was subtle: sub-agents couldn’t trigger skills. VeriContext’s citation enforcement lives as a skill in the parent agent — but when work was delegated, the sub-agent didn’t inherit or execute that skill. So it wrote “clean-looking” docs that were never actually instrumented with VCC tokens.
The documents looked fine on the surface — but parts of them weren’t covered by VeriContext’s guarantees. Delegation had quietly created a verification blind spot.
The fix was adding a PreToolUse hook on the Task tool (the tool agents use to spawn sub-agents). When the master agent creates a doc-related task, the hook detects it and outputs the citation rules. The master agent then injects those rules into the sub-agent’s prompt. No changes to the sub-agent itself — the parent simply propagates the verification contract downward.
Hook fires → detects doc-related task
→ outputs citation rules
→ parent injects them into sub-agent prompt
→ sub-agent writes docs with proper [[vctx:…]] tokens
→ verification actually covers everything
This feels like a broader systems issue: any invariant enforced via parent-level skills disappears unless explicitly propagated. Intercepting the delegation boundary (the tool call itself) turned out to be a clean enforcement point.
Curious how others handle constraint propagation in multi-agent systems — especially when invariants are implemented as skills rather than core behavior.
When I used sub-agents to write documentation, they didn’t include the verification tokens at all. As a result, those sections effectively bypassed verification and slipped through review without any citation checks.
The root cause was subtle: sub-agents couldn’t trigger skills. VeriContext’s citation enforcement lives as a skill in the parent agent — but when work was delegated, the sub-agent didn’t inherit or execute that skill. So it wrote “clean-looking” docs that were never actually instrumented with VCC tokens.
The documents looked fine on the surface — but parts of them weren’t covered by VeriContext’s guarantees. Delegation had quietly created a verification blind spot.
The fix was adding a PreToolUse hook on the Task tool (the tool agents use to spawn sub-agents). When the master agent creates a doc-related task, the hook detects it and outputs the citation rules. The master agent then injects those rules into the sub-agent’s prompt. No changes to the sub-agent itself — the parent simply propagates the verification contract downward.
Hook fires → detects doc-related task → outputs citation rules → parent injects them into sub-agent prompt → sub-agent writes docs with proper [[vctx:…]] tokens → verification actually covers everything
This feels like a broader systems issue: any invariant enforced via parent-level skills disappears unless explicitly propagated. Intercepting the delegation boundary (the tool call itself) turned out to be a clean enforcement point.
Curious how others handle constraint propagation in multi-agent systems — especially when invariants are implemented as skills rather than core behavior.