Context
Filed by course lane 4 (#2902). The lane's audit verdict: planning:interview's lock and auto-detect synthesize-directly paths are a LICENSED exception to the asset-rush failure mode, resting on four structural defenses — lock-invocation-as-confirmation, the STOP-on-gap rule, the auto-guard (a genuinely-user decision is never folded into the Brief silently; unattended it becomes a blocked register row with arbiter: USER-RESERVED), and the mechanical register gate. The first two of those are behavioral rules with no regression coverage — this item turns the audit's "defenses hold" from prose into a gate.
Proposed work
- Eval case A (STOP-on-gap): a
lock invocation over a task with a planted, genuinely-open decision (real tradeoffs, no codebase answer) — assert the skill HALTS and surfaces the gap rather than synthesizing a Brief that fudges it
- Eval case B (auto-guard): an
auto invocation whose context resolves all but one interactive design choice — assert the residue is asked (or, unattended, recorded blocked with arbiter: USER-RESERVED) and never silently captured as an assumption
- Wire per the repo's eval conventions (
evals:design / skill-quality:check validate-evals); planning plugin version bump + CHANGELOG
Acceptance criteria
References
docs/upstream/aihero-core-lanes.md (lane 4 rows)plugins/planning/skills/interview/SKILL.md (Action Router lock, Step 1.5 auto-guard, unattended ladder)
Metadata
| Field |
Value |
| Category |
unspecified |
| Area |
planning, evals |
| Ecosystem |
unspecified |
Context
Filed by course lane 4 (#2902). The lane's audit verdict:
planning:interview'slockand auto-detect synthesize-directly paths are a LICENSED exception to the asset-rush failure mode, resting on four structural defenses — lock-invocation-as-confirmation, the STOP-on-gap rule, the auto-guard (a genuinely-user decision is never folded into the Brief silently; unattended it becomes ablockedregister row witharbiter: USER-RESERVED), and the mechanical register gate. The first two of those are behavioral rules with no regression coverage — this item turns the audit's "defenses hold" from prose into a gate.Proposed work
lockinvocation over a task with a planted, genuinely-open decision (real tradeoffs, no codebase answer) — assert the skill HALTS and surfaces the gap rather than synthesizing a Brief that fudges itautoinvocation whose context resolves all but one interactive design choice — assert the residue is asked (or, unattended, recordedblockedwitharbiter: USER-RESERVED) and never silently captured as an assumptionevals:design/skill-quality:check validate-evals); planning plugin version bump + CHANGELOGAcceptance criteria
References
docs/upstream/aihero-core-lanes.md(lane 4 rows)plugins/planning/skills/interview/SKILL.md(Action Routerlock, Step 1.5 auto-guard, unattended ladder)Metadata