You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I understand that PRs require status:approved before implementation.
📝 Bug Description
On Gentle AI 2.2.4 with OpenCode 1.18.13, both managed Judgment Day judges fail before inspecting an Engram-backed SDD design target. The judge runtimes expose no Engram MCP resources or resource templates, and reads of the exact injected skill and repository source roots are denied by the external-directory policy.
This report records observed evidence only; it does not assert an unconfirmed root cause. The workflow fails closed: no findings or verdict are produced, and dependent SDD task planning remains blocked.
🔄 Steps to Reproduce
Use an Engram-backed SDD change whose design artifact has a stable opaque revision.
Launch the managed blind jd-judge-a and jd-judge-b agents through the supported subagent mechanism.
Ask each judge to read the Engram artifact, verify its revision, load exact injected skills, and inspect bounded repository sources.
Observe both judges return incomplete before inspection because no Engram MCP resources/templates are available and external-directory reads are denied.
Retry once with explicit user authorization and the same immutable target.
Observe the same two failures again.
✅ Expected Behavior
Managed Judgment Day judges should receive the immutable Engram-backed artifact and exact comparison context through a supported, narrowly scoped integration. They should be able to load the required managed skills and permitted repository sources, complete both blind inspections, and return a verdict while unrelated external paths remain denied.
If Engram-backed targets are unsupported, the integration should return a typed preflight result before launching judges, with an actionable supported route that preserves the target and review state.
❌ Actual Behavior
Two bounded attempts were made, each with both blind judges, for four judge runs total. Every run failed before inspection. The repeated envelopes reported:
INCOMPLETE: Engram observation revision could not be read or verified because no MCP resources or resource templates were available.
Required injected skill and repository source reads were denied by the external_directory permission policy.
No judge inspected candidate content. Findings remained empty because inspection never began, and no APPROVE or REJECT verdict was emitted.
Gentle AI Version
2.2.4
Operating System
macOS
AI Agent / Client
OpenCode 1.18.13
📋 Affected Area
Agent Detection
💡 Logs / Error Output
completed_inspection: false
attempts: 2 Judgment Day launches, 2 blind judges per launch
outcomes: 4 incomplete inspections
engram_resources: unavailable
external_directory: denied for injected skills and bounded repository sources
verdict: none
mutation_outcome: none by judge runs
Additional Context
Platform: Darwin arm64; shell: zsh.
Operation shape: read-only Judgment Day over an Engram-backed SDD design, using an opaque target revision and exact injected skill/source roots.
The target revision remained unchanged. Judge runs did not modify repository bytes, staging, lifecycle authority, or the canonical SDD artifact.
Expected dependent action was task planning; it remained blocked.
Diagnostics were privacy-scrubbed: no raw arguments, absolute paths, private repository names, usernames, hostnames, credentials, source contents, diffs, or environment values are included.
Pre-flight Checklist
status:approvedbefore implementation.📝 Bug Description
On Gentle AI 2.2.4 with OpenCode 1.18.13, both managed Judgment Day judges fail before inspecting an Engram-backed SDD design target. The judge runtimes expose no Engram MCP resources or resource templates, and reads of the exact injected skill and repository source roots are denied by the external-directory policy.
This report records observed evidence only; it does not assert an unconfirmed root cause. The workflow fails closed: no findings or verdict are produced, and dependent SDD task planning remains blocked.
🔄 Steps to Reproduce
jd-judge-aandjd-judge-bagents through the supported subagent mechanism.✅ Expected Behavior
Managed Judgment Day judges should receive the immutable Engram-backed artifact and exact comparison context through a supported, narrowly scoped integration. They should be able to load the required managed skills and permitted repository sources, complete both blind inspections, and return a verdict while unrelated external paths remain denied.
If Engram-backed targets are unsupported, the integration should return a typed preflight result before launching judges, with an actionable supported route that preserves the target and review state.
❌ Actual Behavior
Two bounded attempts were made, each with both blind judges, for four judge runs total. Every run failed before inspection. The repeated envelopes reported:
No judge inspected candidate content. Findings remained empty because inspection never began, and no APPROVE or REJECT verdict was emitted.
Gentle AI Version
2.2.4
Operating System
macOS
AI Agent / Client
OpenCode 1.18.13
📋 Affected Area
Agent Detection
💡 Logs / Error Output
Additional Context