
CRUMB
Generative Game Studio · Demo
From a one-line game pitch to a playable draft, with the agents’ work in view and open to intervention.
CRUMB is a multi-agent studio harness for casual game prototyping.
A silent exit made observable, a lost formatter caught at integration
Two cases from the public demo repository. Planner, Builder and Verifier are actors inside the demo, not the coding agents in this view.
May 2, 2026
A run that ended with exit 0 was made to show whether it actually did any work
Under developmentDemo actors
Observation · finding0 transcript events · exit 0
- handoff0 events, yet exit 0→ Tell “ended but did no work” apart from progress (Developer)
PersonDeveloper
ActionTell “ended but did no work” apart from progress
- handoffDiagnostics first→ Pre/post-spawn diagnostics: adapter · prompt present · exit · time · output size (Code changes)
Work sessionCode changes
ActionPre/post-spawn diagnostics: adapter · prompt present · exit · time · output size
- handoff→ commit 6341898 (Artifacts)
- handoffDiagnostic output→ The planner was waiting for input (Demo actors)
ArtifactsArtifacts
Artifactcommit 6341898
Under developmentDemo actors
Observation · findingThe planner was waiting for input
- sent backWaiting for input→ Pass the real goal and a start instruction · reducer check on prompt shape (Code changes)
Work sessionCode changes
ActionPass the real goal and a start instruction · reducer check on prompt shape
- handoffNext run→ Only JSON on stdout, event writer never called (Demo actors)
Under developmentDemo actors
Observation · findingOnly JSON on stdout, event writer never called
- sent backProtocol not injected→ Inject the event-protocol doc into event-emitting actors (Code changes)
Work sessionCode changes
ActionInject the event-protocol doc into event-emitting actors
- handoff→ commit 08ecb07 (Artifacts)
ArtifactsArtifacts
Artifactcommit 08ecb07
- Context
- A real subprocess smoke run ended with no transcript events. Exit 0 alone could not tell whether work was done, so diagnostic records around the run came first; after the first fix, a different failure appeared and was fixed once more.
- Outcome
- The diagnostic records separated “exited successfully but did no work” from progress, and the failure left after the first fix led, through a more specific observation, to the second fix.
- Limit
- The runs of the time were not re-run and their original run records were not retrieved. The raw diagnostic text (output previews and so on) is not included.
Sources and remaining limits
May 3, 2026
Restored the feed formatter lost while integrating two work streams
Work sessionFeed · layout stream
Action#140: feed formatter added
- handoff→ Assemble both streams into the four-pane layout (Integration)
Work sessionScorecard · event detail stream
ActionScorecard work (stale base)
- handoff→ Assemble both streams into the four-pane layout (Integration)
Coding agentIntegration
ActionAssemble both streams into the four-pane layout
- handoffIntegration regression→ The stale base reverted the #140 formatter (Integration)
Coding agentIntegration
Observation · findingThe stale base reverted the #140 formatter
- sent backFormatter missing→ Restore the #140 formatter (Integration)
Coding agentIntegration
ActionRestore the #140 formatter
- handoffCombined diff→ #145 (56f6f1b) · local lint · typecheck · test · build · screen check (Artifacts)
ArtifactsArtifacts
Artifact#145 (56f6f1b) · local lint · typecheck · test · build · screen check
- handoffHandoff for subsequent work→ Handoff record: #142 owned by another stream · #146 CI in progress (Artifacts)
ArtifactsArtifacts
ArtifactHandoff record: #142 owned by another stream · #146 CI in progress
- Context
- Studio work was split into streams such as feed and layout, scorecard and event detail, and theme. After #145 merged, the handoff record for the next stream listed work other streams owned and work still pending.
- Outcome
- #145 combined the scorecard and event-detail work with the layout fix, restoring the feed formatter that had been reverted, and recorded local lint, typecheck, format, test, build and a screen check.
- Limit
- The local checks were not re-run. At the time of the handoff record #146 was still in CI and merged later (109da28), so that snapshot is not drawn as passing. Per-stream models are not recorded.