gitea
05c76852a2
State Enforcement (v2.0):
- .state file as single source of truth for task phase
- Approval gates for research, decomposition, design, test_design
- status.py --transition refuses illegal phase transitions
- status.py --validate-folder detects out-of-order artifacts
- status.py --audit checks all tasks for violations
- status.py --create-task is the only valid way to create tasks
- Pre-v2.0 tasks without .state are UNTRACKED -- all commands refuse them
- New --upgrade command bootstraps .state files for existing tasks
Project Scoping:
- --project flag added to all status.py commands across 16+ files
- _find_project_dir errors instead of silently falling back to ~/.automaton/
- --scope-check marks framework files OUT_OF_SCOPE when working on a project
- Dashboard handlers use stored project_root instead of re-detecting from CWD
- Prompts reference ~/.automaton/scripts/vram_detect.py (not {project}/.automaton/)
Harness Integration:
- status.py --can-edit now supports project-level checks (no --task required)
- --can-edit --file checks file scope without --task
- --json output for machine-readable harness integration
- opencode plugin (plugins/automaton-guard/plugin.ts) intercepts edit/write
- Git pre-commit hook (scripts/git-hooks/pre-commit) blocks commits without task
- Formal integration contract (contracts/harness-integration.md)
Other:
- upgrade.sh delegates to status.py --upgrade instead of manual heuristics
- Phase prompts reference --project {project} for multi-project scoping
- 200 tests passing (14 new)
1.2 KiB
Adversarial Bug Report: Phase-Scoped Prompts
Deep Review
The ALLOWED/FORBIDDEN sections create hard boundaries that prevent phase-skipping. The user override resistance instructions give agents a standard refusal template. The orchestrate.md reduction from 493 to 143 lines is significant and removes state machine duplication.
Potential Issues
-
FORBIDDEN section is advisory only: An agent that ignores the prompt can still perform forbidden actions. The enforcement relies on the agent following instructions.
status.py --validate-foldercatches violations after the fact, but cannot prevent them in real-time. -
Agent can fabricate APPROVED signal: The approval gate says "wait for user approval," but a non-compliant agent could call
status.py --approveitself without waiting. This is mitigated by the spec requirement that--approveis an explicit user action, but a truly adversarial agent could simulate it. -
decompose.md length: At over 150 lines, decompose.md pushes against the prompt discipline target. Not a functional bug but a maintenance concern.