Spaces:
Sleeping
Sleeping
ProofFrame Agent Rules
Language
- User-facing updates should be in Chinese.
- Public product copy, README sections, demo video script, and submission text may be English.
Git
- Work only inside
/Users/junhaocheng/working-dir/ai-competitions/backblaze-proofframe. - Keep this as an independent Git repo. Do not commit changes from the parent
/Users/junhaochengtree. - Current branch:
feature/backblaze-proofframe. - Before editing, confirm
pwdandgit status --short --branch. - Commit by meaningful stages and push to GitHub after each stable stage.
- Never commit secrets, API keys, Backblaze credentials, Devpost cookies, or generated private media.
Hackathon Target
- Primary event: Backblaze Generative Media Hackathon.
- Product: ProofFrame, a provenance-first generative media vault that uses Genblaze-compatible generation and Backblaze B2-compatible storage.
- Winning angle: production-grade media operations, storage/provenance depth, and a demo that feels useful beyond the hackathon.
Agent Bus
- Team id:
proofframe-hackathon-58c62c50. - Controller session:
019f075b-d44b-7f12-9460-1782719ecbaa. - If durable Agent Bus metadata disagrees with this file, treat the path above as the current repo authority and run
python scripts/agent_handoff_check.py. - Roles:
planner: roadmap, scope, task ledger, submission evidence.scout: rules, competitors, deadlines, judging risks.executor: bounded implementation.tester: verification, browser smoke, Docker and release gates.reviewer: claims, secrets, compliance, award readiness.
- Durable Bus storage is not enough. For live role work, visible Codex delivery or a subagent id must also be recorded.
Product Constraints
- MVP must run locally without paid credentials using deterministic mock generation and local file storage.
- B2/Genblaze integration must be real, optional, and fail closed when credentials are absent.
- Every generated asset must have a manifest with prompt, provider, model, checksum, storage pointer, approval state, and license/risk note.
- Keep the demo focused on one creator workflow, not a generic image generator.
Task Ledger
- Canonical machine-readable task file:
tasks.json. - Human overview:
docs/todo.md. - Query and mark tasks with:
python3 scripts/task.py list
python3 scripts/task.py list --status todo
python3 scripts/task.py done T001
python3 scripts/task.py block T010 --note "Needs Backblaze account"