v0.2.0
Diff range: 6f7d1f1a9c7e9416897bab9326477fdae5d7a5f6 → 1105c25b9f16749d3af1505bd7b8b7c62eb22e55
Codex Swarm is now a more “workflow-first” framework: instead of ad-hoc task/docs handling, the swarm runs through a single canonical backlog (tasks.json) that is checksum-protected and managed via python scripts/agentctl.py. Agents are still JSON-defined, but the execution loop is more explicit and repeatable (plan → implement → test → verify/close), with git guardrails (tight allowlists) and an explicit user-approval gate before the final closure commit—making every change more traceable than in v0.1.0.
Highlights
- Introduced
agentctl(python scripts/agentctl.py): the supported, checksum-safe way to managetasks.json, plus git guardrails (ready/verify/start/block/finish/commit). - Added a dedicated TESTER agent and reinforced the CODER → TESTER → REVIEWER workflow for change + coverage.
- Added
tasks.html: a zero-server local UI to browsetasks.jsonin a browser. - Docs overhaul: new
docs/architecture.md(with GitHub-friendly Mermaid diagrams),docs/agentctl.mdquick reference, and a workflow-artifact system underdocs/workflow/. - Improved
clean.shso the repo can be “reset” into a reusable snapshot (while preservingdocs/agentctl.mdas a quick reference).
Added
scripts/agentctl.pyanddocs/agentctl.md(task operations + git commit guardrails)..AGENTS/TESTER.json.tasks.html(local task viewer).docs/architecture.md,docs/workflow/README.md, and initial workflow artifactsdocs/workflow/T-055.md…docs/workflow/T-062.md.
Changed
tasks.jsonis now checksum-backed and intended to be managed only viapython scripts/agentctl.py(manual edits break the checksum).- Commit workflow tightened: a 3-phase cadence (planning → implementation → verification/closure), allowlist-based commit guardrails, and explicit user approval before the final closure commit.
- README / guidelines / contributing docs refreshed;
.gitignoreexpanded (e.g., Python cache artifacts).
Removed
- Legacy task tooling:
scripts/tasks.pyandtasks.md.
Upgrade Notes
- Replace any previous task workflow with
python scripts/agentctl.py quickstartandpython scripts/agentctl.py task .... - Optional: open
tasks.htmlin a browser for a visual task list.