Day to day
After either setup, the work is the same. You steer. Bleeps runs the crew from a GitHub issue to a pull request on develop. Bloops merges when the gates pass. You promote develop to main. The order is the flow. The blocks are the primitives.
The crew
Bleeps is the doer. Bleeps runs this crew and gets the work done.
- Beep
- Plans the work and writes the spec.
- Buzz
- Builds the change in an Orca worktree.
- Blip
- QA lead. Sets up and maintains the tests.
- Echo
- Independent verify. Read-only, and a separate session.
- Chip
- Proactive scout. Files issues for what it finds.
- Bloops
- Merges through a script, then studies the run and improves every agent's skills.
Scripts first: common work gets a script so nobody does it by hand or from scratch. Beep runs scripts/kick-off. Echo writes the verdict with scripts/verify. Bloops merges with scripts/merge-pr and reflects with scripts/bloops-reflect. Chip runs scripts/bleeps-scout. Tests run through bleeps test.
File an issue
A GitHub issue is the ticket. You file it, or Chip files one labelled bleeps:found. Launch an Orca workspace from that issue. Orca runs issueCommand from the shared orca.yaml, with the issue number filled in. Whether someone grooms the issue first is still open.
Spec and tasks
Beep writes specs/<issue>-<slug>/. The folder holds ticket.md and spec.md. spec.md carries a status of draft or approved. design.md is written when the approach is the risky part. Then tasks.md, a checklist. The checklist includes the tests.
Build and verify
Buzz builds, tests, and fixes in an isolated Orca worktree, with a round limit. The number of rounds is not chosen yet. When the change touches a guardrail, the loop pauses for you.
Echo then writes verify.md. Echo is read-only, does not share Buzz's or Blip's session, and does not edit. The verdict that continues is “Ready to ship”. Any other verdict stops the line before a pull request. verify.md records unit, integration, and end-to-end. End-to-end is not applicable only when the project is not a web app.
Blip
Blip is the QA lead. Blip sets up and maintains the unit tests, the integration tests, and the end-to-end tests, and writes the tests for the change. On a web app, end-to-end defaults to Playwright. The role and the onboard, pipeline, and coverage skills are in the repo.
Green tests are a hard gate before auto-merge. A failing test leaves the pull request on its branch. Echo's verdict, the reviews, and guardrail sign-off still have to pass. A missing test command fails bleeps test. It is not a skip.
Pull request and review
The pull request targets develop and the description includes the recap: composes, extends, or adds, plus a risk tier. The agent comments on the issue.
Layered review is CodeRabbit for style and obvious bugs, Devin optionally for an end-to-end pass after its noise is calibrated, and Claude Code for architecture fit. Grok Bot watches the pull-request queue, failing CI, and review nudges. That layer is Phase 2. The bot is planned.
Bloops merges
Bloops merges into develop through scripts/merge-pr. The gates are green tests, Echo’s “Ready to ship”, reviews with no blocking findings, and your sign-off when a guardrail was touched. The target is develop. Never main. Label rules can still keep a ticket from merging.
After a run, Bloops runs scripts/bloops-reflect and proposes improvements to skills, scripts, and AGENTS.md. Those proposals are issues or pull requests. Bloops does not edit those files in place.
scripts/merge-pr reports whether the gates passed. It does not turn on GitHub auto-merge. Whether that setting is on is still an open question. --apply does not merge. High risk waits for you.
After a real merge, the script asks bleeps health for the deployed SHA. scripts/bleeps runs test and typecheck now. health still refuses. A green check is not deployed proof.
You promote main
You merge develop into main. On your project, you are the owner. The crew never touches main.
Guardrails
These five wait for you during the build, and auto-merge stays shut when one of them changed without your sign-off. Risk tier high means the diff touches one of them. The same rule applies to issues Chip filed.
- Auth
- Billing
- Payments
- Data deletion
- Migrations
scripts/guardrails reports a matching path. Bloops's merge script refuses develop when a guardrail was touched without your sign-off.
Friction log
A papercut that will happen again becomes a GitHub issue labeled friction. File what the change left over. Fix a small, in-scope papercut in the current change and mention it. Search open friction issues first, and comment on a match instead of opening a duplicate.
The issue body names:
- What happened.
- What you wanted.
- How to reproduce.
- The cost if it stays.
- What you already looked at.
- Files, when you know them.
Leave out secrets, tokens, and private data. One-off confusion stays out of the log.
Chip is the sweep. Open friction issues that are not labelled friction-skipped join Chip's queue. Chip does not file a second issue for the same gap. friction-skipped holds one issue out until you remove the label. Use it when the fix needs your decision. The crew opens a normal pull request into develop. Nobody merges that work to main.
The patrol
Chip patrols on a schedule. The scans are failing tests, a test that fails once and passes on the retry, CI breaks, typecheck failures, lint, npm audit advisories, TODO and FIXME debt as one issue, coverage gaps, and open friction issues. Chip files bleeps:found for a new finding and starts the same pipeline, under the same guardrails. Open friction issues are not filed again.
The patrol is per project. proactive in orca.project.yaml defaults to on. Set proactive: off and Chip skips that project. Issues already open stay open. scout_schedule defaults to daily and scout_time defaults to 09:00. scout_cap defaults to 2 open pipelines.
The schedule is an Orca automation, not a key inside orca.yaml. After you set scout_provider, this prints the command and then creates it:
python3 scripts/bleeps-scout schedule
python3 scripts/bleeps-scout schedule --installpython3 scripts/bleeps-scout --dry-run prints findings and does not file them. The home server is not running that automation yet. scout_provider stays unset until you choose the agent.