Projects: coordinator + thread sessions merged PRs with red CI and ignored AGENTS.md during multi-thread implementation

Summary

In a Project (coordinator session + multiple thread sessions), a multi-issue server implementation went badly wrong. Several PRs had to be reverted, and the whole task had to restart from task breakdown. The same product works well for single-task investigate-and-fix projects (e.g. a Sentry auto-fix project), but fails badly for coordinated multi-thread implementation.

What happened

• Thread sessions merged 2 PRs while CI was red, even though the project instructions required green CI.
• One PR was merged about 4 seconds after leaving draft. The repo's CI skips the real test shards while a PR is a draft, and the thread treated the draft-time green as final.
• Threads wrote code without reading the repository's AGENTS.md. This broke two documented rules: compatibility with released clients, and running DDL on hot tables alone in its migration run.
• As a result, an old client paired with the new server would have put every new user into an unfinished feature flow.
• Threads reported unverified guesses as facts (e.g. "the client doesn't call this endpoint"), and the coordinator relayed them without checking.
• Outcome: 4 merged PRs reverted, 3 open PRs closed, and the task breakdown restarted from scratch.

Root causes (as analyzed with the user)

• The coordinator only relays messages. It does not supervise or verify thread output.
• Requirements were lost when the coordinator summarized them into thread briefs.
• The harness pushes threads toward action ("bias to action") over reading repo rules and verifying assumptions.
• Merge conditions were reworded across messages, and threads interpreted them loosely.
• The model tier was not the cause: single sessions on the same model follow AGENTS.md fine.

Expected

• Threads always read the repo's AGENTS.md / CLAUDE.md before writing code.
• A hard, harness-level block on merging while any CI check is red or still pending.
• A clear distinction between verified facts and inferences in thread and coordinator reports.
• The coordinator verifies thread claims before relaying them.
• Documentation of which workloads Projects is suited to and which it is not.

Environment

Claude Code on the web, Projects (coordinator + thread sessions), GitHub-connected Python/Alembic backend repo.