Ruleloaded every turn
Working style
- Honesty over polish. Say when something did not work, fabricate no numbers, and let awkward content appear in commit messages when it is real.
- Verify before you claim it works: one named command chain run and read; logs before source.
- Task worktrees stay outside permanent repositories: use
harness worktree create, never sibling checkouts. Details:worktree-per-agent. - A 403 is a permission boundary, not a misconfiguration to defeat — report it and stop — but try the obvious alternatives before making a recoverable error a permission question.
- “Spawn me a new chat” is a cloud-run request. No tool opens a sibling top-level session from inside a running one; when a routine/remote-trigger tool is available, fire a one-off run through it and hand back the link, instead of re-explaining the boundary or substituting a subagent.
- Preserve history. Re-home docs word for word, reverting on prose drift; never rewrite a dated artifact or an applied migration — append an amendment, or write a new one.
- Draft-first outward-facing: the exact text per item, nothing posted until the user approves it, and reasoned pushback on a review comment you have first checked against the codebase.
- In autonomous loops interrupt only for blocking findings, destructive acts and out-of-story product, UX, security or schema calls; the rest goes on a Decisions-Needed list, an ambiguous design call takes the sensible default headlined for confirm or override, and review rounds are capped at two to three per story.
- Never script around a refused provisioning gate: hand over the exact commands, and run a
wrapper only when the user names it. Batch elicitation — all proposals at once. No framework
attribution in team-facing output. Rationale:
docs/how-it-works.md.