| name | fix-all |
| description | Fix all issues — no exceptions, no excuses. Fix every finding uncovered during the work, here and now. "Out of scope," "pre-existing," and "follow-up ticket" are not acceptable framings. Fixes go upstream, everything changed gets a real test, and no reward hacking. |
Mission
Fix every issue uncovered during this work ("finding"). No exceptions, no excuses, no deferrals. "Out of scope," "pre-existing," and "follow-up ticket" are not acceptable framings — if it surfaced, it gets fixed here.
Execution
Parallelization
- Decompose the work and run subagents in parallel wherever tasks are independent.
- Use git worktrees for parallel work. Worktree directories must be prefixed with a stable group id (e.g.
wg-<short-id>/<task>/) so they are trivial to identify, audit, and clean up afterward.
- Subagents inherit every rule in this document. State the rules explicitly when delegating — do not assume a subagent has seen this prompt. In particular, repeat the upstream-fix rules verbatim.
The TODO list must have an item for each finding. If there are X findings, we should have X todos. After you have made your plan go back and re-read the findings we have so far, and ensure nothing was missed in the plan. The double check again.
Commits
- One logical change per commit. Messages explain the why, not just the what.
Testing standards
- Everything you change or add has tests. Period.
- Tests exist to catch bugs. That is their only purpose. A test that cannot fail when the code is wrong has no value and should not be written.
- No reward hacking. Specifically forbidden:
- Weakening assertions to get green.
- Deleting, skipping, or
xfail-ing tests to get green.
- Mocking the system under test.
- Tautological tests, blind snapshot tests, or tests that only verify the code matches itself.
- When a bug is found, the response is fix the bug — never relax the test, never adjust assertions to match buggy output. The standard does not move when it becomes inconvenient.
- For each bug fixed, add the test that would have caught it. If you genuinely cannot, say so explicitly and explain why.
Upstream fixes are non-negotiable
If a problem is caused upstream, fix it upstream. Always. There is no version of this rule that ends with a workaround downstream.
This applies to every layer above us, without exception:
- Vendored or third-party dependencies → patch upstream, land the fix there, then consume it. If upstream is slow, maintain a fork with the fix and consume the fork — but the fix lives upstream-shaped, not as a local monkey-patch.
- Tooling, build systems, frameworks → same rule. Fix at the source.
- Shared libraries owned by other teams → file the fix in their repo. Coordinate, don't route around.
Forbidden downstream workarounds: shims that paper over upstream bugs, conditional branches that exist only because upstream is wrong, "temporary" patches in our tree, comments like // workaround for X instead of a fix in X.
If upstream is genuinely blocked (frozen repo, dead project, hostile maintainer), say so explicitly, fork it, and treat the fork as the new upstream — then fix it there. The rule is not "fix it upstream when convenient." The rule is fix it upstream.
Definition of done
A task is done only when all of the following hold:
- Every issue uncovered has been fixed — not deferred, not ticketed.
- Every upstream-caused problem has been fixed upstream, with our tree consuming the upstream fix. No local workarounds remain.
- All new and modified code has high-value tests of the kind that catch the relevant bug class.
- The full test suite passes locally.
- All commits are atomically scoped.
- Worktrees used for the work are merged or cleanly removed. Nothing orphaned, nothing prefixed
wg-* left lying around.
When rules conflict
- Do the harder, more correct thing.
- If following these rules would require shipping something broken, surface the conflict explicitly. Do not silently lower the bar.
Credit: Isaac Shapira
