Repository navigation
Description
Problem
A coding agent can finish an implementation turn, report success, and move to the next issue even though the previous feature never worked from its real entry point to its intended result. A patch, a passing unit test, or the agent's statement that it is done is not evidence of that connection. Catching problems early and verifying the final connection are complementary, not interchangeable.
Intended workflow
- Before implementation, name the entry point, expected observable result, and what would count as proof. If the expected behavior is undecided, surface the decision instead of locking a guess into a test.
- Implement on Low, run the earliest suitable independent check, and feed confirmed, actionable failures into a bounded Low repair. Rerun the same check.
- If the failure still stands, pass the failing evidence, attempted changes, and acceptance criteria to High for diagnosis: revisit the cause, assumptions, or check boundary. Do not treat High as an automatic code rewrite. The next concrete implementation or repair goes back to Low once the next step is clear. An unavailable or inconclusive check is not a pass and should not blindly trigger retries.
- Run an independent final check of the actual workflow in a representative environment. Record what was exercised, the result, and what remains unverified. Until that evidence exists, report implemented, not yet verified and do not treat dependent work as unblocked. Mock evidence can still be useful, but must be labeled as such.
First acceptance example
- Use a small, reproducible save flow. The UI reports “Saved” and the isolated write test passes, but a reload cannot read the saved note because the pieces are not connected.
- The early check still supplies useful evidence at its own boundary. The final check drives save → reload → read back through the application entry point and prevents a “done” verdict while the note disappears.
- After repair, that same final check verifies the note persists. An existing edit case still passes. An unavailable final check reports “not verified,” not “done.”
- Document the recipe in two layers: a user-readable check that someone can repeat without writing code and a machine-runnable version using existing tooling. Make the first example reusable in another project rather than requiring a new adapter.
Evidence and positioning
The product outcome is fewer issues that look done but were never connected end to end. The outward hook is build faster and spend less. Validate the latter rather than assuming it: with a fixed final acceptance check, compare earlier vs. later checks at the same effort, then compare the Low-first path against a High-first baseline. Record time to detect the first mistake, time and usage to a verified completion (including checks and retries), and cases falsely reported as done. Report scope and avoid general speed or cost claims until the comparison supports them.
Related: #1 (positioning and recipe intake), #2 (actionable diagnostics), #3 (evidence on interruption), #4 (same-check integrity).
Activity