dbtlr finishing-a-task
The through-line from "I think I'm done" to "merged and moved on." Work is not complete until it is verified — by an adversarial agent — and presented to a human. This skill turns that rule into a sequence, orchestrating the skills that own each step rather than doing their jobs itself.
Primary agent / controller only. This drives session-level flow (invokes
adversarial-review, creates PRs, hands off towatching-a-pr). A subagent that built something reports back; the controller finishes the task.
| Thought | Reality |
|---|---|
| "It's obviously finished, I'll just open the PR" | Opening the PR is declaring it done. The workflow runs first, not after. |
| "The change is small — I'll just push to main" | Never push or merge to the main branch — all work ships as a PR. Small changes get a PR and a review too. |
| "I'll review it after I put the PR up" | After the PR is up, the human is already looking. Verify before you present, not after. |
| "I already read it over as I wrote it" | You re-derive your own blind spots. Verification is adversarial and fresh-context — that's adversarial-review, not self-review. |
| "This is just a doc/config tweak" | Then adversarial-review's own proportionality gate will say so in one line. Let it decide the tier or the skip. |
The workflow
Walk it in order. Each gate stops the task from "finishing" on a wrong assumption.
1. Are you actually finished?
If the task isn't complete — tests not written, a TODO still open, the feature only half-built — STOP here. Finishing-a-task is the last step, not a way to skip the work. Come back when the change is genuinely done and its own gates (build, lint, tests, a smoke if it has a runtime surface) are green.
2. Are you opening a PR?
All work ships as a PR — never a direct push or merge to the main branch. So:
- Creating a PR for the user to review → go to step 3.
- Not yet — still local, not sharing → there's nothing to finish; come back when you are.
3. Verify — run adversarial-review to completion
Invoke the adversarial-review skill and let it finish. It runs its proportionality gate (or declares a skip), the suppression scan, delegates the review, and drives the resolution loop until every finding is in a terminal state (fixed / dismissed-with-reason / deferred-to-a-tracked-task), then writes the Adversarial-Review: disposition trailer and presents the table.
Do not continue past this step until that trailer exists on the branch. The trailer is the proof the work was verified; it's what the pre-PR hook (when present) checks, and it's what makes "done" honest. If review surfaced fixes, they're part of the change now — re-run the earlier gates it may have invalidated.
4. Open the PR — opening it is the presentation
When adversarial-review returns, continue straight through and open the PR in the same turn. Do not stop to ask whether to open it, and do not offer to let the human review the code first — the PR is the review surface. Opening it is how the change is "presented to a human"; it is not a step that needs approval. (Merging the PR stays with the human; opening it does not.)
Create the PR (gh pr create), carrying adversarial-review's disposition table into the body — the review's outcome is part of what's presented. adversarial-review produced and presented that table in chat at step 3; you place it in the body here.
(For a declared skip, adversarial-review produced only a skip trailer and no table — put the skip reason in the PR body instead.)
5. Hand off to watching-a-pr
Once the PR is open, invoke watching-a-pr with the PR number. It arms the background watcher and carries the PR the rest of the way — surfacing CI results, routing your inline review comments, and on merge running the merged cleanup and reconciling any stragglers. Your part of finishing-a-task ends here; the watch loop owns the PR until it merges or times out.
What this skill composes
adversarial-review— the verification step (3), not a parallel gate. Both fire at "about togh pr create"; when you're finishing a task, start here and let this workflow invoke it. Reaching foradversarial-reviewdirectly still verifies the change, but skips the finish sequence — the PR presentation (4) and the watch hand-off (5). It owns the review engine, resolution loop, and trailer.watching-a-pr— the post-open loop (5). Owns the watcher, CI/comment routing, and the merge hand-off tomerged.merged— invoked bywatching-a-pron merge; not called directly here.
finishing-a-task holds the sequence and its gates; each skill above owns its own step. The pre-PR enforcement hook (adversarial-review's trailer gate) is the deterministic backstop for step 3 on harnesses with hooks — this skill is the language-first path that should make the hook rarely need to fire.
Harness note
The full loop (hook backstop + background watcher) is a Claude Code capability. On a harness without background execution or hooks — notably Codex, which drives GitHub through its own tool — finishing-a-task runs as far as it can by language: verify via adversarial-review, present the PR, and rely on the human for the watch. The verify-before-present rule is portable; the watch is not.