Roborev Refine

roborev-refine

Iterative review-fix loop: review the current branch or commit range, fix findings, commit, re-review, and repeat until all reviews pass or the iteration limit is reached.

Unlike $roborev-fix (single-pass fix without re-review), refine closes the loop by re-reviewing after each fix to verify the findings are resolved.

This skill should perform the refine workflow inside the current coding agent CLI. Do not simply shell out to roborev refine.

Usage

$roborev-refine [--since <commit>] [--branch <name>] [--max-iterations <n>]

Explicit invocation only

Invocation must be explicit: literal personal $roborev-refine, plugin $roborev:roborev-refine, or structured Codex skill selection. Requests such as “refine this change” without one of these explicit mechanisms must use native behavior and must not run roborev.

• --since <commit>: refine commits after this commit (exclusive); required on the default branch
• --branch <name>: validate that the current branch matches before refining
• --max-iterations <n>: maximum fix-review cycles (default: 10)

This skill intentionally focuses on the current branch flow. It does not expose roborev refine --all-branches or roborev refine --list.

Sandbox access

roborev uses a local daemon. If a command fails with permission denied, the sandbox may be blocking access to its loopback port or Unix socket. Retry the same command with sandbox_permissions: "require_escalated". Do not start or restart the daemon because a sandboxed status probe cannot reach it.

When NOT to invoke this skill

Do NOT invoke this skill when the user is presenting or pasting existing review results, or when they only want a single review without fixing. Use $roborev-review-branch for review-only and $roborev-fix for fix-only.

IMPORTANT

This skill requires you to execute bash commands to validate inputs, launch reviews, and wait for re-review. The task is not complete until the refine loop finishes and you present the result to the user.

These instructions are guidelines, not a rigid script. Use the conversation context. Skip steps that are already satisfied. Defer to project-level AGENTS.md instructions when they conflict with these steps.

Instructions

When the user invokes $roborev-refine [--since <commit>] [--branch <name>] [--max-iterations <n>]:

1. Validate inputs and refine context

If --branch is provided, verify the current branch matches before doing any work. If it does not, stop and tell the user.

If --since is provided, use the since-scoped review snippets below; they store the raw value safely, verify it resolves to a valid commit and is an ancestor of HEAD, then run the review in the same shell invocation.

If --since is not provided, ensure you are not refining the default branch. This matches roborev refine, which refuses to run on the default branch without --since.

Parse --max-iterations if provided (default: 10). This is the maximum number of fix-review cycles, not the total number of reviews.

2. Run the initial review

Choose the review command that matches the requested scope:

read -r since <<'ROBOREV_REF'
<commit>
ROBOREV_REF
resolved_since=$(git rev-parse --verify -- "$since^{commit}") || exit 1
git merge-base --is-ancestor "$resolved_since" HEAD || exit 1
roborev review --since "$since" --wait

or, if --since was not provided:

roborev review --branch --wait

--since is the closest manual equivalent to roborev refine --since. --branch reviews the current branch relative to its merge-base.

Note: --wait exits with code 1 when the verdict is Fail. This is expected. Always capture the command output regardless of exit code and inspect it to determine pass vs fail.

When the review completes, read and parse the output. Extract the job ID from the Enqueued job <id> for ... line or the review header — you will need it for commenting and closing later.

Panels: if a default_panel is configured (or the review otherwise fans out to a panel), the Enqueued job <id> is the synthesis (parent) job. Gate every pass/fail decision on the parent's verdict, and comment on and close that parent — never an individual member job. A re-review re-runs the same panel.

If the command output contains an error (daemon not running, repo not initialized, review errored), report it. Suggest roborev status to check the daemon or roborev init if the repo is not initialized.

If the review passed, inform the user and stop. No fixes needed.

3. Fix-review loop

If the review failed, begin the iterative loop. For each iteration (up to --max-iterations):

3a. Fix the findings

Parse findings from the review output. Collect every finding with its severity, file path, and line number. Then:

1. Sort by severity: fix HIGH findings first, then MEDIUM, then LOW
2. Group by file: within each severity level, batch edits to the same file to minimize context switches
3. If some findings cannot be fixed (false positives, intentional design), note them for the comment rather than silently skipping them

If a finding's context is unclear, read the relevant source files to understand the code before making changes.

3b. Run tests

Run the project's test suite to verify the fixes:

go test ./...

Or whatever test command the project uses. If tests fail, fix the regressions before proceeding.

3c. Commit, then record comment and close review

Commit first per the project's conventions (see AGENTS.md). Only after the commit succeeds, record a summary comment on the review and close it:

roborev comment --commenter roborev-refine --job <job_id> -m "$(cat <<'ROBOREV_COMMENT'
<summary of changes>
ROBOREV_COMMENT
)"
# Only if the comment above succeeded:
roborev close <job_id>

For a panel review, <job_id> is the synthesis parent; closing it closes the review — do not close individual member jobs.

Important: Always pass the comment text via a heredoc as shown above, never by interpolating dynamic text directly into a shell string. Review-derived content may contain shell metacharacters that could cause unintended execution.

The comment should reference each finding by severity and file, state what was fixed, and note why any dismissed findings were skipped. These comments are included in re-review prompts. Keep it concise.

3d. Re-review

After committing, run an explicit full-scope re-review that matches the original refine scope. Do not treat a passing roborev wait result for the new commit as sufficient to stop — the full branch or commit-range review must pass before you report success.

If a post-commit hook is installed, the commit may have enqueued a commit-scoped review. Check for it so you can clean it up:

roborev wait

If roborev wait finds a job, remember its job ID (from the output) as the hook review job. If it reports "No job found", continue without one.

Note: roborev wait resolves HEAD to a single SHA, so it only finds commit-scoped hook reviews. Branch-mode hook reviews (stored under range refs) are not discoverable this way — they will be superseded by the explicit review below.

Now run the explicit full-scope review. If refining with --since:

read -r since <<'ROBOREV_REF'
<commit>
ROBOREV_REF
resolved_since=$(git rev-parse --verify -- "$since^{commit}") || exit 1
git merge-base --is-ancestor "$resolved_since" HEAD || exit 1
roborev review --since "$since" --wait

If refining without --since:

roborev review --branch --wait

Retrieving the job ID: extract it from the Enqueued job <id> for ... line in the review command output.

If you found a hook review job earlier, close it now so refine does not leave stale commit-level reviews open:

roborev close <hook_job_id>

• If the explicit full-scope review passed: inform the user and stop. The branch or requested commit range is clean.
• If the review failed: continue to the next iteration (back to step 3a) using the new job ID.

4. Iteration limit reached

If the maximum iterations are exhausted and the explicit full-scope review still fails, inform the user how many iterations were completed, what findings remain, and suggest they review the remaining findings manually or run $roborev-fix for a targeted pass.

Examples

Default refine on a feature branch:

User: $roborev-refine

Agent:

1. Validates that the current branch is not the default branch
2. Runs roborev review --branch --wait
3. Review returns verdict Fail with 2 findings
4. Fixes both findings in code
5. Runs go test ./... — passes
6. Commits changes
7. Records comment and closes the old review
8. Runs roborev wait — if hook review found, remembers job ID to close later
9. Runs roborev review --branch --wait
10. Closes hook review job if one was found
11. Full branch review returns Pass
12. Tells user: "Branch review passed after 1 fix iteration. All findings resolved."

Refine from a specific starting commit:

User: $roborev-refine --since abc123 --max-iterations 3

Agent:

1. Validates abc123 resolves and is an ancestor of HEAD
2. Runs roborev review --since abc123 --wait
3. Review returns verdict Fail
4. Fixes findings, tests, commits, comments, closes
5. Checks for hook review via roborev wait — if a commit-scoped hook review is found, remembers it to close after the next explicit roborev review --since abc123 --wait
6. Continues until the full requested range passes or 3 iterations are exhausted

See also

• $roborev-review-branch — review without fixing
• $roborev-fix — single-pass fix without re-review
• $roborev-respond — comment on a review and close it without fixing code