Skip to content

Fix Agent ​

Fix agent icon

Review-feedback specialist that reads review comments on open PRs, implements targeted fixes, runs tests and linters, and commits the result.

Setup ​

No additional setup is required beyond the standard fullsend configuration.

How it helps ​

  • Review feedback is addressed quickly — often before the reviewer checks back.
  • Fixes are scoped to exactly what the review requested, reducing churn.
  • The iteration cap prevents the fix and review agents from looping indefinitely.

Triggers ​

The fix agent runs automatically when the review agent submits a "changes requested" review on a same-repo PR (fork PRs are blocked).

It can also be triggered manually with the /fs-fix command.

Commands ​

CommandWhereEffect
/fs-fixPR commentTriggers the fix agent on the PR
/fs-fix-stopPR commentDisables the fix agent for this PR

Requires write-level repository permission (admin, maintain, or write).

The /fs-fix command accepts optional free-text instructions after the command. The text gives you direct control over what to fix:

  • /fs-fix — fix whatever the review agent flagged
  • /fs-fix you forgot to update the docs here
  • /fs-fix the error handling in processItem needs to distinguish between retryable and fatal errors
  • /fs-fix address the concern raised in #42 — same-repo references work (details)

/fs-fix-stop adds the fullsend-no-fix label to the PR, preventing any further automatic fix runs. Manual /fs-fix commands still work. Remove the label or use /fs-fix to re-engage.

Control labels ​

LabelMeaning
fullsend-no-fixPrevents automatic fix runs on this PR. Applied by /fs-fix-stop. Manual /fs-fix commands are unaffected.
needs-humanThe fix agent is approaching its iteration cap and needs human direction. Applied automatically when an automatic fix iteration reaches the warning threshold.

Configuration ​

See Customizing with AGENTS.md and Customizing with Skills.

Variables ​

VariableDefaultEffect
FULLSEND_FORGEgithubSelects the forge platform (github or gitlab). Set automatically by the harness forge block.

How the agent works ​

The fix agent follows a similar pipeline to the code agent, with an additional validation step:

  1. Pre-script validates inputs and checks the iteration cap (preventing infinite fix loops).
  2. Sandbox — the agent reads each review finding, implements targeted fixes, and verifies them against tests and linters.
  3. Validation loop — the output is checked against a schema, with up to 2 retry iterations if the output is malformed.
  4. Post-script pushes the commit and posts a summary comment on the PR.

Signed-off-by trailers ​

Same behaviour as the code agent: a trailer the agent added is removed and the run continues, instead of being discarded.

This matters more on a fix run, because the commits being scanned are not always the agent's. When the agent rebases, the scan widens to everything on the PR branch since it diverged from the target — which on a human's PR includes the human's own commits. Only commits the agent authored are rewritten, so a contributor's DCO sign-off survives with its SHA intact; the rebase is what makes this necessary, since it leaves the bot as committer on a commit the human wrote.

The strip is recorded on the PR summary comment:

text
_Removed a Signed-off-by trailer from 1 agent commit._

Input details ​

Bot-triggered (review agent requests changes):

InputSourceHow it gets there
Review bodyLatest CHANGES_REQUESTED review from the review botPre-fetched on the runner before the sandbox starts, injected as review-body.txt
PR diffForge-specific skill (GitHub: gh pr diff, GitLab: MR changes API)Agent calls this to understand what code changed
Repository checkoutFull repo at PR HEADChecked out on the runner, mounted into the sandbox
Repo conventionsAGENTS.md, CLAUDE.md, CONTRIBUTING.mdRead from the checkout inside the sandbox

Human-triggered (/fs-fix [instruction]):

InputSourceHow it gets there
Human instructionFree text after /fs-fix in the commentExtracted by the workflow, passed as HUMAN_INSTRUCTION env var (up to 10,000 bytes)
PR diffForge-specific skillSame as bot-triggered
Repository checkoutFull repo at PR HEADSame as bot-triggered
Repo conventionsAGENTS.md, CLAUDE.md, CONTRIBUTING.mdSame as bot-triggered
Review body (if any)Prior review bot CHANGES_REQUESTED reviewStill injected as review-body.txt, but human instruction takes precedence

Custom sandbox image ​

The fix agent shares the code agent's sandbox image. If your project uses a custom image, update the image: field in both harness/code.yaml and harness/fix.yaml.

What the agent acts on ​

When triggered by a review: the agent reads the review body, the PR diff, and the full repository checkout.

When triggered by /fs-fix: the agent reads your instruction text, the PR diff, the full repository checkout, and any prior review. When a human instruction is present, it takes precedence over the review body.

What the agent does not read ​

This is worth being explicit about, because the fix agent's scope is narrower than you might expect:

  • Inline PR review comments. The agent reads the consolidated review body, not individual line-level comments. If you need the agent to act on a specific inline comment, copy the relevant text into a /fs-fix instruction.
  • Other PR comments. General discussion comments on the PR are not part of the agent's input. Only the review body and the /fs-fix instruction are read.
  • CI logs and check status. The fix agent does not read CI logs, check run output, or merge readiness indicators. It addresses review feedback, not CI failures. (The code agent handles CI failures during implementation.)
  • Issue body. The fix agent does not read the linked issue. It operates purely on the PR and review context.

The /fs-fix instruction text can contain URLs. Whether the agent can use them depends on where the URL points:

URL typeWorks?Why
Same-repo issue or PR/MR (#123 or full URL)YesResolved via the forge API (GitHub or GitLab)
Same-repo file or commitYesSame mechanism
Cross-repo URLNoAccess is scoped to the target repo only
GitHub GistNoNot accessible from the agent environment
External URL (docs, pastebins, etc.)NoExternal HTTP access is blocked

GitHub may auto-shorten same-repo URLs in rendered comments (e.g., https://github.com/org/repo/issues/2 becomes #2). GitLab does not auto-shorten URLs but the full URL is preserved either way.

If you need the agent to act on external context, paste the relevant content directly into the /fs-fix comment rather than linking to it. The instruction supports multi-line text (up to 10,000 bytes).

Iteration limits ​

The fix agent enforces iteration caps to prevent infinite review-fix loops:

  • Automatic: up to 5 iterations per PR (configurable).
  • Manual (/fs-fix): up to 10 total iterations per PR (configurable), shared across automatic and manual triggers.
  • When an automatic run is approaching its cap, the agent applies the needs-human label.
  • Each /fs-fix comment cancels any in-flight fix run for the same PR and starts a new one.

Multi-forge support ​

The fix agent supports both GitHub and GitLab. The harness forge block selects the platform at runtime via FULLSEND_FORGE. On GitHub, the agent uses gh for API access; on GitLab, it uses curl against the REST API. Scripts dispatch forge-specific operations through fix-ops.lib.sh, and forge-specific skills provide the appropriate CLI recipes.

GitLab-specific variables ​

VariableDescription
PR_URLFull HTTPS URL of the merge request. Used to derive GITLAB_HOST and validate REPO_FULL_NAME.
GITLAB_TOKENPersonal or project access token with api scope.

GitLab host validation ​

gitlab-fix-ops.lib.sh validates GITLAB_HOST against CI_SERVER_HOST, a GitLab CI predefined variable set automatically by the runner. Validation fails closed when CI_SERVER_HOST is not set. The GitLab profile in profiles/fullsend-gitlab-code.yaml must also be updated to allow connections to the host.

Custom network policy ​

If this agent needs to reach hosts beyond the defaults, see the custom network policy guide.

Runtime support ​

Supported runtimes: claude (stable default), pi (experimental). No single-context fallback — full multi-step fix runs on both runtimes.

Effort: high (explicit in the harness; override per run with fullsend run --effort or FULLSEND_EFFORT, values low–max).

Source ​

harness/fix.yaml