Retro Agent

Performs retrospectives on agent workflows — analyzes what happened, identifies improvement opportunities, and proposes changes as issues.
Setup
No additional setup is required beyond the standard fullsend configuration.
How it helps
- Every workflow gets a post-mortem, not just the ones that failed badly enough for someone to notice.
- Improvement proposals are filed as issues with context, so they enter the normal triage/prioritize pipeline.
- Patterns across multiple retros surface systemic issues (e.g., a skill that consistently underperforms).
Triggers
The retro agent runs automatically when a PR is closed (merged or not).
It can also be triggered manually with the /fs-retro command.
Commands
| Command | Where | Effect |
|---|---|---|
/fs-retro | PR or issue comment | Triggers a retrospective analysis |
Requires write-level repository permission (admin, maintain, or write).
The /fs-retro command accepts optional free-text instructions after the command. The text is passed to the agent as high-signal direction about what to focus on:
/fs-retro— general retrospective on the workflow/fs-retro figure out why the review agent approved this and make sure it never happens again/fs-retro the code agent spent 30 minutes on a 2-line fix, what went wrong
Control labels
| Label | Meaning |
|---|---|
ready-for-triage | Applied to proposal issues so they enter the triage pipeline automatically. |
Configuration
See Customizing with AGENTS.md and Customizing with Skills.
Variables
| Variable | Description | Default | Valid values |
|---|---|---|---|
FULLSEND_FORGE | Forge platform. Set automatically by the harness forge.<platform>.env section. | (set by harness) | "github", "gitlab" |
Issue filing allowlist
Cross-repo issue creation is governed by create_issues.allow_targets in config.yaml — the same allowlist used by the triage agent for prerequisite issues. Proposals targeting repos outside the allowlist are skipped with a warning and surfaced in the summary comment so they can be filed manually.
The originating repo (where the PR that triggered the retro lives) is always implicitly allowed.
Multi-forge support
The retro agent supports both GitHub and GitLab. The harness forge.<platform> sections configure platform-specific policies, skills, env vars, and providers. Key differences from single-forge setup:
FULLSEND_FORGEis required. Set automatically by the harnessforge.<platform>.envsection ("github"or"gitlab").- Policy is
policies/base.yamlfor all forges. Network access is provided by profiles (fullsend-gitlab-rofor GitLab). - GitLab uses
curlinstead ofghfor API access. The GitLab profile (fullsend-gitlab-ro) allowscurlfor GitLab API endpoints only. - GitLab host validation —
forge_validate_originating_urlvalidates the host againstCI_SERVER_HOST, a GitLab CI predefined variable set automatically by the runner. Validation fails closed whenCI_SERVER_HOSTis not set. The GitLab profile inprofiles/fullsend-gitlab-ro.yamlmust also be updated.
How the agent works
The retro agent reconstructs the full workflow graph — triage, code, review, fix, and human interactions — by fetching issue and PR timelines, agent run logs, and review threads.
- Pre-script gathers metadata about the originating PR or issue.
- Sandbox — the agent reads the full workflow history, identifies patterns (wasted cycles, missed context, repeated failures), and writes structured proposals. It uses the retro-analysis and finding-agent-runs skills. The agent cannot write files or edit code in the target repo.
- Validation loop — output is checked against a schema, with up to 2 retries.
- Post-script creates issues from the agent's proposals. Proposals whose titles match evidence-for patterns (e.g. "Evidence for #1234: ...") are filtered out and folded into the summary comment as evidence notes instead of being filed as issues.
When triggered via /fs-retro, the human's comment is passed to the agent as high-signal direction about what to focus on.
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). On pi, sub-agents run with the agent model, see Fullsend pi docs on how to define a different model for sub-agents.
Effort: high (explicit in the harness; override per run with fullsend run --effort or FULLSEND_EFFORT, values low–max).
