169. Assess vulnerabilities with a read-only agent before the code agent
Date: 2026-11-11
Status
Accepted
Context
Teams want fullsend to handle vulnerability work: a CVE arrives as a tracker, an alert or an issue, someone decides whether it affects a repository, and if it does, someone fixes it. Many findings need no code change, because the component is absent, the fix is already carried or the vulnerable code cannot be reached. The result is then a justified "not affected" on the tracker, not a pull request. No built-in agent produces that result: triage labels and routes an issue, and code writes a change.
The CVE workflow experiment showed the flow can work. It composes the triage harness with its own agent, result schema and scripts, registers that under the triage name, and keeps a controller in the repository. A second repository would have to copy that code, and a repository that adopts it no longer has regular triage. Dispatch to several targets, recovery and Pi tool selection for this work are being decided separately (#9266, #9264, #9265).
Each step gets its own agent, harness and sandbox (ADR 0020), any harness may name any existing role (ADR 0124), and runs are serialized per harness and work item (ADR 0106). This ADR aims to give any repository a trustworthy verdict on one public vulnerability for one branch, and a fix where one is needed, without orchestration code of its own.
Options
- A skill on the code agent. The smallest change. The code agent's output is a pull request and its role can write code, so every assessment, including the many that end in "not affected", would run with write access and have nowhere to put its answer.
- The triage agent, extended or composed. Triage has the right role. Extending it adds a verdict to a result every adopter shares, on a time budget sized for labelling an issue. Composing over it, as the experiment does, replaces regular triage for the repository.
- A Pi extension that holds the whole workflow. It works on one runtime, which the adopter chooses per agent (ADR 0091), and the checks that gate publishing and coding would run inside the sandbox.
- A read-only assessment agent, then the code agent. Chosen.
Decision
We add a first-party cve agent that assesses one public vulnerability against one branch and changes nothing, and we hand any fix to the existing code agent, so that assessment needs no write access to code and a fix always follows a verdict.
- Role. The agent uses the existing
triagerole with a read-only checkout. It adds no role and no scope, so ADR 0124 is not engaged. It lives infullsend-ai/agents, a repository registers it with anagents:entry (ADR 0058), and its harness trigger dispatches it (ADR 0061). - Start. It starts on a
/fs-cvecomment from a person, or when thecvelabel is added. Triage does not start on a work item labelledcve. - Verdict. The result is affected, not affected, fixed or under investigation, in the terms of CSAF VEX, and each piece of evidence says whether a tool observed it or the model judged it. Host scripts publish it as one comment on the work item, which a later run updates. It is a proposal for a person to review; fullsend does not publish VEX.
- Handoff. An affected verdict records on the work item the vulnerability, the branch, the commit and the files its plan may change. The code agent works on a vulnerability only while that record is current. It opens its change against the assessed branch (ADR 0053), and a host check rejects a change outside those files. By default a person adds
ready-to-code. - One target per run. A run assesses the branch its work item names, from a list the repository gives. Several maintained branches are several work items. How those are created is not decided here; #9266 proposes one.
- Scope. Embargoed vulnerabilities are out of scope. Host scripts stop a run that cannot show its finding is public, before a sandbox exists. Whether a private tracker's text may be used for a public repository is also not decided here (#2513); until it is, the agent by default does not read that text in that case.
- Adopter facts. Maintained branches, vendor VEX sources, readable tracker security levels and label conventions are harness settings (ADR 0162), not new
config.yamlfields.
Forge scope
- GitHub: issues are the work item, including one opened from a Dependabot alert.
- GitLab: covered as the repository, with Jira as the tracker. GitLab issues as the work item are not covered yet, because no adopter files vulnerability trackers there; the agent skips them with a reason.
- Jira: Cloud trackers reach the agent through the Jira poller (ADR 0063).
- Forgejo: fullsend has no Forgejo support yet; nothing to change.
Consequences
- A repository gets vulnerability assessment by registering one agent, and a justified "not affected" needs no write access to code.
- The triage and code harnesses each gain a condition, so a labelled vulnerability reaches a verdict before either acts on it.
- A fix takes two runs with a person between them by default, which is slower than one run and keeps a person on the verdict.
- The check that holds a change to the verdict's plan covers the code agent only; the fix agent, which acts on review findings, needs the same check.
- Finding a tracker's repository and branch, and opening one work item per maintained branch, stay with the adopter until a later decision.
