Repository navigation
Run the Security review on security.harness when it is set #910
Description
This was generated by AI during triage.
A Security review always runs on its Command's Harness, so when that Harness's model refuses security work there is no way to get the review done unattended. Spec run #776 lost its Security review this way, and the Day shift wants to move toward Commands that use more than one Harness.
Evidence
Spec run #776 on stephens, Spec PR #842, 2026-10-20 (UTC):
- 15:51-15:56: the Security review session ran on Codex (gpt-6.1-sol), the Command's Harness from
harness.default = "codex". While it read the repository'sthirdshift-security-auditskill files andconsole/reproductions/session-stop.py, the turn failed with OpenAI's "This content was flagged for possible cybersecurity risk. ... apply for Daybreak access ..." message. - thirdshift classified it correctly as
Codex's cybersecurity safeguard refusal(Safeguard refusals fail a Security run, and each security session's Model is logged #547). As the glossary says, that held the result like an unaddressed Security finding, and the Spec run ended failed at 20:27 after its CI went green. - The Spec adds email/password accounts, login codes, authenticated Stop and a
sudoinstall script, so it is exactly the change the Security review is for, and it reached the Day shift with no Security review.
[security] harness exists, but it chooses only a Security run's Harness. The Security review inside a Run, Spec run or Base fix ignores it.
Decision (Day shift, 2026-10-20)
Option (a): a Security review runs on security.harness when that is set. Update the glossary's one-Harness-per-Command rule accordingly. This is a first step toward letting one Command use different Harnesses for different sessions.
Rejected: (b) retry once on a fallback Harness after a refusal. It adds retry logic and still spends a refused session first.
Related
- A failure report links the most recent session's log, not the one that caused the failure #911: the same failure report linked the wrong Session log.
Activity
JacobStephens2 commentedon Oct 20, 2026
Agent Brief
Category: enhancement
Summary: Run the Security review on
security.harnesswhen it is set, so one Command can use a second Harness for its security sessionCurrent behavior:
Every session of a Command (implement, Spec review, Security review, Resumes and Repairs, and those of its child Runs) runs on the one Harness the Command chose: its
harnessword, elseharness.default, else Claude Code.[security] harnesschooses the Harness only for a Security run. A Security review that the Command's Harness refuses holds the result, and nothing tries elsewhere.Desired behavior:
security.harnesswhen it is set, wherever a Security review runs: a Run, a Spec run on its Spec PR, or a Base fix that asked for one. Its Model and Effort come from that Harness's own[harness.<name>]section, as a Security run's do. Whensecurity.harnessis blank or missing, the Security review uses the Command's Harness, as today.harnessword does not overridesecurity.harnessfor the Security review. The word chooses the Command's Harness; the User config's security choice wins for security sessions. A command-level override for the security session is out of scope..models.json, the Command log's session lines and the Run notification's "Session models" list name the security session's Harness, not the Command's. The "Built with" line keeps naming the Command's Harness. If several Harnesses ran, it may list the security one too, for example· security review on claude · claude-opus-5-5.security.harness.security.harnesswhen set.docs/guide.md's[security]comment and Harness section are updated.security.harnessnow says it covers Security reviews as well as Security runs.Key interfaces:
security.harness, thenharness.default, then Claude Code). The Security review needs a variant that skips the command word.Choiceit passes to the Harness..models.json) and the Run notification's session list.Acceptance criteria:
harness.default = "codex"andsecurity.harness = "claude", a Run's implement and Repair sessions launch Codex. Its Security review launches Claude Code with[harness.claude]'s Model and Effort. Shown with the test fakes for both Harnesses.security.harnessblank or missing, the Security review runs on the Command's Harness (existing tests pass).harness codexword withsecurity.harness = "claude"still runs the Security review on Claude Code..models.json, Command log and Run notification name the Security review's actual Harness.cargo test,cargo clippy --all-targets -- -D warningsandcargo fmt --checkpass.Out of scope:
security.harness.