The Security review in a Run #550

This was generated by AI during triage.

Parent

#427

What to build

Mode decision (2026-10-09)

Jacob chose guidance for the optional Security review after the #549 trial was graded. Use one session to read the relevant security attack-class guidance and review the change with supporting code; do not run the full six-phase audit or delegate auditors for this review. The separate whole-repository Security audit keeps its full-audit workflow. Existing proof-of-concept, private-record, and Self-merge requirements still apply.

Evidence and limits: trial grading and decision. Guidance was cheaper and faster in this sample; detection accuracy and equivalence between modes remain unmeasured. Detailed grades remain private.

Turning it on. The Security review is off by default. It's turned on by security-review on the command, which is passed on to the Runs a command starts, or by review under [security] in the User config. no-security-review forbids it. Setup asks about it.

When it runs. When it's on, Delivery runs the Security review session after the opening session (the implement session, or a Continuation's) and before the push and the Repair loop. A Base fix gets one too.

What the session does. In the chosen mode, it:

• fixes only what it can show with a failing test, kept as a regression test;
• lists the unaddressed findings the change introduced under Security, in the pull request's Unaddressed findings;
• ends on a final line that counts them.

What thirdshift does with the result. If any such finding is left, or the session was refused or cut short, a Merge run doesn't Self-merge. It ends a Failed run whose cause names them, with its pull request left ready for review, as after a Policy refusal. A Run that isn't a Merge run lists the findings and goes on.

Acceptance criteria

• Off by default, and on with the word or the setting. The word overrides the setting, and Setup asks about it.
• The session runs after the opening session and before the Repair loop, and a Base fix runs one.
• A Merge run whose review leaves an introduced finding unaddressed, or is refused, ends a Failed run with its pull request ready, and the cause names why.
• A Run that isn't a Merge run lists the findings under Security in Unaddressed findings.
• The Session prompt is added to the Prompts and skills page.