[JacobStephens2]

Summary

#627, first slice of the #626 security-fix publishing rework: the publishing session stages its proposed public issue text instead of creating the public issue directly, and thirdshift creates the needs-triage top issue from the staged text.

security-fix-publishing session (single public fix)
  final message: Proposed public issue: / Title: <t> / Body: <b>   (no gh issue create)
fixing::publish
  parse_staged_proposal -> GitHub::create_issue_in(repo, title, body, [needs-triage])
  existing checks (open, fresh, needs-triage-only, write-up leakage)
  mark ready (ready-for-agent + security-fix) -> dispatch Run

Spec and private-reuse flows keep the session-created path until staged specs land; the legacy Security fix Ticket: <URL> final line stays as fallback.

Seams under test: the security-fix publishing prompt (prompt::security_fix text); the security-fix publishing seam (security::fixing::publish staged branch plus GitHub::create_issue_in); the prompt catalog (prompts_page generated prompts/ + site/prompts/index.html).

Evidence

• Before: session ran gh issue create itself; no staged-text path existed (parse_staged_proposal did not compile — red: error[E0425]: cannot find function parse_staged_proposal).
  After: cargo test --test security_run a_staged_proposal_is_created_by_thirdshift_not_the_session passes — the fake session never runs gh issue create, thirdshift creates issue 8 from the staged title/body, labels land security-fix + in-progress, and the fix Run dispatches to PR .../pull/1.
• Full suite: cargo test green across all 44 test binaries (incl. 930 unit + 88 security_run), cargo fmt --check clean, cargo clippy --all-targets -- -D warnings clean.
• Review: two-axis review vs issue-626 (reports in .thirdshift-review-pGFnDC/standards.md, spec.md); both axes read all six changed files, none unread.

Merge Danger

Door: two-way — revert the two commits; no data migration, no destructive action.

Blast Radius: security-fix publishing — only the single-public-fix path changes behaviour; legacy session-created path and spec/private flows untouched and still covered by the existing 88 security_run tests.

Unaddressed findings

Standards

• src/security/fixing.rs legacy None arm growth (reviewer: extract if a third path ever joins; fine today). Deferred extraction point, no behaviour involved; refuted by the reviewer's own "today it is fine" plus single-reason scope (fixing changes for one reason only).
• Fixed and not listed here: duplicated create_issue doc (now one line) and STAGED_* marker normalisation (both bare, trim at parse; rendered prompt byte-identical per prompts_page golden test).

Spec

• Finding 1, validation timing (src/security/fixing.rs:85 creates before check_issue_text): rejected staged text still creates a public issue — verified by running the reviewer's a_rejected_staged_proposal_creates_no_public_issue as written, which FAILED with rejected staged text must not create a public issue: Object {"8": String("Private candidate write-up. Private record: ...GHSA-finding-0")}. Declined: pre-creation gating is the exact scope of open sibling #628 ("Validate the staged public issue text before any public creation"), so no new triage issue filed.
• Finding 2, session authority (src/security/fixing.rs legacy None arm accepts Security fix Ticket: <URL>): a session-created ticket is still accepted — verified by running the reviewer's a_session_created_public_ticket_is_refused as written, which FAILED with left: Some(0), right: Some(1) (run succeeded). Declined: enforcement is the exact scope of open sibling #629 ("Remove independent public-issue creation from the publishing session"), so no new triage issue filed.

No changed file was left unread by the review (both axes report none unread).

Closes #627

Built with muse · default model · default effort