fix(agent): publish complete advisory remediation PRs

Goal and scope

Make scheduled phlo-agent maintenance attempt verified, mechanical advisory fixes instead of stopping at an existing issue. Addresses the publishing blocker documented in #1206. This does not upgrade msgpack, waive findings, merge PRs, or change live automation.

Change map and risk

• Run advisory remediation before routine maintenance, including outstanding bot-authored scanner issues. Resolve every affected consumer, verify candidate locks and focused checks, reuse existing remediation PRs, and retain issue-only escalation for blocked fixes.
• Require complete delivery. A writer failure must not produce an arbitrary root-only subset or cite checks from a different local tree as verification of the published PR.
• Increase only the draft-PR JSON request budget from 2.5 million to 20 million characters. Other routes, authentication, restricted paths, the 50-file limit, and PR-body limits remain unchanged.
• Document the workflow and activation requirements. The larger authenticated request budget increases the writer's maximum per-request memory demand.

Verification

• make setup and make check: passed during implementation.
• node --experimental-strip-types --test .amp/plugins/phlo-github/lib.test.ts: 8 passed.
• uv run --locked pytest tests/scripts/test_report_dependency_risk.py -q: 34 passed.
• Writer npm test: 12 passed; npm run typecheck and npm run build: passed.
• Plugin registration and skill discovery passed with webhook credentials omitted, without changing live webhook ownership.
• Codespell, Markdown links, and git diff --check: passed.
• Regression: the multi-consumer draft test failed with HTTP 413 before the limit change and passed afterward. Oversized drafts remain rejected, and other routes retain their smaller limit.
• A local acceptance check submitted all 15 current root/example locks, 12,550,890 JSON characters, and received HTTP 201 for one complete draft. All GitHub traffic used a local fake; no live PR was created by that check.

Live scheduled-agent behaviour and Cloudflare resource usage have not been qualified by these local checks.

Recovery and maintainer decisions

After merge, the writer still needs an explicitly approved deployment. Existing maintenance threads also need an approved replacement/cutover to adopt the new agent definition; preserve their recurrence and timezone. No live schedule, webhook, or deployment changes are included here.

Rollback requires reverting the source changes, redeploying the previous writer, and restoring the previous maintenance configuration if activated. Existing remediation PRs are not deleted or rewritten. There are no data migrations. Merge and activation approval remain pending; no independent reviewer is claimed.

    

MinIO CI blocker follow-up

The installed-provider artifact job rebuilds the checkout's MinIO image. Its Trivy report identified CVE-2026-116587 (GHSA-mgcc-cww5-fp2q) in the explicitly pinned go-jose/v4 v4.1.4 inside the MinIO binary. Upgrade that pin to v4.1.6 and update the corresponding-source notice date. No vulnerability exception, image publication, or deployment is included.

Verification of the rebuilt image:

• The original CI report fails the existing container security policy on this CVE; the rebuilt-image report passes that same policy.
• Trivy 0.72.0 reports go-jose/v4 v4.1.6, zero occurrences of the target CVE, and zero fixable HIGH/CRITICAL findings.
• An isolated disposable MinIO container passed readiness and an authenticated bucket/object write/read roundtrip.
• The bundled corresponding-source archive contains v4.1.6 and the updated notice.
• 101 focused MinIO/container/artifact tests passed (2 integration tests deselected); 17 container-security workflow tests and Dockerfile lint passed.

Fresh PR CI is running for the follow-up commit; the earlier failed run does not verify this head.