fix(#3153): add silent-failure severity escalation guidance to review agent

[fullsend-ai-coder]

Summary

• Adds silent-failure severity escalation guidance to the code-review skill (Section 4, before "Severity anchoring") and the correctness sub-agent, instructing reviewers to escalate severity by one level when a correctness bug produces silent failures (no error, no warning, appears to succeed but produces wrong results)
• Addresses a systematic severity calibration gap where the review agent anchored on triggering probability ("this input is unlikely") rather than failure-mode impact (silent vs loud failure), causing real bugs to be rated [low] when they should be [medium] or higher
• Linked PR feat(cli): add runtime fallback to agents repo for unconfigured agents #2947 as supporting evidence in issue Severity calibration: silent-failure correctness bugs should not be rated [low] #1144

Related Issue

Closes #3153
Related: #1144

Changes

internal/scaffold/fullsend-repo/skills/code-review/SKILL.md

• Added "Silent-failure severity escalation" subsection in Section 4 (Compile findings), before the existing "Severity anchoring (re-reviews)" subsection
• Includes escalation rule, rationale, and concrete examples of silent vs loud failures

internal/scaffold/fullsend-repo/skills/pr-review/sub-agents/correctness.md

• Added matching "Silent-failure severity escalation" subsection before "Technical documentation with correctness surface area"
• Same guidance placed directly in the correctness sub-agent so it applies during pr-review orchestrated reviews

Testing

• Pre-commit hooks pass on both modified files
• Gitlint validates commit message
• scan-secrets confirms no secrets in staged content
• Per issue Evidence for #1144: annotated-tag silent-failure bug consistently rated [low] while human rated [high] #3153 validation criteria: after this change, re-running the review agent on PR feat(cli): add runtime fallback to agents repo for unconfigured agents #2947's 3ce5eb7 diff should rate the annotated-tag finding as [medium] or higher in at least 4/5 trials

Checklist

• Commit message follows COMMITS.md conventions
• No secrets or sensitive data committed
• Pre-commit hooks pass
• Changes are scoped to the issue

[github-actions]

E2E tests did not run

E2E tests run automatically for org/repo members and collaborators on pull requests.

For other contributors, a maintainer must add the ok-to-test label after the latest push.

See E2E testing guide for details.

[github-actions]

Site preview

Preview: https://30f72bd7-site.fullsend-ai.workers.dev

Commit: bad3820e9edf143026b1b762745c544f830b899e

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[ralphbean]

This needs to be done in the fullsend-ai/agents repo. We moved all the agent content there and should be making changes there going forwards. We'll soon remove the old copies in internal/scaffold/fullsend-repo

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 6:38 PM UTC · Completed 6:46 PM UTC
Commit: bad3820 · View workflow run →

[fullsend-ai-retro]

Retro: PR #3156 — Silent-failure severity escalation guidance targeted wrong repo

PR #3156 was closed without merging because the code agent modified files in internal/scaffold/fullsend-repo/skills/ — the old location — when the source of truth for agent content has moved to the fullsend-ai/agents repo (created 2026-06-26). The change itself was well-crafted (88 additions, correct placement, good examples), but was wasted effort because it targeted the wrong location.

Timeline

1. May 19 — Issue #1144 opened proposing silent-failure severity escalation guidance. Triage directed changes to internal/scaffold/fullsend-repo/skills/code-review/SKILL.md (correct at the time).
2. Jun 26 — fullsend-ai/agents repo created; agent content extracted from internal/scaffold/fullsend-repo/.
3. Jul 6 18:14 — Issue #3153 opened with additional evidence for Severity calibration: silent-failure correctness bugs should not be rated [low] #1144.
4. Jul 6 18:19 — Triage ran on Evidence for #1144: annotated-tag silent-failure bug consistently rated [low] while human rated [high] #3153, recommended implementing Severity calibration: silent-failure correctness bugs should not be rated [low] #1144's proposed fix (which referenced the old path).
5. Jul 6 18:27 — Code agent created PR fix(#3153): add silent-failure severity escalation guidance to review agent #3156, modifying the scaffold copies.
6. Jul 6 18:36 — ralphbean closed the PR: "This needs to be done in the fullsend-ai/agents repo. We moved all the agent content there."

Root cause

The codebase has no signal that internal/scaffold/fullsend-repo/ content has migrated. No deprecation markers, no README, no AGENTS.md mention of the agents repo. The files are still present and editable. The code agent followed the triage agent's guidance, which transitively referenced #1144's triage comment pointing to the old paths. This is the same class of problem as #2973 and #2426 but has evolved: the content has moved entirely to another repo, not just changed scope.

Existing coverage

Skipped proposing deprecation markers in the scaffold directory itself because #2973 already covers adding scope boundary guidance to AGENTS.md and inline comments to scaffold skills. However, #2973's scope (generic vs repo-specific content) is narrower than what's needed (content migrated to another repo entirely). Proposal 1 below addresses the gap #2973 does not cover.

Proposals filed

• AGENTS.md: document that agent/skill content has moved to fullsend-ai/agents (in fullsend-ai/fullsend)
• Add silent-failure severity escalation guidance to code-review skill and correctness sub-agent (in fullsend-ai/agents)