feat(#428): instruct triage agent to omit redundant commentary

[fullsend-ai-coder]

Summary

• Add an anti-redundancy rule to the triage agent's "Comment content rules" section in agents/triage.md. The agent must now check whether the issue body or prior comments already convey a point before including it in the triage comment. Self-evident issues should receive minimal commentary focused on net-new value (related issues, test cases, severity, information gaps) rather than a full structured summary restating what the reporter already wrote.
• Update the comment field description in the sufficient action's JSON schema example to reinforce the same principle — focus on information not already present in the issue body.

Context

Issue #428 reports that the triage agent produces verbose comments (~40 lines) that restate information already present in the issue body, adding zero net-new information. The existing "Do NOT echo back raw text" rule (line 338) addresses injection safety but not content redundancy. This change adds the missing behavioral constraint.

Testing

• Verified the one pre-existing test failure (prerequisites-creates-allowed-issue) reproduces identically on the base branch and is unrelated to this change.
• Pre-commit could not run in the sandbox due to a network error (exit code 3). The post-script runs an authoritative pre-commit check on the runner.

---

Closes #428

Post-script verification

• Branch is not main/master (agent/428-triage-no-repeat)
• Secret scan passed (gitleaks — e3293716ee87a1b1a6d07a403ccdeba4d724c9b9..HEAD)
• PR body secret scan passed (gitleaks — no-git)
• Pre-commit hooks passed (authoritative run on runner)
• Tests ran inside sandbox

[rh-hemartin]

/fs-review

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 7:48 AM UTC · Completed 8:02 AM UTC
Commit: 61527d0 · View workflow run →

[fullsend-ai-review]

Review

Findings

Medium

• [protected-path] agents/triage.md — This PR modifies a file under the protected agents/ path. The PR links to issue Triage agent shouldn't repeat #428 and the description explains the rationale for the change (adding an anti-redundancy rule to the triage agent's comment content rules). Human approval is always required for protected-path changes, regardless of context.

Previous run

Review

Findings

Medium

• [protected-path] agents/triage.md — This PR modifies a file under the protected agents/ path. The PR links to issue Triage agent shouldn't repeat #428 and explains the rationale for the change. Human approval is always required for protected-path changes, regardless of context.

Low

• [instruction-completeness] agents/triage.md:339 — The new anti-redundancy rule states "do not produce a full structured summary restating each dimension." The sufficient action's JSON schema (lines 285–296) requires a triage_summary object with mandatory fields (problem, root_cause_hypothesis, reproduction_steps, etc.). While the "Comment content rules" section governs only the comment field (human-readable markdown), the phrasing could be misinterpreted as applying to the triage_summary structured object. Consider adding a clarifying note to distinguish comment content conciseness from structured-field completeness.

---

Labels: PR modifies agents/triage.md, the triage agent definition file

[rh-hemartin]

/fs-fix

[fullsend-ai-coder]

🤖 Finished Fix · ✅ Success · Started 8:11 AM UTC · Completed 8:16 AM UTC
Commit: 61527d0 · View workflow run →

[fullsend-ai-coder]

🔧 Fix agent — iteration 1 (human-triggered)

Addressed the instruction-completeness finding by adding a parenthetical note clarifying that the anti-redundancy rule applies only to the comment field, not to the triage_summary structured object. The protected-path finding is informational and requires human approval for merge.

Fixed (1):

1. instruction-completeness: anti-redundancy rule could be misread as applying to triage_summary structured object (agents/triage.md): Added parenthetical clarification at the end of the anti-redundancy rule: '(This rule governs only the comment field — always populate all triage_summary fields completely regardless of issue clarity.)' This disambiguates the comment content conciseness rule from the structured-field completeness requirement.

Disagreed (1):

1. protected-path: agents/triage.md is under the protected agents/ path: This is an informational finding, not actionable by the fix agent. The PR correctly modifies agents/triage.md as required by issue Triage agent shouldn't repeat #428. Human approval is required for merge — this is enforced by the post-review workflow, not by reverting the change.

Tests: passed

_{Updated by fullsend fix agent}

[rh-hemartin]

/fs-fix add an eval case to test this behavior

[fullsend-ai-coder]

🤖 Finished Fix · ✅ Success · Started 8:19 AM UTC · Completed 8:26 AM UTC
Commit: 15a04c9 · View workflow run →

[fullsend-ai-coder]

🔧 Fix agent — iteration 2 (human-triggered)

Added triage eval case 007-self-evident-bug-redundancy to test the anti-redundancy rule. The case presents a comprehensive bug report where the reporter has already identified root cause, files, and fix. The LLM judge rubric rewards concise net-new-value comments and penalizes restated summaries.

Fixed (1):

1. add an eval case to test this behavior: Created eval case 007-self-evident-bug-redundancy with a comprehensive, self-evident bug report that exercises the anti-redundancy comment rule. The triage_expectations penalize verbose comments that restate the issue body and reward concise comments focused on net-new value (proposed test case, secondary findings). Includes input.yaml, annotations.yaml, and a repo symlink to the shared python-webapp fixture.

Tests: passed

_{Updated by fullsend fix agent}

[rh-hemartin]

/fs-review

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 9:10 AM UTC · Completed 9:22 AM UTC
Commit: bce4241 · View workflow run →

[fullsend-ai-retro]

🤖 Finished Retro · ✅ Success · Started 6:56 AM UTC · Completed 7:08 AM UTC
Commit: bce4241 · View workflow run →

[fullsend-ai-retro]

Retro: PR #429 — instruct triage agent to omit redundant commentary

Overall assessment: This was a clean, well-executed workflow. The code agent produced a surgical first pass, the review agent caught a legitimate ambiguity issue, and the fix agent addressed it effectively. One improvement opportunity identified.

Timeline

1. Issue Triage agent shouldn't repeat #428 filed (Jul 20) — triage agent produces verbose ~40-line comments restating issue content.
2. Triage (Jul 21, run 29833037945) — correctly diagnosed the problem and proposed a test case. Ironically, the triage comment itself was a structured summary that somewhat demonstrated the reported behavior.
3. Code (Jul 24, run 30081729679) — created PR feat(#428): instruct triage agent to omit redundant commentary #429 in ~7 min with surgical changes: anti-redundancy rule in agents/triage.md and updated JSON schema comment description.
4. Review (Jul 27, run 30247422958) — found instruction-completeness (low): anti-redundancy wording could be misread as applying to triage_summary structured fields, not just the comment field. Also flagged protected-path (medium, expected).
5. Fix iteration 1 (Jul 27, run 30248849358) — added parenthetical clarification distinguishing comment from triage_summary. Correctly disagreed with protected-path as informational.
6. Fix iteration 2 (Jul 27, run 30249402627) — human requested /fs-fix add an eval case. Fix agent created eval case 007-self-evident-bug-redundancy with a comprehensive rubric.
7. Re-review (Jul 27, run 30252681895) — confirmed instruction-completeness resolved, correctly dismissed a false-positive punctuation-consistency finding via challenger. Only protected-path remained.
8. Merged (Jul 31) by rh-hemartin after manual approval.

What went well

• Review quality was high. The instruction-completeness finding was a genuine improvement — the parenthetical clarification prevents the anti-redundancy rule from being misinterpreted as suppressing triage_summary structured fields. The challenger correctly removed a false-positive punctuation finding.
• Fix agent handled both structured and freeform instructions well. Iteration 1 addressed the review finding precisely. Iteration 2 responded to a custom human instruction and produced a well-designed eval case with a 1–5 scoring rubric.
• Code agent first pass was clean and surgical. The diff was minimal and correctly scoped.

Existing issue evidence

• Issues #115 and #116 cover the triage and retro agents flagging missing eval coverage, but neither addresses the code agent proactively creating eval cases. The proposal below targets the complementary gap.
• Protected-path noise is well-covered by existing issues (agents #568, #304, #395; fullsend #1392, #2794, and others). No new proposal warranted.

Proposals filed

• Code agent should create eval cases when modifying agent behavior files (in fullsend-ai/agents)