Triage Classifier

Skill: Triage a finding into one of the four tags

Every inline comment posted by an F Prime review agent starts with a reviewer label in brackets (e.g. [Security]) followed by exactly one bolded tag. This skill is the decision procedure for picking the tag. It is referenced by every reviewer agent.

The four tags and their meanings live in the review contract §1. The decision below assumes the agent has already determined that a finding is in-scope; this skill picks the tag.

1. Decision tree

                ┌─────────────────────────────────────────────┐
                │  Is the offending behavior introduced or    │
                │  widened by this PR?                        │
                │  (see the `pr-diff-scoping` skill §1)       │
                └─────────────────────┬───────────────────────┘
                          │ no                  │ yes
                          ▼                     ▼
                  ╔═══════════════╗     ┌──────────────────────┐
                  ║ future work   ║     │ How severe is the    │
                  ║ (preexisting)║     │ behavior?            │
                  ╚═══════════════╝     └─────────┬────────────┘
                                                  │
                                                  ▼
                              ┌─────────────────────────────────────┐
                              │ Severity ladder:                    │
                              │  - blocks merge / safety-critical?  │
                              │    → must fix                       │
                              │  - non-blocking + concrete fix?     │
                              │    → suggestion                     │
                              │  - non-blocking, minor, no fix?     │
                              │    → could fix                      │
                              └─────────────────────────────────────┘

Suggestion-block usage is independent of the tag. A fenced ```suggestion block may attach to any of must fix, suggestion, or could fix whenever a concrete diff is available. See review contract §1 and §9.

2. Severity rubrics

**must fix**

Pick this tag iff the finding is all of:

• Introduced or widened by this PR.
• In-scope for the agent.
• Blocks merge (safety, security, correctness, policy violation, or CI integrity).

**must fix** always blocks merge in the per-agent verdict. Within the CI-safety agents (security, supply-chain), an outstanding **must fix** within CI-safety scope also forces CI safety: No-Go.

**suggestion**

Pick this tag iff:

• Introduced or widened by this PR.
• In-scope for the agent.
• Non-blocking (i.e., the merge can proceed without the fix).
• The agent has a concrete, expressible fix that fits in a fenced ```suggestion block.

The **suggestion** tag MUST be paired with a suggestion block. If no concrete fix is expressible, use a different tag.

**could fix**

Pick this tag iff:

• Introduced or widened by this PR.
• In-scope for the agent.
• Non-blocking.
• Minor — not severe enough to warrant a **suggestion** tier even if a fix were available (e.g., a stylistic nit, a redundant null check, a non-functional naming inconsistency).
• May or may not include a suggestion block.

could fix is NOT the default fallback. If the agent has no concrete fix to express but the finding is severity-worthy, choose **must fix** (blocking) rather than **could fix** (minor).

**future work**

Pick this tag iff:

• The finding is in-scope for the agent.
• The PR did NOT introduce or widen the offending behavior. The offending line is unchanged AND no new caller added by the PR widens its reach (see pr-diff-scoping).
• The agent considered the finding worth recording for future work.

**future work** never blocks merge. It is the agent's mechanism for acknowledging latent issues without burdening the current PR.

3. Anti-pattern: silent downgrading

The agent does not downgrade severity to avoid blocking a PR.

• A safety-critical finding that the agent can't write a suggestion for is still **must fix** — not **could fix**.
• A finding the agent is unsure about is still tagged at the right severity, with a maintainer ping appended (see review contract §4 and maintainer-lookup).

The three agent priorities (review contract §8) apply:

1. Priority 1 — Do not discard or omit a finding.
2. Priority 2 — Prefer suggestions over plain comments when a concrete fix is available.
3. Priority 3 — Be succinct (≤ 6 lines of prose).

Priority 2 never overrides Priority 1, and Priority 3 never overrides either.

4. Low-confidence handling

When the agent's internal confidence rubric (defined per-agent) is low:

1. Pick the appropriate tag on severity grounds — do not downgrade.
2. Append a maintainer ping at the end of the comment body:

   cc @<maintainer1> @<maintainer2> — low-confidence finding, please confirm.
   

3. The lookup procedure lives in the maintainer-lookup skill.

This applies per-finding (not per-comment) — multiple low-confidence findings on the same PR each get their own maintainer ping at the bottom of their respective comments.

5. Worked examples

Example A: ground-reachable assert (must fix)

A new command handler calls FW_ASSERT(arg < MAX_LIMIT) on a ground-controlled command argument. The agent traces arg to a ground input via fprime-ground-input-tracing.

• Introduced by this PR (new handler).
• In-scope for the security agent.
• Safety-critical — the assert is reachable by ground input, which could DoS the spacecraft.
• Concrete fix: replace assert with a validation + error response.

Tag: **must fix**. Include a suggestion block with the proposed validation.

Example B: unverified new GitHub Action (must fix)

A workflow added by the PR uses uses: random-org/some-action@main. The agent cannot determine the trust level of the org. The action's workflow permissions: includes contents: write.

• Introduced by this PR.
• In-scope for the supply-chain agent.
• The privileged surface makes this safety-critical for CI.
• Concrete fix: pin to a specific commit SHA (the agent provides one if it can fetch it; otherwise leaves a (verify the SHA before applying) annotation).

Tag: **must fix** with maintainer ping (low confidence on the exact SHA), suggestion block included.

Example C: redundant comment style (could fix)

A new function has both a brief comment and an inline reiteration of the function name above its definition. Stylistic, no functional impact.

• Introduced by this PR.
• Non-blocking.
• Minor.

Tag: **could fix**. Suggestion block optional (probably attached, since the fix is trivial).

Example D: preexisting latent issue (future work)

The PR refactors a function but does not touch a separate assert that the agent thinks is reachable by ground input. The assert was there before the PR and the PR did not add a new caller.

• Preexisting.
• The PR neither introduced nor widened it.

Tag: **future work**. No suggestion block. Mention in the agent's summary future work column count.

Example E: in-scope but no concrete fix, severity-worthy (must fix)

The PR's new code path leaks resources on an error path. The agent sees the leak but the correct fix depends on what the surrounding component's lifecycle contract is, which the agent can't infer.

• Introduced by this PR.
• Severity-worthy (resource leak on a flight code path).
• No concrete fix expressible.

Tag: **must fix** with maintainer ping. NOT **could fix** (severity drives the tag).

6. One-line summary

introduced ↦ severity-driven (must fix / suggestion / could fix); preexisting ↦ future work; no defaulting; suggestion blocks are orthogonal to tags; low confidence → ping the maintainer at the tagged severity, never below.