Triage

Triage is the LLM-driven step that decides, per ticket, which skills participate in which roles in which pipeline phases. It runs once per pipeline run, produces single-line JSON, and emits the resulting SkillRound / FilterRound commands for the framework to dispatch.

For Discussion-type pipelines (legal-analysis, mad-discussion, init-project, skill-manager, autonomous), triage falls back to a legacy LLM strategy that picks Lead + Participants from available roles — phases don't apply.

Roles

Role	Output	What the skill does
Lead	Plan	Sets the plan that reviewers compare against. One per phase.
Analyst	List of observations	Contributes perspective; no veto.
Reviewer	List of observations	Compares actual code/diff against the lead's plan. Evidence-required.
Filter	Reduced list or single artifact	Reduces a list (drops duplicates, false positives) or synthesizes the final report.

Phases

Structured pipelines (fix-bug, add-feature, security-scan, api-security-scan) declare three phases: Plan, Review, Final.

fix-bug and add-feature run an AgenticExecute step between Plan and Review — the developer agent does the actual code-writing work using the plan from Plan-phase Lead. Review-phase Reviewers then check the resulting diff against the same plan via the {{plan}} template token.

security-scan and api-security-scan have no AgenticExecute (read-only scans); the phases run back-to-back.

Output schema

Single-line JSON:

{"phases":{"Plan":{"lead":"architect","analysts":["tester"],"reviewers":[],"filter":null},"Review":{"lead":null,"analysts":[],"reviewers":["architect"],"filter":null},"Final":{"lead":null,"analysts":[],"reviewers":[],"filter":"reducer"}},"confidence":85,"rationale":"lead=architect:auth-port;analyst=tester:has-tests;-dba:no-db-changes;"}

Validation has two layers:

1. Structural — rationale ≤ 300 chars; no newlines in JSON.
2. Semantic — every assigned skill supports the assigned role (SkillIndexEntry.RolesSupported); every rationale token references a key declared in the cited skill's activation or role_assignment.

A first-attempt failure triggers exactly one retry with a stricter system reminder. A second failure throws.

Rationale token grammar

Tokens are separated by ;. Two forms:

• Positive: <role>=<skill>:<key>; — "I picked <skill> as <role> because of activation/role_assignment key <key>."
• Negative: -<skill>:<key>; — "I rejected <skill> because of <key>."

Roles are lead, analyst, reviewer, filter. Keys must already exist in the cited skill's activation.positive, activation.negative, or role_assignment.<role>.{positive,negative} lists. Invented keys are rejected at validation.

Confidence

The LLM emits a confidence score (0–100). Below 70, the framework auto-downgrades any Blocking=true observation a downstream skill produces to Blocking=false, and logs the downgrade. The triage step itself proceeds — confidence < 50 is a signal but not a stop condition (operator policy may differ).

Ticket label overrides

Two ticket-label patterns are hard overrides applied to the LLM's output before commands are emitted:

• agent-smith:skip:<skill> — drops <skill> from all roles in all phases.
• agent-smith:no-test-adaption — drops the tester skill from all roles.

The LLM is told about these patterns in the system prompt for context; the framework strips the listed skills regardless of what the LLM emits. Operators can use them to disable a noisy skill on a specific ticket without editing skill configuration.

Project concept vocabulary

ProjectMapExcerptBuilder reads .agentsmith/project-map.json (produced by p0110b's AnalyzeProjectHandler) plus a ConceptVocabulary and projects them onto a narrow ProjectMapExcerpt that's fed to the triage prompt. Concept matching happens in the builder via substring match against framework names, primary language, and module paths — the LLM sees a prefiltered concept list, not the raw project map. See concept vocabulary for the file format.

Plan artifact threading

After the Plan phase, the Lead skill's observations are stored in pipeline context as a PlanArtifact. Review-phase skills with a {{plan}} placeholder in their ## as_reviewer body get it substituted at prompt-build time. Reviewers without a Plan-phase lead in the same run see (no plan provided) and run as generic reviewers.