Assess Epic

Assess Epic

Classify the epic and child tickets from durable PM and repository state. This skill decides queues; it does not write specs, implement code, create PRs, or merge branches.

Contract

Inputs

• epic id
• child ticket list
• repo state
• artifact links
• existing labels and comments
• optional run ledger records

Process

• Read the epic and child tickets from the PM system.
• Inspect labels, comments, artifact links, branches, and worktrees.
• Classify each child using the orchestrator-tracked table in epic-orchestrator/state-transitions.md.
• Build ready work and maturity work queues.
• Build the dependency map: it feeds the sprint/waterfall mode decision (§ Scheduling Policy) — the inter-child coupling graph is the discriminator — so record file-surface overlap signals (shared modules, config, schema) alongside ticket dependencies.
• Verify and repair the native relation graph: hard sequencing edges must exist as blocked-by relations on the tickets themselves (dispatch eligibility queries them — ${CLAUDE_PLUGIN_ROOT}/scripts/dispatch-eligible.sh). Add missing relations your analysis implies, flag relations that contradict it, and detect cycles (a cycle is a blocker, not a judgment call). After Gate 1 signoff the relation graph is canonical. Soft parallelism signals stay in the assessment comment — never encode them as blocked-by.
• Treat spec-ready and ready-to-implement as hard gates.
• The outputs feed the Gate 1 signoff package: state map, dependency map, and any child that should carry needs-human-spec.
• Do not create PRs or dispatch lanes from this skill.

Scheduling Policy (Mode)

After the dependency map is built, decide the epic's workflow mode — sprint or waterfall, two scheduling policies over the same lane primitives (sprint interleaves mature/deliver per child; waterfall matures all children, then delivers all):

• Gate 1 delegation wins: if the delegation comment pre-declares or constrains the mode, apply it verbatim.
• Otherwise decide autonomously from the inter-child coupling graph — the discriminator is file-overlap and blocked-by density: tightly coupled children favor waterfall's mature-all-first, independent children favor sprint's interleave. Stay within the delegation. No third human gate.
• Escalated tier: the mode decision is a judgment over the coupling graph binding epic-wide scheduling, not state classification, so it is not taken at this skill's model: haiku main-loop tier. Dispatch a mode-decision worker pinned Capable tier (model: opus on Claude Code; ${CLAUDE_PLUGIN_ROOT}/scripts/hook-require-model-pin.sh enforces the explicit pin) with the coupling graph and delegation as input; it returns the mode plus a one-line coupling rationale.
• Record it: write a # Decision Register Entry with Kind: MODE (keyed by epic id + seam timestamp) carrying the rationale, and apply the mode-sprint/mode-waterfall epic label (the cache; the register entry is truth). See epic-orchestrator/state-transitions.md § Workflow Mode and Driver Priority Table.
• Re-assessment is a standard step, not one-shot: the driver re-runs the mode evaluation at each lane close-out and the boot audit (coupling that looked loose tightens once real diffs exist); mid-epic flips are first-class, each landing a new MODE entry. Those mid-epic re-evaluations are the driver's own inline judgment over the map it already holds — no separate dispatch, since re-reading its own held map at its own session tier is not a fresh gate.
• The chosen mode feeds the Gate 1 package the same way the ready-work queue does.

Evidence

• Record child state map, dependency map, ready work queue, maturity work queue, and blockers.
• Record source links for labels, comments, specs, plans, branches, and worktrees.

Stop Conditions

• Stop on ticket access failure, inconsistent child hierarchy, or missing repository context.
• Stop when a child ticket needs human spec input.
• Stop when dependencies are unclear enough to affect sequencing — an unclear dependency also affects the mode decision (§ Scheduling Policy).