Guru Check Task

---
name: guru-check-task
description: Check the complete current task through one AI-owned semantic round, compact private evidence, and four minimal typed exits.
---

# Guru Check Task

Use this Skill after implementation reaches a reviewable state and before task
commit. Load [references/contract.md](references/contract.md) before acting.

You, the AI currently executing this Skill, are its semantic owner. Follow the
contract's six-step execution order yourself. A worker is an optional evidence
source, never a prerequisite owner: no dispatch, unavailable workers, or a
missing agent ID does not suspend your review. `owner_not_yet_executed` means
continue the review here, not a typed stop or a request for user authorization.

Before searching Docs, code, tests, fixtures, consumers, or history, read
`.trellis/spec/workflow/semantic-retrieval.md` and apply it in the complete
nine-dimension semantic round. A negative coverage conclusion cannot rest on a
single-language zero result, and the search process is not gate artifact data.

Before a proposed implementation mechanism participates in Phase 2 severity,
findings, tests, implementation routing, or planning-stale judgment, invoke
`guru-qualify-solution-mechanism` with `phase2_candidate_set`. A
`mechanism_revision_required` result removes or replaces only that mechanism
and returns here for fresh qualification; it never enters scope confirmation.

Read the approved planning, live diff and dirty paths, implementation, tests,
Docs SSOT, source Issue authority, and applicable repository validation directly. Classify
every candidate through `guru-qualify-normal-scenario:phase2_candidate_set`
before assigning P0-P3 severity, then
review the nine Phase 2 adequacy dimensions. Worker output is ephemeral input to
the owner; it is not a handoff, assignment, raw report, or persisted review
transcript.
For delete, replace, or merge work, apply
`.trellis/spec/workflow/subtraction-first-compatibility.md` and independently
judge `code_subtraction` and `docs_ssot_subtraction`. Check affected deprecated
assets, direct-evolution rationale, current support contracts, and the reason
for production, test, generated/managed, or documentation growth. Unsupported
compatibility or redundant state is a current finding, not future cleanup.
Apply the same maintainability boundary: do not add complexity for excluded
hostile-input, concurrency-stress, unusual-crash, or formal-idempotency cases;
require a named direct consumer for new task-local fields or persistence; and
review every touched non-generated code file at or above 3000 lines for a
mechanical split or small decoupling refactor. Untouched historical large files
are outside this task's scope.

Before this semantic round can pass, consume a fresh
`guru-maintain-architecture-baseline:task_impact_sync(stage=phase2)` result.
Reread the current Architecture Baseline, design constitution, task-local
Architecture change contract, and every applicable project Architecture check.
Phase 2 performs the first semantic before/after satisfaction judgment for the
complete worktree candidate. A failed or unverified mandatory check, new or
worsened deviation, stale identity, authority conflict, or incomplete contract
cannot be hidden as an observation or future cleanup and cannot reach `passed`.
The Architecture owner keeps applicability and route ownership; this Skill
consumes its current result as direct evidence in the existing adequacy review.

The caller forms only candidate refs, observed behavior, locators, and minimal
reproduction clues. `classified` returns to this owner;
`scope_confirmation_required` routes to requirements clarification;
`mechanism_revision_required` returns to implementation for remove/replace and
a full fresh check; `blocked` stops. Rejected candidates cannot produce a
finding, test, implementation route, `planning_stale`, or user question. The
Phase 2 gate stores only this owner's final classifications and six-part witness
for its direct consumer, never qualifier stdout, report, locator, or checkpoint.

Every official worker invocation prompt authorizes approved-plan work only.
For a planning-external observation, the worker must stop before editing,
adding a test, self-fixing, assigning severity, classifying, or choosing a
route, and may return only `candidate_ref`, `observed_behavior`, `locators`, and
`minimal_reproduction_hint`. Reread those locators and run fresh qualification
before continuing the work or dispatching another worker for it. Do not modify
or replace upstream-owned `trellis-*` agent files to enforce this boundary.

After normal-scenario qualification, any implementation-discovery expansion of
scope, risk, owner, state authority, persistence, SDK lifecycle, external
integration, or another architecture boundary immediately invalidates the
Planning-stage Architecture result. Stop before the expanded edit or test and
mandatory invoke
`guru-maintain-architecture-baseline:task_impact_sync(stage=implementation_discovery)`;
only its fresh current route may resume implementation.

The public input only routes initial, finding-fix, or planning re-entry. It does
not author evidence conclusions, findings, the AI gate, or an exit. After the AI
semantic result exists, record and validate only the compact 5.0 owner-private
projection. Its one composite worktree-content token serves only the adjacent
freshness checker; it is not semantic or workflow authority. Return exactly one
of `passed`, `implementation_required`,
`planning_stale`, or `blocked`.

Equivalent formatting or workflow-metadata deltas refresh only direct
dependencies. Real semantic changes, unknown dirty content, reproduced
findings, missing verification, or authority changes still fail closed or route
to the owning step. This package requires the complete compatible Guru Team
preset runtime and is not self-contained or portable.