Guru Create Task Workspace

---
name: guru-create-task-workspace
description: Create or exactly reuse a reviewed issue workspace and Trellis task through an AI-owned semantic gate, ephemeral scoped confirmation, deterministic execution, and four typed exits.
---

# Guru Create Task Workspace

Use this Skill only after `guru-review-change-request:ready`, or as a standalone
invocation that can supply and revalidate the same five prerequisite results.
Load [references/contract.md](references/contract.md) before acting.

Use the reviewed repository root as cwd. The full command and example root is
`.trellis/guru-team/skills/packages/guru-create-task-workspace`; a platform
discovery projection is not the execution package. The normal record input is
the four-field authoring example there plus the actual readiness transition,
not a caller-built private plan.

Perform the semantic forward behavior and AI Review Gate, obtain the
invocation-specific human confirmation without persisting it, then run the
deterministic recorder/executor/checker and return exactly one declared typed
exit. GitHub issue mutation and workspace/task
mutation are mutually exclusive invocations. A created issue always returns
`refresh_review`; it never creates a workspace or task in the same invocation.
Retry recovery reuses one exact post-plan open issue, and mutation execution
reruns the shared base sync once before the first business write.

When a previously returned `created` DTO is lost before task-created attach,
use the `recover_created_result` profile for the exact planning task. Run
`scripts/recover-task-workspace-result.sh --task <task-ref>` and pass its
stdout with the same public input to `scripts/invoke.sh`. This path is
read-only: it requires the current task, branch, worktree, workspace boundary,
and both ignored runtime mappings to close into one identity, and it never
replays issue, branch, worktree, task, or mapping mutation.

The package wrappers require the complete installed Guru Team preset and route
through `run-skill-command`. They are not standalone implementations. Missing,
stale, mismatched, ambiguous, or unconsumed evidence fails closed.

For an ordinary existing issue, author only `naming`, `assignee`,
`side_effects`, and `ai_review_gate`; use the exact input preparation contract
in [references/contract.md](references/contract.md#recorder-input-preparation).
Do not author hashes or reconstruct predecessor private payloads. After the
semantic gate and current confirmation, send `{schema_version, transition,
authoring}` to the recorder; its stdout is the complete existing plan contract.
Drafts and issues carrying created-issue provenance still require the complete
plan compatibility input. The new input never infers missing provenance.

If the user refuses the displayed action, stop before the recorder/executor and
do not generate a plan, result, DTO, or typed exit. After the semantic gate,
confirmation, and owner recorder/executor/checker complete, pipe one closed
call-local workspace envelope to `scripts/invoke.sh --invocation -`. The
envelope contains the current public input, `readiness_current` transition,
owner plan, and checked owner result. The recorder prepares consumer-local
projections from `readiness_current`; complete predecessor owner payloads remain
private. The result checker must be run separately. Public invocation validates
the result schema and its checker status and serializes its typed exit; it does
not re-run live checking or perform an additional issue, worktree, branch, or task
mutation. No predecessor payload, plan, result, or authorization is persisted
for normal transport. Repo-relative plan/result/prerequisite locators remain
compatibility-only until the next breaking Interface migration and must not be
used by normal workflow, production eval, or installed transcript paths.