# Start Task

> Finish implementation or review directly by default, escalating for uncertainty or risk. Load this entry together with independent reads of known task files.

- Skill: `ch4570/start-task` (Agent Skill, multi-file: 51 files)
- Install (CLI): `npx skillmds@latest add ch4570/start-task`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ch4570/start-task/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: ch4570 (https://skillmd.com/u/ch4570)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/ch4570/start-task

---


# Start Task

Contract: `vulpora.start-task-profile/v1`; `vulpora.start-task/v1` applies only to `audit`.

Invoke `$start-task` (Codex, also `/skills`) or `/start-task` (Claude Code) with one non-empty task description.
Accept exactly one leading `--lightweight`, `--standard`, or `--audit` token; otherwise use `auto`. Remove that
token once and preserve the remaining host-parsed description as data. Never evaluate or shell-split task text.
Reject empty or malformed input before discovery.

## Select once

If these instructions are already supplied in context, use them without rereading the entry.
Inspect the relevant repository files and instructions. Run the bundled `scripts/select-execution-profile.js`
with this JSON shape on stdin, adjusting **every signal to observed facts** and the requested profile. Resolve
the script relative to this installed skill; no source-code or reference read is needed to invoke it.

```json
{
  "schema": "vulpora.start-task-profile-input/v1",
  "requested_profile": "auto",
  "task": {
    "scoped_target": true,
    "acceptance_known": true,
    "verification_known": true,
    "material_unknown_count": 0,
    "independent_lane_count": 1,
    "broad_change_scope": false,
    "shared_public_contract_or_schema": false,
    "risk": {
      "destructive_or_irreversible": false,
      "security_or_authorization_boundary": false,
      "credential_access": false,
      "production_data_migration_or_backfill": false,
      "production_deployment_or_multi_service_release": false,
      "external_side_effects_not_easily_reversible": false,
      "regulatory_or_audit_evidence_required": false
    }
  }
}
```

The values above illustrate a clear, reversible task, not defaults for unknown facts. Mark unresolved scope,
acceptance, or verification false and count material unknowns. Count coupled changes as one lane. Any true risk
signal forces `audit`, even with explicit `--lightweight`. Ordinary shared contracts, broad scope, unknowns, or
independent lanes select `standard` automatically. Correct rejected input; never invent a successful decision.
Rerun only when new facts change the signals, escalating before the affected action. Selection grants no authority.

Once inspection supports an expected `lightweight` result, batch the selector, prepared edit, and checks in one
conditional tool call. Print the decision; require selector exit zero and parsed `profile === "lightweight"`
before any write. A failed, malformed, or different decision must stop the batch; follow the selected path before
implementation. Never run selection and writes concurrently. Keep the guard in the tool call, without adding a
wrapper file or another confirmation turn.

## Execute the selected profile

- **Lightweight:** these instructions are complete; no additional workflow references are required. Keep one
  concise brief in session context: goal, scope, decisions, acceptance, verification, risks. The existing progress
  message can serve as the brief. Implement directly with one primary owner; no execution child. Ask only about
  material blockers or new authority; an implementation request already authorizes ordinary in-scope edits.
- **Standard:** read the [standard path](reference/kb/lightweight-path.md). Keep coupled work with one owner;
  use a short plan and batch material questions. Load transport/model references only for actual delegation.
- **Audit:** read the [operating contract](reference/kb/operating-contract.md) and
  [runtime fast path](reference/kb/runtime-fast-path.md), then follow the full v1 contract.

For lightweight/standard, do not create run directories. Do not create a ledger, frozen
DAG, route receipt, hash projection, or formal orchestration report. Standard delegation may create its bounded
transport records. Run the narrowest repository-native checks proving the changed behavior; repeat only after
new changes, failures, or evidence that existing results are insufficient. Inspect the diff, preserve user changes,
and report changed files, observed verification, and material remaining risk. Resolve known in-scope errors before
finishing; if verification is blocked, report the verified subset and exact blocker without claiming completion.