implement-unitlisted

# Implement Unit Shared procedure for building **one** logical unit of work: launch an implementation subagent constrained to working-tree edits, then fork `/commit-merge-push` to land the commit. Handles merge-conflict, pre-commit-hook, and push-rejection recovery. This skill runs in the **caller's thread** — it has no `context:` key — so it can launch subagents via the Agent tool and fork `/commit-merge-push`. Callers (e.g. `/implement`, `/fix-checks`) invoke it once per unit. This skill is the **single canonical home** of the model-selection heuristic. Other skills choosing a model reference this section rather than restating it. ## Parameters The caller supplies: | Parameter | Meaning | |---|---| | `model` | `opus` or `sonnet` — chosen per the heuristic below. | | `scope` | What files/behavior this unit changes, and what is explicitly out of scope. | | `context` | The surrounding plan / issue context the subagent needs to do the work. | | `commit_intent` | The "why" of the change, so `/commit-merge-push` can write a focused commit message. | ## Model-selection heuristic - **`sonnet`** for well-specified, mechanical work: small refactors with a clear diff shape, rote wiring (adding a script to a hook, renaming across files, boilerplate additions), unit-test writing with explicit cases. - **`opus`** for judgment-heavy work: cross-cutting design changes, tricky concurrency / ordering, unfamiliar subsystems, units where the plan itself leaves decisions for imp