do-code

Depends on: mav-bp-error-handling, mav-bp-logging, mav-bp-application-security, mav-local-verification, mav-scope-boundaries, mav-systematic-debugging

Implement a Code Change

Wraps implementation work in a Maverick skill so the workflow report records what was done and the agent applies the project's coding standards before editing. Invoke this skill once per task from inside do-issue-solo / do-issue-guided / do-epic Phase 5. The argument is a one-line task description that gets logged in the report.

This skill is a thin wrapper. The actual edits are made by the orchestrating Claude Code session using its built-in tools (Read, Edit, Write, Bash, Grep). The skill's job is to pin standards and scope, not to perform the edits in a subagent.

Project rules (mandatory)

Before doing anything else, check for repo-specific coding guidance at docs/maverick/skills/do-code/SKILL.md (path relative to the repo root).

• If the file exists, read it fully and treat its contents as binding for this task. It holds project-specific rules that all Maverick coding in this repo must follow. Where it conflicts with generic guidance below, the project file wins.
• If the file is absent, continue without it. Do not create or scaffold the file as part of this task.

Preflight (mandatory)

Run this first. If it exits non-zero, halt and report the stderr output to the user verbatim. Do not proceed.

uv run maverick preflight do-code

The check verifies the project is initialised and uv is on PATH.

Standards to apply

Before editing, read and follow:

• mav-scope-boundaries — implement only what `` asks for. No unrelated cleanup, no opportunistic refactors, no abstractions for hypothetical future requirements. Three similar lines is better than a premature abstraction.
• mav-bp-error-handling — error handling at system boundaries only (user input, external APIs). Don't catch what framework guarantees rule out. No defensive code for scenarios that can't happen.
• mav-bp-logging — log levels, structured fields, and the project's logging conventions.
• mav-bp-application-security — input validation, secret handling, authn/authz patterns. If the change touches any of these, surface the design choice in your verification.

Also read any project-level skills at docs/maverick/skills/ — topic-specific guidance for this codebase that supplements the maverick-wide best practices.

Implementation loop

1. Read before writing. Open the files you're about to change. Grep for existing patterns nearby — the goal is for the edit to look like it was written by whoever wrote the surrounding code.
2. Make the change. Smallest diff that satisfies the task. If you notice an unrelated bug or a tempting refactor, do not fix it in this skill — write it down for a follow-up.
3. Verify per mav-local-verification (lint, typecheck, tests). Run the project's verification commands.
4. If verification fails, diagnose per mav-systematic-debugging and fix. Do not paper over a failing test or silence a lint warning. Do not commit red.
5. Return when (a) the task is done and (b) verification is green.

Rules

• Default to no comments. Identifiers carry the WHAT. Only add a comment when the WHY is non-obvious (hidden constraint, subtle invariant, workaround for a specific bug). No "removed X" / "added Y" / "used by Z" comments — those rot.
• No half-finished implementations. If the task can't be completed cleanly, halt and report rather than leaving partial state.
• No unauthorised destructive actions. Don't git push --force, git reset --hard, rm -rf, drop tables, or delete data. If the task description doesn't explicitly authorise a destructive step, treat it as out of scope and ask.
• Tests are part of the change. If the change is testable, write or update tests in the same task (or wrap that work in do-test as a sibling call before the commit).
• Stay inside the task envelope. When `` is ambiguous, ask for clarification rather than guess.

Return Protocol

do-code is a subroutine of the calling workflow's per-task loop — not a terminal action. Returning from this skill is a hand-back to the caller's next numbered step, not a completion event (#106).

When you return from this skill, do not post a closing summary, do not stop, do not treat verification-green as "task done." The calling workflow's loop still owns, in order:

1. Closing the skill-dispatch interval that wrapped this invocation (uv run maverick report end skill-dispatch … --outcome success).
2. Optionally sibling-calling do-test for tests that weren't folded into this skill.
3. Conventional commit referencing the issue number, then push (with stacked-PR retarget if applicable).
4. Logging the commit row to the workflow report (uv run maverick report commit …).
5. Updating the tasks comment / closing the sub-issue.
6. Heartbeat refresh.

If you find yourself drafting a final summary after returning here, that is the signal: scroll back to the calling workflow's per-task loop and resume from the step immediately after the /do-code invocation.