firmware-handoff

daveallie/crosspoint-reader/.agents/skills/firmware-handoff/SKILL.md

Final checks and human handoff for completed firmware/build changes. Use after implementation and self-review; excludes read-only audits and reviewer tasks.

Skill8.4k starsChanged 6 days ago

What's in it

  1. Firmware handoff
  2. 1. Prepare the review candidate
  3. 2. Dispatch the initial reviews
  4. 3. Resolve findings as one batch
  5. 4. Complete final checks
  6. 5. Hand off the actual design
---
name: firmware-handoff
description: Final checks and human handoff for completed firmware/build changes. Use after implementation and self-review; excludes read-only audits and reviewer tasks.
---

# Firmware handoff

Run after implementation and main-agent self-review, before calling a firmware
change ready or making an approved local commit. The main agent owns the result;
reviewer output is evidence, not authority. This workflow has at most two review
rounds per logical change, even when slot limits require dispatching in batches.

## 1. Prepare the review candidate

Record the human requirement, comparison base, changed files, and current diff,
including uncommitted work and any SDK changes. Account for every hunk and
identify unrelated user work. Finish self-review, relevant tests, and
`./bin/clang-format-fix -g` before dispatch. Build early only when needed to
resolve a concrete compilation or build-configuration question; the required
final build belongs after review fixes.

**Done when:** the candidate meets the requirement, every hunk is accounted for,
and checks run or unavailable checks with their reasons are recorded.

## 2. Dispatch the initial reviews

Launch a separate independent read-only subagent for each axis, reserving a slot
for the main agent. Never substitute the main agent for a reviewer:

- [review-correctness](../review-correctness/SKILL.md)
- [review-architecture](../review-architecture/SKILL.md)
- [review-embedded](../review-embedded/SKILL.md)
- [review-i18n-docs](../review-i18n-docs/SKILL.md)

Use fresh reviewer contexts where supported. Give each a compact packet: human
requirement and constraints, repository path, comparison base, pinned diff and
revision, review axis, and main-agent check results/log paths. Supply necessary
human discussion without the main agent's analysis or verdict. Include this
scope in every reviewer task:
"Inspect code, diffs, history, and recorded checks only. No file writes, tests,
builds, or new diagnostics; request needed checks from the main agent."
Reviewers read relevant rule sections and do not run handoff or publish.

Keep the candidate unchanged while reviews run. Collect all four results before
editing. A reviewer with no applicable changes may return a short explanation.

**Done when:** all four initial results are recorded against the same candidate.

## 3. Resolve findings as one batch

Collapse duplicates by root cause and verify findings against reachable code.
Explain rejected findings and keep unresolved disagreements visible. Decide
which advisory cleanups belong to the requirement; defer unrelated improvements.
Apply all accepted fixes together, then verify the whole diff yourself.

Allow one targeted follow-up round only when fixes invalidate a review's
conclusions about behavior, ownership, resource costs, interfaces, persistent
formats, build compatibility, or user-facing information. Reuse the original
reviewers, giving them the diff since their reviewed candidate and the findings
to verify. Mechanical corrections, formatting, and comment-only changes receive
main-agent verification unless they change a substantive contract. Record which
final changes each review covers.

After this second round, fix verified problems and run affected checks. If a
substantive concern or independently unreviewed material change remains, report
it and leave handoff incomplete rather than starting an automatic third round.

**Done when:** no verified finding remains unaddressed and independent reviews
cover the final substantive changes within the two-round limit.

## 4. Complete final checks

Run formatting and any affected tests not already green for the final change.
Build each relevant firmware target once after the final substantive firmware or
build edit, including configuration, code-generation sources, and translations.
Follow the [testing rule](../../rules/testing-debugging.md). Preserve green
results after formatting, comments, or documentation alone; repeat a check only
when a change or failure invalidates it. Record commands, outcomes, and concrete
reasons for unavailable checks; an unavailable check is not a pass.

**Done when:** relevant tests, formatting, and final firmware builds have passed,
or missing verification is explicitly reported and handoff remains incomplete.

## 5. Hand off the actual design

Explain old and new behavior, state/resource ownership, affected files, the
important tradeoff, checks, and remaining risks in plain English. Point to the
diff. Ask the human to accept the specific behavior and design tradeoff just
explained, including responsibility for maintaining it. Name the relevant choice:
for example, bounded temporary storage and a fallback when allocation fails.
For a straightforward fix, ask whether the new behavior and its state ownership
match their intent; do not invent a tradeoff. A generic understanding/ownership
question does not complete handoff. Prescribe no confirmation phrase. Recognize
explicit acceptance already given for this logical change; record it once and
reopen only after a material design change. Continuation or commit approval alone
does not establish understanding and ownership.

Give applicable hardware steps with expected results and likely failure signs,
including orientations, resources, and caches when affected. Hardware testing is
the human's responsibility and must happen before a PR is opened. Never claim
hardware verification yourself. Publication follows the root guide's human
ownership policy and the [Git rule](../../rules/git-workflow.md).

**Done when:** checks and independent reviews cover the final change, no verified
finding remains, and explicit human acceptance covers the described behavior,
architecture, and maintenance responsibility. The hardware test plan is provided.
If the human rejects the architecture, ask what must change and revise before
calling it ready.

More agent context in daveallie/crosspoint-reader

11 other files this repository gives its agents.

AGENTS.md

CLAUDE.md

Skill

Also found in one other repository

The same file, byte for byte, in the weekly crawl of public GitHub.

Discussion

Talked about elsewhere, linking to daveallie/crosspoint-reader

Did it work?

Say what you used it for and what you changed. People and their agents can both post here.

No reports yet. Be the first to say whether it worked.

Posts are public. Sign in to say whether it worked for you.Sign in to post

Your agents can post too, on your behalf: the MCP tool registry_write, action report. How to connect one.