Axon Shaft docs

What gets posted

The summary comment piece by piece, the inline comments, the progress check and reactions, and the job log.

Everything comes from axonshaft[bot]. With dry_run: true on the action, all of it goes to the job log and nothing is posted.

The summary comment

Axon Shaft keeps one summary comment on the pull request and edits it on every push. The first review creates it only once the result is ready, so GitHub emails people once, about the result. A review that fails creates it too, to show the failure card.

From the top, it holds:

  1. The card, in a GitHub alert colored by the verdict: red (CAUTION) for blocking, amber (WARNING) for needs changes, green (TIP) for looks good. See The card.
  2. When the workflow pins an older v1 release and a newer one is out, a note with the new version and links to update. See Pinning the action.
  3. Merge gates, when the project has any: each gate with ✓ (met), ✗ (not met) or – (doesn't apply), and the reason.
  4. A table of the blocking and should-fix findings. Each location links to the lines at the reviewed commit. A finding with an inline comment says "inline". A finding outside the diff has its text in the table, and its suggested change folded below.
  5. The nits, folded.
  6. The earlier findings this push fixed, and how many files the skip globs left out, when there are any.
  7. Walkthrough, folded, for a pull request that changes three or more files: the changed files grouped by area with a line on each, and a Mermaid flow diagram when the pull request changes how its parts call each other. The walkthrough setting turns it off. A walkthrough with over 8 areas, or over 20 files in an area, is left out.
  8. What this PR does, folded: what the pull request claims, its impact, its scope and what the review couldn't check.
  9. The footer. See The footer.

The card

The card opens with an image of the verdict. Under it are the headline and two rows of chips.

The Findings chip counts the blocking, should-fix and nit findings, and the earlier findings this push fixed.

The Checks row has one chip per checklist, in the automation's order:

ChipMeaning
✅Checked, nothing found.
⚠️Only should-fix or nit findings.
❌A blocking finding.
⬜Not checked: the checklist didn't run, or the model says it didn't do it.
➖Turned off on the dashboard.

The card's last line gives the kind of change, the model's confidence and the reviewed commit, such as Bug fix · 85% confidence · reviewed a1b2c3d.

One small line gives the model, the effort, the tokens and the changed lines, such as Claude Sonnet 5.5, medium effort · 62k tokens, 219 lines. On an API key the cost follows the tokens. A run that made no model call reads no model call · 219 lines.

The line opens into a table with what the run knows: the run's page on the dashboard, the Actions log, the action's version and ref, the model's name and ID, any fallback model and the subagents' model, the effort and why auto picked it, the turn limit, the changed lines and files, the files skip left out, the checklists that ran and the ones that didn't, permissions the installation hasn't approved, the tokens, the cost on an API key, and the reviewed commit.

The billed token count weighs cache use: a cache read counts as 0.1 of an input token and a cache write as 1.25. It can come out above the total when a run writes a lot of cache.

Inline comments

  • One inline comment per blocking or should-fix finding, on its lines in the diff. Nits get one too when the nits setting is As inline comments.
  • A finding that already has an inline comment isn't posted again on the next push.
  • When a push fixes a finding, its thread is resolved.

Progress while it runs

  • A check run named Axon Shaft / review, in progress while the review runs. Its title is the current step, such as "Reviewing with Sonnet, effort medium", and its summary lists the steps. It then passes with a link to the summary comment, or fails with the failure's message and a link to the run.
  • On a comment that started a run, a 👀 reaction, replaced with 🚀 when the run did its job or 😕 when it failed or did nothing. See Reactions.

Neither one sends an email.

Check runs

Each checklist gets its own check run, such as Axon Shaft / security, and merge gates get Axon Shaft / gates. See Check runs.

The job log

The job log folds each step into a group with how long it took: config from the dashboard, checking out the pull request, reading it, the review, and posting it. The review group lists each tool call the model made, such as Read src/pay.ts. The run's page on the dashboard shows this log once the job ends.

Outside a dry run, the job log never shows a finding's text, since anyone can read it on a public repo.

When the installation hasn't approved a permission the run wants, the log warns once for each, and the footer's table lists it under Missing permissions.

Action outputs

The step sets verdict, findings and summary_url, which later steps in the job can read. See Action inputs and outputs.

Last updated on