Human-readable walkthroughs for diffs too large to scan. · balade.dev

balade turns large, agent-generated pull requests into guided, human-readable walkthroughs. It connects explanations to the exact code, validates them against git, and tailors the complete diff for human review. Balade has its own components and review format in a Markdoc file that you can commit to your repo.

You probably already have an AI review pipeline; balade comes after. It organizes, reframes and explains the changes so humans keep a strong understanding of the codebase.

Example output:

Browse a live walkthrough: balade.dev/demo.

Let balade generate draft the walkthrough with your own model (sign in with OpenAI Codex or an Anthropic API key; built on pi.dev). Or install the authoring skill and let your coding agent commit the walkthrough Markdoc file, then render it with balade open.

Agents self-check what they authored with balade check: validation errors say exactly what to fix, so the walkthrough you open is a working one.

• Node.js 22.22.2+, 24.15.0+, or 26+

# cd into your repository and generate with a GitHub PR number or URL
npx balade generate 96

Or if there is an existing walkthrough to visualize

# existing walkthrough file
npx balade open .agents/walkthroughs/pr-96-loan-refactor.md
npx balade check .agents/walkthroughs/pr-96-loan-refactor.md
npx balade build .agents/walkthroughs/pr-96-loan-refactor.md

Run npx balade <command> --help for all flags.

open serves the app from the local repository and launches the default browser. --no-browser prints the URL without launching one; --port selects the port, and --lang en|fr selects the interface language.

During a live review, select text inside a walkthrough section and choose Ask agent. Balade anchors the question to that section and passage, runs a fresh agent against the walkthrough and pinned pull-request diff, and shows the answer in a local thread. Section badges reopen earlier exchanges and each thread accepts follow-up questions. A Clarifications section in the sidebar keeps every pending, answered and failed thread reachable, including when its drawer is closed or several questions are running. If no usable Balade login and model are configured, the first question opens one-time provider and model prompts in the terminal that started balade open, then submits the original question automatically. Run balade agent setup to configure it ahead of time. Clarifications are unavailable in static exports. Run balade agent logout to remove stored Balade logins and exercise first-run setup again. Credentials supplied through the environment remain available.

npx balade agent setup
npx balade agent setup --provider openai-codex --model gpt-5.4
npx balade agent logout

A PR target uses the checked-out walkthrough when available. Otherwise, balade fetches pull/<number>/head and reads the walkthrough from that commit. This doesn't switch branches or modify the checkout.

Discovery scans tracked files matching **/walkthroughs/*.md whose frontmatter contains walkthrough. The default generated path is .agents/walkthroughs/.

Run generation inside the repository clone:

npx balade generate 96

Balade pins the PR head, extracts it under ~/.balade/cache/snapshots/, and gives the model read-only list, search and source tools. The model can't run shell commands or write files. Balade keeps the five most recently used snapshots.

Repository instructions come from the pinned commit. If the PR changes an AGENTS.md or CLAUDE.md, balade ignores that file and prints a warning. Pass --trust-head-instructions after reviewing the change. Files containing a project-context closing tag are rejected.

The first agent-powered run opens a provider and model picker. Balade stores its Pi credentials and model default under ~/.balade/pi/; it doesn't read or modify ~/.pi/agent/. Generation, agent setup, agent logout, and live Q&A share this one configuration.

Anthropic subscription login in third-party tools uses billed extra token usage. It doesn't consume Claude plan limits.

Common options:

npx balade generate 96 --provider openai-codex --model gpt-5.4
npx balade generate 96 --preset odoo
npx balade generate 96 --lang fr
npx balade generate 96 --prompt "focus on the migration; the cache change is the risky part"
npx balade generate 96 --budget low
npx balade generate 96 --dir docs/walkthroughs
npx balade generate 96 --force
npx balade generate 96 --trust-head-instructions
npx balade generate 96 --no-browser
npx balade generate 96 --no-open

--lang controls the authored language during generation. On open and build, it changes only the app interface.

--prompt steers one run with what you already know about the change — which part is risky, what to emphasize, what a previous draft missed. It stacks with --preset and is not recorded in the generated file.

--budget sizes how much the model may inspect. Both sized tiers scale with the pull request's changed-file count: medium, the default, keeps slack for paging and adjacent files, while low allows one read of each kind per changed file — a constrained spend that still yields a walkthrough. high removes the caps entirely.

The default output is .agents/walkthroughs/pr-<number>-<title>.md. Balade decides what happens to an existing walkthrough for the same PR and language before the model runs. One stamped at an older head is refreshed: the run replaces it, even when the new title picks a different filename. One stamped at the current head prompts before replacing; scripts and CI stop instead, and --force skips the question. If a replaced file had uncommitted changes, a copy is kept beside it as <file>.superseded — committed content needs no copy, git already has it. Walkthroughs in another language are left untouched, and --dir writes elsewhere to keep both.

Balade validates the draft and allows up to two model repair turns. If a turn leaves the same diagnostic codes on the same lines, balade stops early. If validation still fails, the draft stays on disk and the command exits with status 1.

On an interactive terminal, the live status names preparation, model generation, short-lived inspection tools, checks and repair turns, with one cumulative elapsed clock for the generation run. Finished tools move into history with a green check or red failure mark and past-tense wording; repeated successful calls become one distinct milestone per authoring turn. Piped runs log each activity start once and the same milestone summary. The final summary reports total time plus preparation, model-turn and check timing. Use --no-open for scripts and CI. Use --verbose to print model-visible text and allowlisted tool calls; provider-hidden reasoning remains hidden.

Generated frontmatter records authoring package version 1.33.0. See the authoring package for the tag catalog, rubric and version policy.

Inside a herdr pane, generate reports its state over herdr's socket API: working while it authors, blocked while a login or model prompt waits for you, done when the walkthrough is ready. No setup is needed; outside herdr the reporting is off.

The frontmatter identifies the PR and commit:

---
walkthrough: 1
title: Loan wizard refactor
pr: 96
commit: 9f3c2ad
meta:
  module: acme_loan
  balade-authoring: 1.33.0
---

commit must be a reachable SHA with 7 to 40 hexadecimal characters. The app shows a stale notice when the PR head moves past it. check reports whether new commits touch referenced content.

The body uses Markdoc:

{% group label="Models" %}
{% section id="allocation" title="The allocation model" %}
What changed and why.

{% code file="src/models/allocation.py" from=41 to=58 expect="def allocate" /%}
{% /section %}
{% /group %}

expect must match part of the range's first line. A mismatch fails validation. A top-level fenced code block renders as read-only text — highlighted when the language is known, plain otherwise, so pseudo-code works. A fence tagged mermaid renders as a diagram. A fence nested in a blockquote or list does not reach the app. The authoring package documents the other blocks and structural rules.

A walkthrough opens with a section whose id is overview, and ends with a {% files /%} block, the full-PR diff browser that lists every changed file with a viewed mark. That block can hold {% filegroup /%} children to group the browser into collapsible sections:

{% files %}
{% filegroup label="Tests" only="**/*.test.ts" /%}
{% filegroup label="UI" only="app/**" /%}
{% filegroup label="Misc" /%}
{% /files %}

A group takes a required label, an optional only glob and an optional status list of A, M, D and R. Groups claim files in authored order: each one takes the changed files its filter matches among those no earlier group claimed, and a group with no filter takes the rest. Files no group claims render after the groups. Grouping splits the full diff, it doesn't filter it, so no changed file disappears from the browser.

Live review marks are stored under .balade/ at the repository root. On the first write, balade adds that directory to .git/info/exclude. Marks are local to the clone and reviewer. Unchanged sections retain their marks after a re-stamp; changed sections reset. The directory mirrors each walkthrough's full repository-relative path, so .agents/walkthroughs/pr-133.md stores marks in .balade/.agents/walkthroughs/pr-133.md.review.json and cannot collide with a same-named walkthrough elsewhere.

Live clarification threads use the matching .qa.json sidecar beside the marks. They belong to one walkthrough commit and are discarded when its PR or stamp changes. Answers are compiled through the same validated block format as the walkthrough, but never modify the walkthrough file or appear in a static export. If the walkthrough changes while it is open, a new question is rejected with a prompt to reload so it cannot attach to a different generation. If the server process stops before an answer finishes, opening the walkthrough again marks that question failed and keeps the thread available for a follow-up.

Static exports store review state in browser localStorage. An export contains:

• the old and new contents of every changed file, including files absent from the walkthrough narrative.
• each full unified diff and the source lines used by code blocks.
• repository, PR, author, branch, commit and walkthrough metadata;
• file paths, author-supplied metadata and validation messages.

Treat the HTML as a copy of the changed repository source. When opened through file://, Chrome may expose its review state to other local file:// pages in the same browser profile. The state contains no source code, but it identifies the repository, PR, commit, walkthrough and changed paths. Serve sensitive exports from a dedicated HTTP origin.

Install the generated authoring skill into the repository:

npx balade skills install

This writes .agents/skills/balade-authoring/SKILL.md. If the repository has a .claude/ directory, it also writes .claude/skills/balade-authoring/SKILL.md. Re-run the command after upgrading balade; if the installed skill is stale, check reports the version mismatch.

Use --out <dir> for another skill layout (other coding agent harnesses). The npm package also includes the rendered skill under dist/skill/.

This workflow validates walkthroughs changed by a pull request:

name: balade check
on:
  pull_request:
    paths: ["**/walkthroughs/*.md"]
jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0 # check resolves blobs at the stamped SHA
      - uses: actions/setup-node@v4
      - run: npx balade check

The workflow requires no write permission and doesn't post to the pull request.

MIT © Philippe L'ATTENTION