What the agent does
The agent reads the plugin's shape before adding to it. In a plugin laid out like obsidianmd/obsidian-sample-plugin, src/main.ts holds MyPlugin extends Plugin, whose onload registers a ribbon icon, a status bar item, three commands (open-modal-simple, replace-selected with an editorCallback, and open-modal-complex with a checkCallback), a SampleSettingTab, a DOM click listener through registerDomEvent and a five-minute registerInterval; src/settings.ts has MyPluginSettings with one string field and the tab that edits it. styles.css is an empty comment. manifest.json names the plugin sample-plugin with minAppVersion 1.0.0 and isDesktopOnly: false. esbuild.config.mjs bundles src/main.ts to main.js as CommonJS targeting ES2021, with obsidian, electron and the CodeMirror packages external, and npm run build runs tsc -noEmit -skipLibCheck first. eslint.config.mts loads eslint-plugin-obsidianmd's recommended rules with the TypeScript project service, tsconfig.json is strict with noUncheckedIndexedAccess, and .github/workflows/lint.yml runs npm ci, the build and npm run lint on Node 20, 22 and 24. AGENTS.md asks for small modules outside main.ts, register* helpers for anything needing cleanup, sentence case in UI text and no network calls.
The feature is a csv code block that renders as a table, which is the shape most custom blocks take: text in, DOM out. The parser goes in src/processors/csv-parse.ts as a pure function with no obsidian import, splitting lines, honouring double-quoted fields that contain commas, and returning rows of strings. The renderer in src/processors/csv.ts exports registerCsvBlock(plugin), which calls plugin.registerMarkdownCodeBlockProcessor('csv', (source, el, ctx) => ...), parses source, and builds a table with el.createEl, a thead from the first row when the new firstRowIsHeader setting is on and a tbody for the rest, each cell set with setText so no block content is ever treated as HTML. An empty block renders a short div with a class and a sentence-case message instead of an empty table. main.ts changes by one call in onload, and settings.ts gains the boolean with a toggle in the tab, defaulting to on.
The parser being pure is what makes it testable without Obsidian. test/csv-parse.test.ts uses node:test and covers a plain row, a quoted field with a comma, trailing blank lines and an empty source, and package.json gets "test": "node --test --experimental-strip-types test/", which runs TypeScript directly on Node 22 and 24 without a transpile step and without a new dependency. Because the lint workflow's matrix still includes Node 20, where that flag does not exist, the agent adds the npm test step to lint.yml with a condition that skips it on the 20.x entry rather than dropping the version from the matrix. styles.css gets rules for the table's class so the plugin's CSS follows Obsidian's theme variables rather than fixed colours.
No bump of minAppVersion is needed, since registerMarkdownCodeBlockProcessor long predates the 1.0.0 the manifest already requires, and the pull request says so. README.md gains a short section with an example block. The agent runs npm ci, npm run build, npm run lint and npm test in its sandbox and checks that main.js stays uncommitted as .gitignore and AGENTS.md require.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Add a custom code block renderer to this Obsidian plugin. Read src/main.ts, src/settings.ts, styles.css, manifest.json, esbuild.config.mjs, eslint.config.mts, tsconfig.json, .github/workflows/lint.yml and AGENTS.md first. Register a processor for csv fenced blocks that renders a table: put a pure parser with no obsidian import in src/processors/csv-parse.ts that handles double-quoted fields containing commas and blank lines, put the renderer in src/processors/csv.ts and call it from onload with one line. Build the DOM with createEl and setText, never innerHTML. Add a firstRowIsHeader setting with a toggle in the settings tab, defaulting to on, and render an empty block as a short message. Add node:test tests for the parser under test/ with a test script using --experimental-strip-types and no new dependencies, and add npm test to lint.yml skipped on Node 20. Add table styles to styles.css using Obsidian's CSS variables and a README section with an example block. Do not change minAppVersion. Run npm ci, npm run build, npm run lint and npm test, and open a pull request.",
"resources": [
{
"url": "https://github.com/your-org/your-app"
}
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
The pull request adds src/processors/csv-parse.ts, src/processors/csv.ts and test/csv-parse.test.ts, adds the setting to src/settings.ts, the one-line registration to src/main.ts, table rules to styles.css, a test script to package.json, a conditional step to lint.yml and a usage section to README.md. Its description shows the build, lint and test output from the sandbox and a sample block with the table it produces, described in words since the sandbox has no Obsidian to screenshot.
Before merging, a reviewer loads main.js into a test vault to see the table in reading view and live preview, decides whether csv is the right language name or a plugin-specific one avoids clashing with other plugins, and whether the header toggle belongs in settings or as a parameter on the block itself.
Run it on every plugin repository
A developer who maintains several plugins from the same template sends this body once per repository from their own server, each a POST /o/{org}/sessions with that repository in resources and a request_id that is a UUID v5 of the repository's full name, so a retried call returns the run already started. Each run reads its own main.ts: a plugin that already has a processors folder gets the new file beside the others, and one whose lint.yml has no Node 20 entry gets the test step without the condition.
Once the block exists, an automation on the Project with a github trigger on pull_request.opened and the prompt "if this pull request changes anything under src/processors, run npm test and npm run lint and comment with the result" starts one run per pull request. The trigger adds the pull request after the prompt, and only work by an owner, member or collaborator fires it.
The use case this job follows, with its automation and trigger: Hand off a long task.
Why run it on Rush?
A code block processor is a small API with a few ways to go wrong that only show up at review time: content pasted in as HTML, a parser tangled with the DOM so nothing can be tested, a minAppVersion bumped for no reason. The agent reads the template's own AGENTS.md conventions and the lint rules the repository already enforces, and writes the feature to pass them.
The build, lint and test output stay with the run, so a reviewer sees that the strict type check and the Obsidian lint rules pass before loading the plugin into a vault. The same request runs across every plugin a developer keeps, each in its own sandbox.
Frequently asked questions
Why keep the parser separate from the renderer?
The renderer needs Obsidian's Plugin and DOM helpers, which only exist inside the app. A parser with no obsidian import runs under node --test, so the quoting and blank-line cases are covered in CI without an Obsidian build. It also keeps main.ts to the lifecycle code AGENTS.md asks for.
Does it need innerHTML or a Markdown library?
No. Each cell is created with createEl and filled with setText, so a comma-separated value that looks like a tag is shown as text. Nothing is parsed as HTML, and the bundle gains no dependency.
Will the test step break the Node 20 job?
The test script relies on --experimental-strip-types, which Node 20 lacks, so the step carries a condition that skips it on that matrix entry. The build and lint still run on all three versions, and the description notes that dropping 20 from the matrix later removes the condition.
Does minAppVersion change?
No. The processor API is older than the 1.0.0 the manifest already requires, and AGENTS.md asks to keep the value accurate rather than high. The versions.json map is untouched because no release is cut by this change.