codemode-tag project configuration

The <codemode> response-format experiment, implemented entirely in this repo. Agents in a project born from this template respond with markdown plus one embedded tag instead of a bare ```ts fence:

Good question! Let me look into it.

<codemode status="Checking your files">
const foo = await itx.doWhatever()
return { abc: foo.bar }
</codemode>

• markdown outside the tag → delivered to chat as the agent's message
• status="..." → the live activity label while the code runs
• tag body → runs as an itx script; the return value drives the next turn

How it works#

The platform births every agent with default response parsing ON and a HIGH debounce (60s) — that window exists exactly so a project's worker can shape the agent before its first turn. worker.ts here is the interpreter:

1. On each agent's birth (agent/created), one atomic batch turns default parsing off (config.interpretResponses: false), replaces ONLY the #output-formatting section of the platform prompt with prompts/agent-system-prompt.md (an agents/context-updated op — the rest of the platform prompt stands), and lowers the debounce to the ordinary 250ms — the done-configuring signal, which releases a held first turn immediately.
2. It injects AGENTS.md as standing context, and re-syncs both it and the prompt on every config-repo commit.
3. On assistant output events (stamped by the agent processor), it parses the tag (codemode-format.ts) and appends the consequences: the script request, the prose as a chat message (marked so it isn't mirrored back into history), the status as agent/summary-updated, or corrective feedback for malformed/multiple tags. It interprets only agents whose parsing flag is off — with the flag on (a birth this worker was too slow for), the platform's own parser owns the turn.
4. On script settlements, it renders the result back as developer context — which is what triggers the agent's next turn.

Everything is public stream events any project member could append; no platform privileges are involved. Iterating on the format — prompt, grammar, rendering — is a commit to this repo. No platform deploy.

Known limits#

• Delivery to the config worker is observation-grade: an event the handler fails on is skipped, not retried forever, so a dropped delivery quietly kills that turn (a new message starts fresh). If the experiment graduates, the promotion path is a real hosted stream processor (createProcessorHost) or platformizing the proven format.
• If this worker is down or slow past the 60s birth window, the agent answers with the platform's fenced-ts defaults — coherent, just not the codemode dialect — until the next deploy's project/worker-updated sweep converts it.
• Slash commands (/example) are platform interpretation and are inert on converted agents.
• Web agents only: slack/telegram/email agent paths are excluded from the conversion and keep the classic fenced format.

Switching an existing project to this template#

There is no first-class "re-template" door yet (configRepoTemplate applies at creation; cli config-repo reset targets only the default template). The wholesale switch is a commit: overwrite /repos/config with this template's files (one multi-file workspace commit — delete what the old config had, write these). The commit auto-redeploys the project worker, whose project/worker-updated sweep appends the parsing-off + debounce config to every existing web agent and syncs the prompt behind it. A turn generated under the fenced prompt but parsed by the codemode parser gets corrective feedback and the loop recovers — accepted experiment caveat.