Skip to content

AGENTS.md

Latest commit

 

History

History
70 lines (50 loc) · 8.97 KB

AGENTS.md

File metadata and controls

70 lines (50 loc) · 8.97 KB
Raw

pai-pro — repo maintainer guide

This file is dev-only. The per-project filmmaking agent's operating manual is at agent-templates/PROJECT_AGENT.md (copied into each project as projects/<id>/PROJECT_AGENT.md).

When you cd pai-pro && codex to work on this repo, this is the file Codex auto-loads. Per-project Codex sessions also load a narrower projects/<id>/AGENTS.md wrapper that points at PROJECT_AGENT.md; for filmmaking work, treat that project-local wrapper as authoritative.

Maintaining this repo

Below is for when you're editing the pai-pro repo itself, not running a project session. Skip if the user is asking for filmmaking work.

Editing principles

A few rules of thumb when modifying source — bias toward caution over speed.

  • Surgical edits. Every changed line should trace to the request. Don't "improve" adjacent code, refactor things that aren't broken, or delete unrelated dead code (mention it, don't remove it).
  • Simplicity first. No speculative features, no abstractions for single-use code, no configurability that wasn't asked for, no error handling for impossible scenarios. If 200 lines could be 50, rewrite.
  • Surface confusion early. State assumptions before implementing. If multiple interpretations of the request exist, present them — don't pick silently. Stop and ask when something's unclear.
  • Goal-driven, verifiable. For non-trivial changes, define what "done" looks like upfront (a test that passes, a curl that returns 200, a visible UI behavior). Loop on that signal, not on intuition.

Spirit borrowed from Karpathy's observations on LLM coding pitfalls.

Architecture

  • server/local_viewer.js — single Node server. Project CRUD, pty spawn for each project's owning agent (cwd = projects/<id>/), canvas file watcher, Socket.IO push to the browser. Routes: /projects (list / create), /projects/:id (bundle), /projects/:id/activate, /projects/:id/positions, /projects/:id/group-frames/..., /projects/:id/nodes/.... Socket events: canvas-state, canvas-positions, title, pending-generations, pty:spawned / pty:output / pty:exit / pty:error.
  • server/cli/*.js — synchronous CLI wrappers (image, video, voice, split, switch_project, reel_stitch). Each prints one { ok, ... } JSON line on stdout; non-zero exit with { ok: false, klass, message } on failure. Shared arg parser + emit helpers in server/cli/_cli.js.
  • server/pai_*.js — PAI media API clients imported by the CLIs:
    • Shared HTTP: pai_client.js (auth, retry policy, classified errors, callGenerate / callSubmit / pollStatus).
    • Image: pai_image_client.js.
    • Image Pro: pai_image_pro_client.js.
    • Video: pai_video_client.js (upstream payload forwarded byte-for-byte; async submit + poll).
    • Voice: pai_voice_client.js (PAI raw tts, body_base64-decoded).
    • Asset uploads: pai_assets_client.js (video-generation-assets raw; chip-UX cache + event-emitter surface — exports paiAssetEvents, snapshotAssetStates, seedAssetCache, uploadReferenceUrl, preuploadReferenceUrl, preuploadCanvasUrl, uploadReferences).
    • local_mirror.js handles the project-side I/O (write bytes, build viewer URLs, resolve refs to data URIs).
  • web/src/ — React + Vite + React Flow + Socket.IO client.
  • skills/* — local skills. ./scripts/setup --agent codex validates the Codex CLI; Codex-owned projects get project-local symlinks under .agents/skills/. Skill-authoring rules live at skills/CLAUDE.md until that subtree gets a Codex-named guide too.
  • agent-templates/PROJECT_AGENT.md — canonical per-project agent operating manual. server/services/projects.js copies it into projects/<id>/PROJECT_AGENT.md at project create time, alongside provider-specific wrapper files.
  • projects/<id>/ — runtime project data. Gitignored. Created via POST /projects or by local_viewer.js's bootstrap on first run. Each contains workflow.json, meta.json, assets/{images,videos,audios,notes,.tmp}/, canvas_positions.json, PROJECT_AGENT.md, and provider-specific files (CLAUDE.md / .claude/ for Claude, AGENTS.md / .agents/ for Codex).

When adding a new media CLI

  1. Add a new pai_<x>_client.js wrapping callGenerate({ model: "<pai-raw-model>", payload, ... }) (sync) or callSubmit + pollStatus (async). Decode the upstream model's response shape and return { bytes, mime, model, durationSeconds, costUsd } so the CLI is decode-agnostic. See pai_image_client.js for the sync template, pai_video_client.js for async.
  2. Add server/cli/generate_<x>.js. Mirror generate_image.js's shape: import the new pai_<x>_client.js, plus local_mirror.js (writeBytesToTmp for in-memory bytes or streamUrlToTmp for a remote URL streamed to disk, plus viewerUrlForLocalPath and buildProviderRefs), _cli.js, _mutate_helper.js; parse args; call the client; stage the output in assets/.tmp/; hand the absolute path to postNodeAddBatch({ ..., tmpPath }) (or postMutation({ op: "addBatch", payload: { nodes: [{ ..., tmp_path }] } }) for multi-node flows); compute the final URL/local_path from the assigned node id + extension; clean up the temp file if the mutation failed or was skipped; print one JSON line including canvas_mutation. On failure print { ok: false, klass, message } and exit non-zero.
  3. Add the model entry to server/model_registry.js and look up getDefault(kind).id in the CLI rather than hardcoding the string. Set hidden: true if the model is internal (not user-facing as a canvas card, e.g. the asset-upload row).
  4. Add a row to the "Media CLIs" table in agent-templates/PROJECT_AGENT.md (and update the Failure-handling table if the CLI surfaces a new class). Existing projects need to re-copy the template to pick up the change — see ### Updating the agent template across existing projects below.
  5. Add a skill skills/<x>-compose/SKILL.md per skills/CLAUDE.md rules. The recipe should pass --ref-source-id (byte refs) and --source-node-id (authorship edge) flags rather than asking the agent to write the node itself.
  6. Add a row to the Skills-routing table at the top of agent-templates/PROJECT_AGENT.md.

When adding a new node type

  1. Update web/src/types/canvas.ts (renderer source of truth). Add a React component to web/src/pages/CanvasPage/nodes.tsx and a NODE_SIZES entry in web/src/pages/CanvasPage/nodeData.ts.
  2. Mirror the type into server/canvas_schema.js: add the data-validator (#<type>Data), the node-validator (#<type>Node), add it to #canvasNode.oneOf, and add a NODE_ID_PREFIX entry + dataValidatorIdByType entry in server/canvas_mutator.js.
  3. Run npm test in server/ — the real <project>/workflow.json validates against doc schema test catches drift.
  4. Update the "Node grammar (what to put in payloads)" section in agent-templates/PROJECT_AGENT.md. If a media CLI emits this type, update the relevant <x>-compose skill recipe.

When changing the agent template

agent-templates/PROJECT_AGENT.md is the canonical source. Keep it lean — push per-tool recipes and reference detail into the relevant skill; this file is the index. Update the Skills-routing table at the top whenever you add or remove a skill. Existing projects keep their copy until manually re-synced.

When changing this file

This file is the maintainer guide. Architecture overview, contributor recipes, debugging notes — keep it focused on the dev experience. Per-agent operating instructions belong in agent-templates/, not here.

Debugging

  • Viewer / spawn / pty: scripts/start.sh runs the viewer; scripts/stop.sh tears it down. The viewer logs to its tmux pane.
  • Per-project sessions: Claude JSONLs live at ~/.claude/projects/<encoded-cwd>/ (encoding maps /, _, . to -); Codex JSONLs live under ~/.codex/sessions/YYYY/MM/DD/. The viewer persists discovered session ids into meta.agent_session_id and asks the provider to resume on refresh.
  • CLI failures: every CLI prints { ok: false, klass, message }. Replay with the same flags to reproduce.
  • Browser ↔ viewer: DevTools → Network → WS frames. Canvas updates fan out as canvas-state (after every mutation); sidecar drag positions as canvas-positions; in-flight generation placeholders as pending-generations; title changes as title. The Home grid does NOT subscribe — it re-fetches on mount.
  • Mutator audit: projects/<id>/mutations.jsonl is an append-only log of every applied mutation (ts, request_id, op, payload, reply). Useful for "who added this node and when".
  • Drive a per-project PTY from outside (skill smoke tests, agent-to-agent prompts): node server/cli/_dev_send_to_pty.mjs --project <id> --text "…" --press-enter --wait-for <regex>. Attaches to the existing pty via Socket.IO, emits keystrokes, captures pty:output, exits on regex match or timeout. The file header explains the two-connection submit pattern (a single connection can drop the trailing \r — keep the bug in mind if writing similar tooling).