Drop the live generator subagent; fix the artifact decoy that broke accept

The first real Claude Code Live run failed, and the subagent was not the cause.

Root cause: progressive publication stages each revision as
`.impeccable/live/artifacts/<id>-r<n>.<source-ext>`, nothing ever deleted them,
and findSessionFile's walker skipped only node_modules/.git/dist/build. It
searches src, app, pages, ... then `.`; a project whose source is not under one of
those (this repo's own site lives in site/pages/) falls through to the `.` walk,
where dot-directories sort before letters. So accept found the artifact instead of
the real file. Two outcomes, both reproduced: where isGeneratedFile returns true
it declines with mode: 'fallback' (what the run hit, after which the agent
hand-carbonized several hundred lines across three stylesheets, including
unrequested drive-by edits); where it returns false, accept writes the variant
into the throwaway artifact and reports handled: true while real source never
changes.

The E2E suite could not have caught this. Every fixture puts source under `src/`,
which is searched before the `.` walk can reach `.impeccable`. Five framework
fixtures and three progressive scenarios pass because of fixture layout, not
because the path works. I read that as evidence and shouldn't have.

- Never search `.impeccable`: it is Impeccable's own state, never project source.
- Retire a session's staged artifacts on accept/discard, so they cannot outlive
  the session and become a decoy for anything else that walks the tree.
- Regression tests use a site/pages layout with artifacts present. All three fail
  against the previous code.

Generator subagent removed, on both harnesses:
The parent must hand-compress the design system into the handoff, and compression
is lossy. Measured on the real run: a 6,826-char handoff carrying exactly one
token reference, after the parent had itself read kinpaku-tokens.css. The subagent
then spent 3 of its first 9 turns hunting DESIGN.md, gave up, and emitted 0
var(--token) uses and 22 raw oklch literals — violating its own spec's "Never
invent raw colors when tokens exist" — including a 1:1 gold-on-gold contrast bug.
Isolation is not a benefit here; knowing the design system is the job. Generation
stays in the main thread, which already holds the tokens and writes them from the
first byte, so carbonize is a move rather than a translation.

Copy edits keep their subagent: applying a known set of ops to a named file is
self-contained, so an isolated context costs nothing. That is the line.

Progressive delivery stays for Codex and Claude Code, main-thread driven. Claude
Code keeps the full benefit because its poll is a background task. Codex's poll
blocks the foreground, so with no subagent the user sees variant 1 early via HMR
but cannot accept it until the trio finishes; that is the cost of the
simplification and it is worth naming.

Prepared with AI assistance under maintainer direction.

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
Paul Bakaus
2026-07-17 18:56:23 -07:00
co-authored by Claude
parent 1b194d9751
commit c7b67b3832
6 changed files with 129 additions and 75 deletions
+3 -3
View File
@@ -17,16 +17,16 @@ Execute in order. No step skipped, no step reordered.
3. Poll loop with the default long timeout (600000 ms). Run `live-poll.mjs` again immediately after every event or `--reply`; Codex runs this one-shot poll in the foreground. Never pass a short `--timeout=`.
The global bar **Impeccable mark** dims and shows a pulsing amber dot when no agent is long-polling `/poll`. Hover the mark for the hint; restart `live-poll.mjs` to reconnect.
4. On `generate`: reuse `event.scaffold` when present; read the screenshot if present; load the action's reference; deliver variants using the harness policy below; `--reply done`; poll again. In Codex and Claude Code, delegate the complete event to `impeccable_live_generator` / `impeccable-live-generator` and resume polling immediately; the generator owns publication and the reply.
4. On `generate`: reuse `event.scaffold` when present; read the screenshot if present; load the action's reference; deliver variants using the delivery policy below; `--reply done`; poll again. Generate in this thread. You already hold the project's tokens, conventions, and file layout; that context is the job, not overhead.
5. On `steer`: read the message and `pageUrl`; do the work (page edits, navigation help, or a short reply in the `--reply` message); `--reply steer_done`; poll again. No pickup ack. The Steer bar unlocks when `steer_done` arrives over SSE.
6. On `accept` / `discard`: the poll script runs `live-accept.mjs`, acknowledges the delivered event, and prints `_completionAck`. Plain accepts/discards are terminal immediately. Carbonize accepts remain recoverable until the foreground task runs `live-complete.mjs --id EVENT_ID`; finish that cleanup before polling again.
7. If interrupted, run `live-status.mjs` or `live-resume.mjs` before guessing. The durable journal replays unacknowledged work after helper restart.
8. On `exit`: run the cleanup at the bottom.
Harness policy:
- **Claude Code**: run the poll as a **background task** (no short timeout). The harness notifies you when it completes, so the main conversation stays free. Do not block the shell. When `generate` arrives and native subagents are available, delegate the complete event to `impeccable-live-generator` with a compact handoff, then start the next poll immediately while that subagent publishes and replies; if unavailable, generate inline with the same contract. Do not paste this full reference into the handoff. The subagent's context is separate from yours, which is the point: a long session's screenshots, references, and variant CSS stay out of the main conversation. Handle Steer, Accept/Discard, manual Apply, carbonize, and Exit in the main conversation.
- **Claude Code**: run the poll as a **background task** (no short timeout). The harness notifies you when it completes, so the main conversation stays free while you generate and publish in it. Do not block the shell.
- **Cursor**: run **one-shot** poll in a **background terminal** with notify on `"type":"(steer|generate|accept|discard|exit)"`. After each event the poll exits; handle it, `--reply`, then start `live-poll.mjs` again. Do **not** use `--stream` on Cursor: incremental stdout notify is slower in practice than exit-based notify (~5s vs sub-second in testing).
- **Codex**: run the default one-shot poll in a **yielded foreground exec session**. Do not suffix it with `&`, use `--stream`, or leave Live without an active foreground poll. When `generate` arrives, delegate to the low-effort `impeccable_live_generator` agent with a compact handoff, then immediately start the next foreground poll while that agent publishes and replies. Do not paste this full reference into the handoff. Handle Steer, Accept/Discard, manual Apply, carbonize, and Exit in the main task; after each handler/reply, restart the foreground poll.
- **Codex**: run the default one-shot poll in a **yielded foreground exec session**. Do not suffix it with `&`, use `--stream`, or leave Live without an active foreground poll. Handle every event in the main task; after each handler/reply, restart the foreground poll.
- **Other harnesses**: one-shot foreground unless you know stdout reliably returns to this session when a shell exits.
Generation delivery policy: