Skip to content

Composer dressing verdict + run surface lab - #9

Draft
arielconti10 wants to merge 5 commits into
mainfrom
ariel/cloud-agent-1783295266127-g5fci
Draft

Composer dressing verdict + run surface lab#9
arielconti10 wants to merge 5 commits into
mainfrom
ariel/cloud-agent-1783295266127-g5fci

Conversation

@arielconti10

Copy link
Copy Markdown
Owner

What

Two playground labs that close the composer question and open the run surface:

Composer dressing lab (/playground/composer-dressing)

Run surface lab (/playground/run-surface)

  • One nine-step run under three autonomy postures ("Asks first" / "Edits freely" / "On its own"), with the posture rendered as an explicit checkpoint policy per action class (reads / edits / commands / boundaries / external).
  • Demonstrates the conservation law: checkpoints relocate from in-flight gates to the post-run receipt — they never disappear. The receipt's checkpoint ledger counts the relocation (decisions made in-flight vs edits applied without review).
  • Two invariants fire at every posture: boundary violations escalate (the "don't touch deps" line, with two choices whose consequences differ downstream), and external side effects gate (the push).
  • The receipt renders on stop as well as success — partial, honest, "nothing was pushed."

Why

Continues the dispatch-contract research: the composer column of the four-surface framework is now settled and documented, and the in-flight + post-run surfaces get their first teaching artifact. The receipt here is the candidate shape for a future run-receipt registry primitive.

Verification

npm run verify passes (lint, typecheck, format, audit-ui, sync-registry check, registry deps, build, registry build). Playground-only changes — no registry components touched. Browser QA at desktop and 390px mobile widths, all three postures plus the stop path.

Open in Web Open in Cursor 

Dispatch-only chrome: receipt and scope rows mount before the first
send, then collapse into a pinned run-fact strip once the run starts.
The state transition itself is the thing under evaluation - no
shipping product does dispatch-only chrome.
Term chrome rejected; composer keeps meters, pickers, and pointers
only. Receipt relocates to the run's opening transcript entry. B-E
kept as the record of what was tried. PRs #6 and #8 closed against
this verdict.
One nine-step run, three postures. The posture is rendered as an
explicit checkpoint policy (reads/edits/commands/boundaries/external),
and the lab demonstrates the conservation law: checkpoints relocate
from in-flight gates to the post-run receipt, never disappear.

Invariants across all postures: boundary violations escalate (the
'don't touch deps' line), external side effects gate (the push).
The receipt renders on stop as well as success, and its checkpoint
ledger counts what was decided in-flight vs applied without review.

Reuses ActionPreview, RunTrace, and UsageMeter; playground-only,
no registry changes.
@vercel

vercel Bot commented Jul 6, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
agentic-craft Ready Ready Preview, Comment Jul 6, 2026 10:40am

Request Review

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant