Skip to content

Propose browser verification + tear-down phase (counsel consensus) - #35

Draft
walker-tx wants to merge 1 commit into
mainfrom
cursor/browser-verification-plan-2335
Draft

Propose browser verification + tear-down phase (counsel consensus)#35
walker-tx wants to merge 1 commit into
mainfrom
cursor/browser-verification-plan-2335

Conversation

@walker-tx

Copy link
Copy Markdown
Member

Summary

Counsel of three frontier agents (Claude Opus 5, GPT 5.6 Sol, Claude Fable 5) produced independent plans, ran a dispute round, and ratified a joint architecture for extending the guide pipeline with browser verification — including the tear-down requirement so flows stay re-validatable.

Plan only — no Playwright harness or doctrine edits are implemented in this PR. Concrete doctrine diffs still need human approval (I8).

The recommendation (one sentence)

Add an opt-in, post-convergence verify-guide phase that deterministically click-throughs converged setup.md in Playwright/Chromium, records a Verification Ledger (confirm/deny/amend/blocked), captures redacted screenshots, always tears down created resources to a re-runnable baseline, and — only on material denies/amends — re-enters the fact flow by amending the Research Dossier then re-rendering affected setup anchors.

What landed

  • proposals/browser-verification.md — full consensus (architecture, credentials/1Password, tear-down, screenshots, local+CI commands, factory labels/exit codes, doctrine impact, phased delivery, open questions)
  • proposals/README.md — proposals index
  • Pointers from FACTORY.md and README.md

Settled disputes (highlights)

Topic Consensus
Placement After draft/review convergence; opt-in guide:verify (not on every draft)
Walk target setup.md order + Dossier locators → deterministic claim plan
LLM ↔ browser No LLM on authenticated browser in v1; harness-only mutations
Confirmations Ledger-only; factual prose unchanged; asset weave + lock rewrite
Tear-down First-class; write-ahead ledger; LIFO; absence-asserted; exit 4 outranks all
Secrets 1Password; human-once sandbox accounts; separate privileged Actions job
First slice GitHub; acceptance = back-to-back double-run with teardown: clean

Human decisions still needed

See §12 of the proposal (Speakeasy sandbox ownership, ToS sign-off, 1Password bootstrap, constitution I1/I9 edits, catalog surfacing of verify status, etc.).

Suggested next step

Approve Phase 0 spike against the GitHub throwaway org (local mise run verify-guide path) before any doctrine or factory wiring.

Open in Web Open in Cursor 

Capture the Opus / GPT 5.6 Sol / Fable counsel consensus for an opt-in
post-convergence verify-guide phase: deterministic Playwright walk of
setup.md, Verification Ledger, redacted screenshots, mandatory tear-down
to a re-runnable baseline, and Dossier-first re-entry for material
denies/amends. Plan only — no pipeline implementation yet.

Co-authored-by: Walker Lockard <walker-tx@users.noreply.github.com>
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.

2 participants