Browser-drive verification — the kit’s rollout draft

Draft adaptation. The pattern itself is specified at the anchor root: lossless-monorepo/context-v/blueprints/Browser-Drive-Verification-For-Agent-Sessions.md. This file is the context-vigilance-kit’s angle on it — the kit operationalizes practices across the whole tree, so its concern is rollout and consistency, not re-stating the mechanics.

The pattern, in one paragraph

UI-bearing work gets a browser-drive rung between scripted E2E and the human walk-through: the agent wires Playwright MCP at project scope (claude mcp add -s project playwright -- npx @playwright/mcp@latest, lands in the committed .mcp.json, loads next session), drives the click-path named in the phase plan (navigate → click → type → assert on accessibility snapshots, screenshots only for visual questions), with reads unrestricted and writes gated to each repo’s designated safe target. claude --chrome is the interactive sibling for watch-me-drive sessions. The drive proves buttons work; the human still judges usable.

What the kit adds — rollout machinery (proposed, unbuilt)

1. A core-MCP registry. One canonical file in this kit (proposed: scripts/core-mcps.json) declaring the tree’s foundational servers and their scope guidance — the same role sync-skills-symlinks.sh plays for skills. Candidate registry as of 2026-07-22:
   • playwright — universal, project scope, every UI-bearing repo
   • chroma — universal, already user-scope laptop-wide + project scope at ai-labs
   • surrealdb — conditional, project scope where the canonical layer is in play (augment-it precedent: scripts/mcp-surrealdb.sh)
2. An idempotent wiring script. Proposed scripts/ensure-core-mcps.sh <repo-path>: merges missing core entries into the target repo’s .mcp.json (create-if-absent, never clobber existing entries, never touch non-core entries) and prints what changed. Run as an opening-habit sibling to the skills sync.
3. The traveling CLAUDE.md block. The canonical instruction block lives in the anchor root CLAUDE.md; every Lossless-owned repo’s CLAUDE.md carries it verbatim. When the block changes, the kit is where the “re-propagate” sweep should live (proposed: scripts/sync-claude-md-blocks.sh, marker-comment-delimited so the sweep is mechanical).
4. Corpus visibility. Because this kit’s ingest scripts feed the Chroma collections, drive scripts and their outcomes (named in plans + changelogs per the pattern) become searchable prior art — “has this surface been browser-driven before” becomes a search-lossless-corpus query, for free.

Open questions (draft status is real)

• Does the registry belong here or in the anchor root’s context-v/? (Leaning here — the kit is the only repo whose job is cross-tree operationalization; the root blueprint stays the human-readable spec.)
• Marker-comment convention for the traveling block (<!-- lossless:core-mcps:start/end -->?) — needed before any sync script can be trusted not to clobber hand-edits.
• Should ensure-core-mcps.sh also verify npx @playwright/mcp is runnable (node present, network) or stay pure-config?

Related

• Anchor-root blueprint (the spec of record): lossless-monorepo/context-v/blueprints/Browser-Drive-Verification-For-Agent-Sessions.md
• ai-labs/augment-it/context-v/loops/Loop-through-Spec-Write-Plans-Implement-Test-Changelog-Commit.md — the loop whose ladder gains the rung
• This kit’s scripts/ingest-all.sh + the four Chroma collections — the corpus that makes drive history searchable
• context-v/skills/sync-skills-symlinks.sh at the anchor root — the opening-habit precedent the wiring script copies