feat: Grok Build first-class harness adapter

[kennymcavoy]

Summary

Grok Build panes become first-class like Claude: composer chrome stripped with an opaque status strip re-surfaced natively, probed dialogs lifted into tappable buttons, a journal adapter, and the Grok slash catalog.

• Chrome (Tier 1) — web/src/lib/harness/grok/: the rounded composer box is located by its bottom status border and stripped; the status run is opaque across models/efforts (never matches Grok 4.6 / (high)). composerReady gates the reply path; the startup screen's bare [stable] chip is matched literally as chrome so a fresh pane accepts its first message. Grok's full-width background paint on blank rows is dropped (no color inference) so the mirror doesn't stripe; glyph rows keep every background.
• Permission cards (prompt-select, family permission) — classified, not layout-pinned: the footer must name the family and count its rows (1/N:select + Tab:next option + Ctrl+o:always-approve + Ctrl+c:cancel), the last row must be the reject, the row above it the one-shot Yes, and every earlier row must prove itself a persistent mode change — those never become buttons. Unclassified rows refuse the whole card. Both live layouts (3-option rm, 4-option file-edit) lift; digits probed per layout (PERMISSION_NOTES.md), including the one-shot-doesn't-persist negative control.
• ask_user_question (family select) — radio cards lift with per-row digits (digit submits directly — probed). The z free-text row is modelled as feedback and locks the buttons whenever the keyboard sits on it; the inner hint's Enter:edit verb is the tell even when the row repaints idle (live-probed: a digit there types into the field). Esc-parked cards send Tab — the footer's own key — before the digit, because a bare digit is silently swallowed in scrollback (probed both ways). Checkbox cards stay raw (a digit submits rather than toggles). Multi-question wizards step through as consecutive cards.
• Plan approval (menu) — footer-named keys only, no digits (ADR 0009); composer gated while the review is up.
• Journal — bridge/journal/grok.ts reads chat_history.jsonl under $GROK_HOME (override COLLIE_GROK_ROOT) through containedRealpath, refusing non-UUID refs.
• Neutral-model additions (called out per your feat: add AGY (Antigravity) harness adapter, interactive prompts, and slash commands #99 note about separate arguments): PromptFeedback.purpose — Grok's z row is "free-text", so the renderer offers no feedback composer and submitPromptFeedback refuses it (the Claude plan-feedback choreography is the wrong recipe); and PromptOption.keyLabel — the badge shows the digit when a recovery key precedes it. Claude behavior unchanged, pinned by the existing claude suites.

The #99 bar

• Exact agent string only: adapterFor stays Object.hasOwn over adapter.agent; grok-build catalog folding lives in canonicalAgent, never in the registry.
• Dated fixture corpus (25 grok--*.txt), notes files with the verified recipes (PERMISSION/ASK/PLAN_NOTES.md), describeAdapterConformance running grok against the claude+omp cohorts and vice versa.
• Live verification: run by me against real Grok Build 1.0.5 panes per the arrangement in feat: add AGY (Antigravity) harness adapter, interactive prompts, and slash commands #99 — every dialog exercised end-to-end through the bridge (buttons pressed via the guarded /keys path, files created/removed on disk as the negative controls), plus stale-signature replays refused with 409.
• No digits synthesized from numbered rows: permission/ask digits are probed specific grammars; the plan menu is driven by the keys its footer names.

Honesty notes

• The six Tier-1 composer files (grok--fresh-idle … grok--user-bubble) are sanitized structure fixtures (plain UTF-8, no ESC); the live 3-SGR-run border shape is pinned separately in grok/markers.test.ts from a dated probe. All dialog captures are byte-faithful format:ansi.
• Deliberately not lifted (each needs its own capture + probe): checkbox asks (digit submits, not toggles), typing a custom z answer from the phone, and Grok's cancel-turn panel.
• Observed variance: after a plan approval Grok sometimes gates the plan's own writes with a permission card and sometimes doesn't — both paths handled.

Suggested CHANGELOG entry

Grok Build gets a first-class harness adapter — strips the composer chrome, re-surfaces the status strip, lifts permission cards, ask_user_question radios/wizards and plan approval into native buttons, and reads the Grok session journal.

Verification

cd web && bun run typecheck clean; full web suite + conformance green; bun test bridge/ green; live pass against real panes as above. Versions untouched at 0.32.0 per the fork-PR rule.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PFH7jQVDVhPGC6FqLetNfd

[AltanS]

Reviewed — same verdict as #124: strong shape, fail-closed classification, and the journal containment is actually stricter than our own adapters (re-running containedRealpath on the cached path is a nice touch). Version files correctly untouched. All gates pass here (root 687, web 2971, both typechecks clean).

Three changes requested before merge:

1. Agent-string match. operator-scope.ts:39 folds grok* → grok for the slash catalog, but harness/registry.ts and bridge/journal/registry.ts match "grok" exactly. If Herdr reports the agent as grok-build on any install, that pane gets the slash palette and nothing else — no chrome strip, no dialogs, no journal. Can you confirm the exact agent string Herdr reports for your Grok Build panes? If it can be grok-build (or varies), widen the harness + journal registries to the same fold the catalog uses.
2. One byte-faithful composer capture. Six of the Tier-1 fixtures are sanitized (no ESC), as you disclosed. markers.test.ts pins the real 3-SGR border, which covers part of the gap, but the chrome strip is regression-guarded against partly synthetic input. One format:ansi capture of a real composer (idle + one with a draft) would close it.
3. scripts/journal-probe.ts — the per-root rework is good but unrelated to the adapter; please split it into its own commit inside this PR so it can carry its own changelog line.

Smaller, take or leave: the free-text entry in prompt-select-block.tsx:63-73 carries five empty-string fields gated only by convention — a discriminated union would turn misuse into a type error.

On our side: Tier-2 means we live-verify the four dialog recipes against a real Grok Build pane before merge. Plan: merge #124 first, then this one after a rebase (six trivial overlaps).

[kennymcavoy]

Review follow-up:

1. Agent string. Confirmed live 2026-08-23: Herdr's kind list is grok, and every Grok Build pane on this host reports agent: "grok" / source: herdr:grok (pane.agent and agent_session.agent). No grok-build variant showed up. Registries stay exact; the grok-build catalog fold in canonicalAgent is unused here. Noted on harness/grok/index.ts.

2. Byte-faithful composer. Replaced grok--fresh-idle.txt and grok--draft-single.txt with format:ansi captures from a sandbox pane (empty idle box, then a testing stuff draft). Darwin temp-dir token is length-preserved. The other four Tier-1 chrome files remain structure fixtures.

3. journal-probe split. fix(journal): probe each root on its own is now its own commit ahead of the adapter.

[AltanS]

Merged — thanks. I did the post-#124 rebase on our side: union of the four both-add files, plus one semantic conflict git missed (main renamed the no-adapter test agent codexAgent → opencodeAgent; grokAgent now builds from that — f1a7dbc). Follow-up 7d8ffb4 makes codex and grok treat each other's corpora as foreign.