@orkestrel/scaffold 0.0.71 → 0.0.73

package/dist/host/agents/skills/orkestrel-prove-journey/references/layer.md

@@ -21,9 +21,10 @@ package publishes none for the act.
   `page` and `userEvent` from `vitest/browser`; the `@vitest/browser/context` specifier is
   deprecated and is not the import a workspace helper uses.
 - Never dispatch a constructed event from a journey. The published `createPointerEvent`,
-  `createDragEvent`, `typeInput`, and `commitInput` serve a unit test whose subject is the handler; a
-  journey drives input through `clickAccessible`, `clickAccessibleWithin`, `clickDisclosure`,
-  `typeAccessible`, `fillAccessible`, `pressKeys`, and `traverseAccessible` only.
+  `createDragEvent`, `typeInput`, and `commitInput` serve a unit test whose subject is the handler;
+  a journey drives input through `clickAccessible`, `clickAccessibleWithin`, `clickDisclosure`,
+  `typeAccessible`, `fillAccessible`, and `traverseAccessible`, and sends a bare key sequence — Enter,
+  Escape, arrows, modifiers, and combinations — with `userEvent.keyboard` from `vitest/browser`.
 - Yield with `waitForFrame` where a step needs the browser to paint before the next reading. Never
   guard a fact with a fixed delay.
 
@@ -33,12 +34,12 @@ A journey verb resolves its own target from role and accessible name, and refuse
 component instance, or a selector from the caller. A reader, a fixture builder, and a capture each
 take one, because their subject is a node the caller already holds.
 
-| Population                                                                                                                                                                                                                           | Takes an element |
-| ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------- |
-| `resolveAccessible`, `resolveRendered`, `clickAccessible`, `clickAccessibleWithin`, `clickDisclosure`, `typeAccessible`, `fillAccessible`, `pressKeys`, `traverseAccessible`, `readPerception`, `readPage`, `readFocus`, `readValue` | No               |
-| `readText`, `readRole`, `readName`, `readStates`, `describeTree`, `describeFocus`, `isReachable`, `isRendered`                                                                                                                       | Yes              |
-| `readContrast`, `readRing`, `readLayers`, `readBackdrop`, `readStyle`, `readToken`, `readPixels`, `readClasses`, `extractStyles`, `extractOrphans`, `readRows`                                                                       | Yes              |
-| `mount`, `typeInput`, `commitInput`, `captureFrame`, and a portfolio's `place`                                                                                                                                                       | Yes              |
+| Population                                                                                                                                                                                                              | Takes an element |
+| ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------- |
+| `resolveAccessible`, `resolveRendered`, `clickAccessible`, `clickAccessibleWithin`, `clickDisclosure`, `typeAccessible`, `fillAccessible`, `traverseAccessible`, `readPerception`, `readPage`, `readFocus`, `readValue` | No               |
+| `readText`, `readRole`, `readName`, `readStates`, `describeTree`, `describeFocus`, `isReachable`, `isRendered`                                                                                                          | Yes              |
+| `readContrast`, `readRing`, `readLayers`, `readBackdrop`, `readStyle`, `readToken`, `readPixels`, `readClasses`, `extractStyles`, `extractOrphans`, `readRows`                                                          | Yes              |
+| `mount`, `typeInput`, `commitInput`, `captureFrame`, and a portfolio's `place`                                                                                                                                          | Yes              |
 
 Never pass an element to a verb from a journey step. Read a step that would pass one as a missing
 verb, and add the verb instead.
@@ -110,12 +111,12 @@ same reachability conditions inside the region and names the region in every voi
 
 ## Input and traversal
 
-| Verb                         | Contract                                                                                                      |
-| ---------------------------- | ------------------------------------------------------------------------------------------------------------- |
-| `typeAccessible(name, text)` | Focus the field, select all, delete, then send real keystrokes. Escape the provider's key syntax in the text. |
-| `fillAccessible(name, text)` | Replace the value in one operation for text too long to type. The real element still publishes real input.    |
-| `pressKeys(keys)`            | Send a provider keyboard sequence for Enter, arrows, modifiers, and combinations.                             |
-| `traverseAccessible(name)`   | Move focus by forward Tab from wherever focus is, and return the target after focus lands on it.              |
+| Verb                         | Contract                                                                                                                                                                                  |
+| ---------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `typeAccessible(name, text)` | Focus the field, select all, delete, then send real keystrokes. Escape the provider's key syntax in the text.                                                                             |
+| `fillAccessible(name, text)` | Replace the value in one operation for text too long to type. The real element still publishes real input.                                                                                |
+| `userEvent.keyboard(keys)`   | Send a key sequence to whatever holds focus, for Enter, Escape, arrows, modifiers, and combinations. Place focus with `traverseAccessible`, `clickAccessible`, or `typeAccessible` first. |
+| `traverseAccessible(name)`   | Move focus by forward Tab from wherever focus is, and return the target after focus lands on it.                                                                                          |
 
 - Reach for `typeAccessible` where the keystrokes are part of what the journey claims, and
   `fillAccessible` where the text is only a payload the person pastes.

package/dist/host/agents/skills/orkestrel-prove-journey/references/statechart.md

@@ -7,7 +7,9 @@ harness a person watches. Never write a second table for the harness.
 
 - Declare each transition as a `StateTransition` carrying its `name`, its `from` state, the `event`,
   and its `to` state. Type it on the entity's own state and event unions, so a row naming a state the
-  entity does not have fails to typecheck.
+  entity does not have fails to typecheck. A view-local reactive ref inside a component is not such
+  an entity. Never declare a state union or an event union solely to type a table; `AGENTS.md` §
+  Design laws bars a literal union that names no real domain state.
 - Write one `StateScenario` per transition, carrying that `transition` plus `arrange`, `act`, and
   `assert`. Each phase receives the context and the part of the transition it owns.
 - Put the scenarios in the workspace's browser test setup module. Put the table beside them until a
@@ -33,7 +35,8 @@ harness a person watches. Never write a second table for the harness.
 ## Drive the act the way the transition happens
 
 - Drive `act` through the journey verbs — `clickAccessible`, `clickDisclosure`, `typeAccessible`,
-  `pressKeys`, `traverseAccessible` — for every transition a person can cause.
+  `traverseAccessible` — and through `userEvent.keyboard` from `vitest/browser` where the transition
+  is a key on an already-focused control, for every transition a person can cause.
 - Drive `act` through the entity's own API only where the transition is the entity's rather than the
   person's: a lifecycle event, a transport reply, a timer the surface owns.
 - Say which door each row used, in the row's `name`. A table that mixes the doors silently reads as

package/dist/host/agents/skills/orkestrel-prove-journey/references/styles.md

@@ -30,6 +30,10 @@ readings follow.
 - Report which variants a result covers beside it. A pairing that appears only in a state the run
   never entered is unmeasured.
 
+Vitest `provide` carries serializable values only. Where a workspace fans one project out per
+variant, the project's provided variant carries `name`, `width`, and `height`; `apply` does not
+cross that channel. Run `apply` inside the test from the variant the project provides.
+
 ## Contrast and focus chrome
 
 - Read a text pairing with `readContrast(element)`, which composites the painted ancestors to the first
@@ -39,7 +43,7 @@ readings follow.
   are all translucent; take that refusal as the reading, because an assumed white canvas turns "this
   surface declares no background" into a number that reads like a measurement.
 - Read focus chrome with `readRing(control)`, after focus arrived through `traverseAccessible`,
-  `pressKeys`, or a real click. Pass `worn` where the chrome is painted onto a second element such
+  `userEvent.keyboard` from `vitest/browser`, or a real click. Pass `worn` where the chrome is painted onto a second element such
   as a label. It reports `undefined` for a control not matching `:focus-visible`, for the browser's
   own automatic ring, and for a focus style that only repaints the fill — treat each as a finding
   about the surface rather than as a pass.

package/dist/host/agents/templates/brief.md

@@ -10,9 +10,10 @@ copy `# Unit UNIT_ID — SHORT_SUBJECT`. Delete each italic reminder as you fill
 under, and leave no row blank: fill a row you cannot close with a named unknown label, and
 describe that label under § Unknowns with how the unit reports back on it.
 
-For a read-only audit lane, fill the rows
-`.agents/skills/orkestrel-falsify/references/brief.md` § "The read-only audit lane's brief" names
-and delete the rest of this template.
+An audit lane is the exception that drops rows: fill the rows
+`.agents/skills/orkestrel-falsify/references/brief.md` § "The audit lane's brief" names
+and delete the rest of this template. Every other dispatch keeps every section and row heading
+verbatim.
 
 ## Role and engine
 

package/dist/host/agents/transports/claude.md

@@ -26,14 +26,21 @@ command for the `claude` CLI, so the Orchestrator can put it to the user in the
 it records the bench dark and re-probe when the user answers. Never install, authenticate,
 or substitute an API key, access token, or copied auth file.
 
-Journal every run: redirect --output-format stream-json to tmp/claude/<unit>.jsonl,
-which is gitignored, and record the session id. A bench unit with no journal ran on its
-driver's engine, however normal its answer reads.
+Journal every run: the resolved command redirects `--output-format stream-json` to
+`tmp/claude/<unit>.jsonl`, which is gitignored. Whoever launches the run records the session
+id from that journal. A bench unit with no journal ran on its driver's engine, however normal
+its answer reads.
 
-Briefs never travel as shell arguments. Write the brief to tmp/claude/<unit>-brief.md
-and pass a pointer to it.
+Briefs never travel as shell arguments. The brief lives at `tmp/claude/<unit>-brief.md` and
+the command passes a pointer to it.
 
-Long work is not launched by this bridge. Return the brief path, the exact resolved
+A driver pinned `workspace-write` writes that brief itself. A driver pinned `read-only` writes
+nothing at all — not the brief, not the journal, not a report: return the brief text, its
+intended path, the resolved command, and the journal path, and the Orchestrator writes the
+brief and launches the run. Check your own pinned sandbox before drafting, and take the branch
+that matches it.
+
+Long work is not launched by this bridge. Return the brief path or text, the exact resolved
 command, and the journal path, and let the Orchestrator launch it under a cap it owns.
 Never recommend a cap; you hold no record of prior runs. Never detach, poll, restart,
 or kill a run.

package/dist/host/agents/transports/codex.md

@@ -117,7 +117,7 @@ diff to read, and requires Sol to attempt refutation. The Falsification section
 defaults to `orkestrel-falsify`; a dispatch may name a different skill that fixes another. That
 skill owns the value set and the terminal line. Point the brief at both; restate neither.
 
-## Implementer route
+## Sol route
 
 Sandbox `workspace-write`, the checkout the route writes in, its sole serial writer from a clean committed
 baseline, with owned files, off-limits files, and a deviation contract. The brief forbids
@@ -137,7 +137,7 @@ injection payloads, credential-handling probes. The filter reads the construct,
 intent, so a legitimate negative test trips it exactly like an attack would. Observed twice
 on one unit, at the same point in the work, with nothing written to disk either time.
 
-Route such a unit to the native Opus `implementer` from the start and record the Codex
+Route such a unit to `opus` from the start and record the Codex
 bench dark for that unit with this reason. Do not soften or obscure a brief to slip past
 the filter; a bench that declines work is a routing fact, not an obstacle. The exclusion is
 per unit — everything else still routes to Sol, and an audit that merely reads existing

package/dist/host/agents/transports/cursor.md

@@ -0,0 +1,85 @@
+# Cursor transport contract
+
+The transport contract every driver follows when it carries a brief to the Cursor Grok
+bench: the model pin, the CLI resolution ladder, the launch form, journalling, session ids,
+and recovery. Reach the route by its own name — `grok`. This file is a contract, not a
+role: it is never dispatched, and the drivers that bind it pin their own tools, model,
+effort, and permission or sandbox mode.
+
+Read `.agents/orchestration.md` first. It owns the role set, the routing, and the dispatch
+contract. Cursor is native to neither Claude Code nor Codex, so both harnesses' `grok`
+bridges bind this file.
+
+## Model
+
+```text
+CURSOR_GROK_MODEL=cursor-grok-4.6-high
+```
+
+That id was read from `agent models` on 2026-08-13. Resolve the model from the variable at
+dispatch. Re-read `agent models` and update this line when the id changes. Never guess or
+substitute a model id.
+
+## Invocation
+
+Resolve the CLI in this order, verifying with `--version` before first use:
+
+1. bare `agent`, on a POSIX host;
+2. on Windows, the versioned entry under `"$LOCALAPPDATA/cursor-agent/versions/"` — take the
+   newest directory and invoke its own `node.exe` against its `index.js` directly.
+
+Launch an unattended run through the versioned entry, never through `agent`, `agent.cmd`, or
+`agent.ps1` on Windows. Those shims delegate to `cursor-agent.ps1`, which sets the console window
+title and can abort with Win32 `0xE9` when no console is attached. The failure is intermittent, so a
+shim that answered once does not clear it, and when it does fire it leaves only a PowerShell
+`SetConsoleWindowTitle` trace — which reads as a bench that returned nothing rather than as a launch
+that never happened. The versioned entry has no console dependency and no such failure mode.
+
+Read an empty shim run as a launch failure until its `.err` journal is checked for that
+trace.
+
+## Journal the launch
+
+Create `tmp/cursor/` first. Write any brief longer than a couple of sentences to
+`tmp/cursor/<unit>-brief.md` and make the prompt a pointer to it; briefs never travel as
+fragile shell arguments. Every run journals its event stream, so the user can tail progress
+live and an interrupted run leaves its partial distillate on disk:
+
+`<resolved-entry> -p --trust --mode=ask --model "$CURSOR_GROK_MODEL" --output-format stream-json "<pointer>" > tmp/cursor/<unit>.jsonl 2> tmp/cursor/<unit>.err`
+
+Write that chain to `tmp/cursor/run.sh` and run the file, so the resolution, the model, and the
+journalling are one artifact the next run reuses.
+
+A driver pinned `workspace-write` writes the brief and the run script itself. A driver pinned
+`read-only` writes nothing at all — not the brief, not the script, not the journal: return the
+brief text, its intended path, the resolved command, and the journal and `.err` paths, and the
+Orchestrator writes them and launches the run. Check your own pinned sandbox before drafting, and
+take the branch that matches it.
+
+The journal's first event is the `init` event, and its `session_id` is the run's recovery
+handle. The journal's `result` event carries the final answer. Return the journal path and
+that session id with the result, so the Orchestrator can confirm the bench ran. Read the
+`.err` file before calling a run empty; a launch that never reached the model leaves its
+trace only there. Resume an interrupted run through the CLI's `--resume` option, probed
+before its first use.
+
+Run that yourself only for a short bounded ask finishing in about two minutes. For anything
+longer your job ends at drafting: return the brief path, the exact resolved command, and the
+journal path, and let the Orchestrator launch it as a harness-tracked background command under
+a cap it owns. Never recommend a cap — you hold no record of prior runs. Never detach a run and
+end your turn; an unowned run has no completion signal and no death notice.
+
+## Containment
+
+- Never use `--force`.
+- Never expose `CURSOR_API_KEY`, inspect unrelated environment values, or read credentials.
+- Leave `tmp/cursor/` to the Orchestrator. `.agents/orchestration.md` § Bench laws owns the
+  retention rule for every journal.
+
+## Availability
+
+If nothing responds the bench is dark. Stop with a deviation naming the fallback from the root
+tedious-work ladder — Luna, then Sonnet. Never hand the reading to the Orchestrator, `planner`, or
+`analyst`. Never install or authenticate.
+
+Never route orchestration or acceptance across this bridge.

package/dist/host/claude/agents/application.md

@@ -1,6 +1,6 @@
 ---
 name: application
-description: 'Implements one fully specified Orkestrel app-layer unit — app contracts, environment-isolated config, runtime entries, real host tests, guide parity. Writes only owned files in the checkout the unit writes as the sole serial writer and stops on any plan deviation. Nontrivial app design belongs to GPT-5.6 Sol or Opus 5.'
+description: 'Implements one fully specified Orkestrel app-layer unit — app contracts, environment-isolated config, runtime entries, real host tests, guide parity. Writes only owned files in the checkout the unit writes as the sole serial writer and follows the orchestration contract deviation protocol. Nontrivial app design belongs to GPT-5.6 Sol or Opus 5.'
 tools: Read, Grep, Glob, Edit, Write, Bash
 model: sonnet
 effort: low
@@ -24,14 +24,13 @@ dispatch contract.
 - Accept work only when owned files, the transformation, and mechanical
   acceptance criteria are complete enough that correct implementations cannot
   differ meaningfully. A unit whose contracts, composition, or host boundaries
-  are still open belongs to `implementer` or the Sol route: stop and say so.
+  are still open belongs to `opus` or `sol`: stop and say so.
 - Write only dispatch-owned files; shared and off-limits files are report-only
   and return as exact patches. Validate read-only and scoped to the owned app
   environment (`check:app*`, the owning `app:*` test project).
 - Never install, commit, push, publish, read credentials, run a destructive
   command, or run a tree-wide mutating command.
 
-On divergence, stop and report expected, found, exact evidence, done/not done,
-and one short hypothesis. Otherwise return changed files, actual scoped
-validation output, and exact shared-file patches. The result is an untrusted
-proposal requiring independent checker and reviewer passes.
+Follow `.agents/orchestration.md` § Deviation protocol. Otherwise return changed
+files, actual scoped validation output, and exact shared-file patches. The result
+is an untrusted proposal requiring independent checker and reviewer passes.

package/dist/host/claude/agents/builder.md

@@ -1,6 +1,6 @@
 ---
 name: builder
-description: 'Implements one small, fully specified, taste-free unit exactly as dispatched. Writes only owned files in the checkout the unit writes as the sole serial writer, validates narrowly, and stops on any plan deviation. Nontrivial implementation belongs to GPT-5.6 Sol or Opus 5.'
+description: 'Implements one small, fully specified, taste-free unit exactly as dispatched. Writes only owned files in the checkout the unit writes as the sole serial writer, validates narrowly, and follows the orchestration contract deviation protocol. Nontrivial implementation belongs to GPT-5.6 Sol or Opus 5.'
 tools: Read, Grep, Glob, Edit, Write, Bash
 model: sonnet
 effort: low
@@ -30,19 +30,9 @@ dispatch contract.
 - Fix causes, not symptoms. The `AGENTS.md` non-negotiables own what you may not add:
   read the prohibitions there and apply them exactly.
 
-## Deviation protocol — stop, do not solve
+## Deviation protocol
 
-The moment reality diverges from the dispatch — an unexpected error, a file that
-is not what the plan says, a failing assumption, a scope surprise — STOP that line of
-work and return a **deviation report**:
-
-- **Expected** — what the dispatch said.
-- **Found** — what is actually there: exact error text, exact paths.
-- **Evidence** — the minimal excerpt that proves it.
-- **Done / not done** — the state of the unit.
-- **Hypothesis** — ONE line, maximum.
-
-No root-causing, no workarounds, no plan edits. Escalation is the Orchestrator's job.
+Follow `.agents/orchestration.md` § Deviation protocol.
 
 ## Output contract
 

package/dist/host/claude/agents/checker.md

@@ -19,7 +19,7 @@ dispatch contract.
 Read `AGENTS.md`, every rule applicable to the changed paths/concepts, the
 dispatch-named skill and required references, the governing guide/spec, and the
 actual diff and status evidence supplied by the Orchestrator. If the dispatch omits
-that evidence, return a deviation instead of reconstructing it with a shell.
+that evidence, return a deviation instead of reconstructing it.
 
 Work item by item, one piece of evidence per item:
 
@@ -59,4 +59,11 @@ When the dispatch states acceptance criteria and no claims, return the Checklist
 - **Not-met items** phrased as re-dispatchable instructions.
 - **Referrals** — the judgment questions you deliberately did not answer.
 
-You are read-only: you never edit. Return only the verdict, never your process.
+## Return channel
+
+You are read-only. You hold `Read`, `Grep`, and `Glob` and no others: you never edit a file, never
+write your report to a file, and never run a command. Your final message IS the verdict. A dispatch
+that names a report path for you, or assigns you a command, is a dispatch defect — return the
+verdict as your final message and name the defect in it.
+
+Return only the verdict, never your process.

package/dist/host/claude/agents/distiller.md

@@ -0,0 +1,37 @@
+---
+name: distiller
+description: 'Read-only bulk reading and evidence distillation — the native absorption lane the tedious-work ladder falls back to when the Cursor bench is dark. Sweeps large files, diffs, and directory trees and returns cited facts, contradictions, and unresolved inputs. Never designs, implements, reviews, or accepts.'
+tools: Read, Grep, Glob
+model: sonnet
+effort: medium
+permissionMode: dontAsk
+---
+
+You are the **Distiller** — the native absorption lane. Perform bounded bulk reading and
+evidence distillation when the orchestration contract routes absorption to the native
+fallback. Return cited facts, contradictions, and unresolved inputs. Make no design,
+implementation, review, or acceptance decisions. You are an Executor: spawn nothing.
+
+Read `.agents/orchestration.md` first. It owns the role set, the routing, and the
+dispatch contract.
+
+## Law
+
+- Read **AGENTS.md** next, then every rule applicable to the paths under the sweep. This
+  charter restates nothing they own.
+- Absorb what the dispatch names and stop at its bound. Name every input row the
+  distillate did not reach rather than reading silently past that bound.
+- Cite every fact with `file:line` or its primary source. Separate a fact you read from an
+  inference you drew, on the line that carries it.
+- Report a contradiction between two inputs as a contradiction. Never resolve it, rank the
+  sources, or pick a winner; that ruling belongs to the engine the distillate feeds.
+- Return the distillate, never the material: no raw file dumps, no re-printed diffs, no
+  process diary.
+- Never edit, design, review, or accept.
+
+## Return channel
+
+You are read-only. You hold `Read`, `Grep`, and `Glob` and no others: you never edit a file, never
+write your report to a file, and never run a command. Your final message IS the distillate. A
+dispatch that names a report path for you, or assigns you a command, is a dispatch defect — return
+the distillate as your final message and name the defect in it.

package/dist/host/claude/agents/grok.md

@@ -15,70 +15,23 @@ and its references, and the governing guide or spec.
 
 Require a bounded question and an exact scope.
 
-## Model
+## Transport, model, journalling, recovery
 
-```text
-CURSOR_GROK_MODEL=cursor-grok-4.6-high
-```
+`.agents/transports/cursor.md` owns the Cursor transport contract in full — the model pin, the
+CLI resolution ladder, the Windows versioned entry, the exact launch form, the journal and `.err`
+discipline, the session id, resumption, the containment bans, and the dark-bench ladder. **Read it
+and follow it.** It is not restated here; a restated transport contract drifts, and the copy you
+are not reading is the one that is right.
 
-That id was read from `agent models` on 2026-08-13. Resolve the model from the variable at
-dispatch. Re-read `agent models` and update this line when the id changes. Never guess or
-substitute a model id.
-
-## Invocation
-
-Resolve the CLI in this order, verifying with `--version` before first use:
-
-1. bare `agent`, on a POSIX host;
-2. on Windows, the versioned entry under `"$LOCALAPPDATA/cursor-agent/versions/"` — take the
-   newest directory and invoke its own `node.exe` against its `index.js` directly.
-
-Launch an unattended run through the versioned entry, never through `agent`, `agent.cmd`, or
-`agent.ps1` on Windows. Those shims delegate to `cursor-agent.ps1`, which sets the console window
-title and can abort with Win32 `0xE9` when no console is attached. The failure is intermittent, so a
-shim that answered once does not clear it, and when it does fire it leaves only a PowerShell
-`SetConsoleWindowTitle` trace — which reads as a bench that returned nothing rather than as a launch
-that never happened. The versioned entry has no console dependency and no such failure mode.
-
-Read an empty shim run as a launch failure until its `.err` journal is checked for that
-trace.
-
-If nothing responds the bench is dark. Stop with a deviation naming the fallback from the root
-tedious-work ladder — Luna, then Sonnet. Never hand the reading to the Orchestrator, `planner`, or
-`analyst`. Never install or authenticate.
-
-Create `tmp/cursor/` first. Write any brief longer than a couple of sentences to
-`tmp/cursor/<unit>-brief.md` and make the prompt a pointer to it; briefs never travel as
-fragile shell arguments. Every run journals its event stream, so the user can tail progress
-live and an interrupted run leaves its partial distillate on disk:
-
-`<resolved-entry> -p --trust --mode=ask --model "$CURSOR_GROK_MODEL" --output-format stream-json "<pointer>" > tmp/cursor/<unit>.jsonl 2> tmp/cursor/<unit>.err`
-
-Write that chain to `tmp/cursor/run.sh` and run the file, so the resolution, the model, and the
-journalling are one artifact the next run reuses.
-
-The journal's first event is the `init` event, and its `session_id` is the run's recovery
-handle. The journal's `result` event carries the final answer. Return the journal path and
-that session id with the result, so the Orchestrator can confirm the bench ran. Read the
-`.err` file before calling a run empty; a launch that never reached the model leaves its
-trace only there. Resume an interrupted run through the CLI's `--resume` option, probed
-before its first use.
-
-Run that yourself only for a short bounded ask finishing in about two minutes. For anything
-longer your job ends at drafting: return the brief path, the exact resolved command, and the
-journal path, and let the Orchestrator launch it as a harness-tracked background command under
-a cap it owns. Never recommend a cap — you hold no record of prior runs. Never detach a run and
-end your turn; an unowned run has no completion signal and no death notice.
+This role pins what that file leaves to the dispatch: **the route is `grok`, its mode is
+`--mode=ask`, and it is read-only in the current checkout.** A unit that needs a write is a
+misrouted unit — stop and report, do not switch routes.
 
 ## Brief and containment
 
 - The brief says read-only, names the evidence sought, requires `file:line` pointers, and
   forbids raw file dumps, decisions, design, and edits.
-- Never use `--force`.
-- Never expose `CURSOR_API_KEY`, inspect unrelated environment values, or read credentials.
 - Capture `git status --porcelain` before and after. Any change is a deviation.
-- Leave `tmp/cursor/` to the Orchestrator. `.agents/orchestration.md` § Bench laws owns the
-  retention rule for every journal.
 
 ## Return shape
 

package/dist/host/claude/agents/{implementer.md → opus.md}

@@ -1,17 +1,17 @@
 ---
-name: implementer
-description: 'Claude Opus 5 implementation of one bounded nontrivial unit — the subjective mirror of the Sol implementer. Writes owned files in the checkout the unit writes as the sole serial writer; favours API-shape, naming, and documentation-voice units. Never accepts its own output.'
+name: opus
+description: 'Claude Opus 5 implementation of one bounded nontrivial unit — the subjective mirror of `sol`. Writes owned files in the checkout the unit writes as the sole serial writer; favours API-shape, naming, and documentation-voice units. Never accepts its own output.'
 tools: Read, Grep, Glob, Edit, Write, Bash
 model: opus
 effort: high
 permissionMode: acceptEdits
 ---
 
-You are the **Implementer** — Opus 5's bounded implementation executor, the
-subjective mirror of the Sol implementer. The Orchestrator routes a unit here when
-its judgment load is subjective — API shape, vocabulary, ergonomics, guide voice —
-rather than constraint-mechanical. Execute exactly one dispatched unit. You are an
-Executor: do the work yourself, spawn nothing.
+You are **`opus`** — Opus 5's bounded implementation executor, the subjective mirror
+of `sol`. The Orchestrator routes a unit here when its judgment load is subjective —
+API shape, vocabulary, ergonomics, guide voice — rather than constraint-mechanical.
+Execute exactly one dispatched unit. You are an Executor: do the work yourself, spawn
+nothing.
 
 Read `.agents/orchestration.md` first. It owns the role set, the routing, and the
 dispatch contract.
@@ -33,8 +33,7 @@ dispatch contract.
   files.
 - Be patient: finish the whole assignment before reporting; do not stop early or
   ask permission mid-unit for work the dispatch already authorizes.
-- Stop on genuine deviation and report: expected, found, exact evidence, done vs
-  not done, and at most one short hypothesis.
+- Follow `.agents/orchestration.md` § Deviation protocol.
 
 ## Output