release: beta.10 wave — complete CLI coverage, agentic combo union, stabilization (integration → main)

.agents/skills/claude-manager/SKILL.md

@@ -9,8 +9,9 @@ description: >
 # Claude Manager
 
 Use this skill when the task involves Claude Code session orchestration rather than package code.
-Claude is the supervisor. OpenHands evaluates. WSL Codex implements harness slices that must be
-visible from Codex Desktop/mobile.
+Claude is the supervisor. WSL Codex implements harness slices that must be visible from Codex
+Desktop/mobile. Evaluation runs on the evaluator lane in `.llm/harness/workflow/lane-policy.md`:
+locally Claude Code + OpenRouter with an open model, and OpenHands for automated cloud runs.
 
 ## Workflow
 
@@ -65,8 +66,10 @@ outcome.
 - Do not use Claude workflows as the default NetScript implementation agent. WSL Codex remains the
   preferred implementation lane because it is daemon-attached, mobile-visible, and cheaper to steer
   slice-by-slice.
-- Keep OpenHands as the evaluator. Claude workflows may prepare evaluator inputs, but they do not
-  replace PLAN-EVAL or IMPL-EVAL.
+- Keep the evaluator a separate session on the evaluator lane from
+  `.llm/harness/workflow/lane-policy.md` — locally Claude Code + OpenRouter with an **open model**, or
+  OpenHands for cloud runs. Claude workflows may prepare evaluator inputs, but they do not replace
+  PLAN-EVAL or IMPL-EVAL, and no session self-certifies.
 - Route every Claude workflow, supervisor, and review session through the canonical lane table in
   `.llm/harness/workflow/lane-policy.md`. Do not reproduce model/effort defaults here or infer a
   paid escalation from workflow prose.
@@ -97,6 +100,7 @@ session should be started.
 
 - [ ] Current branch/worktree is verified.
 - [ ] Harness skill was loaded for harnessed work.
-- [ ] Evaluator surface is OpenHands when evaluating.
+- [ ] Evaluator surface is the lane-policy evaluator route (open model): Claude Code + OpenRouter
+      locally, OpenHands for cloud runs.
 - [ ] Implementation surface is WSL Codex when slice work must be mobile-visible.
 - [ ] Claude remote-control or Codex daemon visibility is proven before claiming phone visibility.

.agents/skills/codex-wsl-remote/SKILL.md

@@ -83,7 +83,7 @@ exit codes: `.llm/tools/agentic/README.md`.
 | **Watch — progress** | Wake on the worktree's next **git** event (commit/ref) — *slice made progress* | `codex-watch.ts --worktree <wt>` (**run inside WSL**) |
 | **Watch — finish** | Wake when the agent's **turn finishes** (rollout `task_complete`) — *agent is idle/done* | `codex-watch.ts --mode turn --thread-id <uuid>` (**run inside WSL**) |
 | **Steer** | Continue/correct an existing thread (never forks a rival) | `codex-resume.ts --thread-id <uuid> --message …` (`--dry-run`) |
-| **Evaluate** | Dispatch an `@openhands-agent` PLAN/IMPL-EVAL; read its verdict | `dispatch-openhands.ts`, `openhands-status.ts` |
+| **Evaluate** | Local PLAN/IMPL-EVAL runs on the open-model evaluator lane (Claude Code + OpenRouter — see `lane-policy.md`); for a **cloud** run, dispatch an `@openhands-agent` and read its verdict | `dispatch-openhands.ts`, `openhands-status.ts` (cloud) |
 | **Merge** | Eval-gated, clean-gated, base-guarded leaf-PR lifecycle | `gh-pr.ts create\|verdict\|merge` |
 
 **The watch distinction is the one that bites you.** `codex-watch` has two signals and they answer

.agents/skills/netscript-harness/SKILL.md

@@ -106,8 +106,9 @@ routing here — defer to that file. The items below are the parts of the contra
   for every implementation tier (B Opus sub-agents, C Workflow-generated slices, D WSL Codex); no
   lane self-certifies. See `workflow/lane-policy.md` for the rule and `workflow/run-loop.md` for the
   step placement.
-- **Evaluator route binding.** Select the opposite-family review route from
-  `workflow/lane-policy.md` and record it in the run. The two-failure eval loop is unchanged.
+- **Evaluator route binding.** Select the evaluator route from `workflow/lane-policy.md` — the
+  open-model evaluator lane for a formal PLAN-EVAL/IMPL-EVAL, or the opposite-family route for
+  ordinary review — and record it in the run. The two-failure eval loop is unchanged.
 - **Tier-D mobile-visibility proof.** A Tier-D (WSL Codex) implementation slice is launched only via
   skills + `.llm/tools/agentic/` (never ad-hoc `wsl.exe`), and only when the run artifacts include
   the WSL worktree path, concrete Codex thread id, daemon-managed `remote-control` proof, and the
@@ -133,13 +134,18 @@ routing here — defer to that file. The items below are the parts of the contra
 - **Self-evaluation** — The evaluator must be a separate session. The generator does not
   self-certify.
 - **Wrong evaluator surface** — the generator session may never evaluate its own output. For a
-  **local-machine run**, PLAN-EVAL/IMPL-EVAL is a **separate local session of the opposite model
-  family** (e.g. a Codex GPT-5.6 session reviews Claude-authored work; a Claude session reviews
-  Codex-authored work), and the **supervisor chooses when to trigger it** — never auto-dispatch a
-  cloud evaluator from a sub-agent. **OpenHands is open-models-only (minimax M3 / Qwen 3.7) and
-  cloud-driven-runs-only; dispatching it with a closed model (Claude/GPT/Gemini) is prohibited (it
-  burns paid OpenRouter credit).** See `.agents/skills/openhands-handoff/SKILL.md` "Routing policy".
-  If no evaluator surface is available, record a blocked launch in `drift.md`.
+  **local-machine run**, PLAN-EVAL/IMPL-EVAL is a **separate local session on Claude Code +
+  OpenRouter** (`claude-openrouter` profile → `claude-print`) running the bound **OPEN-model Qwen
+  evaluation preset** (`qwen/qwen3.7-max`) — an open model is adversarial to both the Claude and
+  Codex families. Minimax M3 remains in the approved open-model set but is not currently bound to a
+  local evaluation preset. **Ordinary (non-formal) review** — the slice review gate, code/PR review — uses a
+  local **opposite-family** session instead (a Codex GPT-5.6 session reviews Claude-authored work; a
+  Claude session reviews Codex-authored work). The **supervisor chooses when to trigger** — never
+  auto-dispatch a cloud evaluator from a sub-agent. **Both evaluator transports are open-models-only;
+  dispatching either with a closed model (Claude/GPT/Gemini) is prohibited
+  (it burns paid OpenRouter credit).** OpenHands remains the default automated **cloud** agent. See
+  `.agents/skills/openhands-handoff/SKILL.md` "Routing policy". If no evaluator surface is available,
+  record a blocked launch in `drift.md`.
 - **Self-certifying a slice** — a green automated gate is not a sign-off. The Tier-A supervisor must
   substantively review the slice before the sign-off commit, for every implementation lane
   (`workflow/lane-policy.md` invariant 2). No lane self-certifies.
@@ -211,11 +217,13 @@ When external docs or examples matter:
 
 ## Evaluator Separation
 
-There are **two** separate-session evaluator passes.
+There are **two** separate-session evaluator passes. Both run on the evaluator lane selected from
+`workflow/lane-policy.md`: locally on **Claude Code + OpenRouter with an OPEN model**, or in the
+cloud on **OpenHands** (also open-models-only). Closed/paid models are prohibited on both.
 
 **PLAN-EVAL** (before implementation):
 
-- Runs in a separate opposite-family session selected from `workflow/lane-policy.md`.
+- Runs in a separate session on the evaluator lane selected from `workflow/lane-policy.md`.
 - Reads `evaluator/plan-protocol.md` + `gates/plan-gate.md`.
 - Reads `research.md`, `plan.md`, and the `## Design` section.
 - Writes `plan-eval.md`.
@@ -224,7 +232,7 @@ There are **two** separate-session evaluator passes.
 
 **IMPL-EVAL** (final pass, after implementation):
 
-- Runs in a separate opposite-family session selected from `workflow/lane-policy.md`.
+- Runs in a separate session on the evaluator lane selected from `workflow/lane-policy.md`.
 - Generator writes `worklog.md`, `context-pack.md`, and `drift.md`.
 - Evaluator reads `.llm/harness/evaluator/protocol.md`, the plan, worklog, context pack, drift, the
   draft-PR commit list + per-slice PR comments (the commit trail), selected archetype, overlays, and
@@ -316,11 +324,11 @@ User says "use harness"
       `supervisor.md`/`drift.md`.
 - [ ] Plan-Gate checklist (`gates/plan-gate.md`) was reviewed.
 - [ ] PLAN-EVAL returned `PASS` before any implementation slice.
-- [ ] PLAN-EVAL used the recorded opposite-family route, or the blocked launch was recorded.
+- [ ] PLAN-EVAL used the recorded evaluator route (open model), or the blocked launch was recorded.
 - [ ] Tier-D (WSL Codex) slices recorded daemon-managed proof, thread id, worktree, and steering
       command.
 - [ ] The slice review gate was performed (Tier-A substantive review) before each sign-off commit;
       no lane self-certified.
 - [ ] Each implementation slice was committed, pushed, and commented on the draft PR.
 - [ ] IMPL-EVAL is a separate session from the generator.
-- [ ] IMPL-EVAL used the recorded opposite-family route, or the blocked launch was recorded.
+- [ ] IMPL-EVAL used the recorded evaluator route (open model), or the blocked launch was recorded.

.agents/skills/openhands-handoff/SKILL.md

@@ -27,19 +27,44 @@ required summary artifacts keep local and cloud agents synchronized.
 
 OpenHands is **not** the evaluator for local runs. Two hard rules:
 
-1. **OpenHands runs OPEN models only** — e.g. `minimax/minimax-m3`, `qwen/qwen3.7-max`. NEVER dispatch
-   OpenHands with a closed/paid model (Claude/`sonnet`, GPT/`gpt`, Gemini/`gemini`). Closed models on
-   OpenHands route through paid OpenRouter/LiteLLM credit and can silently burn the owner's balance —
-   this is prohibited.
+1. **OpenHands runs OPEN models only** — e.g. `minimax/minimax-m3`, `qwen/qwen3.7-max`. NEVER
+   dispatch OpenHands with a closed/paid model (Claude/`sonnet`, GPT/`gpt`, Gemini/`gemini`). Closed
+   models on OpenHands route through paid OpenRouter/LiteLLM credit and can silently burn the
+   owner's balance — this is prohibited.
 2. **OpenHands is for CLOUD-driven runs only** — small GitHub-Copilot-style tasks the owner wants
-   reviewed fully in the cloud with adversarial agents. For any run on the **local machine**, do NOT
-   dispatch cloud OpenHands at all. Use a **local opposite-family adversarial agent** for
-   PLAN-EVAL / IMPL-EVAL / review (e.g. Codex GPT-5.6 reviews Claude-authored work; a Claude session
-   reviews Codex-authored work), launched locally. The **supervisor chooses what to trigger** —
-   sub-agents/implementers must NEVER auto-dispatch a cloud evaluator.
-
-If a local run's harness step calls for PLAN-EVAL/IMPL-EVAL and no local adversarial agent is
-available, record the gap in `drift.md` and let the supervisor decide — do not fall back to OpenHands.
+   reviewed fully in the cloud with adversarial agents. OpenHands remains the **default automated
+   cloud agent**; nothing below changes that. For any run on the **local machine**, do NOT dispatch
+   cloud OpenHands at all — use the local evaluator transport named below. The **supervisor chooses
+   what to trigger** — sub-agents/implementers must NEVER auto-dispatch a cloud evaluator.
+
+### The local evaluator transport (named, 2026-07-13)
+
+A local run's PLAN-EVAL / IMPL-EVAL used to have no named transport — this skill described the gap
+and told you to log it. It is now filled:
+
+**Local PLAN-EVAL / IMPL-EVAL runs on Claude Code + OpenRouter** — the `claude-openrouter` provider
+profile driven via `claude-print` — with the bound **OPEN-model Qwen evaluation preset**
+(`qwen/qwen3.7-max`). Minimax M3 remains in the approved open-model set and is valid for cloud
+evaluation, but its current local preset is workflow-fanout rather than evaluation. **Rule 1 above
+applies verbatim to this lane: OPEN models only; closed/paid
+models (Claude/`sonnet`, GPT/`gpt`, Gemini) are PROHIBITED** — they bill the owner's OpenRouter
+balance and can silently burn it. An open model is neither Claude-family nor Codex-family, so it is
+adversarial to **both** generators, which satisfies the generator-≠-evaluator invariant more
+robustly than a family swap alone.
+
+**Ordinary (non-formal) review** — the slice review gate, code/PR review — still uses a **local
+opposite-family** agent: Codex GPT-5.6 reviews Claude-authored work; a Claude session reviews
+Codex-authored work. Do not conflate it with the formal evaluator pass.
+
+**Capability (verified; drift D-4 amended):** both approved open models return a **real reasoning
+trace** and have a **verified agentic turn** (real tool calls) on this transport. The currently
+bound Qwen evaluation preset can therefore run gates and its `effort` is genuine — not nominal. The zero-reasoning behaviour is
+**specific to GLM 5.2** over OpenRouter (a design-lane model), **not** a client-wide gap: never cite
+"GLM 5.2 · xhigh reasoning" as gate evidence, and do not restate that caveat as a property of the
+transport or of the evaluator lane.
+
+If neither the local transport nor a cloud run can be launched, record the gap in `drift.md` and let
+the supervisor decide — never self-certify.
 
 ## Key Concepts
 
@@ -108,11 +133,11 @@ fallback).
 | `summary-only`     | Upload artifacts only; do not comment.                                                  |
 
 The agent must write `OPENHANDS_SUMMARY_PATH` before exit. For a harness run, the verdict of record
-and its trace live in the **tracked** run dir: `OPENHANDS_RUN_DIR` is the run's `.llm/runs/<run-id>/`
-dir (where the evaluator's `plan-eval.md`/`evaluate.md` is committed), and `TRACE_DIR` is `trace/`
-beneath it (env `OPENHANDS_TRACE_DIR`) for compact trace metadata. Do not reuse legacy
-`.llm/tmp/openhands/summary.md` or `.llm/tmp/run/openhands/…` scratch as the verdict — see
-`.llm/harness/workflow/agent-handoff.md` for the full output contract.
+and its trace live in the **tracked** run dir: `OPENHANDS_RUN_DIR` is the run's
+`.llm/runs/<run-id>/` dir (where the evaluator's `plan-eval.md`/`evaluate.md` is committed), and
+`TRACE_DIR` is `trace/` beneath it (env `OPENHANDS_TRACE_DIR`) for compact trace metadata. Do not
+reuse legacy `.llm/tmp/openhands/summary.md` or `.llm/tmp/run/openhands/…` scratch as the verdict —
+see `.llm/harness/workflow/agent-handoff.md` for the full output contract.
 
 ## Token Rule
 
@@ -138,8 +163,8 @@ and carry a `## SKILL` chapter — see step 2 below) and reads the GitHub token
 in-process env var (`--token-env`, default `GH_TOKEN`), never from a file or argv. It posts exactly
 one trigger comment, respecting the per-PR concurrency-cancel rule.
 `openhands-status.ts --source
-local` needs no token and reads the newest committed trace under the run's `TRACE_DIR`
-(`.llm/runs/<run-id>/trace/`).
+local` needs no token and reads the newest committed trace under the
+run's `TRACE_DIR` (`.llm/runs/<run-id>/trace/`).
 
 ## Workflow
 
@@ -184,26 +209,27 @@ local` needs no token and reads the newest committed trace under the run's `TRAC
   budget during exploration, leaving zero artifacts. Split the task into sequential triggers and/or
   raise `iterations=`, and require the agent to create deliverable files early and grow them
   incrementally.
-- **Eval run mutating `deno.lock`**: a PLAN-EVAL/IMPL-EVAL run's `deno check` / `deno publish
+- **Eval run mutating `deno.lock`**: a PLAN-EVAL/IMPL-EVAL run's `deno check` /
+  `deno publish
   --dry-run` can silently re-resolve and commit `deno.lock` churn (observed: a
   `@opentelemetry/semantic-conventions` downgrade riding into a merged umbrella). Rule: **an
   evaluator must never mutate the lock** — instruct the run prompt to `git checkout -- deno.lock`
   before committing. Diff the lock against the **true base** (the commit the wave forked/synced
   from), not the working baseline — an in-branch "unchanged" claim hides churn that predates the
   first slice. Reconcile-don't-revert mid-wave (Golden Rule 6: never delete the lock or `--reload`
   without approval). The lock is not the only churn: an IMPL-EVAL commit-back can also push
-  scratch/junk files alongside the verdict artifact. **Verify the committed file set before merge** —
-  the intended change is the run's `evaluate.md`/`plan-eval.md` plus any authorized fix, not a
+  scratch/junk files alongside the verdict artifact. **Verify the committed file set before merge**
+  — the intended change is the run's `evaluate.md`/`plan-eval.md` plus any authorized fix, not a
   re-resolved `deno.lock` or stray workspace files; drop anything outside that set.
 - **Trusting the persistent PR summary comment**: the `<!-- openhands-agent-summary -->` comment is
   one per PR and is **not always regenerated** for a new run — it can show a prior run's package,
   phase, and verdict while only its `updated_at` bumps. **The verdict source is the committed run
   artifact** (`plan-eval.md` for PLAN-EVAL, `evaluate.md` for IMPL-EVAL), never the PR comment;
   confirm it names the right run id, slices, and surface before accepting it.
-- **Stacking `@openhands` comments on one PR**: multiple `@openhands` triggers on the same PR
-  cancel all-but-one under the per-PR concurrency group. Post exactly one trigger per intended run.
-  Trigger placement also chooses the checkout: a **PR comment** runs against the PR branch, an
-  **issue comment** runs against `main`.
+- **Stacking `@openhands` comments on one PR**: multiple `@openhands` triggers on the same PR cancel
+  all-but-one under the per-PR concurrency group. Post exactly one trigger per intended run. Trigger
+  placement also chooses the checkout: a **PR comment** runs against the PR branch, an **issue
+  comment** runs against `main`.
 
 ## Reference Files
 

.claude/skills/claude-manager/SKILL.md

@@ -9,8 +9,9 @@ description: >
 # Claude Manager
 
 Use this skill when the task involves Claude Code session orchestration rather than package code.
-Claude is the supervisor. OpenHands evaluates. WSL Codex implements harness slices that must be
-visible from Codex Desktop/mobile.
+Claude is the supervisor. WSL Codex implements harness slices that must be visible from Codex
+Desktop/mobile. Evaluation runs on the evaluator lane in `.llm/harness/workflow/lane-policy.md`:
+locally Claude Code + OpenRouter with an open model, and OpenHands for automated cloud runs.
 
 ## Workflow
 
@@ -65,8 +66,10 @@ outcome.
 - Do not use Claude workflows as the default NetScript implementation agent. WSL Codex remains the
   preferred implementation lane because it is daemon-attached, mobile-visible, and cheaper to steer
   slice-by-slice.
-- Keep OpenHands as the evaluator. Claude workflows may prepare evaluator inputs, but they do not
-  replace PLAN-EVAL or IMPL-EVAL.
+- Keep the evaluator a separate session on the evaluator lane from
+  `.llm/harness/workflow/lane-policy.md` — locally Claude Code + OpenRouter with an **open model**, or
+  OpenHands for cloud runs. Claude workflows may prepare evaluator inputs, but they do not replace
+  PLAN-EVAL or IMPL-EVAL, and no session self-certifies.
 - Route every Claude workflow, supervisor, and review session through the canonical lane table in
   `.llm/harness/workflow/lane-policy.md`. Do not reproduce model/effort defaults here or infer a
   paid escalation from workflow prose.
@@ -97,6 +100,7 @@ session should be started.
 
 - [ ] Current branch/worktree is verified.
 - [ ] Harness skill was loaded for harnessed work.
-- [ ] Evaluator surface is OpenHands when evaluating.
+- [ ] Evaluator surface is the lane-policy evaluator route (open model): Claude Code + OpenRouter
+      locally, OpenHands for cloud runs.
 - [ ] Implementation surface is WSL Codex when slice work must be mobile-visible.
 - [ ] Claude remote-control or Codex daemon visibility is proven before claiming phone visibility.