@kontextmind/kxm 0.7.190 → 0.7.192

package/.claude-plugin/marketplace.json

@@ -11,7 +11,7 @@
       "name": "kxm",
       "source": "./plugins/kxm",
       "description": "Durable workflows, peer agents, and kxm tui",
-      "version": "0.7.190",
+      "version": "0.7.192",
       "category": "development",
       "tags": ["kxm", "multi-agent", "workflows", "mcp"]
     }

package/.kxm/roles/planner.yaml

@@ -3,9 +3,13 @@ id: planner
 purpose: planner
 permission: read-only
 description: Plans the change before implementation.
-# claude-fable is primary. The Pi fallback is a different vendor.
+# claude-fable is primary. claude-claude-sonnet-5-5 is the same-vendor
+# fallback (October 2026 bake-off: Sonnet 126, Fable 110, about 1/5 the
+# list cost). The Pi fallback is a different vendor.
 roster:
   - route: claude-fable
     effort: medium
+  - route: claude-claude-sonnet-5-5
+    effort: medium
   - route: pi-qwen3-8-flash-openrouter
     effort: medium

package/CHANGELOG.md

@@ -225,6 +225,15 @@ All notable user-facing changes are documented here. The project follows [Semant
 
 ### Changed
 
+- **The planner fallback is Claude Sonnet 5.5.** `claude-fable` stays the
+  planner primary. `claude-claude-sonnet-5-5` (`claude` / `claude-sonnet-5.5`,
+  selector `anthropic/claude-sonnet-5.5`, read-only, effort `medium`) is the
+  next route, ahead of `pi-qwen3-8-flash-openrouter`. The October 2026
+  bake-off scored Sonnet 126 and Fable 110 out of 150, at about one fifth
+  the list cost, which was not enough to replace the primary. OpenRouter
+  listed `anthropic/claude-sonnet-5.5` on 2026-10-20 at $2 input and $10
+  output per 1M tokens. See
+  [Planner bake-off, October 2026](docs/evals/planner-bakeoff-2026-10.md).
 - **Writer #2 is OpenRouter Qwen3.7 Plus.** `pi-qwen3-7-plus-openrouter`
   (`openrouter/qwen/qwen3.7-plus`) replaces `pi-qwen3-coder-plus-openrouter`.
   OpenRouter no longer lists `qwen/qwen3-coder-plus`. The live catalog on

package/docs/contributing/assignment-runner.md

@@ -35,7 +35,7 @@ admits each route for specific roles and permissions:
 | Role | Admitted route (harness / model) | Vendor | Permission |
 |---|---|---|---|
 | `writer` | `grok` / `grok-4.7`; relief: `pi` / `openrouter/qwen/qwen3.7-plus`; `agy` / `gemini-4-argon-high`, then `gemini-4-argon-medium`, then `gemini-3.8-flash-medium`, then `gemini-3.8-flash-high` | `xai`; `alibaba`; `google` | `edit` |
-| `planner` | `claude` / `fable` | `anthropic` | `read-only` |
+| `planner` | `claude` / `fable`, then `claude` / `claude-sonnet-5.5`; fallback `pi` / `openrouter/qwen/qwen3.8-flash` | `anthropic`; `anthropic`; `alibaba` | `read-only` |
 | `reviewer-arch` | `claude` / `fable` | `anthropic` | `read-only` |
 | `reviewer-cli` | `codex` / `gpt-6.1-sol`; fallback `pi` / `openrouter/qwen/qwen3.8-flash`, then `codex` / `gpt-5.6-sol` | `openai`; `alibaba`; `openai` | `read-only` |
 

package/docs/contributing/harness-routing-internals.md

@@ -9,7 +9,7 @@ This page records how the KXM repository applies [harness routing](../reference/
 
 The files use one id shape per kind. A route id is `<harness>-<model-slug>[-<provider>]`, with `.` written as `-`. An agent id equals its role. A workflow id is `default`, `land`, or `<role>-only`. An agent step id equals that role.
 
-This checkout's admitted routes are `grok-grok-4-7`, `pi-qwen3-7-plus-openrouter`, `agy-gemini-4-argon-high`, `agy-gemini-4-argon-medium`, `agy-gemini-3-8-flash-medium`, `agy-gemini-3-8-flash-high`, `claude-fable`, `codex-gpt-6-1-sol`, `codex-gpt-5-6-sol`, `pi-qwen3-8-flash-openrouter`, and `pi-glm-5-3-flash-openrouter`. `pi-qwen3-coder-plus-openrouter` stays on disk with `status: retired` because OpenRouter removed `qwen/qwen3-coder-plus` (catalog 404); it is not admitted and not on a roster, and `qwen-openrouter-pi` does not resolve. Agents are `planner`, `writer`, `reviewer-arch`, and `reviewer-cli`. `opus-claude` was removed because `opus` is not an admitted selector. `qwen-token-plan/*` and `zai-coding-cn/*` left the admitted list because those harnesses are not allowlisted; their `prices.yaml` rows stay. Other old ids resolve as aliases. Every roster entry names an effort. `reviewer-arch` dispatches `claude-fable`. #343's architecture critic ran on opus because the justfile recipe review-arch hardcoded that model; the role file on that commit listed only `fable-claude`. The helper now refuses `opus`.
+This checkout's admitted routes are `grok-grok-4-7`, `pi-qwen3-7-plus-openrouter`, `agy-gemini-4-argon-high`, `agy-gemini-4-argon-medium`, `agy-gemini-3-8-flash-medium`, `agy-gemini-3-8-flash-high`, `claude-fable`, `claude-claude-sonnet-5-5`, `codex-gpt-6-1-sol`, `codex-gpt-5-6-sol`, `pi-qwen3-8-flash-openrouter`, and `pi-glm-5-3-flash-openrouter`. `pi-qwen3-coder-plus-openrouter` stays on disk with `status: retired` because OpenRouter removed `qwen/qwen3-coder-plus` (catalog 404); it is not admitted and not on a roster, and `qwen-openrouter-pi` does not resolve. Agents are `planner`, `writer`, `reviewer-arch`, and `reviewer-cli`. `opus-claude` was removed because `opus` is not an admitted selector. `qwen-token-plan/*` and `zai-coding-cn/*` left the admitted list because those harnesses are not allowlisted; their `prices.yaml` rows stay. Other old ids resolve as aliases. Every roster entry names an effort. `reviewer-arch` dispatches `claude-fable`. #343's architecture critic ran on opus because the justfile recipe review-arch hardcoded that model; the role file on that commit listed only `fable-claude`. The helper now refuses `opus`.
 
 ## This checkout's routes and roster
 
@@ -37,11 +37,12 @@ roster:
     effort: medium
 ```
 
-The writer role lists route ids, in order: `grok-grok-4-7`, then `pi-qwen3-7-plus-openrouter`, then `agy-gemini-4-argon-high`, `agy-gemini-4-argon-medium`, `agy-gemini-3-8-flash-medium`, and `agy-gemini-3-8-flash-high`. `grok-grok-4-7` is harness `grok`, model `grok-4.7`, vendor `xai`. An agent file does not declare `harness:`. Argon is writer #3 and #4. Flash medium is the safety net because Argon needs a paid Google AI Pro, Ultra, or Enterprise plan and `agy -p` exits when the pinned model is missing. Flash high stays on the roster so that admitted model file is still used. The planner primary stays `claude-fable`. The CLI critic primary is `codex-gpt-6-1-sol`, then `pi-qwen3-8-flash-openrouter`, then `codex-gpt-5-6-sol`. The architecture critic primary stays `claude-fable`, then `pi-glm-5-3-flash-openrouter`. Argon can serve as a planner or architecture-review fallback; it is not on those rosters. GPT-6.1 Sol Pro is the same model with reasoning mode `pro`. The route id does not carry a reasoning mode, role effort has no `pro` value, and codex is limited to `reviewer-cli`, so that mode is not an architecture fallback.
+The writer role lists route ids, in order: `grok-grok-4-7`, then `pi-qwen3-7-plus-openrouter`, then `agy-gemini-4-argon-high`, `agy-gemini-4-argon-medium`, `agy-gemini-3-8-flash-medium`, and `agy-gemini-3-8-flash-high`. `grok-grok-4-7` is harness `grok`, model `grok-4.7`, vendor `xai`. An agent file does not declare `harness:`. Argon is writer #3 and #4. Flash medium is the safety net because Argon needs a paid Google AI Pro, Ultra, or Enterprise plan and `agy -p` exits when the pinned model is missing. Flash high stays on the roster so that admitted model file is still used. The planner roster is `claude-fable`, then `claude-claude-sonnet-5-5` (`anthropic/claude-sonnet-5.5`), then `pi-qwen3-8-flash-openrouter`. Fable stays the primary. The [October 2026 planner bake-off](../evals/planner-bakeoff-2026-10.md) is why Sonnet is the fallback. The CLI critic primary is `codex-gpt-6-1-sol`, then `pi-qwen3-8-flash-openrouter`, then `codex-gpt-5-6-sol`. The architecture critic primary stays `claude-fable`, then `pi-glm-5-3-flash-openrouter`. Argon can serve as a planner or architecture-review fallback; it is not on those rosters. GPT-6.1 Sol Pro is the same model with reasoning mode `pro`. The route id does not carry a reasoning mode, role effort has no `pro` value, and codex is limited to `reviewer-cli`, so that mode is not an architecture fallback.
 
 `node scripts/kxm.mjs routes list`:
 
 ```text
+admitted anthropic/claude-sonnet-5.5
 admitted anthropic/fable
 admitted google/gemini-3.8-flash-high
 admitted google/gemini-3.8-flash-medium
@@ -54,6 +55,7 @@ admitted openrouter/qwen/qwen3.8-flash
 admitted openrouter/z-ai/glm-5.3-flash
 admitted xai/grok-4.7
 planner claude-fable
+planner claude-claude-sonnet-5-5
 planner pi-qwen3-8-flash-openrouter
 reviewer-arch claude-fable
 reviewer-arch pi-glm-5-3-flash-openrouter
@@ -154,6 +156,8 @@ Pi's own `xai` provider reported `ready` (OAuth) on the capture machine. When th
 
 `claude` was logged out on the capture machine, so the native route failed closed. Pi's `anthropic` provider reported `ready` (OAuth). The `anthropic/fable` row in `prices.yaml` gives a list estimate only when the catalog is dated today.
 
+The planner fallback is the same harness. `claude-claude-sonnet-5-5` is model `claude-sonnet-5.5`, selector `anthropic/claude-sonnet-5.5`, read-only, effort `medium`, and it sits immediately after Fable. `pi-qwen3-8-flash-openrouter` stays third. The [October 2026 planner bake-off](../evals/planner-bakeoff-2026-10.md) scored Sonnet 126 and Fable 110. OpenRouter listed the standard tier on 2026-10-20 at $2 input and $10 output per 1M tokens (cache read $0.10, cache write $2.50). That row is in `.kxm/prices.yaml`. Architecture review does not use Sonnet.
+
 ### Gemini 3.8 Flash
 
 | | Native `agy` | `antigravity` Pi provider | Pi + OpenRouter |

package/docs/evals/planner-bakeoff-2026-10.md

@@ -1,16 +1,18 @@
 # Planner bake-off, October 2026
 
-This page records a scored comparison of Claude Fable 5.1 and Claude Sonnet 5.5 on six KXM planning tasks. It is for maintainers deciding whether the planner primary should stay Fable. Routing is unchanged.
+This page records a scored comparison of Claude Fable 5.1 and Claude Sonnet 5.5 on six KXM planning tasks. It is for maintainers deciding whether the planner primary should stay Fable. Routing was unchanged during the scored run.
+
+The fallback recommended below is now admitted. Fable stays the planner primary. `claude-claude-sonnet-5-5` (selector `anthropic/claude-sonnet-5.5`) is the next planner route, ahead of `pi-qwen3-8-flash-openrouter`. The list price is a `.kxm/prices.yaml` row, verified on OpenRouter on 2026-10-20 at $2 input and $10 output per 1M tokens. The procedure is [Harness routing](../reference/harness-routing.md).
 
 ## Recommendation
 
-Keep Fable as the planner primary. Treat Sonnet 5.5 as a fallback candidate, not a replacement.
+Keep Fable as the planner primary. Treat Sonnet 5.5 as the fallback, not a replacement.
 
 Sonnet scored 126 and Fable scored 110, out of 150, on a blinded judge. Sonnet won four tasks and Fable won two. The two Fable wins include the Argon roster, which is the task closest to an admission decision, and Sonnet's Argon plan admitted only the medium pin and placed it after Flash. Six tasks, one judge, and a transport that is not the KXM planner route are not enough to replace the primary. List price favors Sonnet by about five to one, and a native Claude subscription records both as unmetered, so the invoice may not move.
 
 ## How this run was scored
 
-The planner roster in `.kxm/roles/planner.yaml` is `claude-fable` at effort medium, then `pi-qwen3-8-flash-openrouter`. The native one-shot is `claude -p --model <model> --effort medium --output-format json` with the read-only arguments in `oneShotReadOnlyArgs` for `claude`. There is no command flag that swaps the planner model without editing the role and admitting a route. `kxm routing benchmark` prints placeholder numbers and does not call a model.
+At the time of the run, the planner roster in `.kxm/roles/planner.yaml` was `claude-fable` at effort medium, then `pi-qwen3-8-flash-openrouter`. The native one-shot is `claude -p --model <model> --effort medium --output-format json` with the read-only arguments in `oneShotReadOnlyArgs` for `claude`. There is no command flag that swaps the planner model without editing the role and admitting a route. `kxm routing benchmark` prints placeholder numbers and does not call a model.
 
 This machine has no `claude` binary, no Claude auth directory, and no `ANTHROPIC_API_KEY`. The committed plans are from Cursor subagents pinned to `claude-fable-5-1-thinking-medium` and `claude-sonnet-5-5-medium`. Both received the same prompt. They were told not to read the checkout, because the landed change would leak the answer. The Argon plans still name `pi-qwen3-coder-plus-openrouter` as writer number two, which the current tree no longer has, so they do not look like a read of the solution.
 
@@ -49,7 +51,7 @@ Regenerate scores with `python3 docs/evals/planner-bakeoff-2026-10/score.py --sc
 | Model | List rate per 1M | List cost | Wins |
 | --- | --- | --- | --- |
 | Fable 5.1 | $10 / $50, from `.kxm/prices.yaml` | $0.460 | 2 |
-| Sonnet 5.5 | $2 / $10, from inventory, not `.kxm/prices.yaml` | $0.089 | 4 |
+| Sonnet 5.5 | $2 / $10. The scored estimate read the inventory, because `.kxm/prices.yaml` had no row yet. The admitted row is the same rate. | $0.089 | 4 |
 
 Input tokens are identical because the prompts were identical. Wall time for the wave is 154 seconds, shared.
 
@@ -151,14 +153,14 @@ Both refused a store and refused invented schema versions. Sonnet stopped at cha
 - The judge saw condensations. Weekly is a one-point result and could flip. Argon, the reviewer switch, writer number two, and the journal gap are wider.
 - The judge is one model, GPT-5.6 Sol, which is the CLI critic family. It was blind to the author and not blind to the shipped outcome.
 - Token counts and list cost are estimates. Provider usage was not returned. A Claude subscription would record `costBasis: unmetered` for both models.
-- Sonnet's price is not in `.kxm/prices.yaml`. Quoting it from the inventory is a list estimate, and the price loader would mark that row missing.
+- When these plans were scored, Sonnet's price was not in `.kxm/prices.yaml`. The list estimate came from the inventory. The admitted fallback now has that $2 / $10 row, verified on OpenRouter on 2026-10-20.
 - Cursor subagents are not `claude -p` and not the KXM planner route. Effort was medium on both, matching the roster.
 - Pull request 398, the plan-link repair, was not run. The six tasks are the five roster and docs slices from 401 through 405, plus the open Runtime journal gap.
 - Plans were instructed not to use tools. A subagent that ignored that instruction could have read the landed diff. The Argon plans still describe the pre-change writer number two, which argues against that leak, and it is not proof.
 
 ## Related
 
-- [Harness routing](reference/harness-routing.md)
-- [Configuration reference](reference/config-reference.md)
-- [CLI reference](reference/cli-reference.md)
+- [Harness routing](../reference/harness-routing.md)
+- [Configuration reference](../reference/config-reference.md)
+- [CLI reference](../reference/cli-reference.md)
 - Raw plans, prompts, judge JSON, and `score.py` live in `docs/evals/planner-bakeoff-2026-10/`.

package/docs/reference/cli-reference.md

@@ -1150,6 +1150,7 @@ kxm routes list
 ```
 
 ```text
+admitted anthropic/claude-sonnet-5.5
 admitted anthropic/fable
 admitted google/gemini-3.8-flash-high
 admitted google/gemini-3.8-flash-medium
@@ -1162,6 +1163,7 @@ admitted openrouter/qwen/qwen3.8-flash
 admitted openrouter/z-ai/glm-5.3-flash
 admitted xai/grok-4.7
 planner claude-fable
+planner claude-claude-sonnet-5-5
 planner pi-qwen3-8-flash-openrouter
 reviewer-arch claude-fable
 reviewer-arch pi-glm-5-3-flash-openrouter
@@ -1181,7 +1183,7 @@ kxm routes list --json
 ```
 
 ```text
-{"schema":"kxm.cli-result.v1","ok":true,"command":"routes list","policy":{"schema":"kxm.routes.v2","updatedAt":"2026-10-17T14:30:00.000Z","admitted":["anthropic/fable","google/gemini-3.8-flash-high","google/gemini-3.8-flash-medium","google/gemini-4-argon-high","google/gemini-4-argon-medium","openai/gpt-5.6-sol","openai/gpt-6.1-sol","openrouter/qwen/qwen3.7-plus","openrouter/qwen/qwen3.8-flash","openrouter/z-ai/glm-5.3-flash","xai/grok-4.7"],"disabled":[]},"membership":["planner claude-fable","planner pi-qwen3-8-flash-openrouter","reviewer-arch claude-fable","reviewer-arch pi-glm-5-3-flash-openrouter","reviewer-cli codex-gpt-6-1-sol","reviewer-cli pi-qwen3-8-flash-openrouter","reviewer-cli codex-gpt-5-6-sol","writer grok-grok-4-7","writer pi-qwen3-7-plus-openrouter","writer agy-gemini-4-argon-high","writer agy-gemini-4-argon-medium","writer agy-gemini-3-8-flash-medium","writer agy-gemini-3-8-flash-high"]}
+{"schema":"kxm.cli-result.v1","ok":true,"command":"routes list","policy":{"schema":"kxm.routes.v2","updatedAt":"2026-10-20T22:40:00.000Z","admitted":["anthropic/claude-sonnet-5.5","anthropic/fable","google/gemini-3.8-flash-high","google/gemini-3.8-flash-medium","google/gemini-4-argon-high","google/gemini-4-argon-medium","openai/gpt-5.6-sol","openai/gpt-6.1-sol","openrouter/qwen/qwen3.7-plus","openrouter/qwen/qwen3.8-flash","openrouter/z-ai/glm-5.3-flash","xai/grok-4.7"],"disabled":[]},"membership":["planner claude-fable","planner claude-claude-sonnet-5-5","planner pi-qwen3-8-flash-openrouter","reviewer-arch claude-fable","reviewer-arch pi-glm-5-3-flash-openrouter","reviewer-cli codex-gpt-6-1-sol","reviewer-cli pi-qwen3-8-flash-openrouter","reviewer-cli codex-gpt-5-6-sol","writer grok-grok-4-7","writer pi-qwen3-7-plus-openrouter","writer agy-gemini-4-argon-high","writer agy-gemini-4-argon-medium","writer agy-gemini-3-8-flash-medium","writer agy-gemini-3-8-flash-high"]}
 ```
 
 ### `kxm routes count`

package/docs/reference/config-reference.md

@@ -1082,6 +1082,13 @@ routes:
     roles: [planner, reviewer-arch]
     permissions: [read-only]
     status: admitted
+  claude-claude-sonnet-5-5:
+    harness: claude
+    model: claude-sonnet-5.5
+    vendor: anthropic
+    roles: [planner]
+    permissions: [read-only]
+    status: admitted
   codex-gpt-6-1-sol:
     harness: codex
     model: gpt-6.1-sol
@@ -1098,7 +1105,7 @@ routes:
     status: admitted
 lineup:                         # routes admitted for each role, preference order
   writer: [grok-grok-4-7, pi-qwen3-7-plus-openrouter, agy-gemini-4-argon-high, agy-gemini-4-argon-medium, agy-gemini-3-8-flash-medium, agy-gemini-3-8-flash-high]
-  planner: [claude-fable, pi-qwen3-8-flash-openrouter]
+  planner: [claude-fable, claude-claude-sonnet-5-5, pi-qwen3-8-flash-openrouter]
   reviewer-arch: [claude-fable, pi-glm-5-3-flash-openrouter]
   reviewer-cli: [codex-gpt-6-1-sol, pi-qwen3-8-flash-openrouter, codex-gpt-5-6-sol]
 required_critics:
@@ -1112,6 +1119,11 @@ model_origins:                  # required for every Pi route model
       sha256: 10e11cfa572ef4891e514d229fa8b7f0f355b31e6481315a6dd763bbe44244a1
 ```
 
+The planner lineup is `claude-fable`, then `claude-claude-sonnet-5-5`
+(selector `anthropic/claude-sonnet-5.5`, list price $2 input and $10 output
+per 1M tokens), then `pi-qwen3-8-flash-openrouter`. Fable stays primary.
+The comparison is the [October 2026 planner bake-off](../evals/planner-bakeoff-2026-10.md).
+
 Validated with `validateRosterDocument` from `scripts/roster-policy.mjs`,
 reading evidence from this repository's history. The lineup order is not a
 selection order: the assignment manifest names the harness and model, and the

package/docs/reference/harness-routing.md

@@ -406,6 +406,8 @@ A larger context window on the OpenRouter route does not make OpenRouter an allo
 
 **KXM picks the native route, and fails closed when `claude` is logged out.** Pi's `anthropic` provider may report `ready`, and OpenRouter lists the same model. Neither may stand in, and the brake refuses both. This is the motivating case for the whole rule. The operator runs `claude auth login`, and the run waits until then.
 
+The planner roster keeps `claude-fable` first. The next route is `claude-claude-sonnet-5-5`: harness `claude`, model `claude-sonnet-5.5`, vendor `anthropic`, selector `anthropic/claude-sonnet-5.5`, permission `read-only`, effort `medium`. `pi-qwen3-8-flash-openrouter` stays after that, and it is the other-vendor fallback. Architecture review stays `claude-fable`. The [October 2026 planner bake-off](../evals/planner-bakeoff-2026-10.md) scored Sonnet 5.5 at 126 and Fable at 110 out of 150. Sonnet won four of six tasks. List price favors Sonnet by about five to one, and that sample was not enough to replace the primary, so Sonnet is the fallback. OpenRouter's live catalog on 2026-10-20 (`https://openrouter.ai/api/v1/models`, id `anthropic/claude-sonnet-5.5`) lists $2 input and $10 output per 1M tokens, cache read $0.10, cache write $2.50, context 1,000,000. That standard tier is the `.kxm/prices.yaml` row. The batch id is a different catalog row and is not admitted.
+
 ### Gemini 3.8 Flash
 
 | | Native `agy` | `antigravity` Pi provider | Pi + OpenRouter |

package/docs/reference/workflow-catalog.md

@@ -22,7 +22,8 @@ This catalog names common multi-agent workflows by area and lists, for each role
 
 The catalog is a four-level taxonomy: **Area → Workflow → Stage → Role**. Each area groups related workflows, each workflow lists its stages in order, and each role belongs to one stage, with a short domain description and up to five ranked candidate models.
 
-- **Candidates are dated research.** The lists come from a documentation snapshot taken on 2026-09-06. Prices (USD per million input and output tokens), context windows and capability notes were not re-verified after that date. Verify them before you dispatch work, and read every quality note as a research claim, not a guarantee.
+- **Candidates are dated research.** The lists come from a documentation snapshot taken on 2026-09-06. Prices (USD per million input and output tokens), context windows and capability notes were not re-verified after that date, except for the retired writer model below. Verify them before you dispatch work, and read every quality note as a research claim, not a guarantee.
+- **Writer #2 is Qwen3.7 Plus.** `qwen/qwen3-coder-plus` left the OpenRouter catalog (404) and is not a roster route. Candidate lines that named it now name `qwen/qwen3.7-plus` at the 2026-10-17 list price, $0.32 input and $1.28 output per 1M tokens, 1M context. This repository's writer #2 is the route `pi-qwen3-7-plus-openrouter` (model `openrouter/qwen/qwen3.7-plus`). The writer roster, in order, is `grok-grok-4-7`, `pi-qwen3-7-plus-openrouter`, `agy-gemini-4-argon-high`, `agy-gemini-4-argon-medium`, `agy-gemini-3-8-flash-medium`, `agy-gemini-3-8-flash-high`. [Harness routing](harness-routing.md) is the admission record.
 - **Candidates are not admission.** A model ID here is a research identifier, not a route selector. Your project admits routes in `.kxm/routes.yaml`, and `kxm harness list` shows which harnesses are installed and signed in. [Harness routing](harness-routing.md) explains when to use a native harness and when an aggregator.
 - **Selection is per role.** Areas are navigation only; they never pool unrelated roles into one ranking. Model, harness, platform, modality and required tools are routing attributes, not areas.
 - **Gates stay deterministic.** Where a role is a critic next to a gate, the gate (tests, type checks, linters) decides pass or fail, and the model reviews what a gate cannot check.
@@ -34,7 +35,7 @@ An interactive `kxm init` offers to install the Software Engineering workflows w
 | Tier | Price band | Context window | Example models | Typical work |
 |---|---|---|---|---|
 | **Tier 0: Filter and ingestion** | $0.03 - $0.20 / M | 1.0M - 1.31M | DeepSeek-V4-Flash, Qwen-3.7-Flash, GLM-5.3-Flash, GPT-5.6-Luna | High-throughput telemetry, raw log ingestion, triage, video pre-filtering, and initial inbox classification. |
-| **Tier 1: Workhorse** | $0.30 - $1.00 / M | 262k - 1.05M | Gemini-3.8-Flash, Qwen-3-Coder-Plus, DeepSeek-V4-Pro, Devstral-2512 | Routine code generation, multimodal visual QA, Text-to-SQL, AST diff patching, and fast tool dispatching. |
+| **Tier 1: Workhorse** | $0.30 - $1.00 / M | 262k - 1.05M | Gemini-3.8-Flash, Qwen-3.7-Plus, DeepSeek-V4-Pro, Devstral-2512 | Routine code generation, multimodal visual QA, Text-to-SQL, AST diff patching, and fast tool dispatching. |
 | **Tier 2: Precision and critic** | $1.25 - $4.00 / M | 200k - 1.05M | GPT-5.6-Sol, Grok-4.6, Claude Sonnet 5, OpenAI o3, DeepSeek-R1, Codex 5.3 | Primary code writing, formal contract generation, concurrency race diagnosis, and causal inference. |
 | **Tier 3: Architecture** | $5.00 - $15.00 / M | 1.0M - 1.05M | Claude Fable 5.1, Claude Opus 5, GPT-6-Astra-Pro (escalation only) | Architecture review, executive briefs, high-stakes legal redlining, and RFC review. |
 
@@ -144,7 +145,7 @@ This cross-reference points each software and security workflow at the KXM pages
 1. **`openai/gpt-5.6-sol`** (1.05M ctx | $2.00 / $10.00) — Premier schema contract generator; 100% compliant OpenAPI 3.1, Protobuf v3, and JSON schemas.
 2. **`deepseek/deepseek-v4-pro-0813`** (1.05M ctx | $0.58 / $1.74) — High-throughput Zod/Pydantic validation pipeline synthesis.
 3. **`google/gemini-3.8-flash`** (1.05M ctx | $0.75 / $3.75) — Fast translation of data dictionaries into TypeScript interfaces.
-4. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — 1M-context contract verification across monorepo boundaries.
+4. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — 1M-context contract verification across monorepo boundaries.
 5. **`anthropic/claude-fable-5.1`** (1M ctx | $10.00 / $50.00) — Reserved for mission-critical authentication and security contract definitions.
 
 #### Role: Scaffold Build Specialist
@@ -155,7 +156,7 @@ This cross-reference points each software and security workflow at the KXM pages
 
 1. **`x-ai/grok-4.6`** (500k ctx | $2.00 / $6.00) — Fast repository builder; resolves path aliases, links, and workspace configs cleanly.
 2. **`mistralai/devstral-2512`** (262k ctx | $0.40 / $2.00) — Developer-focused open model; produces clean Makefiles, Taskfiles, and tool configs.
-3. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — Coordinates cross-package build configs and dependency lockfiles in 1M context.
+3. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — Coordinates cross-package build configs and dependency lockfiles in 1M context.
 4. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Complex CI/CD workflows and multi-stage container build optimizations.
 5. **`kwaipilot/kat-coder-pro-v2.5`** (262k ctx | $0.74 / $2.96) — Resilient project scaffolding and bundler configuration.
 
@@ -166,7 +167,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Implements domain entities, SQL queries/migrations, ORM persistence, concurrency primitives, and transactional boundaries.
 
 1. **`x-ai/grok-4.6`** (500k ctx | $2.00 / $6.00) — Strong first-pass implementation that passes deterministic verification gates with minimal rework.
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — 1M-context open-weight anchor; generates performant database layers and CRUD services.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — 1M-context open-weight anchor; generates performant database layers and CRUD services.
 3. **`deepseek/deepseek-v4-pro-0813`** (1.05M ctx | $0.58 / $1.74) — Exceptional database logic, indexing, and edge-case domain implementation.
 4. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Algorithmic precision in Rust, Go, Python, and TypeScript concurrent systems.
 5. **`kwaipilot/kat-coder-pro-v2.5`** (262k ctx | $0.74 / $2.96) — Multi-file backend logic implementation adhering to project coding patterns.
@@ -191,7 +192,7 @@ This cross-reference points each software and security workflow at the KXM pages
 
 1. **`openai/gpt-5.6-sol`** (1.05M ctx | $2.00 / $10.00) — Flawless SDK synthesis, webhook signature validation, and API adapter error handling.
 2. **`x-ai/grok-4.6`** (500k ctx | $2.00 / $6.00) — Robust glue code, rate-limiting backoff algorithms, and token refresh loops.
-3. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — Ingests full third-party API documentation to emit typed wrapper clients.
+3. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — Ingests full third-party API documentation to emit typed wrapper clients.
 4. **`deepseek/deepseek-v4-pro-0813`** (1.05M ctx | $0.58 / $1.74) — Low-cost mapping of disparate vendor payloads into internal domain models.
 5. **`google/gemini-3.8-flash`** (1.05M ctx | $0.75 / $3.75) — High-speed webhook listener and protocol transformation adapter.
 
@@ -310,7 +311,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Generates declarative Terraform/OpenTofu, Pulumi, Kubernetes CRDs, and VPC networking definitions.
 
 1. **`x-ai/grok-4.6`** (500k ctx | $2.00 / $6.00) — Fast, deterministic Terraform/Kubernetes manifest implementation.
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — 1M context whole-infrastructure state dependency resolution.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — 1M context whole-infrastructure state dependency resolution.
 3. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Complex cloud provider networking and peering configurations.
 4. **`google/gemini-3.8-flash`** (1.05M ctx | $0.75 / $3.75) — High-speed validation of cloud resources and linting.
 5. **`mistralai/devstral-2512`** (262k ctx | $0.40 / $2.00) — Lightweight Helm chart and Docker Compose generator.
@@ -366,7 +367,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Authors concise, accurate developer and operator documentation, ensuring clear terminology, correct command flags, and consistent formatting without marketing drift.
 
 1. **`x-ai/grok-4.6`** (500k ctx | $2.00 / $6.00) — Rapid, structured Markdown authoring that follows repository conventions.
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — High-context technical writer; coordinates cross-file references across large documentation sets.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — High-context technical writer; coordinates cross-file references across large documentation sets.
 3. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Fluid technical documentation authoring with clear instructional hierarchy.
 4. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Accurate technical guides, CLI flag references, and runnable examples.
 5. **`deepseek/deepseek-v4-pro-0813`** (1.05M ctx | $0.58 / $1.74) — Cost-effective multi-page documentation generation and reference expansion.
@@ -412,7 +413,7 @@ This cross-reference points each software and security workflow at the KXM pages
 
 1. **`anthropic/claude-opus-5`** (1M ctx | $5.00 / $25.00) — Master design token taxonomist; color space (OKLCH) and scale hierarchy.
 2. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Multi-platform token transformation (W3C Design Tokens to CSS/iOS/Android).
-3. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — Monorepo theme implementation and code generation.
+3. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — Monorepo theme implementation and code generation.
 4. **`google/gemini-3.8-flash`** (1.05M ctx | $0.75 / $3.75) — Fast token validation and contrast ratio auditing.
 5. **`openai/gpt-5.4`** (1.05M ctx | $2.50 / $15.00) — Enterprise design system cross-platform consistency.
 
@@ -455,7 +456,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Apple CoreHaptics (`.ahap` waveforms), Android `VibrationEffect` compositions, JSI hardware bridges, and tactile throttling.
 
 1. **`mistralai/codestral-2508`** (256k ctx | $0.30 / $0.90) — Low-latency, precise hardware API bindings (Swift/Kotlin/C++).
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — High-throughput cross-platform haptic wrapper synthesizer.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — High-throughput cross-platform haptic wrapper synthesizer.
 3. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Fine-grained haptic feedback design and event mapping.
 4. **`openai/gpt-5.4`** (1.05M ctx | $2.50 / $15.00) — Hardware sensor integration and battery lifecycle optimization.
 5. **`openai/gpt-5.6-luna`** (1.05M ctx | $0.20 / $1.20) — Fast AHAP waveform JSON serialization.
@@ -488,7 +489,7 @@ This cross-reference points each software and security workflow at the KXM pages
 
 1. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Flawless asynchronous event loop (Tokio/TEA) and POSIX termios handling.
 2. **`openai/gpt-5.6-sol`** (1.05M ctx | $2.00 / $10.00) — Flagship CLI and sub-process navigation specialist; signal cleanup.
-3. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — High-precision Rust/Go async event handling and state dispatch.
+3. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — High-precision Rust/Go async event handling and state dispatch.
 4. **`anthropic/claude-opus-5`** (1M ctx | $5.00 / $25.00) — Complex modal keybinding architectures (Vim/Emacs emulation).
 5. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Clean command routing and keyboard navigation schemas.
 
@@ -520,7 +521,7 @@ This cross-reference points each software and security workflow at the KXM pages
 
 1. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Gold standard for semantic HTML, ARIA tree hierarchies, and focus management.
 2. **`google/gemini-3.8-flash`** (1.05M ctx | $0.75 / $3.75) — Automated WCAG 2.2 color contrast and visual touch-target auditor.
-3. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — High-throughput AST accessibility linter and automated ARIA fixer.
+3. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — High-throughput AST accessibility linter and automated ARIA fixer.
 4. **`anthropic/claude-opus-5`** (1M ctx | $5.00 / $25.00) — Cognitive load and neurodiverse accessibility compliance evaluation.
 5. **`openai/gpt-5.4`** (1.05M ctx | $2.50 / $15.00) — Screen-reader script simulation and compliance reporting.
 
@@ -663,7 +664,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Emits executable FFmpeg filtergraphs, OpenTimelineIO (OTIO) JSON, and Remotion/DaVinci Resolve Python scripts.
 
 1. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Frontier code generation for complex multi-input FFmpeg filtergraphs and OTIO structures.
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — 1M-context Remotion/React-video and Python automation specialist at low cost.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — 1M-context Remotion/React-video and Python automation specialist at low cost.
 3. **`anthropic/claude-opus-5`** (1M ctx | $5.00 / $25.00) — Multi-track timecode arithmetic, drop-frame calculations, and nested timeline logic.
 4. **`google/gemini-3.8-flash`** (1.05M ctx | $0.75 / $3.75) — Inspects raw asset metadata directly and constructs execution CLI commands.
 5. **`mistralai/devstral-2512`** (262k ctx | $0.40 / $2.00) — Lightweight Bash and Python timeline glue scripts.
@@ -707,7 +708,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Translates natural language questions into performant SQL queries (CTEs, window functions) across enterprise warehouses.
 
 1. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Complex nested SQL queries, CTEs, and window functions.
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — Ingests 1,000-table enterprise data dictionaries in 1M context.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — Ingests 1,000-table enterprise data dictionaries in 1M context.
 3. **`deepseek/deepseek-v4-pro`** (1.05M ctx | $0.63 / $1.25) — High-throughput Text-to-SQL generation.
 4. **`anthropic/claude-opus-5`** (1M ctx | $5.00 / $25.00) — Ambiguous business metric semantic translation.
 5. **`google/gemini-3.8-flash`** (1.05M ctx | $0.75 / $3.75) — Ultra-fast schema resolution and join mapping.
@@ -721,7 +722,7 @@ This cross-reference points each software and security workflow at the KXM pages
 1. **`openai/o3`** (200k ctx | $2.00 / $8.00) — Query execution plan optimization and index utilization logic.
 2. **`deepseek/deepseek-r1-0528`** (164k ctx | $0.50 / $2.15) — High-efficiency algorithmic query rewriting.
 3. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Database engine-specific syntax correction (Postgres/Snowflake/BigQuery).
-4. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — Cost-efficient automated query correction loop.
+4. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — Cost-efficient automated query correction loop.
 5. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Enterprise data governance and query compliance.
 
 #### Role: BI Dashboard Formatter
@@ -751,7 +752,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Writes vectorized, performant Python code for feature engineering, statistical modeling, and ML pipelines.
 
 1. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Complex numerical analysis, PyTorch/Scikit pipelines.
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — High-capacity Python data science code generator with 1M context.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — High-capacity Python data science code generator with 1M context.
 3. **`mistralai/codestral-2508`** (256k ctx | $0.30 / $0.90) — Extremely low-cost Polars/Pandas data pipeline implementation.
 4. **`anthropic/claude-opus-5`** (1M ctx | $5.00 / $25.00) — Complex econometric and Bayesian model formulation.
 5. **`deepseek/deepseek-v4-pro-0813`** (1.05M ctx | $0.58 / $1.74) — Large-scale data ingestion and transformation scripts.
@@ -1033,7 +1034,7 @@ This cross-reference points each software and security workflow at the KXM pages
 *Domain:* Translates trace observations into isolated, runnable reproduction unit/integration tests and scripts.
 
 1. **`x-ai/grok-4.6`** (500k ctx | $2.00 / $6.00) — Fast, repo-accurate test writer; minimal boilerplate without hallucinating dependencies.
-2. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — Ingests full test frameworks; generates parameterized reproduction suites.
+2. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — Ingests full test frameworks; generates parameterized reproduction suites.
 3. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Specialized mocking of network calls, clocks, and internal database states.
 4. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Parameterized property-based regression test harnesses (Hypothesis/QuickCheck).
 5. **`mistralai/devstral-2512`** (262k ctx | $0.40 / $2.00) — Concise, executable unit-test reproductions.
@@ -1066,7 +1067,7 @@ This cross-reference points each software and security workflow at the KXM pages
 
 1. **`openai/gpt-5.3-codex`** (400k ctx | $1.75 / $14.00) — Surgical security patch generator using secure coding primitives.
 2. **`x-ai/grok-4.6`** (500k ctx | $2.00 / $6.00) — Rapid patch application that passes deterministic verification cleanly.
-3. **`qwen/qwen3-coder-plus`** (1M ctx | $0.65 / $3.25) — High-throughput defensive refactoring across large repos.
+3. **`qwen/qwen3.7-plus`** (1M ctx | $0.32 / $1.28) — High-throughput defensive refactoring across large repos.
 4. **`anthropic/claude-sonnet-5`** (1M ctx | $2.00 / $10.00) — Applies defensive validation without breaking valid user flows.
 5. **`bytedance-seed/seed-2.0-code`** (262k ctx | $0.50 / $3.00) — Web and client-side sanitization hardening.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kontextmind/kxm",
-  "version": "0.7.190",
+  "version": "0.7.192",
   "description": "KXM local-first multi-agent orchestration and operator dashboard",
   "type": "module",
   "author": "KontextMind",

package/plugins/kxm/.claude-plugin/plugin.json

@@ -2,7 +2,7 @@
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "kxm",
   "displayName": "KXM",
-  "version": "0.7.190",
+  "version": "0.7.192",
   "description": "Headless multi-agent orchestration, durable workflows, and a live operator dashboard for Pi and Claude Code",
   "author": {
     "name": "KontextMind",

package/plugins/kxm/dist/mcp-server.js

@@ -25462,7 +25462,7 @@ function sessionTokenFixHint(policy) {
 }
 
 // plugins/kxm/src/mcp-server.ts
-var VERSION = "0.7.190";
+var VERSION = "0.7.192";
 var CONFIGURE_PLUGIN = "/plugin configure kxm@kxm";
 var inbox = /* @__PURE__ */ new Map();
 var notifiedInbox = /* @__PURE__ */ new Set();

package/plugins/kxm/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-plugin",
-  "version": "0.7.190",
+  "version": "0.7.192",
   "private": true,
   "type": "module",
   "engines": {

package/plugins/kxm/src/mcp-server.ts

@@ -11,7 +11,7 @@ import { deliverInboxNotification } from "./inbox.ts";
 import type { HubEvent, MessageRecord } from "./protocol.ts";
 import { sessionTokenFixHint } from "./session-token-hint.ts";
 
-const VERSION = "0.7.190";
+const VERSION = "0.7.192";
 const CONFIGURE_PLUGIN = "/plugin configure kxm@kxm";
 const inbox = new Map<string, MessageRecord>();
 const notifiedInbox = new Set<string>();

package/scripts/harness-run.mjs

@@ -156,7 +156,7 @@ export const ROUTES = Object.freeze({
     provider: "anthropic",
     roles: Object.freeze(["planner", "reviewer-arch"]),
     permissions: Object.freeze(["read-only"]),
-    models: Object.freeze(["fable"]),
+    models: Object.freeze(["fable", "claude-sonnet-5.5"]),
     efforts: Object.freeze(["low", "medium", "high", "xhigh", "max"]),
     auth: Object.freeze({ args: ["auth", "status"], loginHint: "claude auth login" }),
   },