omp-llm-role 1.2.0 → 1.3.0

package/docs/dev/testing.md

@@ -137,11 +137,18 @@ passing.
     `filters.preferredMinThroughput: 50` (the registry row and the expanded def
     agree), an explicit `filters.preferredMinThroughput` beats the preset, and
     the schema carries the `roles.<role>.filters.preferredMinThroughput` row
-    (number, min 0, default 0).
+    (number, min 0, default 0). The `tokenProfile` setting (spec #81) defaults to
+    `null`, accepts a record with finite `inputTokens`/`outputTokens` ≥ 0, and
+    rejects a non-record or a negative/non-finite field; the schema carries the
+    `tokenProfile.inputTokens`/`tokenProfile.outputTokens` rows (number, min 0).
 23. **Explorer** (`explorer.test.ts`, `explorer-scopes.test.ts`) — rank deltas;
     contributions sum to `q`; `inverseCardinal` round-trips; Export preserves
     siblings and normalizes nested roles; the scope switcher; the availability
-    overlay.
+    overlay. The per-task price (spec #81): `rankRows`/`explainModel` fill
+    `perTask` from a supplied profile and leave it `null` without one; the
+    `POST /api/pool` toggle returns the gated/pruned dataset in `as-written` mode
+    and the full one in `raw`; the bootstrap payload carries `poolMode` and
+    `tokenProfile`; the explain response carries `leaderFlip`/`thinkingLevels`.
 24. **Project scope** (`project-updater.test.ts`, `project-setup.test.ts`,
     `project-roles-command.test.ts`) — project mode scopes config/state/history/
     lock to `<cwd>/.omp`; the day gate is per scope; the registry write;
@@ -224,6 +231,28 @@ passing.
     from the live model on `turn_start` (the silent-restore case), strips when
     neither source carries a pin, never calls `pi.setModel`, and contains a
     throwing live-model read.
+33. **Usage ledger** (`usage-ledger.test.ts`, `usage-ledger-wiring.test.ts`) —
+    `appendUsageRow` writes one row with the expected fields; `readUsageRows`
+    degrades to `[]` on a missing/malformed file and skips an unparseable or
+    wrongly-shaped line; the cap holds (the newest `USAGE_LEDGER_CAP` rows
+    survive a compaction); `roleTokenProfile` is the role's median per-turn
+    counts (or `null` with no rows) and `resolveTokenProfile` prefers the
+    recorded usage over the setting. The wiring (a fake `pi`/`ctx` with a
+    `sessionManager.getBranch()`) records one row per assistant turn on
+    `turn_end` and `session_shutdown`, dedupes by `ts|model` so the two hooks
+    cannot double-record, writes into `resolveProjectDir() ?? agentDir()`, and
+    contains a throwing branch read.
+34. **Explorer price levers** (`explorer-levers.test.ts`) — `leaderFlip` returns
+    the min-λ threshold over the strictly cheaper challengers, the current
+    `wPrice`, the distance and the challenger; `null` when no cheaper challenger
+    exists; the degenerate case (value leader = quality leader) is handled by the
+    same min formula, not `0/0`. `thinkingLevelEffect` returns one row per level
+    with the engine's own factor and the leader at that level, and `null` for a
+    bare role.
+35. **As-written pool** (`explorer-pool.test.ts`) — `asWrittenPool` applies the
+    catalog/tier gate (`resolveVariant`) and the allowed-providers prune
+    (`providerAllowed`), leaves the routes untouched on a `null` whitelist,
+    yields an empty pool on `tier: "none"`, and never mutates its input.
 
 ## Live checks
 
@@ -244,6 +273,18 @@ passing.
   derived metadata.
 - **Agent removal** — `/remove-agent --name <n>` in a session deletes both
   artifacts and the updater pass removes the config entries.
+- **Explorer pool / per-task / levers** — `/explore-roles` in a session with a
+  recorded usage ledger: confirm the header's pool toggle switches the table
+  between the raw and as-written pools (the model count and the leader change),
+  the `$/task` column renders a value (or `—` with no profile), and the explain
+  panel shows the per-task breakdown and the **Price levers** section (the
+  leader-flip line and the per-level leader table). Verify in headless Chromium
+  (the SPA is not unit-tested).
+- **Usage ledger** — run a real omp turn, then confirm `llm-role-usage.jsonl`
+  gained one row per assistant turn (the model, the role, the provider, the
+  token counts, the cost) in the plugin's state dir (`<project>/.omp` in project
+  mode, `agentDir()` otherwise), and that the explorer's `$/task` column then
+  fills for that role. A session with no completed turn writes no rows.
 - **CI/CD** — push a branch and open a PR (or push to `main`) and `gh run watch`
   the `test` job green; push a tag that disagrees with the package version and
   confirm the `release` job fails in the `notes` guard *before* any publish;

package/docs/dev/writes-and-state.md

@@ -134,6 +134,17 @@ gating — that this plugin never reads (zero occurrences in `src/`, `web/`,
 `tests/`). The plugin's `features.<id>` keys are plugin settings, not omp
 feature flags; do not merge the two.
 
+**`tokenProfile` (spec #81).** A global setting (`tokenProfile.inputTokens` /
+`tokenProfile.outputTokens`, flat dotted keys like the rest) — the per-task
+token profile the explorer's `$/task` column falls back to when a role has no
+recorded usage. `null` (the default) = unset, and the column reads `—`. It is
+validated to `null` or a record whose two fields are finite numbers ≥ 0
+(`tokenProfile: must be null or an object with inputTokens/outputTokens` /
+`tokenProfile.<key>: must be a number >= 0`). It is **display-only**: it never
+feeds the ranking, the reported `$/M` price, or any write. The recorded usage
+takes precedence over it (`resolveTokenProfile`), so once a role has ledger rows
+the setting is ignored for that role.
+
 ## The provider-pin auto-configuration (spec #58)
 
 When a role's **expanded** def sets `preferOwnProvider: true` (the
@@ -209,6 +220,7 @@ settings entry the global behavior is unchanged.
 |---|---|---|
 | `llm-role-state.json` | `{ lastRunDay, managedRoles, roleLastSelector, pluginWrittenChainKeys, previousModelRoles, managedDisabledAgents }` | day gate, managed-role set, last selectors, plugin-owned chain keys, pre-write `modelRoles` snapshot, **managed `task.disabledAgents` names** |
 | `llm-role-history.jsonl` | one JSON row per completed run: `{ ts, trigger, keyMeta{isFreeTier, limitRemaining, creditsRemaining}, decisions[] }` | append-only decision log |
+| `llm-role-usage.jsonl` | one JSON row per assistant turn: `{ ts, role, model, provider, input, output, cacheRead, cacheWrite, cost }` | the per-role usage ledger (spec #81) the explorer's per-task price resolves its token profile from |
 | `llm-role-projects.json` | `{ projects: [{ root, lastUsed }] }` | the plugin's memory of the projects where `/project-roles` was used (the explorer's scope list) |
 | `.llm-role-refresh.lock` | `O_EXCL` create, holds `"<pid> <iso>"` | serializes concurrent session starts; stale (> 60 s) locks are unlinked and retried once |
 | `llm-role-error-reports.json` | `{ reports: [{ fingerprint, label, title, firstSeen, lastSeen, createdDay?, offeredDay?, count, action, issueNumber?, issueUrl? }], dayCount: { day, count } }` | what the opt-in `errorReporting` path filed, offered or suppressed |
@@ -241,6 +253,24 @@ settings entry the global behavior is unchanged.
   that feeds the 3-creations-per-day cap. Nothing is written when
   `errorReporting` is `off`.
 
+- `llm-role-usage.jsonl` is the per-role usage ledger (spec #81,
+  `src/usage-ledger.ts`). Unlike the two ledgers above it is **not** global: it
+  follows the plugin's state location — `<cwd>/.omp` in project mode,
+  `agentDir()` otherwise — so a project's usage stays separate from the
+  user-level profile. The extension records one row per assistant turn on
+  `turn_end` and `session_shutdown` (the backstop for a headless one-turn run),
+  read from `sessionManager.getBranch()` — the same entries the session jsonl
+  persists — into `resolveProjectDir() ?? agentDir()`, deduped by `ts|model` so
+  the two hooks cannot double-record. The file is **append-only** (a crash never
+  corrupts it) and **bounded**: `appendUsageRow` compacts to the newest
+  `USAGE_LEDGER_CAP = 2000` rows via a temp file + atomic rename, so a crash
+  mid-compaction leaves the original intact. Recording is **best-effort** — a
+  failure never breaks a session, and `appendUsageRow` never throws. Reading is
+  tolerant: a missing, unreadable or malformed file degrades to `[]`, and an
+  unparseable or wrongly-shaped line is skipped. The explorer reads the active
+  scope's ledger (a project scope its own `<root>/.omp`), so the per-task column
+  follows the scope.
+
 - `previousModelRoles` is the full snapshot of the last `modelRoles` block
   **before** the plugin changed it — a manual rollback aid; there is no rollback
   command.

package/docs/llm-role-rankings.md

@@ -1,4 +1,4 @@
-llm-stats.com best-fit ranking per omp model role — 408 models, 2026-10-12
+llm-stats.com best-fit ranking per omp model role — 408 models, 2026-10-19
 Value ranking per role: each metric is cardinal-normalized with fixed anchors
 (no ranks): llm-stats index_* affine (v+20)/80 (interval scale, observed −16..+60),
 benchmarks chance-anchored pass rates, throughput log-anchored 10..300 tok/s.
@@ -21,179 +21,179 @@ agon=Design Arena `agents/agon_webapps` Elo from the same endpoint: context
 only, unweighted (22/87 of the designer pool, below the ~35–40% coverage
 bar at which a metric earns weight).
 ★ = Pareto-frontier: no eligible model is both cheaper and better (q).
-Throughput + price: OpenRouter per-provider routes under the default price-based routing (weight 1/price² over stable standard-tier providers; p50 tok/s of routed traffic; $/M 3:1 in:out), throughput 150/408, priced 148; models without OpenRouter throughput or a billed route are not ranked.
+Throughput + price: OpenRouter per-provider routes under the default price-based routing (weight 1/price² over stable standard-tier providers; p50 tok/s of routed traffic; $/M 3:1 in:out), throughput 149/408, priced 148; models without OpenRouter throughput or a billed route are not ranked.
 
 ## @default — Main workhorse: strong general coding-agent quality, sane cost
 eligible: 140 — λ 0.00556 $/quality-point (thinking ×1.857)
 
 |  # | value |     q | model                 | org         |   $/M | tok/s |  ctx | gen 54.7% | code 18.9% | ag 12.6% | tool 10.5% | tput 3.2% |
 | -: | ----: | ----: | --------------------- | ----------- | ----: | ----: | ---: | --------: | ---------: | -------: | ---------: | --------: |
-|  1 | 0.809 | 0.811 | ★ DeepSeek-V4.1-Flash | DeepSeek    | $0.43 |    81 | 1.0M |     0.484 |      0.147 |    0.093 |      0.068 |     0.020 |
-|  2 | 0.797 | 0.809 | GLM-5.3               | Zhipu AI    | $2.24 |    93 | 1.0M |     0.489 |      0.143 |    0.091 |      0.066 |     0.021 |
-|  3 | 0.796 | 0.817 | ★ Muse Spark 1.3      | Meta        | $3.71 |    76 | 1.0M |     0.498 |      0.143 |    0.092 |      0.065 |     0.019 |
-|  4 | 0.778 | 0.784 | MiMo-V2.6-Pro         | Xiaomi      | $1.01 |    28 | 1.0M |     0.478 |      0.142 |    0.090 |      0.065 |     0.010 |
-|  5 | 0.776 | 0.790 | DeepSeek-V4-Pro-0813  | DeepSeek    | $2.49 |    69 | 1.0M |     0.483 |      0.139 |    0.088 |      0.062 |     0.018 |
-|  6 | 0.766 | 0.829 | ★ GPT-5.6 Sol         | OpenAI      | $11.4 |    68 | 1.1M |     0.503 |      0.151 |    0.094 |      0.064 |     0.018 |
-|  7 | 0.765 | 0.809 | GPT-6.1 Sol           | OpenAI      | $7.96 |    35 | 1.1M |     0.506 |      0.143 |    0.093 |      0.056 |     0.012 |
-|  8 | 0.762 | 0.764 | ★ GLM-5.3-Flash       | Zhipu AI    | $0.36 |    65 | 1.0M |     0.472 |      0.124 |    0.087 |      0.064 |     0.017 |
-|  9 | 0.761 | 0.774 | Hy4 preview           | Tencent     | $2.32 |    32 |    — |     0.477 |      0.137 |    0.087 |      0.062 |     0.011 |
-| 10 | 0.759 | 0.808 | Kimi K3               | Moonshot AI | $8.92 |    42 | 1.0M |     0.493 |      0.145 |    0.091 |      0.066 |     0.013 |
+|  1 | 0.807 | 0.809 | ★ DeepSeek-V4.1-Flash | DeepSeek    | $0.37 |    75 | 1.0M |     0.483 |      0.146 |    0.093 |      0.069 |     0.019 |
+|  2 | 0.799 | 0.841 | ★ Gemini 4 Argon      | Google      | $7.65 |    38 |    — |     0.516 |      0.149 |    0.096 |      0.068 |     0.012 |
+|  3 | 0.794 | 0.806 | GLM-5.3               | Zhipu AI    | $2.09 |    73 | 1.0M |     0.488 |      0.142 |    0.090 |      0.067 |     0.018 |
+|  4 | 0.789 | 0.810 | ★ Muse Spark 1.3      | Meta        | $3.71 |    70 | 1.0M |     0.495 |      0.139 |    0.091 |      0.067 |     0.018 |
+|  5 | 0.776 | 0.793 | DeepSeek-V4-Pro-0813  | DeepSeek    | $3.09 |    78 | 1.0M |     0.484 |      0.139 |    0.088 |      0.063 |     0.019 |
+|  6 | 0.776 | 0.781 | MiMo-V2.6-Pro         | Xiaomi      | $1.01 |    29 | 1.0M |     0.476 |      0.141 |    0.089 |      0.066 |     0.010 |
+|  7 | 0.773 | 0.811 | ★ Kimi K3             | Moonshot AI | $6.81 |    52 | 1.0M |     0.494 |      0.144 |    0.091 |      0.067 |     0.015 |
+|  8 | 0.766 | 0.809 | GPT-6.1 Sol           | OpenAI      | $7.89 |    40 | 1.1M |     0.506 |      0.141 |    0.092 |      0.057 |     0.013 |
+|  9 | 0.764 | 0.777 | Hy4 preview           | Tencent     | $2.32 |    37 |    — |     0.478 |      0.137 |    0.087 |      0.063 |     0.012 |
+| 10 | 0.763 | 0.765 | ★ GLM-5.3-Flash       | Zhipu AI    | $0.33 |    66 | 1.0M |     0.473 |      0.123 |    0.087 |      0.065 |     0.018 |
 
 ## @smol — Fast lightweight model: cheap and quick, still competent
 eligible: 140 — λ 0.04091 $/quality-point
 
-|  # | value |     q | model                        | org                       |   $/M | tok/s |  ctx | tput 40% | gen 28.6% | code 17.1% | tool 14.3% |
-| -: | ----: | ----: | ---------------------------- | ------------------------- | ----: | ----: | ---: | -------: | --------: | ---------: | ---------: |
-|  1 | 0.715 | 0.724 | ★ DeepSeek-V4.1-Flash        | DeepSeek                  | $0.23 |    81 | 1.0M |    0.247 |     0.253 |      0.133 |      0.092 |
-|  2 | 0.691 | 0.692 | ★ Ling 3.0 Flash             | InclusionAI               | $0.04 |   169 | 131k |    0.332 |     0.200 |      0.095 |      0.064 |
-|  3 | 0.687 | 0.736 | ★ GLM-5.3                    | Zhipu AI                  | $1.20 |    93 | 1.0M |    0.262 |     0.255 |      0.130 |      0.090 |
-|  4 | 0.672 | 0.690 | DeepSeek-V4-Flash-Vision-Exp | DeepSeek                  | $0.43 |   100 | 1.0M |    0.270 |     0.231 |      0.115 |      0.074 |
-|  5 | 0.658 | 0.666 | GLM-5.3-Flash                | Zhipu AI                  | $0.19 |    65 | 1.0M |    0.221 |     0.246 |      0.113 |      0.086 |
-|  6 | 0.650 | 0.676 | Inkling-Small                | Thinking Machines Lab     | $0.64 |   123 | 524k |    0.295 |     0.207 |      0.104 |      0.070 |
-|  7 | 0.650 | 0.674 | Qwen3.8-27B                  | Alibaba Cloud / Qwen Team | $0.60 |    95 | 262k |    0.265 |     0.231 |      0.108 |      0.071 |
-|  8 | 0.648 | 0.667 | GPT-5.6 Luna                 | OpenAI                    | $0.47 |    80 | 1.1M |    0.245 |     0.229 |      0.117 |      0.076 |
-|  9 | 0.634 | 0.716 | Muse Spark 1.3               | Meta                      | $2.00 |    76 | 1.0M |    0.239 |     0.260 |      0.129 |      0.089 |
-| 10 | 0.634 | 0.689 | DeepSeek-V4-Pro-0813         | DeepSeek                  | $1.34 |    69 | 1.0M |    0.227 |     0.252 |      0.126 |      0.085 |
+|  # | value |     q | model                        | org                   |   $/M | tok/s |  ctx | tput 40% | gen 28.6% | code 17.1% | tool 14.3% |
+| -: | ----: | ----: | ---------------------------- | --------------------- | ----: | ----: | ---: | -------: | --------: | ---------: | ---------: |
+|  1 | 0.706 | 0.714 | ★ DeepSeek-V4.1-Flash        | DeepSeek              | $0.20 |    75 | 1.0M |    0.237 |     0.252 |      0.132 |      0.094 |
+|  2 | 0.687 | 0.713 | Inkling-Small                | Thinking Machines Lab | $0.64 |   167 | 524k |    0.331 |     0.208 |      0.104 |      0.070 |
+|  3 | 0.669 | 0.671 | ★ Ling 3.0 Flash             | InclusionAI           | $0.04 |   140 | 131k |    0.310 |     0.200 |      0.096 |      0.065 |
+|  4 | 0.662 | 0.744 | ★ Muse Spark 1.1             | Meta                  | $2.00 |   127 | 1.0M |    0.298 |     0.245 |      0.113 |      0.087 |
+|  5 | 0.661 | 0.669 | GLM-5.3-Flash                | Zhipu AI              | $0.18 |    66 | 1.0M |    0.223 |     0.247 |      0.112 |      0.088 |
+|  6 | 0.661 | 0.707 | GLM-5.3                      | Zhipu AI              | $1.12 |    73 | 1.0M |    0.233 |     0.255 |      0.129 |      0.091 |
+|  7 | 0.652 | 0.670 | DeepSeek-V4-Flash-Vision-Exp | DeepSeek              | $0.43 |    83 | 1.0M |    0.249 |     0.231 |      0.115 |      0.075 |
+|  8 | 0.652 | 0.733 | Muse Spark 1.2               | Meta                  | $2.00 |   224 | 1.0M |    0.366 |     0.204 |      0.102 |      0.062 |
+|  9 | 0.647 | 0.667 | GPT-5.6 Luna                 | OpenAI                | $0.47 |    80 | 1.1M |    0.244 |     0.229 |      0.117 |      0.076 |
+| 10 | 0.643 | 0.707 | Gemini 3.8 Flash             | Google                | $1.57 |    95 | 1.0M |    0.264 |     0.247 |      0.119 |      0.076 |
 
 ## @slow — Most capable model for hard problems; cost and speed as tiebreakers
 eligible: 140 — λ 0.00263 $/quality-point (thinking ×2.714)
 
 |  # | value |     q | model                 | org         |   $/M | tok/s |  ctx | gen 63.2% | code 18.9% | ag 13.7% | tput 4.2% |
 | -: | ----: | ----: | --------------------- | ----------- | ----: | ----: | ---: | --------: | ---------: | -------: | --------: |
-|  1 | 0.830 | 0.832 | ★ DeepSeek-V4.1-Flash | DeepSeek    | $0.63 |    81 | 1.0M |     0.558 |      0.147 |    0.101 |     0.026 |
-|  2 | 0.828 | 0.842 | ★ Muse Spark 1.3      | Meta        | $5.43 |    76 | 1.0M |     0.575 |      0.143 |    0.099 |     0.025 |
-|  3 | 0.824 | 0.833 | ★ GLM-5.3             | Zhipu AI    | $3.27 |    93 | 1.0M |     0.564 |      0.143 |    0.098 |     0.028 |
-|  4 | 0.812 | 0.856 | ★ GPT-5.6 Sol         | OpenAI      | $16.6 |    68 | 1.1M |     0.580 |      0.151 |    0.101 |     0.024 |
-|  5 | 0.812 | 0.843 | ★ GPT-6.1 Sol         | OpenAI      | $11.6 |    35 | 1.1M |     0.584 |      0.143 |    0.100 |     0.015 |
-|  6 | 0.805 | 0.815 | DeepSeek-V4-Pro-0813  | DeepSeek    | $3.64 |    69 | 1.0M |     0.557 |      0.139 |    0.095 |     0.024 |
-|  7 | 0.799 | 0.803 | MiMo-V2.6-Pro         | Xiaomi      | $1.47 |    28 | 1.0M |     0.551 |      0.142 |    0.097 |     0.013 |
-|  8 | 0.796 | 0.830 | Kimi K3               | Moonshot AI |   $13 |    42 | 1.0M |     0.569 |      0.145 |    0.098 |     0.018 |
-|  9 | 0.787 | 0.796 | Hy4 preview           | Tencent     | $3.39 |    32 |    — |     0.551 |      0.137 |    0.094 |     0.014 |
-| 10 | 0.785 | 0.786 | ★ GLM-5.3-Flash       | Zhipu AI    | $0.53 |    65 | 1.0M |     0.544 |      0.124 |    0.094 |     0.023 |
+|  1 | 0.835 | 0.865 | ★ Gemini 4 Argon      | Google      | $11.2 |    38 |    — |     0.595 |      0.149 |    0.104 |     0.016 |
+|  2 | 0.827 | 0.829 | ★ DeepSeek-V4.1-Flash | DeepSeek    | $0.54 |    75 | 1.0M |     0.558 |      0.146 |    0.100 |     0.025 |
+|  3 | 0.819 | 0.827 | GLM-5.3               | Zhipu AI    | $3.05 |    73 | 1.0M |     0.563 |      0.142 |    0.098 |     0.025 |
+|  4 | 0.818 | 0.833 | ★ Muse Spark 1.3      | Meta        | $5.43 |    70 | 1.0M |     0.572 |      0.139 |    0.099 |     0.024 |
+|  5 | 0.812 | 0.842 | GPT-6.1 Sol           | OpenAI      | $11.5 |    40 | 1.1M |     0.584 |      0.141 |    0.100 |     0.017 |
+|  6 | 0.806 | 0.833 | Kimi K3               | Moonshot AI | $9.95 |    52 | 1.0M |     0.570 |      0.144 |    0.098 |     0.020 |
+|  7 | 0.806 | 0.818 | DeepSeek-V4-Pro-0813  | DeepSeek    | $4.51 |    78 | 1.0M |     0.559 |      0.139 |    0.095 |     0.025 |
+|  8 | 0.796 | 0.800 | MiMo-V2.6-Pro         | Xiaomi      | $1.47 |    29 | 1.0M |     0.549 |      0.141 |    0.097 |     0.013 |
+|  9 | 0.792 | 0.852 | GPT-5.6 Sol           | OpenAI      | $22.9 |    54 | 1.1M |     0.580 |      0.150 |    0.101 |     0.021 |
+| 10 | 0.790 | 0.798 | Hy4 preview           | Tencent     | $3.39 |    37 |    — |     0.552 |      0.137 |    0.094 |     0.016 |
 
 ## @vision — Image understanding: vision index dominates
-eligible: 72 — λ 0.00682 $/quality-point (thinking ×1.857)
+eligible: 73 — λ 0.00682 $/quality-point (thinking ×1.857)
 
 |  # | value |     q | model                 | org                       |   $/M | tok/s |  ctx | vis 54.2% | gen 20.8% | rea 15.6% | code 6.3% | tput 3.1% |
 | -: | ----: | ----: | --------------------- | ------------------------- | ----: | ----: | ---: | --------: | --------: | --------: | --------: | --------: |
-|  1 | 0.728 | 0.731 | ★ DeepSeek-V4.1-Flash | DeepSeek                  | $0.43 |    81 | 1.0M |     0.346 |     0.184 |     0.133 |     0.048 |     0.019 |
+|  1 | 0.733 | 0.786 | ★ Gemini 4 Argon      | Google                    | $7.65 |    38 |    — |     0.385 |     0.196 |     0.143 |     0.049 |     0.012 |
 |  2 | 0.728 | 0.731 | ★ Qwen3.8 Flash       | Alibaba Cloud / Qwen Team | $0.43 |    47 | 1.0M |     0.363 |     0.178 |     0.133 |     0.042 |     0.014 |
-|  3 | 0.717 | 0.772 | ★ GPT-6.1 Sol         | OpenAI                    | $7.96 |    35 | 1.1M |     0.380 |     0.192 |     0.140 |     0.047 |     0.011 |
-|  4 | 0.715 | 0.741 | ★ Muse Spark 1.1      | Meta                      | $3.71 |    48 | 1.0M |     0.371 |     0.179 |     0.135 |     0.041 |     0.014 |
-|  5 | 0.712 | 0.715 | ★ GLM-5.3-Flash       | Zhipu AI                  | $0.36 |    65 | 1.0M |     0.344 |     0.179 |     0.132 |     0.041 |     0.017 |
-|  6 | 0.711 | 0.730 | Gemini 3.8 Flash      | Google                    | $2.79 |    74 | 1.0M |     0.361 |     0.179 |     0.128 |     0.044 |     0.018 |
-|  7 | 0.705 | 0.766 | Kimi K3               | Moonshot AI               | $8.92 |    42 | 1.0M |     0.379 |     0.188 |     0.139 |     0.048 |     0.013 |
-|  8 | 0.705 | 0.712 | Qwen3.8-27B           | Alibaba Cloud / Qwen Team | $1.11 |    95 | 262k |     0.358 |     0.168 |     0.126 |     0.039 |     0.021 |
-|  9 | 0.695 | 0.714 | Gemini 3.7 Flash      | Google                    | $2.79 |    80 | 1.0M |     0.353 |     0.174 |     0.127 |     0.041 |     0.019 |
-| 10 | 0.692 | 0.770 | GPT-5.6 Sol           | OpenAI                    | $11.4 |    68 | 1.1M |     0.369 |     0.191 |     0.142 |     0.050 |     0.018 |
+|  3 | 0.727 | 0.730 | ★ DeepSeek-V4.1-Flash | DeepSeek                  | $0.37 |    75 | 1.0M |     0.345 |     0.184 |     0.134 |     0.048 |     0.018 |
+|  4 | 0.724 | 0.750 | ★ Muse Spark 1.1      | Meta                      | $3.71 |   127 | 1.0M |     0.371 |     0.179 |     0.136 |     0.041 |     0.023 |
+|  5 | 0.722 | 0.768 | ★ Kimi K3             | Moonshot AI               | $6.81 |    52 | 1.0M |     0.378 |     0.188 |     0.139 |     0.048 |     0.015 |
+|  6 | 0.715 | 0.769 | GPT-6.1 Sol           | OpenAI                    | $7.89 |    40 | 1.1M |     0.377 |     0.193 |     0.140 |     0.047 |     0.013 |
+|  7 | 0.714 | 0.734 | ★ Gemini 3.8 Flash    | Google                    | $2.91 |    95 | 1.0M |     0.361 |     0.180 |     0.129 |     0.043 |     0.021 |
+|  8 | 0.713 | 0.715 | ★ GLM-5.3-Flash       | Zhipu AI                  | $0.33 |    66 | 1.0M |     0.344 |     0.180 |     0.133 |     0.041 |     0.017 |
+|  9 | 0.702 | 0.707 | Qwen3.8-27B           | Alibaba Cloud / Qwen Team | $0.72 |    55 | 262k |     0.358 |     0.168 |     0.126 |     0.039 |     0.016 |
+| 10 | 0.697 | 0.717 | Gemini 3.7 Flash      | Google                    | $2.87 |    86 | 1.0M |     0.353 |     0.176 |     0.127 |     0.041 |     0.020 |
 
 ## @plan — Planning: reasoning and long-context coherence, priced so the cheapest capable planner wins
 eligible: 140 — λ 0.01667 $/quality-point (thinking ×1.857)
 
-|  # | value |     q | model               | org                       |   $/M | tok/s |  ctx | gen 80% | lc 15.6% | tput 4.4% |
-| -: | ----: | ----: | ------------------- | ------------------------- | ----: | ----: | ---: | ------: | -------: | --------: |
-|  1 | 0.794 | 0.856 | ★ Muse Spark 1.3    | Meta                      | $3.71 |    76 | 1.0M |   0.728 |    0.101 |     0.026 |
-|  2 | 0.775 | 0.782 | ★ Qwen3.8 Flash     | Alibaba Cloud / Qwen Team | $0.43 |    47 | 1.0M |   0.683 |    0.079 |     0.020 |
-|  3 | 0.758 | 0.765 | DeepSeek-V4.1-Flash | DeepSeek                  | $0.43 |    81 | 1.0M |   0.707 |    0.030 |     0.027 |
-|  4 | 0.758 | 0.804 | ★ Gemini 3.8 Flash  | Google                    | $2.79 |    74 | 1.0M |   0.687 |    0.091 |     0.026 |
-|  5 | 0.749 | 0.788 | ★ Hy4 preview       | Tencent                   | $2.32 |    32 |    — |   0.698 |    0.075 |     0.015 |
-|  6 | 0.747 | 0.794 | Gemini 3.7 Flash    | Google                    | $2.79 |    80 | 1.0M |   0.669 |    0.098 |     0.027 |
-|  7 | 0.738 | 0.744 | ★ GLM-5.3-Flash     | Zhipu AI                  | $0.36 |    65 | 1.0M |   0.689 |    0.030 |     0.025 |
-|  8 | 0.737 | 0.774 | GLM-5.3             | Zhipu AI                  | $2.24 |    93 | 1.0M |   0.715 |    0.030 |     0.029 |
-|  9 | 0.735 | 0.750 | GPT-5.6 Luna        | OpenAI                    | $0.88 |    80 | 1.1M |   0.642 |    0.080 |     0.027 |
-| 10 | 0.725 | 0.742 | MiMo-V2.6-Pro       | Xiaomi                    | $1.01 |    28 | 1.0M |   0.698 |    0.030 |     0.014 |
+|  # | value |     q | model                 | org                       |   $/M | tok/s |  ctx | gen 80% | lc 15.6% | tput 4.4% |
+| -: | ----: | ----: | --------------------- | ------------------------- | ----: | ----: | ---: | ------: | -------: | --------: |
+|  1 | 0.789 | 0.851 | ★ Muse Spark 1.3      | Meta                      | $3.71 |    70 | 1.0M |   0.724 |    0.101 |     0.025 |
+|  2 | 0.774 | 0.781 | ★ Qwen3.8 Flash       | Alibaba Cloud / Qwen Team | $0.43 |    47 | 1.0M |   0.682 |    0.079 |     0.020 |
+|  3 | 0.764 | 0.813 | ★ Gemini 3.8 Flash    | Google                    | $2.91 |    95 | 1.0M |   0.692 |    0.091 |     0.029 |
+|  4 | 0.757 | 0.763 | ★ DeepSeek-V4.1-Flash | DeepSeek                  | $0.37 |    75 | 1.0M |   0.707 |    0.030 |     0.026 |
+|  5 | 0.755 | 0.802 | ★ Gemini 3.7 Flash    | Google                    | $2.87 |    86 | 1.0M |   0.676 |    0.098 |     0.028 |
+|  6 | 0.752 | 0.790 | ★ Hy4 preview         | Tencent                   | $2.32 |    37 |    — |   0.699 |    0.075 |     0.017 |
+|  7 | 0.742 | 0.870 | ★ Gemini 4 Argon      | Google                    | $7.65 |    38 |    — |   0.754 |    0.098 |     0.017 |
+|  8 | 0.740 | 0.746 | ★ GLM-5.3-Flash       | Zhipu AI                  | $0.33 |    66 | 1.0M |   0.691 |    0.030 |     0.025 |
+|  9 | 0.735 | 0.749 | GPT-5.6 Luna          | OpenAI                    | $0.88 |    80 | 1.1M |   0.642 |    0.080 |     0.027 |
+| 10 | 0.735 | 0.769 | GLM-5.3               | Zhipu AI                  | $2.09 |    73 | 1.0M |   0.713 |    0.030 |     0.026 |
 
 ## @commit — Commit messages: cheap and fast with decent general quality
 eligible: 140 — λ 0.04091 $/quality-point
 
-|  # | value |     q | model                        | org                       |   $/M | tok/s |  ctx | tput 41.5% | gen 43.1% | code 15.4% |
-| -: | ----: | ----: | ---------------------------- | ------------------------- | ----: | ----: | ---: | ---------: | --------: | ---------: |
-|  1 | 0.747 | 0.756 | ★ DeepSeek-V4.1-Flash        | DeepSeek                  | $0.23 |    81 | 1.0M |      0.256 |     0.381 |      0.119 |
-|  2 | 0.731 | 0.733 | ★ Ling 3.0 Flash             | InclusionAI               | $0.04 |   169 | 131k |      0.345 |     0.302 |      0.086 |
-|  3 | 0.724 | 0.773 | ★ GLM-5.3                    | Zhipu AI                  | $1.20 |    93 | 1.0M |      0.272 |     0.385 |      0.116 |
-|  4 | 0.715 | 0.732 | DeepSeek-V4-Flash-Vision-Exp | DeepSeek                  | $0.43 |   100 | 1.0M |      0.281 |     0.348 |      0.103 |
-|  5 | 0.695 | 0.719 | Qwen3.8-27B                  | Alibaba Cloud / Qwen Team | $0.60 |    95 | 262k |      0.275 |     0.348 |      0.097 |
-|  6 | 0.694 | 0.702 | GLM-5.3-Flash                | Zhipu AI                  | $0.19 |    65 | 1.0M |      0.229 |     0.371 |      0.101 |
-|  7 | 0.687 | 0.713 | Inkling-Small                | Thinking Machines Lab     | $0.64 |   123 | 524k |      0.307 |     0.313 |      0.093 |
-|  8 | 0.686 | 0.706 | GPT-5.6 Luna                 | OpenAI                    | $0.47 |    80 | 1.1M |      0.254 |     0.346 |      0.105 |
-|  9 | 0.678 | 0.693 | Mercury 2                    | Inception                 | $0.38 |   258 | 128k |      0.397 |     0.234 |      0.062 |
-| 10 | 0.676 | 0.747 | GLM-5.2                      | Zhipu AI                  | $1.72 |   111 | 1.0M |      0.294 |     0.349 |      0.103 |
+|  # | value |     q | model                        | org                   |   $/M | tok/s |  ctx | tput 41.5% | gen 43.1% | code 15.4% |
+| -: | ----: | ----: | ---------------------------- | --------------------- | ----: | ----: | ---: | ---------: | --------: | ---------: |
+|  1 | 0.736 | 0.745 | ★ DeepSeek-V4.1-Flash        | DeepSeek              | $0.20 |    75 | 1.0M |      0.246 |     0.380 |      0.118 |
+|  2 | 0.724 | 0.750 | ★ Inkling-Small              | Thinking Machines Lab | $0.64 |   167 | 524k |      0.344 |     0.313 |      0.094 |
+|  3 | 0.709 | 0.710 | ★ Ling 3.0 Flash             | InclusionAI           | $0.04 |   140 | 131k |      0.322 |     0.302 |      0.086 |
+|  4 | 0.699 | 0.781 | ★ Muse Spark 1.1             | Meta                  | $2.00 |   127 | 1.0M |      0.310 |     0.370 |      0.101 |
+|  5 | 0.697 | 0.779 | Muse Spark 1.2               | Meta                  | $2.00 |   224 | 1.0M |      0.380 |     0.308 |      0.091 |
+|  6 | 0.696 | 0.703 | GLM-5.3-Flash                | Zhipu AI              | $0.18 |    66 | 1.0M |      0.231 |     0.372 |      0.100 |
+|  7 | 0.695 | 0.741 | GLM-5.3                      | Zhipu AI              | $1.12 |    73 | 1.0M |      0.242 |     0.384 |      0.115 |
+|  8 | 0.692 | 0.710 | DeepSeek-V4-Flash-Vision-Exp | DeepSeek              | $0.43 |    83 | 1.0M |      0.259 |     0.348 |      0.103 |
+|  9 | 0.690 | 0.754 | ★ Gemini 3.8 Flash           | Google                | $1.57 |    95 | 1.0M |      0.275 |     0.373 |      0.106 |
+| 10 | 0.685 | 0.704 | GPT-5.6 Luna                 | OpenAI                | $0.47 |    80 | 1.1M |      0.254 |     0.346 |      0.105 |
 
 ## @tiny — Background tasks (titles, memory): cheapest and fastest wins
 eligible: 140 — λ 0.03333 $/quality-point
 
-|  # | value |     q | model                        | org         |   $/M | tok/s |  ctx | tput 58.3% | gen 41.7% |
-| -: | ----: | ----: | ---------------------------- | ----------- | ----: | ----: | ---: | ---------: | --------: |
-|  1 | 0.775 | 0.777 | ★ Ling 3.0 Flash             | InclusionAI | $0.04 |   169 | 131k |      0.485 |     0.292 |
-|  2 | 0.771 | 0.784 | ★ Mercury 2                  | Inception   | $0.38 |   258 | 128k |      0.557 |     0.226 |
-|  3 | 0.766 | 0.768 | Ling 3.0 Flash Fin           | InclusionAI | $0.06 |   132 | 262k |      0.443 |     0.325 |
-|  4 | 0.730 | 0.759 | Gemini 3.5 Flash-Lite        | Google      | $0.89 |   187 | 1.0M |      0.502 |     0.258 |
-|  5 | 0.728 | 0.746 | Gemini 3.1 Flash-Lite        | Google      | $0.56 |   166 | 1.0M |      0.482 |     0.264 |
-|  6 | 0.720 | 0.728 | DeepSeek-V4.1-Flash          | DeepSeek    | $0.23 |    81 | 1.0M |      0.360 |     0.368 |
-|  7 | 0.718 | 0.720 | Laguna XS 2.1                | Poolside    | $0.07 |   174 | 262k |      0.490 |     0.231 |
-|  8 | 0.717 | 0.731 | DeepSeek-V4-Flash-Vision-Exp | DeepSeek    | $0.43 |   100 | 1.0M |      0.394 |     0.337 |
-|  9 | 0.714 | 0.754 | GLM-5.3                      | Zhipu AI    | $1.20 |    93 | 1.0M |      0.382 |     0.372 |
-| 10 | 0.713 | 0.716 | Nemotron 3 Nano (30B A3B)    | NVIDIA      | $0.09 |   187 | 262k |      0.502 |     0.213 |
+|  # | value |     q | model                         | org                       |   $/M | tok/s |  ctx | tput 58.3% | gen 41.7% |
+| -: | ----: | ----: | ----------------------------- | ------------------------- | ----: | ----: | ---: | ---------: | --------: |
+|  1 | 0.769 | 0.782 | ★ Mercury 2                   | Inception                 | $0.38 |   255 | 128k |      0.555 |     0.227 |
+|  2 | 0.765 | 0.831 | ★ Muse Spark 1.2              | Meta                      | $2.00 |   224 | 1.0M |      0.533 |     0.298 |
+|  3 | 0.764 | 0.786 | ★ Inkling-Small               | Thinking Machines Lab     | $0.64 |   167 | 524k |      0.483 |     0.303 |
+|  4 | 0.743 | 0.745 | ★ Ling 3.0 Flash              | InclusionAI               | $0.04 |   140 | 131k |      0.453 |     0.292 |
+|  5 | 0.738 | 0.756 | Gemini 3.1 Flash-Lite         | Google                    | $0.56 |   176 | 1.0M |      0.492 |     0.265 |
+|  6 | 0.726 | 0.793 | Muse Spark 1.1                | Meta                      | $2.00 |   127 | 1.0M |      0.435 |     0.358 |
+|  7 | 0.717 | 0.768 | Gemini 3.6 Flash              | Google                    | $1.54 |   132 | 1.0M |      0.443 |     0.325 |
+|  8 | 0.707 | 0.713 | DeepSeek-V4.1-Flash           | DeepSeek                  | $0.20 |    75 | 1.0M |      0.345 |     0.368 |
+|  9 | 0.705 | 0.707 | Laguna XS 2.1                 | Poolside                  | $0.07 |   161 | 262k |      0.477 |     0.231 |
+| 10 | 0.702 | 0.743 | Qwen3-235B-A22B-Thinking-2507 | Alibaba Cloud / Qwen Team | $1.21 |   177 |    — |      0.493 |     0.250 |
 
 ## @task — Subagents: agentic + tool calling, moderate cost sensitivity
 eligible: 140 — λ 0.00747 $/quality-point (thinking ×1.857)
 
 |  # | value |     q | model                 | org                       |   $/M | tok/s |  ctx | ag 29.9% | tool 20.7% | code 18.4% | gen 25.3% | tput 5.7% |
 | -: | ----: | ----: | --------------------- | ------------------------- | ----: | ----: | ---: | -------: | ---------: | ---------: | --------: | --------: |
-|  1 | 0.752 | 0.755 | ★ DeepSeek-V4.1-Flash | DeepSeek                  | $0.43 |    81 | 1.0M |    0.220 |      0.134 |      0.142 |     0.224 |     0.035 |
-|  2 | 0.730 | 0.746 | GLM-5.3               | Zhipu AI                  | $2.24 |    93 | 1.0M |    0.214 |      0.130 |      0.139 |     0.226 |     0.038 |
-|  3 | 0.721 | 0.748 | Muse Spark 1.3        | Meta                      | $3.71 |    76 | 1.0M |    0.217 |      0.128 |      0.139 |     0.230 |     0.034 |
-|  4 | 0.709 | 0.716 | MiMo-V2.6-Pro         | Xiaomi                    | $1.01 |    28 | 1.0M |    0.212 |      0.128 |      0.138 |     0.221 |     0.018 |
-|  5 | 0.701 | 0.720 | DeepSeek-V4-Pro-0813  | DeepSeek                  | $2.49 |    69 | 1.0M |    0.207 |      0.122 |      0.135 |     0.223 |     0.033 |
-|  6 | 0.698 | 0.701 | ★ GLM-5.3-Flash       | Zhipu AI                  | $0.36 |    65 | 1.0M |    0.206 |      0.125 |      0.121 |     0.218 |     0.032 |
-|  7 | 0.688 | 0.692 | Qwen3.8 Flash         | Alibaba Cloud / Qwen Team | $0.43 |    47 | 1.0M |    0.199 |      0.127 |      0.124 |     0.216 |     0.026 |
-|  8 | 0.683 | 0.700 | Hy4 preview           | Tencent                   | $2.32 |    32 |    — |    0.205 |      0.123 |      0.133 |     0.221 |     0.019 |
-|  9 | 0.674 | 0.759 | ★ GPT-5.6 Sol         | OpenAI                    | $11.4 |    68 | 1.1M |    0.222 |      0.126 |      0.146 |     0.232 |     0.032 |
-| 10 | 0.671 | 0.738 | Kimi K3               | Moonshot AI               | $8.92 |    42 | 1.0M |    0.215 |      0.131 |      0.141 |     0.228 |     0.024 |
+|  1 | 0.750 | 0.753 | ★ DeepSeek-V4.1-Flash | DeepSeek                  | $0.37 |    75 | 1.0M |    0.219 |      0.136 |      0.141 |     0.223 |     0.034 |
+|  2 | 0.726 | 0.742 | GLM-5.3               | Zhipu AI                  | $2.09 |    73 | 1.0M |    0.213 |      0.131 |      0.138 |     0.225 |     0.033 |
+|  3 | 0.716 | 0.743 | Muse Spark 1.3        | Meta                      | $3.71 |    70 | 1.0M |    0.215 |      0.132 |      0.134 |     0.229 |     0.033 |
+|  4 | 0.710 | 0.767 | ★ Gemini 4 Argon      | Google                    | $7.65 |    38 |    — |    0.228 |      0.134 |      0.145 |     0.238 |     0.022 |
+|  5 | 0.707 | 0.715 | MiMo-V2.6-Pro         | Xiaomi                    | $1.01 |    29 | 1.0M |    0.211 |      0.129 |      0.137 |     0.220 |     0.018 |
+|  6 | 0.702 | 0.725 | DeepSeek-V4-Pro-0813  | DeepSeek                  | $3.09 |    78 | 1.0M |    0.208 |      0.124 |      0.135 |     0.224 |     0.035 |
+|  7 | 0.700 | 0.703 | ★ GLM-5.3-Flash       | Zhipu AI                  | $0.33 |    66 | 1.0M |    0.206 |      0.127 |      0.120 |     0.218 |     0.032 |
+|  8 | 0.692 | 0.743 | Kimi K3               | Moonshot AI               | $6.81 |    52 | 1.0M |    0.215 |      0.132 |      0.140 |     0.228 |     0.028 |
+|  9 | 0.688 | 0.691 | Qwen3.8 Flash         | Alibaba Cloud / Qwen Team | $0.43 |    47 | 1.0M |    0.198 |      0.127 |      0.124 |     0.216 |     0.026 |
+| 10 | 0.687 | 0.705 | Hy4 preview           | Tencent                   | $2.32 |    37 |    — |    0.205 |      0.124 |      0.133 |     0.221 |     0.022 |
 
 ## @advisor — Advisor/watchdog: deep reasoning over long context
 eligible: 140 — λ 0.01250 $/quality-point (thinking ×1.857)
 
 |  # | value |     q | model                 | org                       |   $/M | tok/s |  ctx | gen 74.3% | lc 12.6% | tput 13.1% |
 | -: | ----: | ----: | --------------------- | ------------------------- | ----: | ----: | ---: | --------: | -------: | ---------: |
-|  1 | 0.790 | 0.837 | ★ Muse Spark 1.3      | Meta                      | $3.71 |    76 | 1.0M |     0.677 |    0.082 |      0.078 |
-|  2 | 0.757 | 0.763 | ★ DeepSeek-V4.1-Flash | DeepSeek                  | $0.43 |    81 | 1.0M |     0.657 |    0.024 |      0.081 |
-|  3 | 0.754 | 0.789 | ★ Gemini 3.8 Flash    | Google                    | $2.79 |    74 | 1.0M |     0.638 |    0.073 |      0.077 |
-|  4 | 0.752 | 0.758 | ★ Qwen3.8 Flash       | Alibaba Cloud / Qwen Team | $0.43 |    47 | 1.0M |     0.635 |    0.063 |      0.060 |
-|  5 | 0.746 | 0.774 | ★ GLM-5.3             | Zhipu AI                  | $2.24 |    93 | 1.0M |     0.664 |    0.024 |      0.086 |
-|  6 | 0.746 | 0.780 | Gemini 3.7 Flash      | Google                    | $2.79 |    80 | 1.0M |     0.621 |    0.079 |      0.080 |
-|  7 | 0.733 | 0.737 | ★ GLM-5.3-Flash       | Zhipu AI                  | $0.36 |    65 | 1.0M |     0.640 |    0.024 |      0.072 |
-|  8 | 0.731 | 0.742 | GPT-5.6 Luna          | OpenAI                    | $0.88 |    80 | 1.1M |     0.597 |    0.065 |      0.080 |
-|  9 | 0.724 | 0.753 | Hy4 preview           | Tencent                   | $2.32 |    32 |    — |     0.649 |    0.060 |      0.044 |
-| 10 | 0.723 | 0.755 | DeepSeek-V4-Pro-0813  | DeepSeek                  | $2.49 |    69 | 1.0M |     0.656 |    0.024 |      0.074 |
+|  1 | 0.783 | 0.830 | ★ Muse Spark 1.3      | Meta                      | $3.71 |    70 | 1.0M |     0.673 |    0.082 |      0.075 |
+|  2 | 0.767 | 0.803 | ★ Gemini 3.8 Flash    | Google                    | $2.91 |    95 | 1.0M |     0.643 |    0.073 |      0.087 |
+|  3 | 0.755 | 0.790 | ★ Gemini 3.7 Flash    | Google                    | $2.87 |    86 | 1.0M |     0.628 |    0.079 |      0.083 |
+|  4 | 0.754 | 0.759 | ★ DeepSeek-V4.1-Flash | DeepSeek                  | $0.37 |    75 | 1.0M |     0.657 |    0.024 |      0.078 |
+|  5 | 0.752 | 0.757 | Qwen3.8 Flash         | Alibaba Cloud / Qwen Team | $0.43 |    47 | 1.0M |     0.634 |    0.063 |      0.060 |
+|  6 | 0.737 | 0.764 | ★ GLM-5.3             | Zhipu AI                  | $2.09 |    73 | 1.0M |     0.663 |    0.024 |      0.076 |
+|  7 | 0.735 | 0.831 | ★ Gemini 4 Argon      | Google                    | $7.65 |    38 |    — |     0.700 |    0.079 |      0.051 |
+|  8 | 0.735 | 0.739 | ★ GLM-5.3-Flash       | Zhipu AI                  | $0.33 |    66 | 1.0M |     0.642 |    0.024 |      0.073 |
+|  9 | 0.731 | 0.760 | Hy4 preview           | Tencent                   | $2.32 |    37 |    — |     0.649 |    0.060 |      0.050 |
+| 10 | 0.730 | 0.741 | GPT-5.6 Luna          | OpenAI                    | $0.88 |    80 | 1.1M |     0.597 |    0.065 |      0.080 |
 
 ## @designer — Design work: visual/UX judgement on image-capable models
-eligible: 86 — λ 0.01024 $/quality-point (thinking ×1.857)
+eligible: 87 — λ 0.01024 $/quality-point (thinking ×1.857)
 
 |  # | value |     q | model                 | org         |   $/M | tok/s |  ctx | gen 31.3% | code 7.2% | vis 21.7% | tput 18.1% | web 21.7% | agon |
 | -: | ----: | ----: | --------------------- | ----------- | ----: | ----: | ---: | --------: | --------: | --------: | ---------: | --------: | ---: |
-|  1 | 0.781 | 0.785 | ★ DeepSeek-V4.1-Flash | DeepSeek    | $0.43 |    81 | 1.0M |     0.277 |     0.056 |     0.139 |      0.111 |     0.202 |    — |
-|  2 | 0.751 | 0.789 | ★ Muse Spark 1.3      | Meta        | $3.71 |    76 | 1.0M |     0.285 |     0.054 |     0.125 |      0.108 |     0.216 | 1289 |
-|  3 | 0.735 | 0.763 | Gemini 3.8 Flash      | Google      | $2.79 |    74 | 1.0M |     0.269 |     0.051 |     0.144 |      0.106 |     0.193 | 1231 |
-|  4 | 0.727 | 0.756 | Gemini 3.7 Flash      | Google      | $2.79 |    80 | 1.0M |     0.262 |     0.048 |     0.141 |      0.110 |     0.195 | 1215 |
-|  5 | 0.721 | 0.724 | ★ GLM-5.3-Flash       | Zhipu AI    | $0.36 |    65 | 1.0M |     0.270 |     0.047 |     0.138 |      0.100 |     0.169 |    — |
-|  6 | 0.713 | 0.742 | Gemini 3.6 Flash      | Google      | $2.79 |   101 | 1.0M |     0.245 |     0.042 |     0.140 |      0.123 |     0.192 | 1188 |
-|  7 | 0.697 | 0.707 | MiMo-V2.6-Pro         | Xiaomi      | $1.01 |    28 | 1.0M |     0.273 |     0.054 |     0.127 |      0.055 |     0.197 |    — |
-|  8 | 0.688 | 0.770 | GPT-6.1 Sol           | OpenAI      | $7.96 |    35 | 1.1M |     0.289 |     0.055 |     0.152 |      0.066 |     0.207 |    — |
-|  9 | 0.688 | 0.779 | Kimi K3               | Moonshot AI | $8.92 |    42 | 1.0M |     0.282 |     0.055 |     0.152 |      0.076 |     0.214 | 1285 |
-| 10 | 0.678 | 0.795 | ★ GPT-5.6 Sol         | OpenAI      | $11.4 |    68 | 1.1M |     0.288 |     0.058 |     0.148 |      0.102 |     0.200 | 1184 |
-focus: website — coverage 125/408 (30.6%) ok; dispersion 0.992 ok; composition ok; freshness 0.0mo ok; trust unknown; modality unknown; maintenance unknown; provenance unknown; cross-source unknown
+|  1 | 0.771 | 0.775 | ★ DeepSeek-V4.1-Flash | DeepSeek    | $0.37 |    75 | 1.0M |     0.277 |     0.056 |     0.138 |      0.107 |     0.197 |    — |
+|  2 | 0.744 | 0.774 | Gemini 3.8 Flash      | Google      | $2.91 |    95 | 1.0M |     0.271 |     0.050 |     0.144 |      0.119 |     0.189 | 1236 |
+|  3 | 0.741 | 0.779 | ★ Muse Spark 1.3      | Meta        | $3.71 |    70 | 1.0M |     0.283 |     0.053 |     0.125 |      0.103 |     0.214 | 1296 |
+|  4 | 0.730 | 0.760 | Gemini 3.7 Flash      | Google      | $2.87 |    86 | 1.0M |     0.265 |     0.047 |     0.141 |      0.114 |     0.192 | 1223 |
+|  5 | 0.727 | 0.765 | Muse Spark 1.1        | Meta        | $3.71 |   127 | 1.0M |     0.269 |     0.048 |     0.148 |      0.135 |     0.165 |    — |
+|  6 | 0.725 | 0.755 | Gemini 3.6 Flash      | Google      | $2.87 |   132 | 1.0M |     0.244 |     0.042 |     0.140 |      0.137 |     0.190 | 1194 |
+|  7 | 0.721 | 0.724 | ★ GLM-5.3-Flash       | Zhipu AI    | $0.33 |    66 | 1.0M |     0.270 |     0.047 |     0.138 |      0.101 |     0.168 |    — |
+|  8 | 0.718 | 0.788 | ★ Kimi K3             | Moonshot AI | $6.81 |    52 | 1.0M |     0.283 |     0.055 |     0.152 |      0.088 |     0.211 | 1292 |
+|  9 | 0.702 | 0.781 | Gemini 4 Argon        | Google      | $7.65 |    38 |    — |     0.295 |     0.057 |     0.154 |      0.070 |     0.204 |    — |
+| 10 | 0.698 | 0.709 | MiMo-V2.6-Pro         | Xiaomi      | $1.01 |    29 | 1.0M |     0.272 |     0.054 |     0.127 |      0.056 |     0.199 |    — |
+focus: website — coverage 127/408 (31.1%) ok; dispersion 0.992 ok; composition ok; freshness 0.0mo ok; trust unknown; modality unknown; maintenance unknown; provenance unknown; cross-source unknown
 
 # Suggested settings.modelRoles (best-fit #1 per role, resolved via the omp catalog).
 # Selectors are openrouter/<id>; the omp-llm-role plugin writes this block daily.
 modelRoles:
   default: "openrouter/deepseek/deepseek-v4.1-flash"
   smol: "openrouter/deepseek/deepseek-v4.1-flash"
-  slow: "openrouter/deepseek/deepseek-v4.1-flash"
-  vision: "openrouter/deepseek/deepseek-v4.1-flash"
+  slow: "openrouter/google/gemini-4-argon"
+  vision: "openrouter/google/gemini-4-argon"
   plan: "openrouter/meta/muse-spark-1.3"
   commit: "openrouter/deepseek/deepseek-v4.1-flash"
-  tiny: "openrouter/inclusionai/ling-3.0-flash"
+  tiny: "openrouter/inception/mercury-2"
   task: "openrouter/deepseek/deepseek-v4.1-flash"
   advisor: "openrouter/meta/muse-spark-1.3"
   designer: "openrouter/deepseek/deepseek-v4.1-flash"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omp-llm-role",
-  "version": "1.2.0",
+  "version": "1.3.0",
   "type": "module",
   "description": "omp plugin: daily model-role updater driven by llm-stats/OpenRouter rankings",
   "license": "MIT",
@@ -78,6 +78,16 @@
           "auto"
         ]
       },
+      "tokenProfile.inputTokens": {
+        "type": "number",
+        "description": "Input tokens per task for the per-task price (used when the role has no recorded usage)",
+        "min": 0
+      },
+      "tokenProfile.outputTokens": {
+        "type": "number",
+        "description": "Output tokens per task for the per-task price (used when the role has no recorded usage)",
+        "min": 0
+      },
       "features.endpointCeilings": {
         "type": "boolean",
         "description": "Drop routes that cannot call tools and routes whose output ceiling is below 16 384 tokens",

package/src/agent-create.ts

@@ -424,8 +424,10 @@ function coveredValues(models: readonly Model[], metric: string, fill: number):
   return values;
 }
 
-/** Linear-interpolated quantile (R type 7) of a value list. */
-function quantile(values: readonly number[], p: number): number {
+/** Linear-interpolated quantile (R type 7) of a value list — the repo's one
+ * quantile convention: the focus-axis dispersion stats and the usage ledger's
+ * per-turn median (`quantile(values, 0.5)`) both read it. */
+export function quantile(values: readonly number[], p: number): number {
   const sorted = [...values].sort((a, b) => a - b);
   const h = (sorted.length - 1) * p;
   const lo = Math.floor(h);

package/src/engine.ts

@@ -313,6 +313,19 @@ export type OpenRouterEndpointPages = Record<string, OpenRouterEndpointRecord[]>
 
 export type OrEnrichment = { tput: number; latency: number | null; price: number | null; thinking: boolean; routes: OpenRouterEndpointRecord[] };
 
+/** The route's input (prompt) $/M under an assumed cache-hit rate: the input term
+ * `routeEffectivePrice` blends into the billed price. A hit rate `h` blends the
+ * cache-read input in, `h·cacheRead + (1−h)·input`; with `h = 0` or no cache-read
+ * price this is the listed input price, and a route with no input price has none —
+ * missing data keeps the full input price, never a silent discount. The single
+ * accessor for the input term: the billed blend, the output-price inverse
+ * (`routeOutputPrice`) and the explorer's per-task price all read it. */
+export function routeInputPrice(route: OpenRouterEndpointRecord, cacheHitRate = 0): number | null {
+  if (route.weightPrice == null) return null;
+  if (cacheHitRate <= 0 || route.cacheReadPrice == null) return route.weightPrice;
+  return cacheHitRate * route.cacheReadPrice + (1 - cacheHitRate) * route.weightPrice;
+}
+
 /** The effective billed $/M of one route under an assumed cache-hit rate. The route's
  * billed 3:1 price is `price = (3·input + output)/4` on the listed input price; a hit
  * rate `h` blends the cache-read input into the input term,
@@ -320,15 +333,33 @@ export type OrEnrichment = { tput: number; latency: number | null; price: number
  * (the output term is unchanged). With `h = 0`, no cache-read price, or no input price
  * this is the route's listed `price` — missing data keeps the full input price, never a
  * silent discount. Shared by the blend (`blendRoutePool`), the pin path (`rankRole`) and
- * `roleMetricValue`. */
+ * `roleMetricValue`; the input term itself is `routeInputPrice`. */
 export function routeEffectivePrice(route: OpenRouterEndpointRecord, cacheHitRate = 0): number | null {
   if (route.price == null) return null;
   if (cacheHitRate <= 0 || route.cacheReadPrice == null || route.weightPrice == null) return route.price;
   const input = route.weightPrice;
-  const effInput = cacheHitRate * route.cacheReadPrice + (1 - cacheHitRate) * input;
+  const effInput = routeInputPrice(route, cacheHitRate) as number;
   return route.price + (3 * (effInput - input)) / 4;
 }
 
+/** The output (completion) $/M implied by a billed 3:1 blend and the input price
+ * behind it: the inverse of `billed = IO_PRICE_RATIO·(3·input + output)`, i.e.
+ * `output = billed/IO_PRICE_RATIO − (1−IO_PRICE_RATIO)/IO_PRICE_RATIO · input`.
+ * The 3:1 ratio lives here (and in `IO_PRICE_RATIO`), not in the callers; dividing
+ * by `IO_PRICE_RATIO` (2⁻²) is exact, so this is the plain `4·billed − 3·input`. */
+export function blendOutputPrice(billed: number, input: number): number {
+  return billed / IO_PRICE_RATIO - ((1 - IO_PRICE_RATIO) / IO_PRICE_RATIO) * input;
+}
+
+/** The route's output (completion) $/M: the inverse blend of its own billed and
+ * input prices (see `blendOutputPrice`). Null when either of them is missing. */
+export function routeOutputPrice(route: OpenRouterEndpointRecord, cacheHitRate = 0): number | null {
+  const billed = routeEffectivePrice(route, cacheHitRate);
+  const input = routeInputPrice(route, cacheHitRate);
+  if (billed === null || input === null) return null;
+  return blendOutputPrice(billed, input);
+}
+
 /** Expected price/throughput/latency of one request over a route pool under the
  * router's default price-based load balancing: P(route i) ∝ 1/input_price_i² — the
  * router sorts on the input (prompt) price alone, so the weight basis is