omnilane 0.42.9 → 0.45.0

package/docs/aa-model-coverage-2026-09-05.json

@@ -35,7 +35,7 @@
   },
   "counts": {
     "aaRows": 643,
-    "catalogEntries": 113,
+    "catalogEntries": 114,
     "namedAliases": 18
   },
   "namedAliases": [
@@ -1034,6 +1034,48 @@
       ],
       "mappingReason": "AA base row is the closest family record; provider effort semantics may differ"
     },
+    {
+      "vendor": "grok",
+      "access": "cli",
+      "catalogModel": "grok-4.7",
+      "lifecycle": "active",
+      "disposition": "default",
+      "routeRefs": [
+        {
+          "lane": "hardest-coding",
+          "position": "fallback",
+          "effort": "high"
+        },
+        {
+          "lane": "consult",
+          "position": "fallback",
+          "effort": "high"
+        },
+        {
+          "lane": "coding-overflow",
+          "position": "primary",
+          "effort": "high"
+        },
+        {
+          "lane": "hard-judgment",
+          "position": "fallback",
+          "effort": "high"
+        },
+        {
+          "lane": "taste-final",
+          "position": "fallback",
+          "effort": "high"
+        },
+        {
+          "lane": "live-search",
+          "position": "primary",
+          "effort": "high"
+        }
+      ],
+      "mappingStatus": "no-aa-row",
+      "aaSlugs": [],
+      "mappingReason": "Released 2026-09-21, after this v4.2 inventory was taken, so none of its 643 rows is this model; config/aa-model-policy.json scores it from AA v4.3.2 as grok-4-7 (xhigh) and grok-4-7-high. Its routeRefs are routing.yaml as of 2026-09-22; every other entry's are as of this inventory"
+    },
     {
       "vendor": "grok",
       "access": "cli",

package/docs/model-capabilities-2026-09.md

@@ -1,10 +1,201 @@
 # Model capabilities — September 2026 snapshot
 
-External records are dated per section. The 2026-09-05 decision below is the
-current basis for `routing.yaml`; the earlier 2026-09-02 v4.1.1-era snapshot is
-kept as historical evidence rather than relabelled as v4.2.
+External records are dated per section. The 2026-09-22 decision below is the
+current basis for `routing.yaml`. The 2026-09-05 v4.2 decision and the
+2026-09-02 v4.1.1-era snapshot are kept as historical evidence; their scores
+are on other scales and are not comparable with the v4.3.2 figures.
+
+## Current routing decision — 2026-09-22 (AA v4.3.2)
+
+`routing.yaml` was rewritten on this date, not amended. Each lane names the
+measurements that match its kind of work, and its candidates are listed
+best-first on them. Because a caller may only dispatch to a target scoring at or
+below its own ceiling, each chain also steps down through the score range, so
+that whatever a controller's ceiling, the first candidate it can reach is the
+best one it can reach. The `score` column below is the registry's ceiling value.
+
+Sources. Registry scores: `docs/reports/aa-v4.3.2-extract-2026-09-22.json`.
+Lane measurements: `docs/reports/aa-v4.3.2-lane-metrics-2026-09-22.json`, a second
+capture of the same AA page the same day; every index value in the two files is
+identical. The page served at that capture carried no hallucination-rate field,
+so that one field comes from a capture about an hour earlier, as the file's
+`supplement` block records. The tables are generated, not typed:
+`scripts/aa_rebaseline.py lanes --extract docs/reports/aa-v4.3.2-lane-metrics-2026-09-22.json`.
+
+What the measurements are. AA's own v4.3 note says Terminal-Bench moved to 4.0 and
+AutomationBench joined the index. Terminal-Bench 2.1 is kept for reference only:
+the frontier sits between 0.87 and 0.91 on it, so it no longer separates anyone,
+while 4.0 spreads the same models from 0.2 to 0.6. Briefcase is AA's expert
+grading of finished business documents, with separate Elo scores for analytical
+quality and for presentation. `mlcrOverall` is carried in AA's page data without
+a caption; it is read here as a harder long-context benchmark because it sits
+beside AA-LCR and spreads the models that AA-LCR cannot separate. That reading is
+unverified.
+
+**hardest-coding**
+
+| # | candidate | score | Terminal-Bench 4.0 | Terminal-Bench 2.1 | SciCode | hallucination rate |
+|---|---|---|---|---|---|---|
+| 1 | codex gpt-6-astra xhigh | 52 | 0.596 | 0.891 | 0.557 | 0.483 |
+| 2 | claude claude-fable-5-1 xhigh | 53 | 0.551 | 0.910 | 0.609 | 0.705 |
+| 3 | codex gpt-6-astra high | 51 | 0.540 | 0.899 | 0.554 | 0.448 |
+| 4 | claude claude-fable-5-1 high | 51 | 0.520 | 0.899 | 0.587 | 0.688 |
+| 5 | codex gpt-6-astra medium | 50 | 0.495 | 0.895 | 0.542 | 0.465 |
+| 6 | claude claude-opus-5 high | 48 | 0.460 | 0.876 | 0.554 | 0.612 |
+| 7 | codex gpt-6-astra low | 46 | 0.419 | 0.880 | 0.541 | 0.469 |
+| 8 | codex gpt-5.6-sol xhigh | 44 | 0.247 | 0.895 | 0.571 | 0.919 |
+| 9 | grok grok-4.7 high | 46 | 0.247 | not published | 0.578 | 0.324 |
+| 10 | grok grok-4.6 high | 44 | 0.212 | 0.884 | 0.565 | 0.343 |
+| 11 | gemini gemini-3.8-flash-high | 41 | 0.197 | 0.876 | 0.566 | 0.552 |
+
+**bulk-mechanical**
+
+| # | candidate | score | Terminal-Bench 4.0 | Terminal-Bench 2.1 | hallucination rate | minutes / task | index run cost ($) |
+|---|---|---|---|---|---|---|---|
+| 1 | codex gpt-6-astra low | 46 | 0.419 | 0.880 | 0.469 | 1.3 | 1537 |
+| 2 | codex gpt-5.6-sol high | 42 | 0.207 | 0.873 | 0.912 | 3.2 | 1487 |
+| 3 | gemini gemini-3.8-flash-high | 41 | 0.197 | 0.876 | 0.552 | 3.3 | 1623 |
+| 4 | claude claude-opus-5 medium | 45 | 0.343 | 0.861 | 0.607 | 5.1 | 2732 |
+
+**triage**
+
+| # | candidate | score | index | index run cost ($) | minutes / task |
+|---|---|---|---|---|---|
+| 1 | codex gpt-5.6-luna high | 32 | 32.1 | 108 | 1.6 |
+| 2 | gemini gemini-3.8-flash-low | 33 | 33.5 | not published | not published |
+| 3 | claude claude-sonnet-5 low | 24 | 24.3 | 653 | 2.5 |
+| 4 | claude claude-haiku-4-5 | 17 | 16.9 | 524 | 2.6 |
+
+**hard-judgment**
+
+| # | candidate | score | HLE | GPQA | CritPt | Briefcase analytical Elo | hallucination rate |
+|---|---|---|---|---|---|---|---|
+| 1 | claude claude-fable-5-1 xhigh | 53 | 0.587 | 0.934 | 0.311 | 1971 | 0.705 |
+| 2 | codex gpt-6-astra xhigh | 52 | 0.546 | 0.963 | 0.314 | 1724 | 0.483 |
+| 3 | claude claude-opus-5 max | 51 | 0.549 | 0.932 | 0.291 | 1965 | 0.608 |
+| 4 | codex gpt-6-astra high | 51 | 0.531 | 0.949 | 0.289 | 1703 | 0.448 |
+| 5 | claude claude-opus-5 xhigh | 50 | 0.544 | 0.937 | 0.277 | 1958 | 0.595 |
+| 6 | claude claude-opus-5 high | 48 | 0.528 | 0.937 | 0.283 | 1855 | 0.612 |
+| 7 | grok grok-4.7 high | 46 | 0.423 | not published | 0.180 | 1967 | 0.324 |
+| 8 | grok grok-4.6 high | 44 | 0.429 | 0.949 | 0.171 | 1690 | 0.343 |
+| 9 | gemini gemini-3.8-flash-high | 41 | 0.478 | 0.953 | 0.183 | 1151 | 0.552 |
+
+**taste-final**
+
+| # | candidate | score | Briefcase overall Elo | Briefcase presentation Elo | GDPval |
+|---|---|---|---|---|---|
+| 1 | claude claude-opus-5 max | 51 | 1673 | 1560 | 0.604 |
+| 2 | claude claude-fable-5-1 xhigh | 53 | 1669 | 1473 | 0.610 |
+| 3 | claude claude-opus-5 xhigh | 50 | 1649 | 1492 | 0.588 |
+| 4 | grok grok-4.7 high | 46 | 1644 | 1506 | 0.597 |
+| 5 | codex gpt-6-astra xhigh | 52 | 1544 | 1503 | 0.508 |
+| 6 | claude claude-opus-5 high | 48 | 1573 | 1446 | 0.540 |
+| 7 | grok grok-4.6 high | 44 | 1546 | 1519 | 0.553 |
+| 8 | gemini gemini-3.8-flash-high | 41 | 1202 | 1200 | 0.456 |
+
+**consult**
+
+| # | candidate | score | index |
+|---|---|---|---|
+| 1 | codex gpt-6-astra xhigh | 52 | 52.4 |
+| 2 | claude claude-fable-5-1 xhigh | 53 | 53.2 |
+| 3 | grok grok-4.7 high | 46 | 46.3 |
+| 4 | grok grok-4.6 high | 44 | 44.3 |
+| 5 | gemini gemini-3.8-flash-high | 41 | 40.9 |
+
+**ui-draft**
+
+| # | candidate | score | MMMU-Pro | Terminal-Bench 4.0 |
+|---|---|---|---|---|
+| 1 | codex gpt-6-astra high | 51 | 0.864 | 0.540 |
+| 2 | claude claude-opus-5 high | 48 | 0.824 | 0.460 |
+| 3 | codex gpt-6-astra low | 46 | 0.846 | 0.419 |
+| 4 | gemini gemini-3.8-flash-high | 41 | 0.856 | 0.197 |
+
+**long-context**
+
+| # | candidate | score | mlcrOverall | AA-LCR | index run cost ($) |
+|---|---|---|---|---|---|
+| 1 | claude claude-opus-5 high | 48 | 0.594 | 0.790 | 4332 |
+| 2 | claude claude-opus-5 medium | 45 | 0.561 | 0.820 | 2732 |
+| 3 | claude claude-opus-5 low | 39 | 0.539 | 0.813 | 1561 |
+| 4 | codex gpt-5.6-terra max | 42 | 0.317 | 0.830 | 2501 |
+| 5 | gemini gemini-3.8-flash-high | 41 | 0.217 | 0.813 | 1623 |
+
+**fast-agentic**
 
-## Current routing decision — 2026-09-05
+| # | candidate | score | AutomationBench | minutes / task | first answer token (s) |
+|---|---|---|---|---|---|
+| 1 | codex gpt-6-astra low | 46 | 0.591 | 1.3 | 2 |
+| 2 | gemini gemini-3.8-flash-medium | 40 | 0.609 | not published | not published |
+| 3 | codex gpt-5.6-sol medium | 39 | 0.513 | 2.1 | 5 |
+| 4 | claude claude-opus-5 low | 39 | 0.518 | 2.5 | 3 |
+
+**live-search**
+
+| # | candidate | score | hallucination rate | knowledge (omniscience) |
+|---|---|---|---|---|
+| 1 | grok grok-4.7 high | 46 | 0.324 | 30.9 |
+| 2 | grok grok-4.6 high | 44 | 0.343 | 30.5 |
+| 3 | gemini gemini-3.8-flash-high | 41 | 0.552 | 29.6 |
+| 4 | claude claude-opus-5 medium | 45 | 0.607 | 31.0 |
+
+**coding-overflow**
+
+| # | candidate | score | Terminal-Bench 4.0 | SciCode |
+|---|---|---|---|---|
+| 1 | grok grok-4.7 high | 46 | 0.247 | 0.578 |
+| 2 | grok grok-4.6 high | 44 | 0.212 | 0.565 |
+| 3 | gemini gemini-3.8-flash-high | 41 | 0.197 | 0.566 |
+| 4 | kimi kimi-k3 | not scored | — | — |
+| 5 | qwen qwen3-coder-plus | not scored | — | — |
+| 6 | opencode - | not scored | — | — |
+
+Why each lane is ordered this way:
+
+- **hardest-coding** follows Terminal-Bench 4.0. Astra xhigh leads it and has the
+  lowest hallucination rate among the top rows. Fable xhigh leads the saturated
+  2.1 benchmark and SciCode, and beats Fable max on 4.0 (0.551 against 0.520) at
+  about two thirds of max's run cost, so max is in no chain. Astra loses far less
+  than any other family as effort drops (low still scores 0.419, where Opus is
+  down to 0.343 by medium and Sol high sits at 0.207), which is why it supplies
+  most of the rungs.
+- **bulk-mechanical** wants routine coding per minute and per dollar. Astra low
+  costs what Sol high costs and is ahead on both coding benchmarks, in under half
+  the time, with roughly half the hallucination rate. Sol high remains for
+  callers whose ceiling is below Astra low.
+- **triage** is bought by the run. Luna high's run cost is a small fraction of any
+  other row that still reads code. Sonnet low outscores Haiku for similar money.
+- **hard-judgment** follows HLE and CritPt for reasoning and Briefcase analytical
+  Elo for written analysis; hallucination rate decides the second opinion. Fable
+  xhigh leads both reasoning measures. Astra xhigh is the independent family and
+  the least likely of the top rows to assert something false. Opus keeps its
+  analytical Elo down the effort ladder (1855 at high) where Astra does not
+  (1703). Grok 4.7's analytical Elo equals Opus max's, but its reasoning scores
+  are among the lowest in the chain, so it stays below the Opus rows.
+- **taste-final** has no benchmark, and nothing here measures Chinese phrasing.
+  The nearest evidence is Briefcase. Opus max is level with Fable overall and
+  clearly ahead on presentation. Grok 4.7 is graded above every Astra row overall,
+  so it precedes Astra. Final prose still needs human review.
+- **ui-draft** follows MMMU-Pro first and Terminal-Bench 4.0 second. Fable has no
+  published MMMU-Pro result and is left out rather than assumed.
+- **long-context** follows `mlcrOverall`, where Opus leads every other family by
+  a wide margin and keeps that lead down to low effort; AA-LCR is shown to make
+  the saturation visible. Fable max scores higher still (0.711) but at three
+  times Opus high's run cost, and it is the only Fable row with a published value.
+- **fast-agentic** follows AutomationBench and time per task. Astra low is the
+  fastest row that still scores near the top. Flash medium scores as high and is
+  the cross-vendor row; Flash low falls to 0.365. Haiku scores 0.032 and is gone.
+- **live-search**: Grok is the only native X source, and both releases hallucinate
+  less than either fallback. Opus medium replaces Sonnet because Sonnet's effort
+  rows have a negative omniscience score, more false assertions than true ones.
+- **coding-overflow** has no Codex row by design. Grok 4.7 edges 4.6 on
+  Terminal-Bench 4.0 and SciCode.
+- **consult** lists each vendor's strongest generally reachable configuration.
+
+The capability boundaries listed under 2026-09-05 still hold.
+
+## Previous routing decision — 2026-09-05 (AA v4.2)
 
 ### Comparable evidence and limits
 

package/docs/native-executor.md

@@ -88,6 +88,65 @@ workdir is task context, not an OS boundary. Requests for hard `read-only`,
 native jobs: `auto` stays on the same resolved model through CLI, while forced
 native fails closed.
 
+### Generating the capability file
+
+`omnilane native-context [--workdir DIR]... [--mode advise|work]... [--harness NAME]
+[--inherits-caller-runtime] [--out FILE]` writes a capability file for the
+harness it runs under and prints the path. Vendor, current model and current
+effort are read the way `omnilane whoami` reads them; nothing is inferred from
+installed CLIs. It emits one capability row: the caller's own model at its own
+effort. A caller whose effort is unrecorded gets `efforts: ["unverified"]`, which
+matches no lane target, so that file serves `--inherit` only. Rows for other
+models are the host's to add, from its real agent-tool contract.
+
+`inherits_caller_runtime: true` is written only when the host passes
+`--inherits-caller-runtime`. It is the host's statement that its sub-agent tool,
+given no model override, runs the caller's model and effort.
+
+What this release has and has not verified about each harness's sub-agent tool:
+
+| Harness | Sub-agent tool | Status |
+|---|---|---|
+| Codex | `collaboration.spawn_agent`; a model override requires `fork_turns: "none"` or a bounded count | documented above from the tool contract; inheritance without an override is host-asserted |
+| Claude Code | `Agent` tool; optional `model` override | documented (code.claude.com/docs/en/sub-agents, read 2026-09-20). Model resolves in this order: the per-invocation `model` parameter, the agent definition's `model` frontmatter, `CLAUDE_CODE_SUBAGENT_MODEL`, then the main conversation's model. Effort: the definition's `effort` frontmatter overrides the session level and its default is "inherits from session". So a host may pass `--inherits-caller-runtime` only when it spawns with no `model` argument, the agent type's definition sets neither `model` (other than `inherit`) nor `effort`, and `CLAUDE_CODE_SUBAGENT_MODEL` is unset; the built-in general-purpose agent meets the definition part. Two full `--inherit` cycles on 2026-09-20 under those conditions reported the parent's exact model. The worker cannot see its own effort, so completion reports `runtime.effort: "unknown"`; effort inheritance rests on the documentation, not on an observation |
+| Grok Build | `spawn_subagent`; agent types `general-purpose`, `explore`, `plan` and user-defined ones | model: documented in the subagent guide embedded in grok 1.0.34 — "By default a subagent inherits the parent session's model"; only `[subagents.models].<agent>` in the config (highest priority) or the agent definition's `model` overrides that, and the bundled `general-purpose` definition says `model: inherit`. Effort: a definition may carry an `effort` override; no statement of the default was found, so effort inheritance is not verified. No `--inherit` cycle has been run in Grok |
+| Antigravity | — | `agy` 1.2.7 exposes `--agent`, `--model` and `--effort` for the session and its help names no sub-agent spawning surface; not verified, and no `--inherit` cycle has been run |
+
+### Inherited worker (`--inherit`)
+
+`dispatch.sh --inherit --native-context FILE <lane> "<task>"` plans a native
+worker spawned with no model override. `aa_policy.decide_inherited` allows it
+with code `native-inherited-allowed`: the worker runs the caller's runtime, so
+its score equals the caller's and the downward rule holds by construction, with
+or without a recorded effort. No target configuration is resolved
+(`target_config_id: null`, `target: {"inherit": true, ...}`), the transport
+overlay is not consulted, and there is no CLI fallback: a host that cannot take
+it gets `native-inherit-unavailable` with `failed_gate: "native-capability"` and
+the `omnilane native-context` command. The plan carries `inherit: true`,
+`effort: "inherited"`, `worker_contract.model_override: false`,
+`inherit_caller_runtime: true` and `satisfies_lane_target: false`. The capability
+file must set `inherits_caller_runtime: true`, name the caller's vendor and
+model, and hold a new-agent row for that model matching mode, workdir, tools,
+`shared-inherited` and `single-shot`; effort is deliberately not matched.
+
+When no caller identity reaches the gate, the file's `vendor` and `current_model`
+stand as the host's statement and the decision is
+`native-inherited-unverified-caller` (`caller_kind: "model-unverified"`,
+`caller_identity_verified: false`, `caller_identity_source: "host-asserted"`). It
+has no `effective_ceiling`, publishes neither `aa-authorizer.json` nor
+`aa-child-context.json`, and `jobs status` shows `caller_identity_verified: false`
+and `satisfies_lane_target: false`. A file with no `current_model` is refused as
+`missing-caller-context`. `omnilane native-context --vendor V --model M` writes
+such a file only when the identity cannot be read or resolved, marks it
+`caller_identity_verified: false`, and exits 2 if the statement contradicts an
+identity it can read. This exists for a caller that cannot be read, for example a
+codex command wrapped in `;`, `&&` or a pipe, where the shell codex started stays
+between codex and the command and carries no `CODEX_THREAD_ID` of its own; run
+alone, the same command is read. Lane dispatch from an unread caller is refused.
+Completion checks vendor, model and harness; `runtime.effort` is whatever the
+host observed. A human operator, an unidentified caller, `--vendor`/`--model`/
+`--effort`/`--target-config`, and every CLI-only lifecycle are refused.
+
 Native supports only a caller-supervised single task. `--background`, explicit
 `--live` / `--single-shot`, `--thread`, `sysops`, explicit/environment whole-job
 or idle watchdogs, vote/multi-round and `exec` arbitration paths stay CLI or

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnilane",
-  "version": "0.42.9",
+  "version": "0.45.0",
   "description": "One routing table, every harness — classify subtasks into lanes and delegate each lane through a compatible caller-owned native agent or vendor CLI.",
   "bin": {
     "omnilane": "bin/omnilane"

package/plugin.json

@@ -1,6 +1,6 @@
 {
   "$schema": "https://antigravity.google/schemas/v1/plugin.json",
   "name": "omnilane",
-  "version": "0.42.9",
+  "version": "0.45.0",
   "description": "One routing table, every harness: classify subtasks into lanes and delegate through compatible caller-owned native agents or vendor CLIs with exact-AA downward policy and supervised jobs."
 }

package/routing.yaml

@@ -12,33 +12,35 @@
 # deliberately carry no numbers: they state WHY a lane is ordered the way it is, which
 # stays true for months, while the numbers move every few weeks. Change an ordering and
 # you update the doc; a figure going stale should never need a routing-table edit.
-# (Audited 2026-07-12; re-audited 2026-07-25, 2026-08-02, 2026-08-03, 2026-09-02, and 2026-09-05.)
-# defaults follow Artificial Analysis data, 2026-09
-# snapshot. Verified against AA site records + vendor pricing pages: Intelligence &
-# Coding indexes and 7:2:1 blended prices all match (AA field price1mBlended7To2To1);
-# coding cost-per-task is chart-only (not independently reconstructed). Prices are
-# standard short-context API tier — on subscription CLIs treat $ as relative ranking.
+# (Audited 2026-07-12; re-audited 2026-07-25, 2026-08-02, 2026-08-03, 2026-09-02 and
+# 2026-09-05; rewritten 2026-09-22 on Artificial Analysis Intelligence Index v4.3.2.)
+# How a chain is ordered: each lane names the measurements that match its kind of work,
+# and candidates are listed best-first on those. A caller may only dispatch to a target
+# scoring at or below its own ceiling, and a candidate above the ceiling (or one this
+# host has not proven) is skipped, so every chain also steps down through the score
+# range: whatever the caller's ceiling, the first candidate it can reach is the best
+# one it can reach. Vendors alternate where the evidence allows, so a host with one
+# subscription still gets a full chain. Prices are standard short-context API tier — on
+# subscription CLIs treat $ as relative ranking.
 # Your own job outcomes (~/.omnilane/jobs/) outrank these priors; edit lanes to match.
 
-hardest-coding: claude claude-fable-5-1 max | codex gpt-6-astra xhigh | grok grok-4.6 high | gemini gemini-3.8-flash-high - # correctness-first coding: Fable max leads the same-condition comparison; Astra xhigh is the default Codex quality/cost tradeoff; explicit --effort max remains available; Grok and Flash keep cross-vendor depth
-bulk-mechanical: codex gpt-5.6-sol high | gemini gemini-3.8-flash-high - | claude claude-sonnet-5 high # endurance and migration work stays on proven Sol high; Flash 3.8 refreshes the fast middle fallback; Sonnet preserves cross-vendor depth
-triage: codex gpt-5.6-luna high | gemini gemini-3.8-flash-low - | claude claude-haiku-4-5 - # first-pass filtering favors the low-cost Luna row; Flash low and Haiku remain cheap cross-vendor fallbacks
-hard-judgment: claude claude-fable-5-1 xhigh | codex gpt-6-astra xhigh | grok grok-4.6 high # strongest same-condition judgment row first, then an independently verified Codex family and Grok; this lane is not a controller selector
-taste-final: claude claude-fable-5-1 xhigh | codex gpt-6-astra xhigh | grok grok-4.6 high | gemini gemini-3.8-flash-high - # final prose and style still require human taste review; general quality evidence orders the candidates without claiming an aesthetic benchmark
-consult: codex gpt-6-astra xhigh | claude claude-fable-5-1 xhigh | grok grok-4.6 high | gemini gemini-3.8-flash-medium - # direct named-model consultation chain; keep --vendor to pin the requested family and prevent fallback
-ui-draft: codex gpt-5.6-sol high | claude claude-fable-5-1 xhigh | gemini gemini-3.8-flash-high - # UI drafts require a design system or references; measured coding/agent evidence supports the order but does not prove visual taste
-long-context: gemini gemini-3.8-flash-medium - | codex gpt-5.6-terra max | claude claude-opus-5 medium # Flash medium leads long-document synthesis; Terra and Opus stay because context capacity alone does not prove task quality
-fast-agentic: gemini gemini-3.8-flash-low - | codex gpt-5.6-luna high | claude claude-haiku-4-5 - # low-latency tool loops favor Flash low; Luna and Haiku retain cross-vendor fallback depth
-live-search: grok grok-4.6 high | gemini gemini-3.8-flash-high - | claude claude-sonnet-5 high | off # Grok remains the native X/web choice; Flash and Sonnet provide generic web-search fallback, not equivalent social context
-coding-overflow: grok grok-4.6 high | gemini gemini-3.8-flash-high - | kimi kimi-k3 - | qwen qwen3-coder-plus - | opencode - - | off # explicit quota-relief lane; keep all existing non-Codex fallbacks and do not infer unverified Qwen aliases from another harness
+hardest-coding: codex gpt-6-astra xhigh | claude claude-fable-5-1 xhigh | codex gpt-6-astra high | claude claude-fable-5-1 high | codex gpt-6-astra medium | claude claude-opus-5 high | codex gpt-6-astra low | codex gpt-5.6-sol xhigh | grok grok-4.7 high | grok grok-4.6 high | gemini gemini-3.8-flash-high - # ordered on the hard terminal-coding benchmark, the one coding measurement the frontier has not saturated, with scientific coding and hallucination rate as tie-breakers: Astra xhigh leads it and hallucinates least among the top rows; Fable xhigh leads the saturated coding benchmarks and beats Fable max on the hard one for less, so max is explicit-only; the Astra ladder holds up far better than any other family as effort drops, which is why it supplies most rungs
+bulk-mechanical: codex gpt-6-astra low | codex gpt-5.6-sol high | gemini gemini-3.8-flash-high - | claude claude-opus-5 medium # endurance work wants routine-coding accuracy per minute and per dollar: Astra low matches Sol high's cost and beats it on both coding benchmarks, on speed and by a wide margin on hallucination; Sol high stays for callers below Astra low's score; Flash and Opus medium are the quick rows of their vendors
+triage: codex gpt-5.6-luna high | gemini gemini-3.8-flash-low - | claude claude-sonnet-5 low | claude claude-haiku-4-5 - # first-pass filtering is bought by the run, not by the answer: Luna high costs a small fraction of any other row that can still read code; Flash low is the cheap cross-vendor row; Sonnet low outscores Haiku for similar money and Haiku stays for hosts that have not proven Sonnet's effort rows
+hard-judgment: claude claude-fable-5-1 xhigh | codex gpt-6-astra xhigh | claude claude-opus-5 max | codex gpt-6-astra high | claude claude-opus-5 xhigh | claude claude-opus-5 high | grok grok-4.7 high | grok grok-4.6 high | gemini gemini-3.8-flash-high - # ordered on expert-level reasoning and on graded analytical quality of written work, with hallucination rate deciding who gives the second opinion: Fable xhigh leads both reasoning measures and equals Fable max for less; Astra xhigh is the independent family and the least likely of the top rows to assert something false; Opus holds its analytical quality down the effort ladder where Astra loses it; Grok is the third family, 4.7 ahead on analytical quality; this lane is not a controller selector
+taste-final: claude claude-opus-5 max | claude claude-fable-5-1 xhigh | claude claude-opus-5 xhigh | grok grok-4.7 high | codex gpt-6-astra xhigh | claude claude-opus-5 high | grok grok-4.6 high | gemini gemini-3.8-flash-high - # no benchmark measures taste, and none here covers Chinese phrasing; the nearest evidence is expert grading of finished documents, overall and on presentation: Opus max is graded best on presentation and level with Fable overall; Grok 4.7 is graded above every Astra row on finished documents, so it sits ahead of Astra; final prose still needs human review
+consult: codex gpt-6-astra xhigh | claude claude-fable-5-1 xhigh | grok grok-4.7 high | grok grok-4.6 high | gemini gemini-3.8-flash-high - # direct named-model consultation: each vendor's strongest generally reachable configuration, with the older Grok behind the newer for a lane dispatch on a host that has not proven 4.7; keep --vendor to pin the requested family. --vendor takes the vendor's first segment and never falls through, so on a host that has not proven grok-4.7 ask for --model grok-4.6
+ui-draft: codex gpt-6-astra high | claude claude-opus-5 high | codex gpt-6-astra low | gemini gemini-3.8-flash-high - # UI drafts need a design system or reference images, so the order follows visual understanding first and hard coding second: the Astra rows read images best and code best; Flash reads images nearly as well; Fable has no published visual result and is left out rather than assumed; none of this proves visual taste
+long-context: claude claude-opus-5 high | claude claude-opus-5 medium | claude claude-opus-5 low | codex gpt-5.6-terra max | gemini gemini-3.8-flash-high - # the established long-context benchmark is saturated and separates nobody; on the harder one Claude leads every other family by a wide margin at every effort, and Opus keeps that lead down to low effort, which makes it both the best and the cheap choice; Terra and Flash are the other vendors' best rows; context capacity alone proves nothing
+fast-agentic: codex gpt-6-astra low | gemini gemini-3.8-flash-medium - | codex gpt-5.6-sol medium | claude claude-opus-5 low # tool loops need automation accuracy and a short time per step: Astra low is the fastest row that still scores near the top on automation; Flash medium scores as high and is the cross-vendor row (Flash low falls well behind it); Sol medium and Opus low are the quick rows below them; Haiku is gone because it scores close to nothing on automation
+live-search: grok grok-4.7 high | grok grok-4.6 high | gemini gemini-3.8-flash-high - | claude claude-opus-5 medium | off # Grok is the only native X/web source, newest proven release first, and both releases hallucinate less than any fallback; Flash and Opus medium provide generic web search, not equivalent social context; Opus medium rather than Sonnet because Sonnet's effort rows assert more false facts than true ones on the knowledge benchmark
+coding-overflow: grok grok-4.7 high | grok grok-4.6 high | gemini gemini-3.8-flash-high - | kimi kimi-k3 - | qwen qwen3-coder-plus - | opencode - - | off # explicit Codex-quota relief, so no Codex row: Grok 4.7 edges 4.6 on the hard terminal-coding benchmark and on scientific coding; keep all existing non-Codex fallbacks and do not infer unverified Qwen aliases from another harness
 arbitrate: off - - # opinion panel remains opt-in because each voter and round consumes quota
                                                 # Enable: `arbitrate: vote codex,claude,grok -` (any 1-4 of codex/claude/grok/gemini)
                                                 # Debate round (each voter rebuts the others): set the effort field to 2.
                                                 # Custom gate: `arbitrate: exec /path/to/script -`
-# Claude Fable 5.1 is in the taste and hardest-coding defaults because it leads
-# Opus 5 on every Artificial Analysis axis at the same effort. It is not in
-# bulk or triage: it prices at twice Opus 5 per token and consumes the most
-# subscription quota per turn. Fable leads hard-judgment; Opus remains
-# explicitly selectable via --vendor claude --model claude-opus-5.
-# Astra defaults to xhigh; use --vendor codex --effort max when explicitly needed.
+# Fable max and Astra max are in no chain: on the measurements used above, max is never
+# more than marginally ahead of the same model's xhigh row and on hard terminal coding it
+# is behind, while xhigh takes markedly less time and money.
+# Both remain selectable with --vendor … --effort max.
 # No automatic risk escalation or failure-triggered effort upgrade is added.