lidge-jun · lidge-jun · Sep 23, 2026 · Sep 22, 2026 · Sep 22, 2026 · Sep 22, 2026
@@ -0,0 +1,122 @@
+# 260923 grok-4.7 parity — plan
+
+grok-4.7 shipped on 2026-09-21 and already answers through xAI, Devin, Command Code and Cursor, but OpenCodex has no
+registry entry for it: the picker shows it without a context window, reasoning ladder, image input, Fast row or
+Responses wire, and cost estimates are unavailable. This unit gives grok-4.7 the same declarations grok-4.6 carries,
+using values measured with real grok-4.7 calls (010_probe-evidence.md) and xAI's published model page, and applies
+them to the other providers that serve it where their evidence supports each declaration.
+
+## Loop spec
+
+- Archetype: satisfy-spec, single work-phase (wp1), one PABCD cycle, one PR to dev.
+- Trigger: user request 2026-09-23 "grok-4.7 모델 피커 컨텍스트 fast 와이어를 실제 토큰응답으로 ... grok-4.6과 같이 패치하고 다른 프로바이더들에도 적용하는 pr".
+- Goal: grok-4.7 has grok-4.6-equivalent picker/context/effort/image/Fast/wire/price metadata on xAI, plus Devin,
+  Command Code, Cursor, OpenCode Go and bundled gateway metadata where evidenced.
+- Non-goals: changing default or sidecar models (web-search defaults stay grok-4.6); grok-4.7-build-fast; GitHub
+  Copilot wire pin and OpenCode Go hosted web_search strip for 4.7 (no probe possible, not configured locally);
+  merge, release, service restart. The user forbade local tests: no bun test, typecheck or build runs locally.
+- Verifier: static only locally — `git diff --check`, JSON parse of edited JSON, `rg` roster consistency (every
+  grok-4.6 xAI-block key has a grok-4.7 sibling), byte-equality of regenerated metadata via the generator script
+  (codegen, not a test); then exact-head hosted CI on the PR (typecheck + 4 test shards + file-size + layout).
+- Stop: PR open, independent review has no unresolved blocker, exact-head CI reported.
+- Memory artifact: this unit (000/010), goalplan add-grok-4-7-to-opencodex-with-the-same-first-cl.
+- Expected terminal outcomes: DONE (PR open, CI green or failures fixed); BLOCKED if push refused.
+- Escalation: a CI failure that needs a local run to diagnose, or a design dispute that requires changing a default.
+- HOTL bounds: tools = repo edits, gh, live proxy probes already done; write scope = files listed below; no token or
+  time budget was set by the user.
+
+## Architect consultation
+
+Handle 01a0caad-1195-74b3-8338-e7a354313b5d (Banach, gpt-6-sol). Proposal D1–D8. Dispositions:
+
+- D1 accept (xAI declarations, grok-4.7 ahead of 4.6 in XAI_MODELS).
+- D2 amended in revision 3 (see Audit round 1 synthesis): toggle set unchanged; 4.7 gets the OAuth Responses default
+  through modelWireDefaults only.
+- D3 accept (Devin roster, 500k, measured low..max ladder, default medium).
+- D4 accept: add xai/grok-4.7 and xai/grok-4.6 to COMMAND_CODE_IMAGE_MODELS; the 4.6 negative is contradicted by the
+  same two-path grid probe the header demands.
+- D5 amend: OpenCode Go wire/efforts/default ARE mirrored — opencode.ai/docs/go lists "Grok 4.7 grok-4.7
+  https://opencode.ai/zen/go/v1/responses @ai-sdk/openai", the same documented evidence the 4.6 pin (#3394) used.
+  The Go web_search strip and the Copilot Responses pin stay 4.6-only (unprobed; recorded as follow-ups).
+- D6 amend: Cursor's live GetUsableModels roster (explorer 01a0caad-9fb2-70a1-9cf5-120ad392775f) lists
+  grok-4.7-{low,medium,high,xhigh} and the same with -fast, no cursor- prefix, no max. Mirror with no wirePrefix and
+  keep the prefix condition 4.5/4.6-only. Context: Cursor's API reports no window; 4.6's 500k is likewise the model's
+  published window, so 4.7 gets 500k from xAI's page and the measured xAI limit.
+- D7 accept for xAI prices; OpenRouter's distinct prices arrive through the regenerated bundled metadata rather than
+  a new overlay (4.6 has no OpenRouter overlay either). Devin-cli gets a derived row only if DEVIN_GROK equals xAI's
+  list price, labeled derived like the GPT-6 rows.
+- D8 accept.
+
+Reflection: see "Reflection" below.
+
+## File change map (dependency order)
+
+1. src/providers/registry/model-seeds.ts — XAI_MODELS: insert "grok-4.7" before "grok-4.6".
+   COMMAND_CODE_IMAGE_MODELS: add "xai/grok-4.6" and "xai/grok-4.7" with the probe note; drop xai/grok-4.6 from the
+   verified-negative header list (both mentions).
+2. src/providers/registry/entries-core.ts, xai block: modelSupportsServiceTier, modelWireDefaults (oauth, responses
+   inbound), modelInputModalities, preserveReasoningContentModels, modelReasoningEfforts [low..xhigh],
+   modelDefaultReasoningEfforts high, modelContextWindows 500_000 — each with a grok-4.7 sibling of 4.6; comments cite
+   devlog/_plan/260923_grok47_parity/010_probe-evidence.md. Devin block: add "grok-4-7" after "grok-4-6" in models.
+   OpenCode Go block: modelWireDefaults, modelReasoningEfforts, modelDefaultReasoningEfforts for grok-4.7.
+3. src/adapters/devin/live-models.ts — DEVIN_MODEL_CONTEXT_WINDOWS "grok-4-7": 500_000; DEVIN_MODEL_EFFORTS (if it
+   has per-model entries) "grok-4-7": low..max, default medium if a default map exists.
+4. (removed in revision 3 — see Audit round 1 synthesis; xai-responses-opt-in.ts is unchanged)
+5. src/usage/expected-prices.ts — xai grok-4.7 base {2,6,0.5,0}, priority 2x rule list gains grok-4.7, >=200k
+   UNIFORM_DOUBLE row with confirmedPriorityRelation lower-bound; devin-cli grok-4-7 conditional (D7).
+6. src/adapters/cursor/{catalog.ts,effort-map.ts,discovery.ts} — "grok-4.7" capability (displayName "Cursor Grok
+   4.7", CONTEXT_500K, no wirePrefix, regular+fast low..xhigh), tiers for "grok-4.7" and "grok-4.7-fast", heuristic
+   window 500_000 for grok-4.7 ids. Verify the Fast path emits flattened grok-4.7-<effort>-fast (accepted live) and
+   not the bare grok-4.7-fast (not_found live).
+7. scripts/model-metadata.source.json + src/generated/model-metadata.ts — add grok-4.7 rows beside each existing
+   grok-4.6 row for providers whose current models.dev entry lists 4.7 (xai, opencode-go, opencode, openrouter, kilo,
+   vercel, zenmux if present), copying that provider's live models.dev record; regenerate with
+   scripts/generate-model-metadata.ts.
+8. Tests (hosted CI runs them): tests/providers/provider-registry-parity.test.ts:1259 default-effort map;
+   tests/service/service-tier-capability.test.ts:125; tests/usage/usage-cost.test.ts:939; an xAI
+   wire-default case for 4.7 (OAuth Responses inbound resolves openai-responses; explicit modelAdapters Chat wins); command-code vision assertion
+   (tests/providers/command-code-provider.test.ts:219 flips 4.6 to image-capable); cursor effort/Fast wire-id cases for
+   4.7; opencode-go Responses wire case for 4.7. codex-catalog.test.ts is at its cap: no edits there.
+9. Docs: docs-site guides/codex-app-models.md model table (+ locales), reference/configuration/providers.md xAI Responses
+   default note if it lists models (the opt-in toggle list stays 4.5/4.6); structure/providers/xai-grok.md Fast set, structure/transports/responses.md:418,
+   structure/providers/cursor.md grok row.
+
+## Acceptance
+
+- A1 every xai-block map that names grok-4.6 also names grok-4.7 with the measured value (rg check).
+- A2 Cursor 4.7 wire ids equal the live roster: regular grok-4.7-<e>, Fast grok-4.7-<e>-fast, no cursor- prefix.
+- A3 (removed in revision 3): toggle set unchanged; xai-transport and management toggle tests stay as they are.
+- A4 generated metadata byte-matches the generator output (codegen run + model-metadata-sync test in CI).
+- A5 hosted CI green at the PR head, or failures diagnosed and fixed.
+
+
+## Reflection
+
+Architect 01a0caad-1195-74b3-8338-e7a354313b5d on revision 1: MISALIGNED (narrow), D1–D8 all mapped. Gaps and
+dispositions:
+
+- OpenRouter >=200k band rule: rebutted. `src/usage/expected-prices.ts` carries no OpenRouter context tier for any
+  model, including grok-4.6 whose OpenRouter entry publishes the same kind of override band. Adding one only for 4.7
+  would create a new, inconsistent pattern; it belongs in a separate change that covers OpenRouter tiers as a whole.
+  Recorded as a follow-up.
+- Broken evidence pointer (010 -> 010_plan.md): fixed to 000_plan.md, and the Cursor Fast success / bare-id
+  rejection recorded in 010_probe-evidence.md.
+- Missing Reflection section: this section (revision 2).
+
+
+## Audit round 1 synthesis (revision 3)
+
+Reviewer 01a0cab4-38a2-7ec0-8de2-02c1982a4345: FAIL, 2 High, both caused by D2 (4.7 joining the Responses toggle).
+
+Root cause: `XAI_RESPONSES_OPT_IN_MODELS` is the scope of a legacy compatibility switch (dashboard copy
+"Grok 4.5 and 4.6", management write path provider-routes.ts:449, v1 migration). grok-4.6's actual Responses
+default comes from `modelWireDefaults` (entries-core.ts:264), and that is the declaration parity requires.
+
+D2 amended (before -> after): before, 4.7 joins the toggle set and the migration gets a separate legacy list; after,
+the toggle set, its migration, the GUI copy and the management tests stay unchanged, and 4.7 receives the same OAuth
+Responses default through `modelWireDefaults` only. A user who wants 4.7 on Chat sets
+`modelAdapters["grok-4.7"]="openai-chat"`, which always wins (the same escape hatch the Go/Copilot pins document).
+Consequences: blocker 1 (xai-transport.test.ts:85, management-provider-validation.test.ts:4102/4123) and blocker 2
+(gui/src/i18n copy, ProviderAuthPanel mixed state) no longer arise; plan step 4 is removed, and so is acceptance A3.
+Non-blocking note folded: DEVIN_STATIC_MODELS (src/adapters/devin/live-models.ts:18) gains "grok-4-7".
+The proposed "grok-4-6" fallback entry was removed because its Devin-specific ladder was not measured.
@@ -0,0 +1,51 @@
+# Live probe evidence — grok-4.7 (2026-09-23 KST)
+
+Mechanics: POST /v1/responses on the running proxy (127.0.0.1:10100, ocx 2.62.0), then `ocx logs --json` for the
+matching attempt (adapter, credentialSource, reasoningWireField/Value, tierOutcome, usage). xAI traffic used the Grok
+OAuth lane (credentialSource "grok-oauth"); no API key was involved. The Responses-wire and `--fast` rows used a
+temporary `providers.xai.modelAdapters["grok-4.7"]="openai-responses"` plus
+`modelSupportsServiceTier["grok-4.7"]=true` override, applied through the attested provider reload
+(`notifyRunningProxy("xai")`, the path `ocx login` uses) and removed the same way afterwards; the restored maps were
+compared against a pre-probe backup. Scratch scripts lived in the gitignored `.tmp/`.
+
+## xAI grok-4.7 (Grok OAuth)
+
+| Probe | Chat wire (provider default) | Responses wire (temporary override) |
+|---|---|---|
+| effort low / medium / high / xhigh | 200, sent as `reasoning_effort` | 200 on all four |
+| effort max | 400 `Invalid reasoning effort.` | 400 `Invalid reasoning effort.` |
+| effort none | 200, but 640 reasoning tokens: the proxy omits the field and the model still reasons | not probed |
+| image, user message (3x3 random color grid, 180x180 PNG) | 9/9 | — |
+| image, tool result (same grid inside function_call_output) | 9/9 | — |
+| caller `service_tier: "priority"` | 200, wire service-tier priority, response tier priority | 200, applied/confirmed, response tier priority |
+| `xai/grok-4.7--fast` | — | 200, fastOutcome applied, confirmation confirmed, response tier priority |
+| 530,000-word prompt | 400 `context_length_exceeded`: "531243 tokens > 500000 tokens" | — |
+
+Upstream model name on the Responses wire is `grok-4.7-build` (grok-4.6 reports `grok-4.6-build` the same way).
+
+Billing parity: the Responses `cost_in_usd_ticks` fits these per-token rates exactly across every probe, for both models:
+default input 6800, cached input 1700, output 20400 ticks; priority input 40000, cached 10000, output 120000 ticks.
+grok-4.6 probed in the same window produced identical rates (e.g. 83 uncached + 128 cached input, 58 output =
+1,965,200 ticks). The OAuth subscription is not per-token billed, so these ticks are recorded as parity evidence only;
+the key-auth prices below come from xAI's published page.
+
+Published (docs.x.ai/developers/models/grok-4.7, read 2026-09-23): 500,000 context; reasoning effort
+low/medium/high (default)/xhigh, reasoning cannot be disabled; text+image input; $2.00 input, $0.50 cached,
+$6.00 output per 1M; prompts over 200k tokens $4.00 / $1.00 / $12.00; Responses and Chat Completions.
+models.dev `xai/grok-4.7`: output limit 500,000 (same as grok-4.6), released 2026-09-21.
+
+## Other providers
+
+| Provider | Evidence | Result |
+|---|---|---|
+| devin (`grok-4-7`) | live probe 200 at low and xhigh; tool-result grid 9/9; proxy /v1/models from Devin's live catalog: context 500000, input text+image, efforts low/medium/high/xhigh/max, default medium | exposes |
+| command-code (`xai/grok-4.7`) | live probe 200; grid 9/9 on user-message and tool-result paths; COMMAND_CODE_TEXT_ONLY_MODELS is empty and the logs show no vision-sidecar request, so the route read the image natively | exposes, native image |
+| cursor (`grok-4.7`) | live probe 200 through the cursor adapter; live GetUsableModels lists grok-4.7-{low,medium,high,xhigh} and the same ids with -fast (no cursor- prefix, no max); probes: grok-4.7-low and grok-4.7-xhigh-fast accepted, bare grok-4.7-fast rejected not_found (see 000_plan.md D6) | exposes |
+| opencode-go / opencode-zen | public `/zen/go/v1/models` and `/zen/v1/models` list `grok-4.7` | listed (not configured locally, not probed) |
+| openrouter | public API `x-ai/grok-4.7`: 500000 ctx, max completion 450000, $1.6/$4.8/$0.4, >=200k $3.2/$9.6/$0.8, text+image+file | listed (not probed) |
+| github-copilot | models.dev `grok-4.7`: ctx 500000, input 372000, output 128000 | listed (not configured locally) |
+| kilo, vercel | models.dev lists kilo `x-ai/grok-4.7` and vercel `spacexai/grok-4.7` | listed |
+
+Command Code `xai/grok-4.6` is included in `COMMAND_CODE_IMAGE_MODELS`: it read the grids 9/9
+(user message) and 8/9 (tool result) without a vision sidecar. The registry accepts native image
+input; 8/9 remains the measured limitation on the tool-result path.
@@ -143,8 +143,8 @@ approximation fondée sur un ancien modèle d'entrée.
 | Connexion Codex (ligne Daybreak transférée explicitement) | `openai/gpt-daybreak-blue-latest` uniquement lorsque l'entrée `customModels` exacte est configurée sur le fournisseur canonique `openai`. Elle conserve l'identifiant Daybreak transmis et utilise l'instantané de capacités Sol épinglé (contexte de 372 000 jetons ; compactage automatique à 334 800 jetons). |
 | OpenAI (clé API) | Exactement dix lignes avec espace de noms : `gpt-5.5`, `gpt-5.6`, Sol/Terra/Luna, les trois identifiants virtuels `*-pro` et les deux alias Daybreak (contexte de 1 050 000 jetons ; entrée maximale de 922 000 jetons pour les dix) |
 | OpenRouter | `openrouter/openai/gpt-5.6-sol`, `openrouter/openai/gpt-5.6-terra`, `openrouter/openai/gpt-5.6-luna` (1 050 000) |
-| Cursor | Le repli statique comprend `cursor/gpt-5.6-sol`, `cursor/gpt-5.6-terra` et `cursor/gpt-5.6-luna` (1 000 000), ainsi que des lignes ordinaires/rapides pour Grok 4.5 et 4.6 (500 000) ; 4.6 ajoute `xhigh`, et la découverte dynamique propre au compte détermine quelles lignes restent visibles. |
-| xAI | La découverte dynamique fait autorité. Le catalogue de secours comprend `xai/grok-4.6` et utilise `xai/grok-4.5` par défaut ; les deux ont une fenêtre de 500 000 jetons. Grok 4.6 propose `low` / `medium` / `high` / `xhigh` (valeur amont par défaut : `high`), tandis que Grok 4.5 s'arrête à `high`. |
+| Cursor | Le repli statique comprend `cursor/gpt-5.6-sol`, `cursor/gpt-5.6-terra` et `cursor/gpt-5.6-luna` (1 000 000), ainsi que des lignes ordinaires/rapides pour Grok 4.5, 4.6 et 4.7 (500 000) ; 4.6 et 4.7 ajoutent `xhigh`, et la découverte dynamique propre au compte détermine quelles lignes restent visibles. |
+| xAI | La découverte dynamique fait autorité. Le catalogue de secours comprend `xai/grok-4.6` et `xai/grok-4.7` et utilise `xai/grok-4.5` par défaut ; les trois ont une fenêtre de 500 000 jetons. Grok 4.6 et 4.7 proposent `low` / `medium` / `high` / `xhigh` (valeur amont par défaut : `high`), tandis que Grok 4.5 s’arrête à `high`. |
 
 Les entrées GPT-5.6 épinglées préservent exactement l'échelle amont. Sol et Terra proposent les niveaux de
 `low` à `ultra` ; Luna s'arrête à `max`. Sol utilise `low` par défaut, contre `medium` pour Terra et Luna.

@@ -622,12 +622,13 @@ Cursor est géré séparément comme adaptateur expérimental. `adapter: "cursor
 le sélecteur **Ajouter un fournisseur** du tableau de bord comme entrée expérimentale de la configuration locale,
 avec les métadonnées du catalogue statique de repli de Cursor. Lorsqu'un jeton d'accès Cursor est configuré,
 opencodex utilise le transport HTTP/2 direct de Cursor. Sa liste de repli intégrée comprend `gpt-5.6-sol` /
-`terra` / `luna` (contexte de 1M), les variantes ordinaires et Fast de Grok 4.5 et 4.6 (500K), ainsi que
-`kimi-k3` (262K) ; la découverte en direct détermine celles qui restent visibles pour le compte. Grok 4.6 expose
-`low` / `medium` / `high` / `xhigh` sous les deux formes, tandis que 4.5 s'arrête à `high`. Les requêtes Fast
-envoient le modèle Grok de base correspondant avec des paramètres `effort` et `fast=true` `requested_model`
-distincts ; les identifiants aplatis `cursor-grok-{version}-{effort}-fast` servent uniquement à la découverte et
-à la sélection. Cursor ne fournit Kimi K3 qu'avec des identifiants de protocole suffixés par l'effort ;
+`terra` / `luna` (contexte de 1M), les variantes ordinaires et Fast de Grok 4.5, 4.6 et 4.7 (500K), ainsi que
+`kimi-k3` (262K) ; la découverte en direct détermine celles qui restent visibles pour le compte. Grok 4.6 et 4.7 exposent
+`low` / `medium` / `high` / `xhigh` sous les deux formes, tandis que 4.5 s'arrête à `high`. Pour Grok 4.5 et 4.6,
+les requêtes Fast envoient le modèle de base avec des paramètres `effort` et `fast=true` distincts dans
+`requested_model` ; leurs identifiants aplatis `cursor-grok-{version}-{effort}-fast` servent uniquement à la découverte
+et à la sélection. Grok 4.7 figure sans préfixe `cursor-` et envoie directement `grok-4.7-{effort}-fast`.
+Cursor ne fournit Kimi K3 qu'avec des identifiants de protocole suffixés par l'effort ;
 `cursor/kimi-k3` expose donc une échelle `low` / `high` / `max` avec `max` par défaut, conformément à la valeur
 par défaut documentée de l'API du modèle. L'exécution native read/write/delete/ls/grep/shell/fetch pilotée par
 le serveur Cursor est désactivée par défaut, car elle contourne le parcours d'approbation et le bac à sable de