Add GPT-5.6 capability sheets (sol/terra/luna)

docs/MODEL-NOTES.md

@@ -287,6 +287,13 @@ checks and raw logs support — no vibes, no worker self-reports.
 
 ## gpt-5.6-sol (codex)
 - 2026-07-15 ringer-self-update run (3 serial tasks, direct-repo-edit mode): code-fix baseline-test repair 1/1 first-try (61k tokens, 1.6m); code-feature self-update mechanism (git fetch/ff-pull/re-exec + HUD staleness restart + 20-test suite) 1/1 first-try at high effort (153k, 8.1m); code-feature signal-contract (all 3 scoreboard surfaces + canonical-route lint enforcement) passed on retry (358k, 13.7m) — attempt 1 died on stale old-column assertions in pre-existing tests it hadn't finished updating; the retry prompt's injected FAIL list was enough to close it out. Lesson: when a task rewrites a display contract, name every test file asserting the old contract in the spec's ownership list AND tell it to update them FIRST.
+- 2026-07-13 reasoning-effort surfaces (dated observation): codex CLI on this
+  box accepted `model_reasoning_effort=max` for gpt-5.6-sol. As of 2026-07-16
+  the Codex config reference types the field `minimal|low|medium|high|xhigh`
+  (no `max`), while OpenRouter metadata lists supported_efforts including
+  `max` for ALL three 5.6 tiers. The surfaces disagree — per-tier effort
+  ceilings are unsettled; the capability sheets carry only the per-surface
+  documented facts.
 - 2026-07-09 code-feature/code-fix (ringside-overhaul): 4/4 first-try — a ringer.py logging change with tests, a 265-line stdlib backfill CLI (atomic rewrite, dry-run, idempotence all check-verified), a ~1500-line single-file HTML redesign (running-now pills + worker-card grid + multi-expansion refactor, 30KB patch, node --check + contract greps + unittest), and a render-gating change where it correctly UPDATED tests asserting the old behavior instead of gaming the check. Medium/high reasoning, 65–120k tokens/task.
 - Same day, different session (bench-harness-patches, code-fix): 0.29 first-try over 7 tasks on a Next.js/Turbopack harness. Spec and check quality dominate model choice — see the scoreboard before generalizing either number.
 
@@ -309,6 +316,13 @@ checks and raw logs support — no vibes, no worker self-reports.
 
 ## gpt-5.6-luna (codex)
 - 2026-07-09 code-feature (unlock-ai guide-format conversion, strict type-contract check): 1/1 first-try, 42.6k tokens, 80s. Followed a multi-file TS pattern precisely at $1/$6 pricing. Good candidate for mechanical codegen/docs lanes; audition in adjacent types.
+- 2026-07-14 long-context (dated observation, no run): Luna's nominal window
+  is 1.05M, same as Sol/Terra, but a search that day found no tier-by-tier
+  long-context recall benchmark from OpenAI or any third party — per-tier
+  long-context behavior is unverified in both directions. Routing judgment,
+  not scoreboard evidence: until run evidence accumulates, the conservative
+  slot for Luna is short-context, low-stakes, latency-sensitive work; treat
+  long-context Luna assignments as exploration, not proven routing.
 
 ## opencode / z-ai glm-5.2 (via openrouter)
 - 2026-07-09 (aicred-invoice-downloads, 4 code-fix tasks + 1 follow-up, worktrees+npm ci checks): systematic attempt-1 NO-OP — all 4 parallel workers produced zero edits and no summary on first attempt, then completed cleanly on attempt 2 after retry-prompt injection (34k-69k tokens each). Follow-up single task passed attempt 1. Suspect first-invocation session warm-up in opencode-sandboxed under parallel spawn; budget for 2 attempts on parallel GLM batches. Output quality on Next.js/Stripe route+test work: solid, spec-faithful, one boss-caught design gap (used user-scoped supabase client where RLS demanded service role — spec didn't say explicitly; say it explicitly).

registry/model-capabilities/gpt-5.6-luna.toml

@@ -0,0 +1,139 @@
+[model]
+key = "gpt-5.6-luna"
+display = "GPT-5.6 Luna"
+vendor = "OpenAI"
+release_date = "2026-07-09"
+identity_verification = "OpenAI's developer model page for gpt-5.6-luna lists it as a member of the GPT-5.6 family (Sol/Terra/Luna) with a 1,050,000 context window, 128,000 max output tokens, reasoning token support, and a Feb 16, 2026 knowledge cutoff (fine-tuning listed as Not supported). GA of the GPT-5.6 family is reported as 2026-07-09 by third-party coverage (single source); the model page carries the Feb 16 2026 knowledge cutoff but no in-page GA date."
+
+[api]
+endpoint_families = ["responses", "chat-completions", "batch"]
+auth = "In this setup, Codex CLI signs in with ChatGPT/OAuth plan credentials and uses ChatGPT workspace permissions, plan access, and ChatGPT-managed Codex entitlements. Codex CLI can also sign in with an OpenAI API key; API-key usage is billed through the OpenAI Platform account at standard API rates instead of included ChatGPT plan credits."
+streaming = "Supported. API streaming is available with stream=true on Responses and Chat Completions; streaming responses use server-sent events. Family-level (GPT-5.6) behavior; the per-tier model page does not separately enumerate a streaming flag."
+
+[caching]
+supported = true
+mode = "implicit"
+how_enabled = "Prompt caching works automatically for eligible requests with no code changes; prompt_cache_options.mode defaults to implicit (the mode field records this default). For GPT-5.6, explicit cache breakpoints are additionally available in both the Responses and Chat Completions APIs, and prompt_cache_key must be set to get the more reliable cache matching. Caching starts at 1024 prompt tokens (cached_tokens is zero for requests under 1024); repeated prompt prefixes, tools, and structured-output schemas can contribute to the cached prefix."
+ttl = "For GPT-5.6, prompt_cache_options.ttl sets a MINIMUM cache lifetime; the only supported value is 30m (also the default) — a cached prefix remains eligible for reuse for at least 30 minutes and may be retained longer. The prompt_cache_retention field used by earlier families (24h on GPT-5.5) is deprecated for GPT-5.6 and later."
+read_pricing_per_m = 0.10
+write_pricing_per_m = 1.25
+min_cacheable_prefix_tokens = 1024
+
+[tool_calling]
+format = "OpenAI Responses API tools and function calls with JSON Schema tool parameters. Chat Completions function calling is also supported. Structured Outputs are supported, including strict JSON Schema mode with strict=true. Family-level GPT-5.6 behavior inherited from the OpenAI platform; not separately re-verified on the per-tier model page."
+parallel = true
+strict_json_schema = true
+quirks = [
+  "For reasoning models using function calling through the Responses API, pass back reasoning items, function call items, and function call output items during tool loops.",
+  "Prompt-only JSON mode can still produce incomplete JSON when output is cut off by max_output_tokens or a content filter; use Structured Outputs for schema adherence.",
+]
+
+[reasoning]
+params = ["reasoning.effort", "text.verbosity", "max_output_tokens", "model_reasoning_effort", "model_reasoning_summary", "model_verbosity"]
+defaults = "The gpt-5.6-luna model page confirms reasoning token support but does not enumerate an effort scale. Effort scales are surface-specific: the Codex CLI config reference types model_reasoning_effort as minimal|low|medium|high|xhigh (xhigh is model-dependent), and OpenRouter endpoint metadata lists supported_efforts [max, xhigh, high, medium, low, none] with default medium for openai/gpt-5.6-luna. OpenAI publishes no per-tier effort ceiling; Ringer passes effort per-task via engine_args [\"-c\",\"model_reasoning_effort=...\"]."
+max_output_interaction = "max_output_tokens limits total generated tokens, including reasoning tokens, visible output tokens, and non-visible formatting tokens. If the cap is exhausted during reasoning, the response can be incomplete before any visible output is produced."
+
+[limits]
+context_window = 1050000
+max_output_tokens = 128000
+rate_limits = "Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request (per the gpt-5.6-luna model page). ChatGPT-plan Codex OAuth usage is governed by plan entitlements, not API billing tiers; per-tier long-context API rate-limit tiers are not published on the model page."
+
+[openrouter]
+upstream_providers = ["openai"]
+require_parameters_behavior = "not applicable; this registry entry is for OpenAI direct/Codex OAuth access. gpt-5.6-luna is also listed on OpenRouter as openai/gpt-5.6-luna (see sources); OpenRouter routing would pass the model slug through the opencode engine's model field."
+
+[pricing]
+prompt_per_m = 1.00
+completion_per_m = 6.00
+cache_read_per_m = 0.10
+cache_write_per_m = 1.25
+as_of = "2026-07-16"
+
+[[sources]]
+claim = "model: gpt-5.6-luna context window"
+url = "https://developers.openai.com/api/docs/models/gpt-5.6-luna"
+accessed = "2026-07-14"
+quote = "1,050,000 context window"
+
+[[sources]]
+claim = "model: gpt-5.6-luna max output tokens"
+url = "https://developers.openai.com/api/docs/models/gpt-5.6-luna"
+accessed = "2026-07-14"
+quote = "128,000 max output tokens"
+
+[[sources]]
+claim = "pricing: gpt-5.6-luna input, cached input, cache writes, and output per 1M"
+url = "https://developers.openai.com/api/docs/pricing"
+accessed = "2026-07-16"
+quote = "Input Cached input Cache writes Output ... gpt-5.6-luna $1.00 $0.10 $1.25 $6.00"
+note = "The pricing page's GPT-5.6 rows carry an explicit Cache writes column (older families show '-'). The $1.25/M cache-write price is independently corroborated by OpenRouter endpoint metadata (input_cache_write = 0.00000125 for openai/gpt-5.6-luna, accessed 2026-07-16)."
+
+[[sources]]
+claim = "limits: gpt-5.6-luna large-prompt (>272K) pricing rule"
+url = "https://developers.openai.com/api/docs/models/gpt-5.6-luna"
+accessed = "2026-07-14"
+quote = "Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request"
+
+[[sources]]
+claim = "model: gpt-5.6-luna knowledge cutoff"
+url = "https://developers.openai.com/api/docs/models/gpt-5.6-luna"
+accessed = "2026-07-14"
+quote = "Feb 16, 2026 knowledge cutoff"
+
+[[sources]]
+claim = "model: gpt-5.6-luna fine-tuning not supported"
+url = "https://developers.openai.com/api/docs/models/gpt-5.6-luna"
+accessed = "2026-07-16"
+quote = "Fine-tuning Not supported"
+
+[[sources]]
+claim = "reasoning: Codex CLI reasoning-effort scale"
+url = "https://developers.openai.com/codex/config-reference"
+accessed = "2026-07-16"
+quote = "minimal | low | medium | high | xhigh"
+note = "The config reference types model_reasoning_effort with this enum ('xhigh is model-dependent') and does not include 'max'."
+
+[[sources]]
+claim = "reasoning: OpenRouter supported reasoning efforts for openai/gpt-5.6-luna"
+url = "https://openrouter.ai/api/v1/models"
+accessed = "2026-07-16"
+quote = "\"supported_efforts\":[\"max\",\"xhigh\",\"high\",\"medium\",\"low\",\"none\"],\"default_effort\":\"medium\""
+
+[[sources]]
+claim = "tool_calling: strict structured outputs use JSON Schema strict true (platform, version-independent)"
+url = "https://developers.openai.com/api/docs/guides/structured-outputs"
+accessed = "2026-07-16"
+quote = "strict: true"
+
+[[sources]]
+claim = "caching: GPT-5.6 cache writes are billed at 1.25x the uncached input rate"
+url = "https://developers.openai.com/api/docs/guides/prompt-caching"
+accessed = "2026-07-16"
+quote = "On GPT-5.6 models and later model families, cache writes are billed at 1.25× the uncached input token rate and reported in `cache_write_tokens`."
+note = "Direct OpenAI source (replaces an earlier search-relayed quote from openai.com/index/gpt-5-6, which 403s to direct fetch). 1.25x of $1.00 input = $1.25/M, matching the pricing page's Cache writes column."
+
+[[sources]]
+claim = "caching: explicit cache breakpoints and 30m minimum cache life"
+url = "https://developers.openai.com/api/docs/guides/prompt-caching"
+accessed = "2026-07-16"
+quote = "use `prompt_cache_options.ttl` to set the minimum lifetime of all breakpoints written by the request. The only supported value is `30m`, which is also the default."
+note = "The guide also states: 'you can mark the end of a reusable prompt prefix with an explicit cache breakpoint. Breakpoints are available in both the Responses API and Chat Completions API.', documents `prompt_cache_options.mode` with implicit as the default, requires prompt_cache_key for the more reliable matching, and deprecates prompt_cache_retention for GPT-5.6 and later."
+
+[[sources]]
+claim = "caching: 1024-token floor for cached_tokens"
+url = "https://developers.openai.com/api/docs/guides/prompt-caching"
+accessed = "2026-07-16"
+quote = "For requests under 1024 tokens, `cached_tokens` is zero."
+
+[[sources]]
+claim = "release: GPT-5.6 family (Sol/Terra/Luna) GA date (third-party, single source)"
+url = "https://www.digitalapplied.com/blog/gpt-5-6-sol-terra-luna-public-ga"
+accessed = "2026-07-14"
+quote = "GPT-5.6 went to general availability on July 9, 2026 — across ChatGPT, Codex, and the OpenAI API"
+note = "release_date records the public GA date, not first availability (pre-GA access via Codex existed for the family; see docs/MODEL-NOTES.md, 'GPT-5.5 (codex) — attribution caveat'). Do not use release_date as a lower bound for scoreboard row attribution."
+
+[[sources]]
+claim = "openrouter: gpt-5.6-luna listing and cache-write price metadata"
+url = "https://openrouter.ai/api/v1/models"
+accessed = "2026-07-16"
+quote = "\"id\": \"openai/gpt-5.6-luna\", \"name\": \"OpenAI: GPT-5.6 Luna\" ... \"input_cache_read\": \"0.0000001\", \"input_cache_write\": \"0.00000125\""