credentials.yaml

# credentials.yaml — top-level credential registry for the quorum eval harness.
# Names must match ^[a-z0-9_]+$.
# Each entry is parsed by src/contracts/credential.ts (CredentialSchema).
# harnesses: lists the runtime-family name(s) this credential serves.

opus:
  # Direct-API opt-out. Pinned to claude-opus-4-8 (not the floating `opus` alias)
  # so the direct-vs-Bedrock comparison partner can't drift when Anthropic
  # repoints the alias, matching the sonnet5/sonnet46 pinning convention.
  model: claude-opus-4-8
  api: anthropic
  api_key_env: ANTHROPIC_API_KEY
  harnesses: [claude]

opus_bedrock:
  # Default Claude coding-agent path: Opus 4.8 via the Bedrock Mantle endpoint
  # (native Anthropic shape, bearer key, no SigV4). Region us-east-1 (Mantle is
  # In-Region-only; only us-east-1 serves opus-4-8/sonnet-5 in the US). Bare id
  # `anthropic.claude-opus-4-8` is the canonical Mantle model id.
  # max_concurrency 6: quota probed 2026-07-16 (20M in / 4M out TPM, no RPM row; PRI-2650).
  model: anthropic.claude-opus-4-8
  api: mantle
  auth: bedrock-bearer
  api_key_env: AWS_BEARER_TOKEN_BEDROCK
  region: us-east-1
  harnesses: [claude]
  max_concurrency: 6

opus5:
  # Direct-API Opus 5: the Claude Code partner column for copilot_opus5 —
  # same model, each harness on its vendor-native routing (Anthropic API vs
  # GitHub). Bare id: the Claude 5 family ships undated ids.
  model: claude-opus-5
  api: anthropic
  api_key_env: ANTHROPIC_API_KEY
  harnesses: [claude]

opus5_bedrock:
  # Opus 5 via the Bedrock Mantle endpoint, mirroring opus_bedrock's shape.
  # Model id probed live 2026-08-06: HTTP 200 in us-east-1, response.model
  # claude-opus-5; near-miss ids 404. max_concurrency 4 (not opus_bedrock's
  # 6): Service Quotas gives 20M in / 2M out TPM — output is HALF of
  # opus-4-8's 4M — and observed appliance concurrency has never exceeded 4.
  model: anthropic.claude-opus-5
  api: mantle
  auth: bedrock-bearer
  api_key_env: AWS_BEARER_TOKEN_BEDROCK
  region: us-east-1
  harnesses: [claude]
  max_concurrency: 4

sonnet:
  model: sonnet
  api: anthropic
  api_key_env: ANTHROPIC_API_KEY
  harnesses: [claude]

haiku:
  model: claude-haiku-4-5-20251001
  api: anthropic
  api_key_env: ANTHROPIC_API_KEY
  harnesses: [claude]

sonnet5:
  # Claude Sonnet 5. Pinned to the stable generation id `claude-sonnet-5` (the
  # bare id is canonical; Anthropic publishes no dated snapshot for it) rather
  # than the floating `sonnet` alias, so the benchmarked model can't silently
  # repoint to a later Sonnet and break longitudinal comparability.
  # COST NOTE: priced by obol ≥0.7.0 at litellm's INTRO rate ($2/$10 per MTok,
  # window through 2026-08-31; standard is $3/$15). obol has no time-windowed
  # rates, so dollar figures for runs captured after the window closes will be
  # understated until a post-window prices refresh picks up the standard rate.
  # Runs captured on obol 0.6.0 remain est_cost_usd: null (frozen at capture).
  model: claude-sonnet-5
  api: anthropic
  api_key_env: ANTHROPIC_API_KEY
  harnesses: [claude]

sonnet46:
  # Claude Sonnet 4.6 — the previous-generation Sonnet, pinned as the control for
  # `sonnet5` (Sonnet 5) comparisons so the benchmarked model can't drift with the
  # floating `sonnet` alias. Unlike sonnet5, obol's bundled rate table includes
  # claude-sonnet-4-6, so this column reports real coding cost.
  model: claude-sonnet-4-6
  api: anthropic
  api_key_env: ANTHROPIC_API_KEY
  harnesses: [claude]

opencode_gpt5:
  # OpenAI gpt-5.5. base_url is OpenAI's first-party host (api.openai.com), which
  # the opencode translator routes to opencode's BUILT-IN `openai` provider (by
  # name, no custom npm) instead of a custom 'quorum' block: the built-in provider
  # uses @ai-sdk/openai, which sends `max_completion_tokens` for gpt-5.x reasoning
  # models (they reject `max_tokens`) and self-reports cost. A non-OpenAI
  # openai-chat endpoint (GLM, ollama) keeps the custom 'quorum' + compatible shim.
  model: gpt-5.5
  api: openai-chat
  base_url: https://api.openai.com/v1
  api_key_env: OPENAI_API_KEY
  harnesses: [opencode]
  # max_concurrency 5: first-party OpenAI (api.openai.com), same endpoint and
  # quota pool as openai_responses; parallelize the opencode column to match.
  max_concurrency: 5

serf_default:
  # serf.yaml model: anthropic/claude-sonnet-4-6
  model: claude-sonnet-4-6
  api: anthropic
  api_key_env: ANTHROPIC_API_KEY
  harnesses: [serf]
  max_concurrency: 1

kimi_default:
  # src/agents/kimi.ts:45 — DEFAULT_KIMI_MODEL_ENV.KIMI_MODEL_NAME = 'kimi-for-coding'
  # Auth: OAuth-or-env. Primary path is OAuth (seeded from ~/.kimi-code);
  # KIMI_MODEL_API_KEY enables the api-key path. Default here is oauth.
  model: kimi-for-coding
  auth: oauth
  harnesses: [kimi]

openrouter_kimi_k27_code:
  # OpenRouter-hosted Kimi coding flagship via OpenAI-compatible chat; added
  # 2026-08-06 after the Moonshot direct-API credits exhausted mid-battery.
  # NOT poolable with kimi-CLI (kimi-for-coding) cells -- different serve.
  model: moonshotai/kimi-k2.7-code
  api: openai-chat
  base_url: https://openrouter.ai/api/v1
  api_key_env: OPENROUTER_API_KEY
  harnesses: [pi, opencode]
  max_concurrency: 5

openrouter_glm_5_2:
  # Default Pi path: OpenRouter-hosted GLM 5.2 via OpenAI-compatible chat.
  # The provider/model choice lives in checked-in config; only the key belongs in
  # the local/appliance credential bundle as OPENROUTER_API_KEY.
  model: z-ai/glm-5.2
  api: openai-chat
  base_url: https://openrouter.ai/api/v1
  api_key_env: OPENROUTER_API_KEY
  harnesses: [pi, opencode, hermes]
  # max_concurrency 5: pi's api-key path reaches GLM 5.2 through OpenRouter.
  # Dial back if 429s surface as indeterminates.
  max_concurrency: 5
  compat:
    thinking_format: zai

pi_default:
  # Native pi OAuth (host login seeded from ~/.pi/agent). Provider PINNED to
  # openai-codex (OpenAI gpt-5.5 via the ChatGPT subscription) so eval runs use a
  # reproducible provider instead of inheriting the mutable host settings.json
  # defaultProvider. Requires an openai-codex login in the host auth.json. This is
  # pi's checked-in DEFAULT (see coding-agents/pi.yaml); openrouter_glm_5_2 remains
  # an opt-in alternative.
  provider: openai-codex
  model: gpt-5.5
  auth: oauth
  harnesses: [pi]
  max_concurrency: 1

codex_sub:
  # Codex uses ChatGPT subscription auth (auth_mode: chatgpt, no OPENAI_API_KEY).
  # The adapter does not pass a --model flag; model selection is subscription-driven.
  # CONCERN: model field is a placeholder (no model baked into adapter).
  model: gpt-4o
  auth: subscription
  api: openai-responses
  harnesses: [codex]

gemini_default:
  # Gemini CLI defaults its own model; the adapter does not pin one.
  # CONCERN: model field is a placeholder (no default model baked into adapter).
  model: gemini-2.5-pro
  api: gemini
  api_key_env: GEMINI_API_KEY
  harnesses: [gemini]
  max_concurrency: 1

# GLM dev endpoint is dual-protocol (chat-completions AND /v1/responses), but a
# credential carries ONE wire protocol (limiterKey = base_url|api). Split per
# protocol: pi/opencode used chat-completions; codex 0.141 only speaks responses.
# Both are opt-in benchmark credentials (no agent defaults to them).
openrouter_hermes4:
  # KNOWN-BROKEN as of 2026-07-23: no OpenRouter hermes-4 endpoint (405b or
  # 70b) supports the `tools` parameter, and the hermes CLI is tool-driven —
  # every API call 404s with "No endpoints found that support tool use"
  # (experiments/2026-07-23-hermes-target-bringup.md). Kept as a documented
  # disproof; Hermes-4 validation needs the Nous Portal API + key instead.
  model: nousresearch/hermes-4-405b
  api: openai-chat
  base_url: https://openrouter.ai/api/v1
  api_key_env: OPENROUTER_API_KEY
  harnesses: [hermes]
  max_concurrency: 5

glm_5_2_chat:
  model: glm-5.2-fp8
  api: openai-chat
  base_url: https://oak-receiver-hear-homework.trycloudflare.com/v1
  api_key_env: GLM_API_KEY
  harnesses: [pi, opencode]
  max_concurrency: 2
  compat:
    thinking_format: zai

glm_5_2_responses:
  model: glm-5.2-fp8
  api: openai-responses
  base_url: https://oak-receiver-hear-homework.trycloudflare.com/v1
  api_key_env: GLM_API_KEY
  harnesses: [codex]
  max_concurrency: 2
  compat:
    thinking_format: zai

# Local ollama (OpenAI chat-completions at localhost:11434/v1) — live validation
# target for the pi/opencode translators. ollama ignores the API key; OLLAMA_API_KEY
# is a dummy placeholder in .env. Opt-in (no agent defaults to it).
ollama_local:
  model: gemma4:e4b
  api: openai-chat
  base_url: http://localhost:11434/v1
  api_key_env: OLLAMA_API_KEY
  harnesses: [pi, opencode]

# OpenAI's real Responses API — exercises codex's api-key/responses translator
# path (wire_api=responses) against a live endpoint, standing in for the dead GLM
# endpoint. Opt-in (no agent defaults to it).
# max_concurrency 5: the earlier no-response hangs (first seen at 4) were OpenAI
# quota exhaustion (billing), not concurrency throttling — resolved by topping up
# the account. Dial back if genuine 429s surface as indeterminates.
openai_responses:
  model: gpt-5.5
  api: openai-responses
  base_url: https://api.openai.com/v1
  api_key_env: OPENAI_API_KEY
  harnesses: [codex]
  max_concurrency: 5

openai_responses_56sol:
  # GPT-5.6 Sol comparison partner for openai_responses (gpt-5.5), following the
  # sonnet5/sonnet46 pinning convention. Pinned to the explicit gpt-5.6-sol id —
  # bare `gpt-5.6` is a floating alias that routes to Sol (no dated snapshot
  # published as of 2026-07-13). Sol prices identically to gpt-5.5 ($5/$0.50/$30
  # per MTok, 272k long-context tier at $10/$45); priced by obol ≥0.8.0
  # (as_of 2026-07-14). Runs captured on obol 0.7.0 remain est_cost_usd: null —
  # cost is frozen at capture time.
  # Shares openai_responses's limiterKey (same base_url|api), so both columns
  # draw from ONE 5-slot pool and one 429 latch; keep max_concurrency equal —
  # on a limiterKey collision the alphabetically-first credential's cap wins.
  model: gpt-5.6-sol
  api: openai-responses
  base_url: https://api.openai.com/v1
  api_key_env: OPENAI_API_KEY
  harnesses: [codex]
  max_concurrency: 5

antigravity_default:
  # DORMANT: no antigravity translator consumes this; it carries the scheduler cap
  # (PRI-2203: Code Assist 429 protection) + satisfies quorum check. model is a
  # placeholder (antigravity is Code Assist / Gemini-based; confirm exact model
  # if needed — the adapter does not bake one in).
  # Auth sourced from src/agents/antigravity.ts:67-101 — AGY_OAUTH_CREDENTIAL_FILES
  # (oauth_creds.json, google_accounts.json) seeded from ~/.gemini; OAuth path.
  model: gemini-code-assist
  auth: oauth
  harnesses: [antigravity]
  max_concurrency: 1
  launch_spacing_seconds: 30

# Copilot credentials: model reaches the launcher's --model flag via the
# $COPILOT_MODEL_SH substitution. Auth sourced from
# src/agents/copilot.ts — resolveCopilotAuthEnv() resolves
# COPILOT_GITHUB_TOKEN / GH_TOKEN / GITHUB_TOKEN / gh auth token; GitHub
# token auth path (fine-grained PAT with the Copilot Requests permission).
# All copilot credentials share one base_url so the scheduler folds them
# into a single limiter pool: one GitHub Copilot subscription serves them.
copilot_default:
  # claude-opus-5 is the Copilot CLI's own default model; pinned explicitly
  # so casual runs are labeled with the model they actually exercised.
  model: claude-opus-5
  auth: oauth
  base_url: https://api.githubcopilot.com
  harnesses: [copilot]
  max_concurrency: 1

copilot_gpt56_sol:
  model: gpt-5.6-sol
  auth: oauth
  base_url: https://api.githubcopilot.com
  harnesses: [copilot]
  max_concurrency: 1

copilot_gpt56_luna:
  model: gpt-5.6-luna
  auth: oauth
  base_url: https://api.githubcopilot.com
  harnesses: [copilot]
  max_concurrency: 1

copilot_opus5:
  model: claude-opus-5
  auth: oauth
  base_url: https://api.githubcopilot.com
  harnesses: [copilot]
  max_concurrency: 1

copilot_mai_flash:
  # The CLI's embedded id carries the -picker suffix; verified live 2026-08-05
  # (session events report currentModel mai-code-1-flash-picker verbatim).
  model: mai-code-1-flash-picker
  auth: oauth
  base_url: https://api.githubcopilot.com
  harnesses: [copilot]
  max_concurrency: 1