You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The repo owner has directed that the shipped defaults for consult lane models move forward — not just the config mechanism from #1286. Every adopter that hasn't configured consult.models should get these lanes by default:
Lane
Current shipped default
New shipped default
claude
claude-opus-4-6 (consult/index.ts:558)
claude-opus-5
codex
gpt-5.4 @ medium reasoning (consult/index.ts:417-419)
gpt-5.6-sol @ medium
gemini
agy default, no --model
unchanged
Live-probe evidence (both ids functionally verified — do not re-litigate from cached catalogs)
Per this repo's standing rule, model-id existence claims require a live probe, never a cached table:
claude-opus-5 — probed 2026-07-28 via claude --model claude-opus-5; the model self-identifies. Confirmed.
gpt-5.6-sol — probed 2026-07-29 via codex exec --model gpt-5.6-sol; returned a correct completion (11.7k tokens). Confirmed.
gpt-5.6 (plain) and gpt-5.6-codex — both probed 2026-07-29 and REJECTED: The '<id>' model is not supported when using Codex with a ChatGPT account. The -sol suffix is load-bearing; do not "simplify" the id.
Scope (BUGFIX/AIR-sized)
Flip the two hardcoded ids in packages/codev/src/commands/consult/index.ts (claude lane claudeQuery model; codex lane startThread model). Keep codex reasoning effort at medium.
CODEX_PRICING (consult/index.ts:386) hardcodes gpt-5.4 per-1M rates. Update to gpt-5.6-sol's published rates, verified against OpenAI's current pricing page at implementation time. If published per-1M rates for this id cannot be verified, record cost_usd as null for the lane rather than computing from gpt-5.4 rates (matches consult: configurable per-lane models and per-review-type lane selection #1286's cost rule: a confidently wrong number is worse than none).
Docs: codev/resources/commands/consult.md + skeleton counterpart; the "Multi-Agent Consultation" section of CLAUDE.md/AGENTS.md names GPT-5.4 — update BOTH files identically, and mirror every framework change across codev/ and codev-skeleton/.
A smoke test or CI-safe unit assertion that the shipped defaults are the ids above (so a future drive-by "correction" of the unusual -sol id fails a test instead of silently landing).
#1286 refactors these exact lines into config-driven resolution with the hardcoded values as fallback defaults. Whichever PR lands second rebases; the semantic is identical either way — these ids become the fallback when no consult.models config is present. If this issue lands first, #1286's "default preservation" test vectors must assert the NEW ids.
Out of scope
Default lane composition (which lanes run) — unchanged.
Decision (owner-directed, 2026-07-29)
The repo owner has directed that the shipped defaults for consult lane models move forward — not just the config mechanism from #1286. Every adopter that hasn't configured
consult.modelsshould get these lanes by default:claude-opus-4-6(consult/index.ts:558)claude-opus-5gpt-5.4@mediumreasoning (consult/index.ts:417-419)gpt-5.6-sol@medium--modelLive-probe evidence (both ids functionally verified — do not re-litigate from cached catalogs)
Per this repo's standing rule, model-id existence claims require a live probe, never a cached table:
claude-opus-5— probed 2026-07-28 viaclaude --model claude-opus-5; the model self-identifies. Confirmed.gpt-5.6-sol— probed 2026-07-29 viacodex exec --model gpt-5.6-sol; returned a correct completion (11.7k tokens). Confirmed.gpt-5.6(plain) andgpt-5.6-codex— both probed 2026-07-29 and REJECTED:The '<id>' model is not supported when using Codex with a ChatGPT account.The-solsuffix is load-bearing; do not "simplify" the id.Scope (BUGFIX/AIR-sized)
packages/codev/src/commands/consult/index.ts(claude laneclaudeQuerymodel; codex lanestartThreadmodel). Keep codex reasoning effort atmedium.CODEX_PRICING(consult/index.ts:386) hardcodes gpt-5.4 per-1M rates. Update to gpt-5.6-sol's published rates, verified against OpenAI's current pricing page at implementation time. If published per-1M rates for this id cannot be verified, recordcost_usdas null for the lane rather than computing from gpt-5.4 rates (matches consult: configurable per-lane models and per-review-type lane selection #1286's cost rule: a confidently wrong number is worse than none).codev/resources/commands/consult.md+ skeleton counterpart; the "Multi-Agent Consultation" section of CLAUDE.md/AGENTS.md names GPT-5.4 — update BOTH files identically, and mirror every framework change acrosscodev/andcodev-skeleton/.-solid fails a test instead of silently landing).Coordination with #1286 (in flight)
#1286 refactors these exact lines into config-driven resolution with the hardcoded values as fallback defaults. Whichever PR lands second rebases; the semantic is identical either way — these ids become the fallback when no
consult.modelsconfig is present. If this issue lands first, #1286's "default preservation" test vectors must assert the NEW ids.Out of scope