feat(routing): tier parity when Codex drives (GPT-6 Astra as reasoning tier)

[pacphi]

Problem

When Claude drives, reasoning work runs on Claude's reasoning-tier model (claude-opus-5-5), and the implementation/testing escalation steps up a tier (gpt-6-sol → claude-opus-5-5).

When Codex drives, the seed mirrors those routes, but Codex had no reasoning tier:

• Reasoning work fell back to the workhorse (gpt-6-sol).
• The escalation stepped sideways, claude-sonnet-5 → gpt-6-sol. Both cost $2/$10, so escalating gained nothing.

ADR-0006 had recorded this as a known consequence: "the mirror is coarse where host model tiers don't line up".

Change

• gpt-6-astra takes the Codex reasoning tier, paired with claude-opus-5-5.
• Catalog entries can list pairs: extra tiers a model answers for when the other host has no model of that tier. Astra pairs with Claude's flagship, so Fable 5.1 pins still mirror to Astra rather than to the workhorse.
• Claude-led defaults are unchanged.

Tier	Claude	Codex
reasoning	claude-opus-5-5	gpt-6-astra
balanced	claude-sonnet-5	gpt-6-sol
fast	claude-haiku-4-5-20251001	gpt-6-luna

Codex-led seed:

Activity	Before	After
architecture, design, security-analysis, debugging	codex:gpt-6-sol	codex:gpt-6-astra
implementation, testing	claude:claude-sonnet-5 → codex:gpt-6-sol	claude:claude-sonnet-5 → codex:gpt-6-astra
specification, review, release	codex:gpt-6-sol	unchanged
security-scan, docs, packaging	Claude Sonnet / Haiku	unchanged

Trade-off (stated in ADR-0006 and PROVIDERS.md)

Matching tiers doesn't mean matching prices. Astra is $10/$50 per million tokens; Opus 5.5 is $4/$20. Why Astra anyway:

• OpenAI positions Astra for "the hardest end-to-end work" (Codex models).
• Anthropic reports Opus 5.5 at Fable 5.1 level, which is Astra's class (models overview).

The cost: Codex-led reasoning uses more Codex plan allowance. We considered two alternatives:

• Sol at higher reasoning effort. Routes can't carry an effort level yet, so this would need adapter changes.
• Sol leading, with Astra only for escalation. This gives reasoning work less than Claude-led routing does.

Docs

• ADR-0006: Updated note; replaces the "mirror is coarse" consequence.
• ADR-0002: tier names and pairs.
• PROVIDERS.md: tier table and the Codex-led seed.
• MODEL-PRICING-AUDIT.md.

Verification

• pnpm run check: exit 0. Tests: 4327 pass, 0 fail, 6 skipped. Lint warnings are the same as on main.
• New tests in routing-primary.test.mjs fail on main and pass here. They cover:
  • reasoning work on each host's reasoning model;
  • escalation stepping up a tier whichever host drives;
  • tier pairs mirroring both ways, including Fable 5.1 → Astra.
• Routes seeded before this change keep their model. They show as diverged until ak x host refresh.

🤖 Generated with Claude Code