Executor tiers (T0–T4)

Every issue gets a tier during planning. The tier picks the executor + model variant. Cheapest model that can actually do the job.

Tier	Executor	Typical work	Est. cost
T0	Gemini flash	Typo, copy tweak, hide a column	cents
T1	Codex mini / Claude sonnet medium	Simple CRUD, <150 LoC	low
T2	Claude sonnet high	Multi-file, moderate complexity	moderate
T3	Claude opus	Architecture, migration, cross-cutting	high
T4	Claude opus + brainstorm phase	Research, RFC, ambiguous spec	highest

Routing rules

vibe-plan estimates tier from:

• Scope: LoC estimate, # of files, # of repos
• Novelty: exists-in-codebase pattern vs. new architecture
• Ambiguity: is the spec crisp? If not → T4 with brainstorm

Overrides in .vibe-flow.yaml:

tier_overrides:
  "ISSUE-42": T3
  "ISSUE-17": T0

Fallback chain

If an executor fails (rate-limit, timeout, unrecoverable error), vibe-ship falls back down the chain:

fallback_chain:
  - opus-4.6
  - sonnet-4.6-high
  - sonnet-4.6-medium

Never falls back below the issue’s original tier for review — max(tier, T2) is used as reviewer model regardless.

Cost budget

cost_budget_usd: 50
max_opus_per_wave: 2

Before starting a wave, vibe-flow sums estimated cost. If it exceeds budget, it asks you to confirm, suggest downgrades, or abort.