Model tiering

Tier ladder — start at the lowest tier a task class has passed calibration for¶

Tier	Runs	Use for
0 script	`bin/*.py`, no model	fetch, paginate, dedupe, filter, count — any fixed command list
1 `haiku`	one shot, no tool loop, input prefetched to a file	mechanical classification, extraction, formatting
2 `sonnet`	main thread, default specialists	orchestration, judgment, tests, docs, lint
3 `opus`	reasoning specialists	high-risk items, adversarial review, architecture, ambiguous cross-file design

Escalate one tier at a time, one retry per tier, only on a checkable trigger: missing or schema-invalid envelope · failed completeness count (items out ≠ items in) · ## Confidence below 0.85 · the same failure signature twice · an item flagged high-risk. The retry reuses the brief file plus the prior output path, never the transcript. No trigger, no escalation — "looks hard" is not one.
De-escalate only on evidence: a foundry:calibrate result at the cheaper tier at least equal to the current one on the same inputs (agents/curator.md model policy). Haiku 5.5 bills 5× above 100K prompt tokens — keep Haiku runs under 100K.
Effective tier = the agent's frontmatter pin unless the call passes Agent(model=…); the override is how one call site drops a tier without touching the agent (§Parallel Spawn Ceilings in agent-spawn.md counts it).

No model pin on inline skills¶

A model: on a skill without context: fork replaces the session model for the rest of that turn and the next turn reverts, so the whole cached prompt prefix is rewritten when the model changes — measured $1.4 to $1.5 per /oss:resolve run in an Opus session, of which removing the pin saves about $0.4 (the former Sonnet turns now run on Opus); the large saving needs the session itself on Sonnet. model: belongs on agents and context: fork skills only. The session picks its model once; never switch the main-thread model mid-session. Measure with /foundry:audit --efficiency --measured (Phase B3, check R1).

Delegate lean — what a spawn receives¶

Every subagent starts at a floor of about 61K tokens (the global instructions included) that no prompt trims, so pay it once per batch, not per item. Keep the rest small:

Brief in a file, not the prompt body: goal, acceptance test, input paths with line ranges, constraints, output path, envelope schema. No session history, prior reasoning or inlined file contents.
Script the deterministic prefix. A spawn prompt that lists five or more literal gh/git commands is a bin/ script not yet written; the agent then reads the file it produced.
Bound tool output (--jq, head, line ranges): every result is re-read on every later call.
Envelope back, full output to a file; never fork; the fewest tools the agent type allows.
Batch items per agent up to about 55 calls (claude-config.md §Agent/Skill Spawn Discipline).
State a tool budget for classification-only spawns ("Read and Write only"). Unbounded, they drift into repository archaeology — measured: 25 calls, 61K to 143K context, for work that needed about 8.

Author checklist — before shipping a skill or agent¶

Inline skill: no model:. Agent or context: fork skill: model: set.
Every Agent() spawn has a tier on purpose — the pin, or model= — and Opus only where a trigger above names it.
No spawn prompt lists five or more literal CLI calls; the script exists in bin/ and is wired from the skill.
Every spawn prompt names its input file, output path, envelope and tool budget.
Bash blocks and AskUserQuestion gates follow the turn-batching and idle-gate rules (claude-config.md §Turn Batching, compaction.md).
Measured, not guessed: an inline model: pin change is checked with /foundry:audit --efficiency --measured before and after (it keeps a baseline and prints the delta): no model switch inside the skill's own span, pre-gate median not higher. The audit asks which projects to read: choose the repositories where the skill actually runs. Spawn tiers, briefs and scripts are measured with foundry:profile before and after.