Hints
Advisory fields capture authoring intent about the runtime resources an artifact ideally consumes:
effort_hint: low | medium | high | max
model_class_hint: nano | small | medium | large | frontier
Both are advisory only. Podium does not enforce them. Adapters and SDK consumers translate to host-native knobs (model selector, extended-thinking flag, retry budgets) where supported; otherwise ignored.
Applicable to types agent, skill, and command. Ingest lint warns when set on other types.
effort_hint
Hint about the reasoning budget the artifact ideally consumes:
| Value | Meaning |
|---|---|
low |
Quick, single-pass. No extended thinking. Tight token budget. |
medium |
Standard. May involve a small amount of extended thinking. Moderate token budget. |
high |
Deep reasoning. Extended thinking encouraged. Generous token budget. |
max |
Maximum effort. Extended thinking, retry loops, validator strictness. |
Use low for skills like greeting handlers or simple lookups, where time-to-first-token matters more than depth. Use max for agents that do open-ended investigation, multi-step planning, or anything where the right answer matters more than throughput.
model_class_hint
Hint about the model capability tier:
| Value | Approximate tier (industry-standard) |
|---|---|
nano |
Smallest tier (Haiku, GPT-4o-mini, Gemini Flash, etc.). |
small |
Cost-optimized tier. |
medium |
Standard mid-tier (Sonnet, GPT-4o, Gemini Pro mid-tier). |
large |
Higher-capability tier. |
frontier |
Top-of-line model (Opus, GPT-5, Gemini Ultra, etc.). |
The mapping to specific models is harness- and deployment-dependent. The tiers are author-facing labels; deployment configuration decides which model name backs each tier.
Advisory framing
These fields are advisory:
- Authors set them based on what the artifact ideally wants.
- Consumers (harness adapter, SDK caller) decide whether and how to honor them.
- Hosts without a notion of “model class” or “effort” ignore the fields.
- Podium itself never validates that a deployment has a model matching the hint.
Concretely: an artifact with model_class_hint: frontier does not fail to load when a deployment lacks a frontier-tier model. The host runtime makes the routing decision; the hint is one signal among many (cost budget, availability, user override).
Adapter support
No built-in adapter currently translates these fields. Custom SDK consumers that build their own routing logic can read the hints from the manifest and route accordingly. Adapter-level honoring lands as adapters add the capability.
Why these fields are advisory
Authors don’t know what models the consumer’s deployment has. A skill author at one company doesn’t know whether the consuming team has access to frontier models, whether their cost budgets allow it, or whether the runtime they use respects per-call model selection.
Recording author intent without enforcement keeps the artifact portable across deployments. Consumers with the relevant runtime knobs can opt to honor the hint; consumers without those knobs ignore it without breaking. The artifact still works either way.
Lint behavior
effort_hintormodel_class_hinton acontext,rule,hook, ormcp-serverartifact: lint warning (“hints apply to types: agent, skill, command”).- Both fields are optional. Missing fields produce no warnings.