Tianyi Cui
5.1 KiB
RFC: Package Model Experience contract
Status: implemented
Problem
A package README can explain APIs and runtime mechanics without answering the question that dominates an agent harness's behavior and cost: what from this package reaches a model request, under which conditions, and how long those tokens remain. The omission is especially hard to audit in a plugin architecture. A consumer may turn a backend result into a tool message, a policy plugin may replace success with an error, compaction may remove old history, and an agent-scoped registration may change one agent's prompt or schemas while leaving every other agent unchanged. Reading only the nominally model-facing packages therefore misses real context effects, while reading source across every dependency is too expensive for routine review.
Decision
Every workspace package README ends with the canonical Model Experience section, immediately before ## Known Limitations and Deferred Work; a package on the no-limitations allowlist ends with Model Experience itself. Packages with direct, multi-surface, conditional, capped, or lifetime effects use the three-column table. Each row identifies a concrete request surface, says what the relevant model literally receives and when, and classifies the token effect. Stable source literals are quoted verbatim, with named placeholders only for interpolated values; summaries are reserved for data-dependent payloads, provider-owned text, or schemas too large to reproduce. The default subject is the conversation model; a package that invokes an auxiliary model, such as a summarizer or search provider, names that request separately. Agent-scoped visibility is stated where it changes which agent receives a contribution. Prompt text and tool schemas are described separately whenever configuration or scoping can hide one without the other.
Every package participates. A package with no model-context effect, or one simple effect rendered entirely by another package, can join the verifier's audited sentence allowlist. It then uses exactly one sentence beginning None, as or Indirectly, through instead of stretching a negative fact across a three-column table. Implementations that shape results, caps, history, lifetimes, or more than one request surface keep the table even when they add zero direct prompt tokens.
verify-package-readme-model-experience discovers packages from packages/*/*/package.json, requires one sibling README and the canonical final-section order, and validates one of two package-classified bodies outside fenced code. An allowlisted package must carry exactly one sentence with its assigned prefix; every other package must carry only the exact three-column header and at least one complete row. It runs in doc-sync and the parallel gate runner. The check owns package classification, structural presence, shape, and order; implementation review owns coverage and the truth of the prose.
Alternatives considered
- Document only packages that register prompts or tools — rejected because backends, policy plugins, adapters, persistence, scoping, and compaction change the content or lifetime of tokens without owning a model-facing schema.
- Generate one central context-cost catalog from source — rejected because an AST can find registrations but cannot infer semantic conditions such as history retention, output truncation, parent-versus-child visibility, or an auxiliary model boundary. The package README is the implementation-local contract; a central copy would add another drift surface.
- Require numeric token counts — rejected because exact counts depend on the selected model tokenizer, adapter serialization, configuration, and runtime data. The stable contract is the growth shape: fixed per request, conditional per call, retained, replaced, capped, or zero-direct.
- Allow zero-impact packages to omit the section — rejected because absence is ambiguous between an audited zero and forgotten documentation. One explicit sentence is cheap and mechanically distinguishable.
- Require the full table for audited zero or simple indirect packages — rejected because it spreads one fact across three cells and encourages repetitive prose. A gated sentence preserves explicit coverage without the ceremony.
- Convention without a gate — rejected because a repo-wide contract must also cover every future package; review memory cannot reliably detect an omitted README section.
Consequences
A reviewer can start at any package and see its contribution to the conversation model, child models, and auxiliary calls without reconstructing the full plugin graph. Token-budget work can distinguish repeated request overhead from data-dependent history, and agent-scoped changes have an explicit documentation checkpoint. Package authors pay for one small table or one classified sentence and must update it whenever model-visible behavior changes. Tables deliberately do not promise provider-exact token counts; measurements remain model- and workload-specific, while the documented growth and visibility contract stays stable.