tea-context

Tokio-free prompt modules, context-provider contracts, deterministic compilation, provenance, trust labels, budgets, diagnostics, and inspection for tea-rs.

The package is tea-context; Rust code imports it as tea_context. Its normal dependency graph contains no model provider, tool executor, JSON Schema compiler, filesystem, process, network, database, or async runtime.

Prompt modules and segments

Each PromptSegment carries a canonical segment ID, non-empty bounded content, PromptProvenance, a TrustLevel, a CacheScope, an optional ConflictClaim, and a BudgetBehavior (Required, Truncate, or Omit). Segments are grouped into non-empty PromptModule values with a canonical module ID, a fixed PromptAuthority, and a bounded numeric PromptPriority.

PromptAuthority fixes high-to-low precedence: Kernel, Organization, Product, Workspace, Tool, Skill, Session, UserAddition. Numeric priority only breaks ties within one authority.

Provider contract

ContextProvider is object-safe and returns a project-owned boxed standard Future. It receives an immutable ContextRequest snapshot of the active profile, session, run, canonical active tools, and bounded metadata. Built-in providers perform no I/O: ToolHintProvider renders active ToolSpec::prompt_hint() guidance, WorkspaceInstructionProvider carries caller-supplied documents, SessionSummaryProvider inserts an optional durable summary, and SkillMetadataProvider exposes metadata with one explicit @skill <id> invocation syntax.

Compiler

PromptCompiler is pure and synchronous. It flattens modules, sorts by authority, descending priority, module ID, and source order, deduplicates exact segment duplicates, resolves conflict keys by precedence, enforces byte and conservative token budgets (including separators), and renders accepted segments joined with \n\n.

• Exact duplicates across modules are emitted once with a diagnostic.
• Reusing a segment ID with different content is a compile error.
• A lower-precedence claim never overrides a protected winner; replaceable winners produce shadow diagnostics.
• Equal-precedence conflicts with different content fail closed.
• Required overflow fails; Omit emits a diagnostic; Truncate uses a UTF-8 boundary and a fixed [truncated] marker only when it fits.
• Fully optional prompts may compile to an empty string with diagnostics.

Inspection

CompiledPrompt exposes exact text, byte count, conservative estimated tokens, ordered diagnostics, and one PromptInspectionEntry per unique input segment with provenance, trust, cache scope, disposition, exact byte range, rendered bytes, and estimated tokens. Byte ranges slice exact output content; every output byte maps to an included segment or a separator.

Trust labels describe origin for inspection and downstream policy; they do not sanitize content or claim prompt-injection prevention.