01 — TinyAgents Crate Audit (vendor/tinyagents @ 1.5.0 + ac73382)

~41.5k non-test LOC (~60k with tests); 54 integration tests, 15 examples. Deps: tokio (sync/time/macros), reqwest (rustls/json/stream), serde, sha2; optional rusqlite (sqlite), rhai (repl). No provider-specific SDKs.

The crate's own goal.md is a pre-written gap list from OpenHuman's migration perspective and corroborates most findings below.

1. Module map

harness/ — the agent runtime. Largest: middleware/ (~5.3k, trait + ~18 built-ins incl. retry/timeout/fallback/budget/tool-policy/contextual selection/approval/redaction/prompt-cache guard), providers/ (~3k: MockModel + one real OpenAiModel), agent_loop/ (~2.3k), model/ (~1.9k: ChatModel, ModelRequest/Response, ModelStream, StreamAccumulator, ModelRegistry), observability/ (~1.8k journals/ status/Langfuse), subagent/ (~1.2k), events/ (~1.2k), plus store/tool/ summarization/embeddings/steering/cache/usage/cost/no_progress/testkit.
graph/ (~19k) — durable typed graphs: compiled runtime, checkpoints (memory/file/sqlite), orchestration TaskStore + tools, todos, goals, map_reduce, subagent/subgraph nodes, recursion, export, testkit.
registry/, language/ (.rag), repl/ (.ragsh, feature-gated).
Ergonomics gap: graph surface is fully re-exported at crate root, but core harness types (AgentHarness, ChatModel, ModelRequest, OpenAiModel, middleware, message types) require deep harness::… paths (lib.rs).

2. The harness loop (what's good)

AgentHarness::run_loop (harness/agent_loop/mod.rs:293-673): cancellation checkpoint → steering drain → deadline/call caps → build request → before_model → wrap-onion model call with retry + fallback chain (:766-859) + per-call tokio::time::timeout budget (:902-917) + response cache (:685-756) → after_model → usage/events → MiddlewareControl (StopWithFinal/Interrupt) → tools → repeat. Unknown-tool policy Fail | ReturnToolError | Rewrite (:563-618). Sub-agents nest as tools with a depth cap. Rich extension points; the loop shape itself is fixed.

3. Defects & gaps (file:line), grouped

Reasoning / thought tokens — scaffolding exists, last mile unwired

ContentBlock = Text | Json | Image | ProviderExtension (harness/message/types.rs:21-30) — no Thinking/Redacted variant; no persisted home for Anthropic thinking blocks or their signatures.
OpenAI SSE path hardcodes reasoning: String::new() per delta (providers/openai/mod.rs:772); never parses reasoning_content or o-series reasoning.
StreamAccumulator::finish drops accumulated reasoning — final message built from text + tool chunks only (model/mod.rs:683-710).
Loop's middleware-facing ModelDelta built with content + tool_call only, dropping reasoning (agent_loop/mod.rs:980-984) — the exact sdk-gaps §3 item blocking ThinkingForwarder deletion.
Usage.reasoning_tokens (usage/types.rs:29) and CostTotals.reasoning_cost (cost/mod.rs:69) exist but are always 0: convert_usage maps only prompt/completion/total/cached (openai/mod.rs:708-719); completion_tokens_details.reasoning_tokens isn't even in the wire struct (openai/types.rs:235-256).
No signature/redacted_thinking handling anywhere → correct multi-turn extended-thinking + tool-use vs native Anthropic is currently impossible.

Streaming — real SSE, but observation-only and single-level

OpenAI streaming is genuine incremental SSE (openai/mod.rs:1020-1073, line buffering :874-896, chunk folding :753-804); tool-call arg fragments stream as ToolCallDelta (:796-799).
invoke_streaming returns a completed AgentRun, not a Stream (agent_loop/mod.rs:195-241); deltas reachable only via EventSink/ middleware (:929-994).
Sub-agents never stream: SubAgentTool::call → SubAgent::invoke non-streaming path (subagent/mod.rs:535, :179-189). No parent/root attribution on deltas.
No explicit tool-call started/completed stream events.
MockModel::stream replays a completed invoke (model/types.rs:536-549) — mock-backed tests aren't truly incremental.

Performance — correctness-first cloning in the hot loop

Full history cloned per turn: ModelRequest::new(messages.clone()) (agent_loop/mod.rs:356); full request cloned again per retry/fallback attempt (:804, :937).
Tool schemas rebuilt every iteration: .with_tools(self.tools.schemas()) (:356).
Provider translate_request re-clones every message + tool schema and re-serializes tool args per call (openai/mod.rs:404-449, 570-584).
Response-cache key = serde round-trip + canonicalize + SHA-256 over the entire request per cached call (cache/mod.rs:101-106).
StreamAccumulator::push_tool_chunk linear scan per fragment (model/mod.rs:643-650).
Tools execute strictly serially (agent_loop/mod.rs:526-659); parallel_tool_calls is a capability flag only.

Providers — one adapter, many presets

Only OpenAiModel is real; anthropic()/deepseek/groq/xai/ openrouter/together/mistral/ollama are base-URL presets on the Chat Completions wire (openai/mod.rs:309-383).
No Anthropic Messages API: zero cache_control/ephemeral hits in providers; no thinking config; Usage.cache_creation_tokens always 0 (only cache_read_tokens from prompt_tokens_details.cached_tokens, openai/mod.rs:713-716).
Prompt-cache shaping is metadata-only: cache_segments / prompt_fingerprint / cacheable_prefix_ids() (model/types.rs:385-405) + PromptCacheGuardMiddleware (middleware/mod.rs:19-60) never reach any wire field — the spec's "extreme prompt caching" (docs/spec/README.md:258-278) is observability, not request shaping.
Spec explicitly plans feature-flagged native openai/anthropic/ollama (docs/spec/README.md:279-284; slot documented at providers/types.rs:322-337).

Durability / storage

TaskStore is in-memory/JSONL only; no lifecycle history/replay (goal.md §4/§11) — what running_subagents.rs (1931 LOC) needs to shrink.
SQLite ownership: bundled rusqlite 0.40 vs app-owned sqlite → needs a trait-first/connection-adapter path (goal.md §5).
Store has no compare-and-set (single-writer constraint documented in the old plan).

4. Test coverage holes

No tests for: native Anthropic/prompt caching/thinking (no impl), reasoning end-to-end (only hand-fed accumulator test, tests/e2e_reasoning_and_selection.rs:57-98), parallel tool execution, caller-consumable streaming, sub-agent stream propagation. Live-provider tests exist behind env keys; conformance suites exist for graph/checkpoint.

5. Ranked improvement opportunities (effort: S≈1-2d, M≈3-5d, L≈1-2w, XL multi-week)

Reasoning end-to-end — L — doc 02. Unblocks ThinkingForwarder deletion + real reasoning cost accounting.
Hot-loop de-cloning + cache-key + schema caching — M — doc 04. Biggest runtime win for OpenHuman's long transcripts.
Streaming out of the harness + up from sub-agents — L — doc 03. Unblocks tool_progress.rs/progress projection deletions (C4).
Native Anthropic Messages provider + cache_control shaping — L→XL — doc 05. Highest-leverage provider gap.
Parallel tool execution — M — doc 04 §4.
Usage completeness (OpenAI reasoning tokens) + catalog pricing — S→M — doc 02 §5; feeds old-plan C3 budget flip.
Durable TaskStore (SQLite/JSONL lifecycle + replay) — L — doc 06 §4.
SQLite dependency flexibility — M — goal.md §5.
Crate-root re-export ergonomics — S.

7.6 KiB Raw Blame History