Hand the host-agnostic half of workflows to tinyflows

Cargo.toml

@@ -45,7 +45,7 @@ tinyhumans-sdk = { path = "vendor/openhuman/vendor/tinyhumans-sdk" }
 # openhuman host consumes it — one shared pin, one shared upgrade cadence.
 # `mock` is a normal (not dev-only) feature: dry-run authoring uses the engine's
 # deterministic capability stand-ins in ordinary builds.
-tinyflows = { version = "0.6", features = ["mock"] }
+tinyflows = { version = "0.6", features = ["mock", "host-caps", "store"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 toml = "0.8"
@@ -120,16 +120,24 @@ unicode-width = "0.2"
 tinyagents = { path = "vendor/openhuman/vendor/tinyagents" }
 tinychannels = { path = "vendor/openhuman/vendor/tinychannels" }
 tinycortex = { path = "vendor/openhuman/vendor/tinycortex" }
-# OpenHuman's extracted TinyMemory core names this contract by version.  The
-# vendored OpenHuman manifest patches it, but path dependencies do not inherit
-# their own patch tables, so Medulla must bind it at this workspace root too.
+# `tinymemory-core` names `tinycortex-api` by version requirement while this
+# graph depends on tinycortex by path, so without this the resolver goes to
+# crates.io for it and fails outright — the crate is not published. The
+# vendored OpenHuman manifest patches it too, but path dependencies do not
+# inherit their own patch tables, so Medulla must bind it at this workspace
+# root as well.
 tinycortex-api = { path = "vendor/openhuman/vendor/tinycortex/api" }
 tinyflows = { path = "vendor/openhuman/vendor/tinyflows" }
 tinyjuice = { path = "vendor/openhuman/vendor/tinyjuice" }
 # Load-bearing even though no crate in this workspace links tinyplace any more:
 # the vendored OpenHuman core still does, and without the redirect Cargo would
 # resolve the published crate instead of the vendored tree.
 tinyplace = { path = "vendor/openhuman/vendor/tinyplace/sdk/rust" }
+# Not a tiny* crate, but patched for the same reason the others are: OpenHuman
+# vendors its own copy and this table is the only one Cargo reads, so an entry
+# missing here silently resolves the published crate instead of the vendored
+# tree.
+motosan-ai-oauth = { path = "vendor/openhuman/vendor/motosan-ai-oauth" }
 # NOT copied from OpenHuman: its whisper-rs-sys git patch. Cargo fetches a git
 # patch source during resolution even when the patched crate is absent from the
 # graph, which would put a network fetch in the critical path of every

scripts/init-submodules.sh

@@ -38,14 +38,23 @@ git submodule update --init --depth 1 vendor/openhuman
 # root manifest must stay in lockstep with its `[patch.crates-io]` table;
 # tinybus and tinymemory are unpublished direct paths from OpenHuman's manifest.
 git -C vendor/openhuman submodule update --init --depth 1 \
+  vendor/motosan-ai-oauth \
   vendor/tinyagents \
   vendor/tinybus \
   vendor/tinychannels \
   vendor/tinycortex \
   vendor/tinyflows \
   vendor/tinyhumans-sdk \
   vendor/tinyjuice \
-  vendor/tinyplace \
-  vendor/tinymemory
+  vendor/tinymemory \
+  vendor/tinyplace
 
-echo "Submodules initialized (OpenHuman core + its nine required dependencies)."
+# Deliberately NOT recursive. Three of these vendor crates of their own —
+# tinyflows has its own `tinyagents`, tinymemory its own `tinyagents`, `tinybus`
+# and `tinycortex` — and none of those nested copies is ever resolved: the root
+# patch table redirects each name to the copy beside OpenHuman, so the graph
+# holds one of each. Initializing them recursively would clone several hundred
+# megabytes that nothing links, and would put two checkouts of one crate on
+# disk for anyone reading the tree.
+
+echo "Submodules initialized (OpenHuman core + its ten vendored crates)."

src/sdk/src/daemon/task_loop/workflow/handle.rs

@@ -18,8 +18,7 @@ use crate::protocol::{TaskFrame, TaskFrameKind, WorkflowAdvert, WorkflowInputAdv
 use crate::workflows::bridge::trigger_input;
 use crate::workflows::evolve::{EvolveConfig, EvolveSession, EvolveTrigger};
 use crate::workflows::{
-    run_workflow_versioned, FileWorkflowStore, RunContext, RunStatus, StoreWorkflowResolver,
-    WorkflowStore,
+    run_workflow_versioned, RunContext, RunStatus, StoreWorkflowResolver, WorkflowStore,
 };
 
 use super::super::super::types::{DaemonRuntime, FrameAttachments, CAPACITY_REJECTION_PREFIX};
@@ -72,7 +71,10 @@ impl DaemonRuntime {
     /// resolves it.
     fn workflow_store(&self) -> Arc<dyn WorkflowStore> {
         let cwd = std::path::Path::new(&self.inner.config.workspace);
-        Arc::new(FileWorkflowStore::discover(&self.inner.config.env, cwd))
+        Arc::new(crate::workflows::store::discover(
+            &self.inner.config.env,
+            cwd,
+        ))
     }
 
     /// Run the workflow a `task` frame named, replying with its outcome.

src/sdk/src/flow_engine/README.md

@@ -4,7 +4,7 @@ The adapter seam between Medulla and the `tinyflows` workflow engine.
 
 ## Contents
 
-- [`caps/`](./caps/) — Assembling the capability bundle the engine runs against.
+- [`caps/`](./caps/) — Assembling the capability bundle the engine runs against. The host-agnostic implementations moved to `tinyflows::caps::host`; what is left is the harness seam.
 - [`execute.rs`](./execute.rs) — Driving the engine: compile, run, resume, simulate.
 - [`harness_choice.rs`](./harness_choice.rs) — Which harness and model an `agent` node runs on.
 - [`harness_choice_tests.rs`](./harness_choice_tests.rs) — Unit tests for harness and model selection.
@@ -13,7 +13,7 @@ The adapter seam between Medulla and the `tinyflows` workflow engine.
 - [`observability_tests.rs`](./observability_tests.rs) — Tests for the run observer.
 - [`settings.rs`](./settings.rs) — What the capability adapters are allowed to do, and where they keep things.
 - [`settings_tests.rs`](./settings_tests.rs) — Tests for capability settings and the config that produces them.
-- [`tests/`](./tests/) — The offline capability-seam suite. `mod.rs` holds the shared fixtures; submodules cover the `medulla:shell` tool, harness/model selection, the state store, the tool invoker, and the HTTP capsule.
+- [`tests/`](./tests/) — The offline capability-seam suite. `mod.rs` holds the shared fixtures; submodules cover the `medulla:shell` tool, harness/model selection, the tool invoker, and dry-run simulation. The state store's and HTTP capsule's own cases moved to `tinyflows::caps::host` with the code.
 
 ## Maintenance
 

src/sdk/src/flow_engine/agent_evidence/tests.rs

@@ -95,7 +95,9 @@ fn prompt_queue_is_bounded_before_the_run_finishes() {
 
     let input = serde_json::to_vec(&steps[0].input).unwrap();
     assert!(input.len() <= 64 * 1024 + 256);
-    assert_eq!(steps[0].input.as_ref().unwrap()["_medullaTruncated"], true);
+    assert!(tinyflows::store::is_truncated(
+        steps[0].input.as_ref().expect("an input")
+    ));
 }
 
 #[test]

src/sdk/src/flow_engine/caps/README.md

@@ -5,14 +5,20 @@ Assembling the capability bundle the engine runs against.
 ## Contents
 
 - [`agent.rs`](./agent.rs) — `agent` nodes, run on a real harness.
-- [`code.rs`](./code.rs) — The `code` node's runner — refusing by default.
 - [`dispatch.rs`](./dispatch.rs) — Handing a workflow node's instruction to a harness.
-- [`http.rs`](./http.rs) — Outbound HTTP for `http_request` nodes, behind a host allowlist.
-- [`mocks.rs`](./mocks.rs) — Capability stand-ins for dry runs.
 - [`mod.rs`](./mod.rs) — Assembling the capability bundle the engine runs against.
-- [`state.rs`](./state.rs) — Durable key/value state for stateful workflows.
 - [`tools.rs`](./tools.rs) — `tool_call` dispatch across two namespaces.
 
+## What is no longer here
+
+The capability implementations with no Medulla in them — the out-of-process
+script runner and its path policy, the `code` and `shell` runners, the file
+state store, the allowlisted HTTP client, and the dry-run stand-ins — moved to
+`tinyflows::caps::host`, behind that crate's `host-caps` feature. Every host
+embedding the engine needs them, and each one that wrote them itself rewrote the
+same subtle parts. [`mod.rs`](./mod.rs) re-exports them, so a call site in this
+crate still names one place.
+
 ## Maintenance
 
 Keep this index synchronized when responsibilities move. Put shared data structures in `types.rs`, focused unit tests in `tests.rs` or a sibling `_tests.rs`, and preserve the module-level Rust documentation as the API source of truth.