refactor(core_host): port the embedded core onto openhuman_core::Harness

[senamakel]

Draft — blocked on tinyhumansai/openhuman#5668.
This branch's vendor/openhuman gitlink points at a commit that only exists on that PR's branch, so submodule init cannot resolve it from the canonical repo and CI will fail to fetch. Before this is ready: #5668 merges → re-point vendor/openhuman at the resulting main commit → mark ready for review.

Summary

• Ports the embedded-core host onto openhuman_core::Harness and the new typed embed::Core::agent() / auth() facades from openhuman#5668.
• Deletes core_host/auth.rs and its tests (222 lines). Every caller uses core.auth(); medulla's private AuthState projection becomes a re-export of the core's.
• Removes all four std::env::set_var calls that configured the core. CoreSettings::resolve(env, config, home) returns a value; boot(settings) takes it.
• Replaces the hand-built turn JSON with harness.turn(..).session(..).cwd(..).model(..).route(..).origin(..).on_progress(..).send(), deleting reply_text and scoped_origin along with both private copies of the RpcOutcome envelope heuristic.
• Bumps vendor/openhuman by 1047 commits (v0.63.7-era → v0.63.17 + #5668), with the manifest and API updates that requires.

Net on the ported files: −854 / +553 lines.

Problem

Running one agent turn through the embedded core took ~2,100 lines of glue across 17 ordered steps, and the sharp edges were all in this repo's own comments:

• Four std::env::set_var calls, sequenced before a constructor they do not appear in, replicated at six call sites. Process-global, inherited by every child, and increasingly hostile in newer Rust editions.
• execution.rs:240 documented that AgentChatParams carries no rename_all, so the hand-built params depended on an unmarked naming coincidence — an upstream rename would fail silently at runtime.
• Two divergent copies of the {result, logs} envelope unwrapper (auth.rs strict, execution.rs loose).
• A private AuthState mirror of the core's response, decoded by hand out of that envelope.
• core_host/tests.rs needed a process-wide mutex and had to clear four environment variables around every case, because the functions under test wrote to the environment.

Solution

Settings are values, not environment variables. CoreSettings { workspace, action_dir, backend_url } is resolved once and handed to boot. The precedence is unchanged and still ours — an operator who exported OPENHUMAN_WORKSPACE still wins — but it is a parameter now instead of a convention about call ordering.

One nuance worth reviewing: the lazy boot path genuinely has no caller to hand settings to (a workflow agent node runs deep inside the engine, and the dispatch signature belongs to every harness). That is replaced by a typed OnceLock<CoreSettings> in core_host. Still process-global — same problem — but the blast radius changes: a cell this crate owns affects this process's core, where OPENHUMAN_WORKSPACE affected every library in the process and every child it spawned.

The TUI still writes those variables into its child-process env snapshot, and should: OpenHuman's picker entry and the harness CLIs are spawned processes. What changed is that we no longer mutate this process's environment to talk to a library living inside it. backend_url also collapses two bindings into one — with no OPENHUMAN_MEDULLA_BASE_URL override, the core's Medulla client falls through to the same api_url.

Two things deliberately not folded in:

• with_turn_cwd is still scoped by hand. It is read by Medulla's process-global lifecycle hooks, not by the core, so there is nothing upstream that could take it as an argument.
• boot_for_auth keeps using CoreBuilder rather than Harness. It is a short-lived core for one question, and claiming the process's single harness slot for it would refuse a later real boot in the same process.

The vendor bump is unrelated churn but unavoidable, since the new API lives upstream. It cost:

Change	Why
tinyflows 0.6 → 0.8	vendored engine version
new [patch."https://github.com/tinyhumansai/tinymemory"]	mirrors openhuman's own; without it tinymemory-api resolves as both a git copy (via tinycortex) and a path copy, producing a wall of E0308s about two MemoryTaint types
dropped stale tinyjuice patch	that vendor directory no longer exists
6 new NodeKind variants covered in 4 matches	Spawn/Scatter/Gather/Gate/Approval/Void
2 new Capabilities fields	tasks set to the engine's TokioTaskRunner so spawn/gate genuinely overlap; approvals: None until we have a review surface to push at
5 new SDK socket error variants	listed by name, not swept into a catch-all, so a future variant is a compile error

Commit history is the auto-commit hook's granular checkpoints, left as-is rather than squashed.

Impact

• Behaviour: no intended change. Same precedence, same presets (DomainSet::embedded() / ServiceSet::embedded()), same one-core-per-process rule — now enforced from both sides, since Harness::build returns AlreadyRunning rather than silently sharing process-global state.
• Testing: core_host/tests.rs went 295 → 189 lines and lost its environment mutex entirely — the cases are independent values now and run in parallel. That the lock could be deleted is the point of the change, not a side effect.
• Risk: the vendor bump is the larger half of this diff and touches workflow-graph rendering and SDK error mapping. Both are covered by the existing suite.

Validation

• cargo test --workspace — 70 test binaries, 0 failures
• cargo clippy --workspace --all-targets — clean
• cargo fmt --all --check — clean
• Baseline confirmed green at the pre-bump pin before any change, so every error resolved here was attributable.

Related

• Depends on: feat(embed): run the OpenHuman harness as a one-call library API openhuman#5668 (must merge first)
• Follow-up: re-point vendor/openhuman at main once #5668 lands, then mark ready

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: dde8a94d-1091-4140-94a7-969fd30ead53

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}