[tinyagents-slim] harness-tools: new tinyagents-harness-tools crate (todos, goals, plan_exit, use_skill, clarification, time, subagent tools) #57

Parent: tinyhumansai/openhuman#8088

Context

A new crate holds every default tool tinyagents defines, together with the harness-level functionality behind them. Tool infrastructure stays in tinyagents-harness: rules, deferral, discover, select, schema prep and the timeout wrapper. This is the tool implementation side.

WaitTool and WaitLoopTool are deleted, not moved; OpenHuman drops them. Delete the unused exports on the way: goal_tools, register_goal_tools, todo_tools, register_todo_tools, GoalProgress and TurnOutcome.

Goals gaps to close here, since OpenHuman's host goals adapter is deleted by senamakel/openhuman#209:

• A GoalObserver (on_updated(&ThreadGoal), on_cleared(thread)) that fires on resume, clear, account and complete, not only on tool calls.
• A turn-prep helper: load, auto-resume if paused, active_goal_context_block, and GoalBudgetStopHook::for_goal. This is what the host inlines today in session_host/runtime_session/prelude.rs:213-251.
• A correct budget-meter throttle. The host version compares one turn's delta against 5% of the budget instead of tracking usage since the last publish, so the meter goes stale. Make account_turn atomic: the host reads, then accounts, and races. Take elapsed time as a Duration, not whole seconds; the host truncates sub-second turns to 0.

tinyagents-hooks' goals feature repoints at this crate.

Scope

• Create crates/tinyagents-harness-tools with features todos, goals, subagents, skills (default: all but subagents).
• Move the rows above with their tests, and port the e2e_graph_goals/e2e_graph_todos integration tests and the goals_and_todos example.
• Add GoalObserver, the turn-prep helper and the fixed throttle/atomic account_turn, each test-first.
• Repoint tinyagents-hooks (stop/goal_budget.rs) at the new crate.
• Leave no re-export shims in graph or harness; graph is deleted by [tinyagents-slim] graph: delete the tinyagents-graph crate #61.

Acceptance criteria

• Test: many small turns publish a meter update once cumulative usage since the last publish crosses the threshold.
• Test: concurrent account_turn calls produce a consistent status transition and exactly one observer event per transition.
• Test: GoalObserver fires on resume, clear, account and complete.
• tinyagents-harness-tools depends only on tinyagents-harness, tinytools and (feature subagents) tinyagents-orchestration. dependency_boundary.rs is updated to assert it.

Done when: the rows above are gone from their source directories and the gates below pass.

Agent brief

Depends on: none · Blocks: #61 · #65 · senamakel/openhuman#209

Research first

• Can the subagent tool shells move without dragging SubAgentJobRegistry along? If not, harness-tools re-exports the registry type behind subagents and the registry stays in orchestration.

Files to consider

• Read: crates/tinyagents-graph/src/{todos,goals,thread_locks}/, crates/tinyagents-harness/src/{tools,tool/plan_exit.rs,tool/packs}/
• Read: OpenHuman crates/openhuman-core/src/agent/goals/runtime.rs (the throttle bug at ~170-180, the non-atomic account at 131-148)
• Change: Cargo.toml (workspace members), crates/tinyagents-hooks/Cargo.toml, crates/tinyagents-integration-tests/

Workflow: implementing-an-epic step; follow openhuman's migrate-to-submodule skill for the release and the OpenHuman gitlink bump

Subagents: plan-writer on Opus for the crate layout, then tdd-implementer on Sonnet, then code-reviewer before the PR.

Gates: cargo check --workspace --all-targets · cargo clippy --workspace --all-targets -- -D warnings · cargo test --workspace (in vendor/tinyagents)

Deep dive (orchestration audit)

Per-file detail of what moves from orchestration: tinyhumansai/openhuman#8088 (comment).

• Correction to the dependency rule: the moved subagent shells import tinyagents_tasks::{TaskStore, OrchestrationTaskFilter, DetachedTaskRegistryError} and tinyinference_llm::Message (jobs.rs:16,26). harness-tools therefore depends on tinyagents-tasks and tinyinference-llm too, behind subagents.
• Make these registry methods pub: wait_owned, cancel_owned, get_owned, try_get_owned, try_list_owned, send_message_owned, send_message_with_request_id, register_driver_job, mark_running, mark_driver_result, driver_job_for_key. Also make DriverJobRegistration and SubagentRequest::resume() public.
• sessions_tests.rs:217,360,414 and jobs_tests.rs move with the shells. A dev-dependency from orchestration back onto harness-tools would be circular.
• Plan review and plan mode are a separate leaf, [tinyagents-slim] harness-tools: plan review and plan mode (request_plan_review, review store, run-mode store) #70, built on this crate.