Repository navigation
Conversation
…agents-runtime/src/tool_registr Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
…tests.rs Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The README now explains that sub-agent turns are leaf turns by default and that hosts must opt in via ToolAdmission::nested_spawn_depth, noting the runtime's allowlist and depth-ceiling checks and the host's remaining responsibility. The re-export list and a test assertion were reformatted without behaviour changes. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Update the module and field documentation to describe that spawn and delegate tools are withheld from sub-agent runs by default rather than never handed out, and that a host may grant bounded nested spawning for a turn. The new wording also notes that granted tools must still pass the ordinary allowlist and that the host enforces the depth ceiling when a child executes. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Split the terminal-child case into its own test that asserts a finished child routes through the fallback with thread and agent metadata while leaving the stale handle undrained. The previous assertions mixed this scenario with the unknown, unowned, and terminal rejection checks, which made the fallback behaviour hard to see. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Rename two steering tests so their names describe what they actually assert, and adjust the terminal transition test to expect a fallback selection instead of an unknown error. The unused sender binding is prefixed with an underscore to silence the warning. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Terminal tasks with retained metadata are no longer rejected as unknown; they now fall through to the host fallback like tasks without a live steering handle. This makes steering idempotent for finished tasks and updates the docs to match. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Introduce a RecordedSteer type alias for the captured fallback payload and import Mutex directly, replacing the inline tuple type. Also reflow the status send and closure body to satisfy rustfmt. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
SubAgentBudget, SubAgentOutput and SubAgentPolicy now live in tinyagents-orchestration, and tinyagents-graph re-exports them so existing call sites keep working. The policy logic moved with the types, letting the orchestration layer own subagent policy without depending on the graph crate. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
…tions The invocation types module now only declares the shared state and registries used by driver-backed asynchronous subagent jobs, with the harness-based sub-agent, session, and tool types removed. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Job creation now takes the parent run and thread identifiers directly instead of a JobLink, and the unused inline creation helpers were removed so only driver-backed jobs remain. The settle guard was renamed to JobSettleGuard to reflect that it is no longer specific to inline jobs. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The driver subagent tool no longer accepts a mode parameter, so callers can no longer force inline execution and the delivery target is always consulted. The inline job guard was renamed to JobSettleGuard to match its broader role. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The multi-target delegation adapter and its supporting policy, admission, and inline test modules were removed from the subagent invocation layer. The delegation flow is now handled elsewhere, so these host-owned tool implementations and their tests are no longer needed. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The subagent module no longer re-exports the step and delegation helpers, and the module docs now describe only the lifecycle traits and driver. This narrows the public surface to the pieces hosts actually consume. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The orchestration crate no longer exposes the parallel, status, teams, topology, and workflow modules, and the fleet tool set helpers in the subagent role module are gone. Admission is now documented as enforced solely by the async driver, and status mapping imports come from tinyagents-tasks. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The job registration now takes an optional parent thread id instead of a required string, so callers without a parent thread can register jobs without fabricating one. Unused imports were dropped alongside the change. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Drops the mark_result, mark_result_applied, and mark_budget_overrun methods from the job registry along with the now-unused IncompleteKind import, since these paths are no longer called. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The driver tool now reads the registered job directly instead of waiting on ownership or spawning a detached task, and awaiting_input is treated as a successful status. This removes the duplicated completion plumbing so the adapter only reports the current job state. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The test now exercises a background request that has no delivery target and asserts the job is spawned and tracked in the registry rather than run inline. It waits on the job separately and checks the completed status and output. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Delete the parallel, teams, workflow, subagent, status, and topology modules along with their tests, examples, and integration tests. These components are being dropped from the orchestration crate, leaving the remaining surface to be rebuilt without them. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The async spawn tool is now exposed as `spawn_subagent`, so the spawn tool allowlist and the orchestrator fixture were updated to match. The subagent module also re-exports `SubAgentOutput` from its policy module. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Condense the orchestration README to describe the host-neutral policy and lifecycle surface, dropping the long inventory of tool wrappers that now live in tinyagents-harness-tools. The subagent README is trimmed to match the same framing. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The subagent session and tool dispatch cases were removed from the mock capabilities suite, leaving the fuzz matrix to exercise only regular tools. This trims the test surface to the capabilities still under test. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The fuzz-style matrix test was reduced to a single loop over the lookup flag, dropping the always-false subagent dimension and its nested loop. The scenario label and doc comment were updated to match the narrower coverage. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
…ties test Reindent the mock matrix test body and correct the expected tool count to only account for the lookup tool, since the subagent flag is not part of this scenario. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Documentation and doc comments now point at `DriverSubAgentTool` in `tinyagents-harness-tools` instead of the removed `SubAgent`/`SubAgentTool` types in `tinyagents-orchestration`, and the orchestration subagent README drops the stale agent-step section in favour of a short note on the temporary graph re-exports. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The status tool result now only returns a successful JSON result for completed, running, and queued statuses, so awaiting_input falls through to the error branch. This makes the tool report awaiting_input as a failure rather than a normal state. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Rename and rewrite the concurrent driver alias tests so they assert the async dispatch path returns a running job snapshot immediately and coalesces followers onto the same task id, rather than blocking until the shared result completes. The replay test now uses a single async tool with a completion handler and verifies the snapshot is returned before the job settles. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Introduce a RegistryTestHelpers trait and JobLink fixture builder so registry tests register jobs through the same driver path used by DriverSubAgentTool, with helpers for marking results and budget overruns. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Adjust the test helper to match the result by reference and simplify the budget overrun call to the new single-argument signature, keeping the race-with-cancel test aligned with the current registry API. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The spawn tool name list now contains only spawn_subagent, dropping spawn_parallel_agents and spawn_worker_thread so that scope detection no longer treats those retired names as spawn tools. A test pins the roster and the detection behaviour for the removed names. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The module documentation now states that every accepted call starts a background job and returns its status immediately, with waiting and completion delivery handled as separate host operations. This makes the non-blocking contract of the tool adapter explicit. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Add a test that runs a detached subagent whose policy times out while the executor is still blocked, asserting the job is recorded as incomplete with a timeout kind. Also drop the delivery target from an unrelated alias pause test so it no longer relies on a parent thread. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The driver tool documentation now states that every call starts a detached job and returns its task reference immediately, with the delivery target only controlling host completion callbacks. The README was updated to drop the reference to dropped inline calls, since inline execution no longer occurs. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The orchestration crate no longer references tinyagents-runtime in its dev-dependencies, so the entry was removed to keep the manifest accurate. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The driver subagent tool no longer accepts a mode field, so the test fixtures and schemas that still passed mode were updated to match the current parameter surface. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Replace references to the removed `SubAgentTool` and `SubAgentSession` with the async `DriverSubAgentTool` and host-planned `SubagentDriver`, which return a job reference instead of running child work inline. The runtime comparison pages and steering README are updated to match the new delegation model. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Update Cargo.lock to reflect the current dependency graph, adding tinyagents-orchestration and dropping tinyagents-graph and tinyagents-runtime from the affected packages. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Replaced fully qualified crate paths with the already imported SubAgentOutput, SubAgentPolicy, and SubAgentIncompleteKind names, and dropped an unused serde_json::Value import. No behaviour change. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
…nd imports The orchestration crate now depends on tinyagents-runtime for its tests, and the subagent test modules import IncompleteKind directly instead of aliasing SubAgentIncompleteKind. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Remove the unused RunConfig import from the subagent node types and have the registry test helper record a completed SubagentOutcome for successful driver results so the fixture matches real invocation behaviour. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The in-crate test_support module and its ledger_in fixture were no longer referenced by any test, so the module declaration and the now-unused IncompleteKind import in policy_tests were dropped. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The boundary tests module is no longer gated behind cfg(test), so it is compiled in all builds. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Move the cfg(test) attribute above the boundary_tests module declaration so the test-only module is only compiled during test builds. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Reword the SubAgentNode deprecation note to point at SubagentDriver and note the node's removal in a follow-up migration, and correct the module docs to describe the actual dependency direction between orchestration and graph. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The fixture metadata literals already specify every field, so the functional update to the default value was unnecessary. Removing it keeps the test setup consistent with the struct's current shape. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The job registry's insert_job took a long positional argument list, so the parameters are now grouped into a JobInsert struct that is destructured on entry. This keeps call sites readable and makes the metadata fields explicit at the point of construction. An unused test module declaration was also dropped. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Reformat the subagent driver tool and its tests, and simplify the invocation and policy modules without changing behaviour. The restructuring keeps the job registry and status mapping logic easier to follow ahead of further subagent work. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Removed the Cargo test target declaration for live_subagents, which is no longer needed now that the test is gated elsewhere. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The README, hosting guide, and LangChain parity notes now describe child execution as host-planned through `SubagentDriver`, where every call starts an asynchronous job and returns a stable job id. This removes references to the retired inline/background mode split and the `invoke_hosted_in_parent` path so the docs match the current orchestration surface. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The crate-level docs referenced the `multimodal` module with an intra-doc link that fails to resolve when the feature is disabled, so the link is replaced with plain text pointing at the module. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the ⚙️ Run configuration
You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
The driver tool no longer references SubagentOutcomeKind, so the import was dropped to keep the module free of unused-import warnings. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The driver subagent tool test now records the planner request and verifies the tool returns a running job that later completes with the driver result, rather than expecting the output inline. This covers the asynchronous spawn-and-settle behaviour and checks the task key and input passed to the planner. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
The driver subagent tool test now expects the job agent to be "subagent" rather than "worker", matching the name the driver actually assigns. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
Wrap the recorded request assertions in an explicit block so the mutex guard is released at the end of the scope instead of relying on an explicit drop, keeping the test's locking behaviour unchanged while making the guard's lifetime clearer. Auto-committed-on: cuttlefish Co-authored-by: Medulla <medulla@tinyhumans.ai>
senamakel
marked this pull request as ready for review
October 22, 2026 05:29
Tiny Sweeper review
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
senamakel
merged commit Oct 22, 2026
17c085d
into
tinyhumansai:epic/8088-tinyagents-slim
13 checks passed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Makes
spawn_subagentthe single subagent tool and removes the legacy synchronousSubAgentAPI, inline mode, delegation, teams, parallel fan-out, and orchestration workflows. Driver-backed calls now return a job immediately even without a delivery target; hosts wait through the owned job registry separately. Completion delivery remains conditional on its target.Moves the policy types from graph into orchestration and reverses that dependency while the graph crate remains for its separate deletion leaf. Preserves owned-child waiting, pause/resume, cancellation, attempt tracking, and panic settlement through
JobSettleGuard. Rewrites surviving job tests onto the driver; graph-node tests move intact into the graph crate.Refs senamakel#58. Part of tinyhumansai/openhuman#8088.
This PR is stacked on #519 and remains a draft until that dependency lands on the epic base branch. The final 3.0 release and downstream adoption are separate epic leaves.
Validation:
The retained and rewritten behavior cases are mapped in
tinyagents-58-test-migration.jsonin the epic evidence ledger. Failed compile/lint attempts remain preserved there.