Repository navigation
refactor(orchestration,cron): pass the DispatchCtx to running-subagent, busy and completion readers (#7801) - #7862
Merged
Merged
Conversation
Refresh the vendored tinyagents and tinybox dependencies to their current revisions. Auto-committed-on: dragonfly
The tinybus vendored dependency is no longer used now that the event bus has been replaced by the in-tree implementation, so the submodule entry was dropped to keep the vendor tree clean. Auto-committed-on: dragonfly
Cancellation, registry, resolve, roster, steering, and wait paths for running subagents were reworked so that cancelling a subagent reliably stops it and waiting observes the correct terminal state. The vendored tinycomputer dependency was updated to match. Auto-committed-on: dragonfly
Vendors the tinyhumans SDK so the project can depend on it without pulling it from an external source at build time. Auto-committed-on: dragonfly
Deleting a thread now also tears down any subagent sessions it spawned, so their runtime state and control handles do not outlive the parent thread. Previously these sessions were left behind and could keep running after the thread was gone. Auto-committed-on: dragonfly
The channel_ops module in the web chat ops directory is no longer referenced anywhere, so it has been deleted to keep the crate tidy. Auto-committed-on: dragonfly
Adds a tool that spawns a subagent asynchronously and returns immediately so the parent agent can continue working while the child runs. This lets orchestration fan out work without blocking on each subagent's completion. Auto-committed-on: dragonfly
Adds orchestration tools that let an agent close a running subagent, send it steering input, and wait for it to finish. These give the parent agent explicit control over subagent lifecycle instead of relying on implicit completion. Auto-committed-on: dragonfly
Background completions now record which owner is responsible for each completion, and subagent control consults that ownership when routing results. This prevents completions from being delivered to the wrong target when multiple owners are active. Auto-committed-on: dragonfly
Adds a tool that spawns subagents asynchronously and executes them, letting orchestration hand off work without blocking the caller. The vendored tinymcp dependency is updated to support the new execution path. Auto-committed-on: dragonfly
Refresh the vendored tinymemory and tinysearch dependencies to their current upstream revisions. Auto-committed-on: dragonfly
Vendors the tinyagents library so the project can rely on it without fetching it at build time. Auto-committed-on: dragonfly
Update the vendored tinybox and tinybus submodules to their latest commits. Auto-committed-on: dragonfly
Update the vendored tinycomputer submodule to the latest commit. Auto-committed-on: dragonfly
Background delivery now drains queued messages during an active run instead of waiting for the session to go idle, and the busy guard coordinates with the drain so deliveries are not dropped or duplicated. This keeps relay and web chat replies flowing while the agent is still working. Auto-committed-on: dragonfly
Refresh the vendored tinyhumans-sdk, tinymcp, tinymemory, and tinysearch dependencies and adjust the tenant runtime to match their updated APIs. Auto-committed-on: dragonfly
The background completion handler now drains and forwards every queued result instead of stopping after the first one, so completions that arrive in a burst are no longer silently lost. Auto-committed-on: dragonfly
Thread purge and CRUD operations now check the busy guard before proceeding, and web chat channel and start-chat flows respect the same guard. This prevents deleting or mutating threads while an agent is still running against them. Auto-committed-on: dragonfly
Background subagent completions are now routed back to the parent orchestrator so their results are surfaced instead of being dropped. The registry tracks running subagents to support this delivery. Auto-committed-on: dragonfly
Reworked the host profile lookup to reduce branching and make the resolution path easier to follow. Behaviour is unchanged. Auto-committed-on: dragonfly
Split the runtime bootstrap logic into smaller helper functions to make the startup sequence easier to follow and test. Behaviour is unchanged. Auto-committed-on: dragonfly
Moved the relay channel operations out of the provider file into their own ops module so the provider stays focused on wiring and the request handling logic is easier to find and test. No behaviour change. Auto-committed-on: dragonfly
Scheduled runs now route their completion output back to the session that created them instead of a generic delivery target, so replies land in the conversation the user actually started. Delivery ownership is tracked per run to keep concurrent schedules from clobbering each other's output. Auto-committed-on: dragonfly
Transcript persistence now lives in the session store instead of the runtime session, so the store owns all session data and the runtime session no longer reaches into storage directly. Auto-committed-on: dragonfly
…on is empty, and the stat is Could you paste the diff (or at least the stat and the changed hunks)? Once I have it I'll produce the Conventional Commits message in the exact format you specified. Auto-committed-on: dragonfly
…on is empty, and the stat is Could you paste the diff (or at least the stat and the relevant hunks) for `crates/openhuman-core/src/agent/session_import/live.rs`? Once I can see the `+`/`-` lines I'll produce the Conventional Commits message. Auto-committed-on: dragonfly
Introduce a bus that lets security device components publish and subscribe to events, giving the devices module a shared channel for coordinating state changes. Auto-committed-on: dragonfly
…ery tests Updated the background completion and delivery test suites to pass an explicit dispatch context into the functions under test, matching the new signatures. This keeps the tests compiling against the context-aware API without changing the behaviour they assert. Auto-committed-on: dragonfly
Add tests exercising tenant-scoped runtime behaviour to lock in the isolation guarantees between tenants. Auto-committed-on: dragonfly
The turn gate was left held when a relay task errored out, blocking subsequent turns for the profile. The gate is now released on the failure path so later turns can proceed. Auto-committed-on: dragonfly
The policy failure tests now reference the dispatch context through the openhuman_core crate path and the affected calls have been reformatted to satisfy rustfmt. No test behaviour changed. Auto-committed-on: dragonfly
…profile Background completion, delivery, and running-subagent bookkeeping now resolve threads, sessions, and delivery slots through the caller's profile so one user's Stop, delete, or busy turn can no longer gate or reach another user's identically named thread. Tests were updated to exercise the per-profile scoping. Auto-committed-on: dragonfly
The background completion, delivery, and running subagent tests now obtain their dispatch context from a local cx() helper instead of calling DispatchCtx::ambient() at each site. This keeps the tests dispatching under the calling task's own context while removing the repeated boilerplate. Auto-committed-on: dragonfly
Moved the thread and session lookup helpers out of background_completions into a dedicated thread_index module so the parent file stays focused on completion recording and delivery. Behaviour is unchanged. Auto-committed-on: dragonfly
Moved the per-thread delivery slot and its in-flight tracking into a dedicated slot module so the background delivery file stays focused on orchestration. Behaviour is unchanged. Auto-committed-on: dragonfly
Removed an unused DispatchCtx import and reflowed the tool_policy field type and a filter closure to satisfy rustfmt. No behaviour changes. Auto-committed-on: dragonfly
Split the runtime session logic into smaller helper functions to make the session host easier to follow and test. Behaviour is unchanged. Auto-committed-on: dragonfly
Split the runtime session logic into smaller helper functions to make the session host easier to follow and test. Behaviour is unchanged. Auto-committed-on: dragonfly
Adds a tool that spawns a subagent asynchronously and returns immediately so the parent agent can continue working while the child runs. This lets orchestration fan out work without blocking on each subagent's completion. Auto-committed-on: dragonfly
Adds a release-mode test suite for background completions alongside the existing tests, covering the same behaviour under optimised builds so regressions that only surface with release settings are caught. Auto-committed-on: dragonfly
The cancellation-related sub-agent tests were extracted from the shared running_subagents test file into a dedicated cancel_tests module, wired in via a path attribute. This keeps the growing test suite organised without changing any test behaviour. Auto-committed-on: dragonfly
…ry slot key Move the late-cancel doc comment onto the test that actually exercises it and add a pending_ids helper for release tests. The delivery slot key becomes pub(super) so sibling modules can read it. Auto-committed-on: dragonfly
Reformatted the workspace re-export list and collapsed multi-line assert! calls in the workspace tests onto single lines. No behaviour change. Auto-committed-on: dragonfly
…tions and subagents Adds coverage asserting that background completions and running subagents stay scoped to their owning tenant, so cross-tenant leakage in orchestration state is caught by tests. Auto-committed-on: dragonfly
…spatch Move the check that alice cannot cancel bob's worker ahead of the concurrent dispatch section so the isolation guarantee is verified on the initial state, and replace the trailing assertion with one confirming alice's own worker is gone after cancellation. Auto-committed-on: dragonfly
The tenant subagent tests now build their dispatch context with DispatchCtx::ambient() instead of process_default(), matching the context the controller actually runs under. Auto-committed-on: dragonfly
…tions and delivery Adds coverage for background completion and delivery paths to verify that tenant scoping is enforced. These tests guard against cross-tenant leakage when background work is dispatched. Auto-committed-on: dragonfly
Replace fully qualified `crate::core::runtime::DispatchCtx` paths with the already imported `DispatchCtx` alias and collapse the resulting multi-line calls, keeping the tenant isolation assertions unchanged. Auto-committed-on: dragonfly
Add tests asserting that busy marks and cron completion ownership are keyed by the explicit dispatch context rather than any ambient task scope, so one profile's in-flight turn or Stop cannot affect another's thread and a job's owner is resolved from the explicit session agent alone. Auto-committed-on: dragonfly
Device RPC handlers now validate that the request id is present and non-empty before dispatching, returning an error instead of proceeding with an unusable identifier. Auto-committed-on: dragonfly
Reordered use statements so local module imports follow the public re-exports, and reformatted an over-long function signature in the delivery slot to satisfy rustfmt. No behaviour changes. Auto-committed-on: dragonfly
The in-flight delivery registry now lives in the parent background_delivery module instead of the slot submodule, so it can be shared by other delivery code paths. Behaviour is unchanged. Auto-committed-on: dragonfly
…wners Background delivery now records which slot and completion owner each deferred result belongs to, so results are routed back to the right caller instead of being matched by position. The busy guard and profile surface were adjusted to consult the new ownership data when deciding whether a session can accept more work. Auto-committed-on: dragonfly
The `delivering` helper from the slot module was imported under `cfg(test)` but is no longer referenced, so the import was dropped to keep the module free of dead code. Auto-committed-on: dragonfly
# Conflicts: # crates/openhuman-core/src/agent/session_store/mod.rs # crates/openhuman-core/src/agent/session_store/mod_tests.rs # crates/openhuman-core/src/agent/subagent_host/ops/graph/dispatch.rs # crates/openhuman-core/src/agent/subagent_host/ops/graph/transcript.rs # crates/openhuman-core/src/agent/subagent_host/ops/graph_tests.rs # crates/openhuman-core/src/cron/scheduler/origin_delivery.rs # crates/openhuman-core/src/cron/scheduler/origin_delivery_tests.rs
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. |
Tiny Sweeper review
|
Contributor
|
Note Currently processing new changes in this PR. This may take a few minutes, please wait... ⚙️ Run configuration
📒 Files selected for processing: 77
Comment |
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
1 participant
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.
Summary
Part of #7801 (child of #7523): moves the running-subagent state and the next batch of
current_tenant()/profile_key()readers off the ambientCoreContextand onto an explicit&DispatchCtx. No behaviour change intended: an explicit context resolves to exactly what the ambient read returned.running_subagents/*):admit_spawn,register,owner_key,control_admits,admit_workspace,resolve,wait,steer,steer_control,roster,cancel_*,caller_workspaceall take actx. Tools usedispatch_ctx_for_tool/ the run context'sdispatch_ctx(); the RPC controls and thread delete/purge/stop use the handler'sctx.profile_key()is gone.tenant::profile_key_in(&Tenant, id)replaces it. The background-completion maps (background_completions),busy_guard::TurnBusy/is_busy/clear_busy_for_thread, and theDeliverySlotkey from the caller's context.CompletionTargetcaptures the spawning context so a detached child records its completion under its own tenant (it runs on a spawned task).completion_owners::note,recover_on_boot/open,run_draintake the ctx.current_tenant()readers: relay inbound (channel_relay_inbound) andturn_gate::acquire,croncompletion owner / origin delivery /deliver_*/cron_run/run_job_now,session_store::current/current_agent_key_or, the devices bus anddevices_create_pairing.CronJobRun Now: the background taskcron_runspawns now carries the caller's context (it ran on a baretokio::spawnwith no scope before, so in SaaS the delivery saw no tenant). Embed'sCronholds the runtime's context forrun_now.background_completions/thread_index.rs,background_delivery/slot.rsand two test modules so the touched files stay under the layout limit. No process static moved (delivering()stays inbackground_delivery.rs).Merged
upstream/main(including #7859, which reworked the origin transcript store): kept itsagent_key_for/origin_transcriptsand threaded the ctx intodeliver_to_origin.Tests
New regression tests (sibling
*_tests.rs), each with no ambient scope anywhere and across awaits / spawned tasks /join!:running_subagents_tenant_tests::saas_controls_follow_the_explicit_context_not_the_tasks_scope(SaaS child process): a worker is reachable through its own profile's context and refused through another's; each cancel removes only its own worker.background_completions_tenant_tests::the_explicit_context_keys_completions_without_any_ambient_scope: two profiles on the same thread id never see each other's queue, stop or session index.background_delivery_tenant_tests::busy_marks_are_keyed_by_the_explicit_context: busy marks and Stop are per-profile.cron::completion_owner_tests::the_explicit_context_names_the_owner_with_no_task_scope.Existing tenant tests were pinned to the ambient read:
control()now passesDispatchCtx::ambient()(as the RPC server builds it from the scope), and two helpers that evaluatedDispatchCtx::ambient()outsideCoreContext::scopenow passDispatchCtx::new(ctx).cargo test -p openhuman --lib: 9526 passed, 1 failed (memory::lifecycle::prefetch::tests::slow_engine_results_survive_the_manual_pre_turn_cutoff, a timing test, passes alone, untouched code).saas_mode_e2e(19),saas_profile_isolation_e2e(3), embedsaas_profiles(1),agent_lifecycle,flow_approvalspass.cron_host_deliveryis the known pre-existing embed failure.json_rpc_e2e.Gates
cargo checkdefault and--workspace --all-targets; product features (product-features.sh --core) all targets;--no-default-features -p openhuman; app manifest;cargo build -p openhuman-cli --bin openhuman-core;pnpm rust:layout; agent-runtime boundary (baseline untouched); process statics.scripts/ci/check-saas-ambient.mjsreports sites in files this PR does not touch, same as upstream main.Refs #7801, #7523