Skip to content

feat(delegation): async background subagents via delegate_task(background=true) - #40946

Merged
teknium1 merged 4 commits into
mainfrom
hermes/hermes-b1109a1a
Jun 15, 2026
Merged

feat(delegation): async background subagents via delegate_task(background=true)#40946
teknium1 merged 4 commits into
mainfrom
hermes/hermes-b1109a1a

Conversation

@teknium1

@teknium1 teknium1 commented Jun 7, 2026

Copy link
Copy Markdown
Contributor

Summary

delegate_task(background=true) now dispatches a subagent that runs in the background and returns a handle immediately — the user and the model keep working while it runs, and the full result re-enters the conversation as a new turn when the subagent finishes.

It rides the same completion-queue rail as terminal background processes (notify_on_complete), so we never splice a result mid-loop. The re-injected message is a self-contained block carrying the original task source (goal, the context the parent supplied, toolsets, role, model, dispatch time, status) plus the result — because by the time it lands, the agent may be deep in unrelated context and won't remember why the subagent existed.

Changes

  • tools/async_delegation.py (new): daemon-executor registry; capacity cap (rejects, doesn't queue); pushes a rich type="async_delegation" event onto the shared process_registry.completion_queue; interrupt_all() for shutdown.
  • tools/delegate_tool.py: background param + single-task dispatch branch (captures session_key on the parent thread, hands _run_single_child to the dispatcher as a runner). Batch async rejected in v1.
  • tools/process_registry.py: format_process_notification renders the self-contained task-source block for the new event type.
  • gateway/run.py: dedicated _async_delegation_watcher drains + injects results into the originating session (covers the idle case), session_key → routing enrichment, shutdown interrupt of dangling delegations.
  • hermes_cli/config.py: delegation.max_async_children (default 3).

Why this shape

The existing completion path (completion_queue → CLI _pending_input / gateway synthetic internal=True MessageEvent) surfaces results as a new turn when the agent is idle, never between a tool result and an assistant message. That's load-bearing — it keeps message-role alternation legal and the prompt cache intact (hard invariant: never mutate past context). Async delegation reuses that rail rather than reaching into a running loop, so it inherits the idle-drain wiring, de-dup, and crash-recovery checkpointing already present in cli.py and gateway/run.py.

v1 is single-task only; async batch (N handles, partial completion) is a follow-up.

Validation

Result
13 targeted tests (tests/tools/test_async_delegation.py) pass
137 existing delegate tests pass (sync path intact)
Non-blocking proof delegate_task(background=true) returns in 2ms while the child is gated closed; a sync impl would hang 5s
CLI E2E dispatch → background run → completion event on shared queue → rich re-injection block
Gateway E2E session_key parsed → routable SessionSource → injected; foreign watch events requeued untouched
Rebased onto current origin/main clean

Test plan

  • delegate_task(goal=..., background=true) returns {status: dispatched, delegation_id, mode: background} immediately.
  • When the subagent finishes, an [ASYNC DELEGATION COMPLETE — deleg_xxxx] message re-enters the chat with the original goal/context + result.
  • At capacity (max_async_children), further async dispatches are rejected with a fall-back-to-sync hint.
  • background=true + multi-item tasks → rejected ("single-task only").

Infographic

async-background-delegation

@github-actions

github-actions Bot commented Jun 7, 2026

Copy link
Copy Markdown
Contributor

🔎 Lint report: hermes/hermes-b1109a1a vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10924 on HEAD, 10914 on base (🆕 +10)

🆕 New issues (9):

Rule Count
invalid-argument-type 6
invalid-assignment 1
unresolved-import 1
unresolved-attribute 1
First entries
tools/delegate_tool.py:2273: [invalid-argument-type] invalid-argument-type: Argument to function `_run_single_child` is incorrect: Expected `str`, found `Unknown | str | None | list[str]`
tools/delegate_tool.py:2286: [invalid-argument-type] invalid-argument-type: Argument to function `dispatch_async_delegation` is incorrect: Expected `str | None`, found `Any | None | str | list[str]`
tools/delegate_tool.py:2285: [invalid-argument-type] invalid-argument-type: Argument to function `dispatch_async_delegation` is incorrect: Expected `str`, found `Any | str | None | list[str]`
hermes_cli/cli_commands_mixin.py:246: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to `def interrupt_all(reason: str = "shutdown") -> int`
tools/async_delegation.py:65: [invalid-argument-type] invalid-argument-type: Argument to bound method `SimpleQueue.put` is incorrect: Expected `_WorkItem[Any]`, found `None`
tests/tools/test_async_delegation.py:12: [unresolved-import] unresolved-import: Cannot resolve imported module `pytest`
tools/delegate_tool.py:2287: [invalid-argument-type] invalid-argument-type: Argument to function `dispatch_async_delegation` is incorrect: Expected `list[str] | None`, found `(Any & ~AlwaysFalsy) | (str & ~AlwaysFalsy) | list[str] | None`
tests/tools/test_delegate.py:2272: [invalid-argument-type] invalid-argument-type: Argument to function `delegate_task` is incorrect: Expected `bool | None`, found `str | Unknown`
tools/async_delegation.py:82: [unresolved-attribute] unresolved-attribute: Object of type `AbstractSet[Thread]` has no attribute `add`

✅ Fixed issues: none

Unchanged: 5747 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

@teknium1
teknium1 force-pushed the hermes/hermes-b1109a1a branch from 4202a8e to d8754a7 Compare June 7, 2026 02:59
@alt-glitch alt-glitch added type/feature New feature or request P3 Low — cosmetic, nice to have comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint comp/gateway Gateway runner, session dispatch, delivery tool/delegate Subagent delegation labels Jun 7, 2026
@teknium1
teknium1 force-pushed the hermes/hermes-b1109a1a branch from 67397d8 to 1c00cb6 Compare June 10, 2026 16:22
teknium1 added 4 commits June 15, 2026 07:32
…ound=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.
CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.
Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.
@teknium1
teknium1 force-pushed the hermes/hermes-b1109a1a branch from 1c00cb6 to b47e65d Compare June 15, 2026 14:33
Codename-11 added a commit to Codename-11/hermes-agent that referenced this pull request Jun 15, 2026
Codename-11 added a commit to Codename-11/hermes-agent that referenced this pull request Jun 15, 2026
Codename-11 added a commit to Codename-11/hermes-agent that referenced this pull request Jun 15, 2026
@chadsm-sys

Copy link
Copy Markdown

Follow-up patch opened: #46838

Summary:

  • Preserves raw background process/delegation notification events alongside formatted CLI text.
  • Adds regression coverage for raw event preservation and backward-compatible plain input handling.

Validation from sandbox:

.venv/bin/python -m pytest tests/test_cli_process_notification_input.py tests/tools/test_process_registry.py tests/gateway/test_background_process_notifications.py -q
104 passed in 3.71s

I could not push directly to this PR head branch with the current account, so #46838 targets this PR's branch.

@teknium1
teknium1 merged commit c66ecf0 into main Jun 15, 2026
30 checks passed
@teknium1
teknium1 deleted the hermes/hermes-b1109a1a branch June 15, 2026 20:33
Codename-11 added a commit to Codename-11/hermes-agent that referenced this pull request Jun 15, 2026
…ation"

This reverts commit 500dc0f, reversing
changes made to 70f93c5.
Codename-11 added a commit to Codename-11/hermes-agent that referenced this pull request Jun 15, 2026
rdnot added a commit to rdnot/hermes-agent-medical-research that referenced this pull request Jun 15, 2026
…kfile-fix.yml)

Upstream advanced from 5bfed0f (Jun 15) to 3e7e9b2 (Jun 16). Major upstream
changes since last merge:

- feat(skills): add optional payments skills (Stripe Link, MPP, Projects) NousResearch#31343
- feat(delegation): async background subagents via delegate_task(background=true) NousResearch#40946
- fix(desktop): read HERMES_HOME from Windows registry when env is stale NousResearch#46772
- fix(desktop): let explicit model switches escape broken config providers NousResearch#46796
- fix(dump): report effective terminal backend in hermes debug
- fix(inventory): deduplicate models between user-defined and aggregator providers
- fix(doctor): recognize nvidia as vendor-slug-accepting provider
- fix(discord): cap slash commands at Discord's 100-command limit
- fix(teams): package Microsoft Teams SDK as an installable extra NousResearch#46764
- fix(dashboard): scope chat sidebar model card to selected profile NousResearch#46665
- fix(deps): declare websockets as core dep + relax dev setuptools pin NousResearch#46744
- fix(desktop): pin @assistant-ui/store so the cluster shares one tap
- fix(install): make npm install -g packages reachable on PATH
- feat(honcho-setup): replace deployment-shape prompt with gateway-gated identity tree
- refactor(honcho): canonicalize identity-mapping on pinUserPeer, migrate legacy key
- mattermost: preserve thread-local delivery hygiene
- ci: rip out nix ci for now
- ...plus desktop Electron changes, web UI updates, website docs, honcho docs

Fork customizations preserved (all 22 verified intact):
- run_agent.py: max_iterations: int = 200
- agent/conversation_loop.py: _force_final_threshold, _build_tools_summary,
  _current_turn_tool_calls, _pending_tool_summary, tool_summary flow
- agent/turn_finalizer.py: 'tool_summary' key in result dict
- tools/web_tools.py: tiered fetcher (curl_cffi→scrapling→httpx),
  MAX_OUTPUT_SIZE=10000, _get_extract_backend, max_result_size_chars=500_000,
  LLM-summarization disabled, NCBI/PubMed/PMC hardening
- tools/budget_config.py: 400_000 / 500_000
- gateway/run.py: HERMES_MAX_ITERATIONS, "200"
- hermes_cli/config.py: max_turns: 200
- cli.py: result.get('tool_summary') display
- gateway/config.py: Platform.LINE = "line"
- gateway/authz_mixin.py: Platform.LINE: "LINE_ALLOWED_GROUPS"
- plugins/platforms/line/adapter.py: reply_only_mention, build_source

Conflict resolution:
- .github/workflows/nix-lockfile-fix.yml: removed by upstream.
  Was fork-only (disabled with if:false, requires APP_ID/APP_PRIVATE_KEY
  secrets the fork does not have). Upstream deletion is correct.

Backup branch: backup-main-pre-merge-20260616-063126
pib-hermes-bot pushed a commit to Partners-in-Biz/hermes-agent that referenced this pull request Jun 16, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
alanbratu pushed a commit to alanbratu/hermes-agent that referenced this pull request Jun 17, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
zmlgit pushed a commit to zmlgit/hermes-agent that referenced this pull request Jun 17, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
al3xar pushed a commit to al3xar/hermes-agent that referenced this pull request Jun 17, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
T02200059 pushed a commit to T02200059/hermes-agent that referenced this pull request Jun 18, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
agt-user pushed a commit to agt-user/hermes-agent that referenced this pull request Jun 25, 2026
…utor + completion queue

New tools/async_delegation.py daemon-executor registry:
- dispatch() runs subagents in background threads, returns immediately
- Capacity capped by delegation.max_async_children (default 3)
- Completion events pushed to process_registry.completion_queue

Modified files:
- delegate_tool.py: added background=False param, early-dispatch branch
- process_registry.py: format_process_notification handles async_delegation type
- gateway/run.py: _async_delegation_watcher drains queue between idle turns
- hermes_cli/config.py: delegation.max_async_children setting

Partially implements NousResearch#40946 (v1: single-task only, no persistent state).

Co-authored-by: Luna Selene <luna@hermes>
agt-user pushed a commit to agt-user/hermes-agent that referenced this pull request Jun 25, 2026
- When at max_async_children capacity, dispatch() now queues the item
  (status=queued) instead of rejecting it.
- Promotion thread auto-dispatches queued items as slots free up (FIFO).
- _promotion_cv Condition variable signals the promotion loop.

New 'hermes delegation' subcommand:
- hermes delegation list [--json]  — running + queued overview
- hermes delegation status <id> [--json] — per-delegation details

New helpers in tools/async_delegation.py:
- count_queued(), get_queued_details(), get_running_details(), get_detail()

Partially implements NousResearch#40946 (v1).

Co-authored-by: Luna Selene <luna@hermes>
waefrebeorn pushed a commit to waefrebeorn/slermes that referenced this pull request Jul 2, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
Methodician added a commit to Methodician/hermes-agent that referenced this pull request Jul 4, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
habarmc1223-sudo pushed a commit to habarmc1223-sudo/hermes-agent-fluxmem that referenced this pull request Jul 8, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
santhreal pushed a commit to santhreal/hermes-agent that referenced this pull request Jul 13, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
Gravezzz pushed a commit to Gravezzz/hermes-agent that referenced this pull request Jul 21, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
leewenjie pushed a commit to leewenjie/hermes-agent that referenced this pull request Aug 7, 2026
…ound=true) (NousResearch#40946)

* feat(delegation): async background subagents via delegate_task(background=true)

delegate_task(background=true) dispatches a subagent that runs in the
background and returns a handle immediately, so the user and model keep
working while it runs. The full result — plus the original task source —
re-enters the conversation as a new turn when the subagent finishes,
riding the same completion-queue rail as terminal background processes.

- tools/async_delegation.py: daemon-executor registry, capacity cap,
  rich self-contained completion event pushed onto the shared
  process_registry.completion_queue (type='async_delegation').
- delegate_tool.py: background param + single-task dispatch branch;
  batch async rejected (v1).
- process_registry.py: format_process_notification renders the rich
  task-source block (goal/context/toolsets/model/status/result).
- gateway/run.py: dedicated _async_delegation_watcher drains + injects
  results into the originating session (idle + post-turn), session_key
  routing enrichment, shutdown interrupt of dangling delegations.
- config: delegation.max_async_children (default 3).

Reuses the existing idle-drain wiring rather than mutating a running
agent loop, preserving message-role alternation and prompt-cache
invariants. 13 targeted tests; CLI + gateway paths E2E-verified.

* test(delegation): make async non-blocking tests environment-independent

CI 'test (5)' flaked on a cold, 8-worker runner: the first
delegate_task(background=true) call measured 2.27s of one-time setup
(config load + child-agent construction + imports), tripping the
elapsed < 1.0 wall-clock assertion. That assertion was testing setup
overhead, not blocking.

Replace the wall-clock thresholds with the real invariant: dispatch
returns while the child is still gated (active_count == 1, completion
queue empty), which a synchronous impl could not do. Keep only a loose
4s sanity backstop well under the runner's 5s gate.

* fix(delegation): harden async background delegation

Follow-up review fixes:
- Detach background child from parent._active_children at dispatch —
  otherwise parent-turn interrupts (Ctrl+C, mid-turn steering), cache
  evicts (release_clients), and session close (/new) kill/close the
  detached subagent mid-run, defeating the point of background mode.
  Lifecycle is owned by the async registry's interrupt_fn.
- Make the capacity check atomic with the record insert (TOCTOU: two
  concurrent dispatches could both pass active_count() and exceed the cap).
- TUI dedup: key async_delegation events by delegation_id — the
  fallthrough keyed them all as ("", type), suppressing every completion
  after the first in the desktop/TUI status feed.
- CLI /stop now interrupts running background delegations and /agents
  lists them (they live outside the process registry and were invisible).
- Drop stray unbalanced ']' line from the re-injection block and the
  unused _ASYNC_DEFAULT import.

Tests: detach-at-dispatch + concurrent-capacity race added (15 total in
test_async_delegation.py); 137 delegate + 140 process-registry/notify/watch
+ 7 TUI dedup tests pass.

* fix(delegation): harden async background completion drains
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint comp/gateway Gateway runner, session dispatch, delivery P3 Low — cosmetic, nice to have tool/delegate Subagent delegation type/feature New feature or request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants