Two named custom providers may share one gateway URL with different
credentials. `_resolve_child_credential_pool` matched custom pools by URL
only, so a child inherited whichever pool was registered first for that
endpoint. Pass the child's `requested_provider` (and the parent's) into
`get_custom_provider_pool_key(provider_name=...)`, which already prefers a
name match over URL order.
Second half of the class fixed by #117839; grafted from PR #89021 onto the
split `delegate_tool_config` layout (#45763).
A finite `hermes chat -q` / `--oneshot` run has no later session in its HERMES_HOME
to learn for, yet it ran the full interactive self-improvement loop. Measured over
21 one-shot benchmark trajectories: 7 skills created and a bundled one patched
mid-task, 37 of ~215 tool calls on skill_view/skill_manage, skill text = 34% of all
tool-result bytes fed back into context, plus reviewer subagents spawned on the
agent's own diff (one task: 5 delegations, 62 subagent API calls, each re-paying a
cold system prompt).
Keyed on the existing HERMES_SINGLE_QUERY_SESSION marker (approval gate, delegation
dispatcher), so interactive and gateway sessions are byte-identical:
* agent/oneshot_footprint.py (new sibling): skill_manage is pruned from the tool
set; the ## Skills block keeps the index + skill_view but drops the record/patch/
offer-to-save coaching and the "load process skills for work you already know"
push (SKILLS_GUIDANCE follows because it is gated on skill_manage).
* delegation.oneshot_max_children (default 2, 0 = unlimited): total children a
one-shot run may spawn; past it delegate_task returns a tool error telling the
model to finish inline.
* requesting-code-review skill: reviewer/fixer subagents (Steps 5 and 7) are
interactive-only; one-shot applies the checklist inline.
Live: `hermes chat -q "list tools starting with skill_"` on the portal — base
"skill_manage, skill_view, skills_list", fix "skill_view, skills_list"; a real
AIAgent under a temp HERMES_HOME shows the interactive prompt unchanged (6643 chars
both) and the one-shot prompt without skill_manage / offer-to-save.
`preview_threshold_tokens` restated the resolve -> floor -> compute -> cap chain that `update_model`
runs; two copies of the trigger math drift the next time a step is added — the exact bug class
#83450 fixes (the guard quoting a number the compressor will not install). `_derive_trigger` is the
single pure derivation; the auxiliary-summariser ceiling stays in `_apply_threshold_tokens_cap`
because it is per-runtime, not per-model.
The startup banner names the cap only when it set the trigger; on windows where the ratio already
sits below it "(capped at 256,000)" was noise. Comments no longer repeat the default literal.
delegation.compression_threshold_tokens and _apply_child_compression_cap both described a 1M-window
child compacting at 500K "like its parent"; with compression.threshold_tokens defaulting to 256K the
parent (and so the child, via the min of both caps) compacts at 256K. preview_threshold_tokens now
says it ignores the auxiliary-summariser ceiling, which only the post-switch probe can know.
A turn's hard interrupt fans out only to `_active_children`; background
delegate_task units are detached from the parent at dispatch
(`_dispatch_background` / honor_parent_interrupt=False), so /stop left them
running to completion and their result arrived minutes later as a parked
wake. Every stop surface now also calls
`tools.async_delegation.interrupt_for_session` for the session's units:
- gateway: `_interrupt_and_clear_session` (busy /stop, /new share it — /new
already did this in `_handle_reset_command`, the second request is
idempotent) and the idle `_handle_stop_command` tail, which replied
"No active task to stop." while a background child was running; it now
stops them and replies "Stopped".
- tui_gateway `session.interrupt` (Desktop Stop / TUI stop) — own UI sid +
spawner id only, so a viewer tab never kills gateway work.
- acp_adapter `cancel`.
- CLI `/stop` already used `interrupt_all` (process-wide); unchanged.
The stop recurses: depth>0 delegations are always synchronous
(`_model_background_value`), so the child's hard interrupt reaches its
workers through its own `_active_children` fan-out, and each level's
interrupted partial result rolls up as that child's completion.
An interrupted child's entry now carries what it actually had: the loop's
`final_response` is the "Operation interrupted." placeholder (also appended
as the closing assistant row), so `_build_result_entry` takes the child's
last real assistant text as `summary` and keeps the placeholder as `error`.
The unit finalizes normally and re-enters at once as its completion notice
(status=interrupted, "Partial output: ...", "Subagent Task Interrupted" on
TUI/Desktop) instead of the chat waiting for the child's budget to run out.
Docs: delegate_task description, tools/AGENTS.md, delegation.md,
gateway-session-lifecycle.md.
Part of #114456
kanban_complete tool collapsed every complete_task refusal into
'unknown id, stale run, or already terminal'. When a parent reopens
mid-run, a worker with genuinely done work is refused for a reason the
message hides, and the run burns (then crashes, re-queues, and repeats
deterministic work). Name the unsatisfied parents (id, status) from a
read-only task_links mirror of _parents_satisfied so the worker and the
operator complete the gates instead of re-running; unknown-id and
terminal refusals keep the generic text, and the dependency invariant
itself is untouched (no force path).
Complementary to the CLI-surface PRs #110323/#110330/#110334, which
touch hermes_cli/kanban_db.py + hermes_cli/kanban.py; this change is
tool-surface only (tools/kanban_tools.py).
Also surface the delegate_task handoff contract at spawn time: children
cannot close tracked work and must return findings for the parent to
apply, so a delegated card no longer burns a full run before learning
it can never close (#113373 ask 1).
Tests: tests/tools/test_kanban_complete_parents_tool.py (6) — parents
named, deterministic multi-parent order, unknown/terminal generic,
happy path, description contract. Proven red on unfixed sources.
On a finite runtime (`hermes chat -Q`, the Bot Chat one-shot delivery child) delegate_task falls
back to a synchronous batch and had no bound at any layer: delegate_task is exempt from the
sequential tool deadline, delegation.child_timeout_seconds defaults to off, and the heartbeat
only stopped itself on its stale verdict ("let gateway timeout fire") — a watchdog a -Q turn does
not have. A worker wedged after its final answer therefore held the parent turn, the Bot Chat
session lease, and the process forever.
_Heartbeat.settled is set on the stale verdict and _ChildRun.await_child waits on it (the
worker future done-callback sets it too), so the wedge ends through the existing timeout /
deferred-close path with a structured `timeout` entry naming the stalled progress.
Slim redo of the mechanism proposed in #109762 (@fangliquanflq, earliest) and #113275
(@JoaoMarcos44); their quarantine subsystem, 50 ms polling loop and collect-grace were dropped.
Fixes#109749
Co-authored-by: fangliquanflq <fangliquan@qq.com>
Co-authored-by: joaomarcos <joaomarcosdias444@gmail.com>
When a child's final answer still missed its output_schema after the one
bounded retry, the result entry flipped to status=failed with the error
"Final answer does not satisfy the declared output_schema" — the completion
line printed ✗ and orchestrators read a finished audit as a failure. Five
audits of 413-4103 s were lost this way in the Sep 10-14 retrospective and
the parent had to mine the live transcripts; in four of them the "violation"
was a ```json fence around a valid array, which the candidate extractor
sliced to its first..last object.
Now: status stays completed, `summary` is the child's raw final text,
`schema_valid: false` + `schema_errors` carry the verdict and a `schema_note`
says the text is unvalidated; the sync completion line shows ⚠ with the
reason. The extractor tries the earliest-opening bracket span and keeps the
first that parses (fenced arrays validate). The OUTPUT CONTRACT the child
sees now says "ONLY the JSON value — no prose, no code fence" and what a miss
costs. One bounded retry is unchanged.
Subagents can now SEE images. Each delegate_task task accepts an optional
images list (max 8; local paths or http(s) URLs). Vision-capable children
receive native image_url content parts on their goal turn (local files as
data URLs, remote URLs verbatim); non-vision children get
[Image attached at: ...] hints plus a vision_analyze pointer. Routing
reuses agent.image_routing (decide_image_input_mode /
build_native_content_parts), so agent.image_input_mode governs delegation
exactly like inbound gateway images.
Best-effort by contract: malformed images arrays fail the call loudly
before any child spawns; unreadable paths are skipped with a log line;
any exception in the forwarding path degrades to the text-only goal.
Adapted from RooCodeInc/Roomote#1796 / #1767 (Fast agent forwards bounded
current-turn attachments to delegated coding tasks).
_resolve_child_fallback_chain re-implemented hermes_cli.fallback_config's entry
validation to log per-index warnings, wrapped a pure function in try/except, and
carried two names (routing_cfg / fallback_cfg) for one argument. It now delegates
to get_fallback_chain() and keeps the single warning for "no usable routes";
the facade re-export is dropped (import from the defining module). Tests collapse
to the parametrized decision table plus the pin-derivation, real-config-loader,
review-ownership and activation invariants (22 -> 20 cases, 417 -> ~230 lines).
A child's background processes are killed at its teardown and their
notify_on_complete notices are suppressed in the parent, yet the child's
terminal result still said `notify_on_complete: true` and the parent's
delegation notice said nothing about processes left behind. Orchestrators
believed "CI watcher running" and waited on a completion that could never
arrive (recurring in the Sep 7 campaign sessions).
- process_manage(action="handoff", session_id, data="<purpose>"), children
only: process_registry.transfer_ownership flips owner_task_id/task_id/
session_key to the parent under the registry lock, so the completion is
stamped with the parent's owner at exit, passes the parent's sa- filter,
and is reaped by the parent, not the child. Cap 3 per child; an exited,
foreign, or non-child request is a tool error. The purpose rides the
event as handoff_note and renders in the parent's notice.
- Child terminal(background=True, notify=True) now returns
notify_on_complete=false plus a note: wait, kill, or hand off.
- _ChildRun.account_background_processes records handed_off_processes and
orphaned_processes on the result before cleanup kills the leftovers; the
parent's delegation block renders both.
Three orchestrator failures traced through the Sep 7 gpt-6-astra campaign sessions:
1. delegation.independent_completions (new, default false). #104299 made every
ungrouped task its own completion message, so a 15-task call woke the
orchestrator up to 15 times; one chain received 132 notices and answered
130 of them with "already incorporated". A multi-task call now returns as
ONE consolidated message unless the flag is on; `group` is inert until then.
2. Queued units were killed before they started. Units of one call share a
pool slot but the executor was still sized by slots, so with 15 units live
a new unit queued behind a full pool; the stale monitor's clock ran from
dispatch, interrupted it at 450 s, and the child exited `interrupted 0.02s`
when its thread finally came up (13 such lanes in one session). The
executor now grows to the number of live units and the stall clock arms
when the runner actually starts.
3. The tool text said "do not wait or poll — just continue" without saying
that completions are delivered only BETWEEN turns. A model that never ends
its turn (one 203-minute turn, 717 API calls) never received 40 finished
results. Tool description, dispatch note and completion header now say to
finish independent work, give a one-line status, and end the turn.
Slimmed after review: the 200K default is dropped. Children compact at the same
0.50 x window ratio trigger as their parent (500K on a 1M model). Reasons:
- the run this came from happened at 0.85 (850K); main was already at 0.50, so
the real delta against main was 500K -> 200K, not 850K -> 200K;
- a replay of the run's 22,489 logged calls (evals/postmortem, cap sweep) put
200K-400K caps within 5% of each other in cost once cache prefixes are intact,
because the write price dominates and the cap only trims read volume;
- every compaction is a chance to lose detail, and the accuracy side was never
measured; at 500K a 1M child compacts roughly never.
What stays: the reviewer's finding that the value was coerced, not validated
(YAML true -> int 1 -> a one-token trigger; "200k" -> silently off). Values are
validated: int >= 16000 enables the cap, 0/false/null/unset = off, anything else
is warned and ignored. Docs and config comment restated accordingly.
A background delegate_task call used to be ONE async unit: the runner joined on
every child and a single consolidated message re-entered the conversation when
the SLOWEST finished. Fifteen independent PR reviews therefore waited on the
fifteenth before the parent could act on the first.
Each task now carries an optional `group`. `_units_of` partitions the call's
children into units — one per distinct group, one per ungrouped task — and each
unit is dispatched to the async registry on its own, so its results re-enter the
conversation as soon as THAT unit is done. Tasks that must be compared or merged
share a group and still return together.
Capacity is unchanged: every unit of one call joins the first unit's pool slot
(`slot_key` in `async_delegation._dispatch`), so splitting never consumes more of
`delegation.max_concurrent_children` than the call did. Unit ids suffix the
call's id (`deleg_xxxx-1`, `-2`, …) so live transcripts stay under one dir; the
completion block names the group and notes that sibling units report separately;
`active_task_count` counts a unit's own tasks.
A delegated child copies the parent's prompt_caching.cache_ttl. The 1h tier
is priced for a person who steps away between turns (2x write vs 1.25x for
5m); a subagent calls every few seconds for minutes and is gone, so it paid
2x on every tool result and never collected the retention. Live measurement
(Sep 5, 40 concurrent Fable 5.1 children via OpenRouter, 1h markers on the
wire): cache writes billed at $20/M against $12.51/M for the same run at
5m, ~60% of a write-dominated bill.
_apply_child_cache_ttl runs right after child construction: 1h -> 5m,
5m stays, disabled stays disabled, parent untouched. Tests: unit (markers
on the wire drop ttl; parent's 1h layout still differs) and the real spawn
path with cache_ttl: 1h in a temp HERMES_HOME.
Independent review: a YAML `true` coerced to int 1 and gave every child a
one-token compression trigger; "200k" silently disabled the default cap.
Values are now validated: an int >= 16000 is used, 0/false/null disable on
purpose, anything else warns and falls back to the 200K default so a typo
never costs money. Config comment and docs say it caps the compaction
trigger, not a hard request-size limit.
Test: true / "200k" / 5 -> default; 0 / false / None -> disabled; valid ints
pass.
A delegate_task child inherits the compression threshold as a RATIO of the
model window. On a 1M-window model at the run's configured 0.85 that is an
850K-token trigger: in the 1,393-agent refactor run 1,373 of 1,375 children
never compressed once, 62% of all API calls carried >150K of context, and the
calls above 200K carried ~$10.9k of the $19.3k bill (58% of it cache WRITES,
i.e. re-sending a 300-800K prefix on every call). A sawtooth replay of the
logged calls with a 200K cap / 65K floor cuts context spend by ~49% (~$7k).
Children are brief-driven and disposable; they re-read their brief and the
files they touch, so a large window buys them little. New
delegation.compression_threshold_tokens (default 200000) is applied to the
child's ContextCompressor right after construction as the lower of it and
any global compression.threshold_tokens; the parent's own trigger is
untouched. 0 disables the subagent-specific cap. The compressor applies
threshold_tokens_cap on first window resolution, so this is byte-equivalent
to the user having set compression.threshold_tokens for the child.
Live through the real spawn path (_build_child_agent, real imports, temp
HERMES_HOME, 1M-window model): main child trigger 500,000 / branch 200,000;
parent 500,000 on both.
Tests (3): default caps a 1M child at 200K; the cap is the lower of the
delegation and global values and never raises a small-window child's
trigger; 0 disables and an already-resolved trigger is re-clamped.
Docs: delegation.md, configuration.md.
Every PLUGIN-COMPAT __getattr__ now calls hermes_cli.plugin_compat.warn_once(facade, name, target) before
resolving, emitting a HermesPluginCompatWarning (FutureWarning) once per process per name: old path, new
path, removal target. Importing a facade for its live API stays silent; only resolving a moved name warns.
COMPAT_MANIFEST.md documents the warning and how to silence it during migration.
Verified the runtime never routes through a pointer: every entry point (run_agent, cli, hermes_cli.main,
gateway.run, tui_gateway.server, web_server, model_tools + tool discovery, hermes_state, cron.scheduler,
browser_tool, mcp_tool, kanban, auth) imports clean and `hermes doctor` runs end to end with the warning
promoted to an error.
Also restores the check_compat_pointers CI step to .github/workflows/lint.yml, which a0be177aac dropped
when the compat layer was regenerated (the lint script itself was present; the workflow step was not).
hermes_cli/plugin_compat.py, tests/test_plugin_compat_warning.py and the two-line insert per facade are
part of the compat layer and go away with it.
The Sep 2026 decomposition (PR #102117) makes internal import paths a non-API: names now live in
the focused modules that define them. This commit is the ONLY thing keeping the old paths alive,
so external plugins have time to update. It is deliberately a single, unsquashed commit:
git revert <this sha>
removes every shim, stub and manifest at once on the announced date. Nothing in-tree may depend on
these pointers: scripts/check_compat_pointers.py (wired into lint.yml) fails CI if it does.
What it adds (see COMPAT_MANIFEST.md, compat_manifest.json):
- 332 facade modules get one delimited `PLUGIN-COMPAT` block appended at the end of the file
- 1,172 moved names resolved lazily via a module `__getattr__` (PEP 562) — never a top-level import,
so no import cycles; facades that already had `__getattr__` get a chained one
- 592 third-party/stdlib names the old modules used to expose, with their original import statements
- 266 public definitions that had been deleted as unused, restored byte-for-byte from the pre-decomposition
tree (+40 private helpers and 16 imports pulled in only because a restored definition needs them)
- 3 deleted modules recreated as re-export stubs (gateway/startup_watchdog, hermes_cli/observability/
relay_runtime, tools/environments/modal_utils)
- private names (`_x`) get no pointer: they were never API (3,792 skipped)
Verified: all 335 touched modules import under a fresh HERMES_HOME and every manifest name resolves;
the lint reports zero in-tree uses; ruff clean; targeted suites unchanged.
hermes_state.py: delete every '# noqa: F401 (re-exported...)' import block (hermes_state_common/errors/guard/
readpool/sessions/fts/dbfile/wal/repair/registry + agent.context_compressor _DB_PERSISTED_MARKER_KEY); keep
only the names hermes_state.py itself uses, without noqa.
hermes_state_registry.py: drop get_shared_session_db/release_shared_session_db/close_shared_session_dbs
aliases; every caller (gateway/, tools/, tui_gateway/, cron/, mcp_serve, run_agent, tests) now imports
acquire/release/close_all/release_or_close from hermes_state_registry.
hermes_state_titles.py: drop set_auto_title_if_empty shim (title_generator keeps its getattr fallback).
Re-remove shim-only names restored by 34abf954bd: latest_user_message_row_id (tests call
latest_message_row_id(key, role='user'); role-targeting assertions kept) and get_session_activity (tests
build the snapshot via agent.session_activity.build_activity_snapshot over db.get_session(sid)).
hermes_state_wal._log_once resolves its dedupe sets as module globals instead of via hermes_state;
hermes_state_repair helpers call module globals directly (tests patch hermes_state_repair.<name>).
Frozen updater surface untouched (update_cmd_maint imports only SessionDB from hermes_state).
For each issue anchor present in BASE 63279301bc non-test .py and absent on HEAD, the BASE comment/docstring block was re-attached at the HEAD location of the code it explained (matched by the distinctive code line / enclosing def). Sentences already covered by an existing HEAD comment were deduped; the issue number always survives. Insert-only: no code lines changed.
A fan-out of N in-process subagents used to add one sleeping daemon
thread per delegated child (delegate heartbeat, 30s) and one or two per
active turn (durable turn-lease refresher; turn-liveness watchdog). A
profiled session with ~130 children was carrying ~1000 threads. All
of these timers now run on a single process-wide daemon thread.
- agent/periodic_scheduler.py (new): heap-ordered periodic scheduler on
one Condition-driven daemon thread. schedule(fn, interval) -> handle;
handle.cancel(wait=) blocks for an in-flight run like the old join.
A callback returning False stops itself; a raising callback is logged
at debug and rescheduled, so one bad timer cannot kill the rest.
- tools/delegate_tool.py: _heartbeat_loop body -> _heartbeat_tick,
scheduled at _HEARTBEAT_INTERVAL; stale-cycle closure state and
idle/in-tool thresholds unchanged; cancel(wait=5) in finally where the
stop-event + join(5) lived.
- run_agent.py: _refresh_durable_turn_lease body scheduled at
_lease_refresh_interval; lease-lost / refresh-error interrupt paths
and the stop-event fencing are unchanged; the join(timeout=1.0) is now
cancel(wait=1.0) so the interrupt clear still runs after any in-flight
tick.
- agent/turn_liveness.py: TurnLivenessWatchdog.make_thread/start ->
schedule(); the poll body is _tick(), same sampling state machine.
Bench (evals/fanout_resource_bench.py, 30 children / 10 worktrees,
ok=30/30 both): peak threads 168 -> 132. At peak the old tree held 30
"Thread-N (_heartbeat_loop)" threads; the new one holds zero plus one
"hermes-periodic-scheduler".
