Child escalations (coder plan approvals, deviation gates, user__* calls
from subagents) previously expired after the escalating agent's
escalation_timeout, defaulting to 300 s, and 0 timed out instantly. In
autonomous runs the orchestrator often could not reply within that
window, so approvals silently fell through to the best-judgment
fallback.
- Extract await_escalation_reply as the wait seam: 0 awaits the oneshot
with no deadline; nonzero keeps tokio::time::timeout with the same
timeout message. A dropped sender still resolves to the existing
cancelled reply either way.
- DEFAULT_ESCALATION_TIMEOUT_SECS and default_escalation_timeout() -> 0,
so graph agents (which inherit the default via from_graph) are covered.
- Document the new semantics in config.agent.example.yaml and drop the
5-minute-timeout claim from the default spawn instructions.
- Tests: paused-clock unlimited wait (3600 s), nonzero expiry message,
dropped-sender completion, default pin, and a grep pin that the stale
timeout wording is gone from src/, assets/, and the example config.
Plan: PLAN-review-retry §C (D3).
- agent__spawn accepts an optional variables object (string->string);
values overlay the parent-inherited agent variables per key, with
inheritance preserved as the fallback layer
- unknown variable names, non-string values, and non-object shapes
return self-correcting status:error JSON listing the agent's
declared variables
- spawned children can no longer trigger the interactive variable
prompt: a new interactive gate in Agent::init_agent_variables routes
missing required variables to the existing bail path while leaving
CLI/REPL prompting and --info skip semantics unchanged
- agent__task_create persists variables on the task, agent__task_list
echoes them, and task auto-dispatch threads them into the spawn
- tool schemas and built-in spawn docs updated; unit tests extended
in agents.rs and taskqueue.rs
Semaphore::new panics above tokio's MAX_PERMITS, so a fixed or templated
map cap (or settings.max_concurrency) that large crashed the run instead
of erroring. The validator rejects both at load, resolve_max_concurrency
rejects every arm at run time, and the executor checks the graph setting
before building its own semaphore, so validation mode makes no difference.
Also corrects run_child_agent_with_graph_inputs' contract comment: inputs
for a config-only agent are rejected, not ignored.
LlmNodeExecutor::execute now takes the graph's abort signal and threads
it through run -> run_with_retries -> run_chat_loop, replacing the
loop-private signal, so an in-flight model call is interrupted and every
turn starts with an abort check that bails 'llm node aborted' before any
client is created.
run_llm_function gains an optional abort signal: the synchronous tool
wait loop kills the child and returns a 'Tool call ... aborted' error
within one poll interval when it is set. The tool dispatcher passes the
session abort; the agent _instructions runner passes None.
A graph run now observes user interrupts wherever it is: GraphExecutor::run
spawns a run-scoped bridge that maps SIGINT or the enclosing session abort
onto the graph's own abort flag (only when a session signal exists; a
pre-set session trips a synchronous check). The frontier join selects on
the abort and, like the timeout path, cancels every in-flight branch task,
failing the run with 'Graph ... aborted during super-step'.
Branch and map item tasks are cancellation-owned via TaskCancelGuard, so
dropping a map task's future no longer detaches its item chains. Peer
retirement is a Drop guard (PeerRetireGuard) constructed before the
semaphore acquire and dropped explicitly right after step(), so the
success-path ordering is unchanged while task abort, panic, and the
pre-step abort return all still retire the identity.
A child agent spawned from a graph node now carries its parent's session
abort signal, so a user interrupt is observable from every sync wait in
the agent tree. Background agent__spawn children are unchanged.
run_child_agent_with_graph_inputs also guards the child's own supervisor
with a disarmable CancelOnDrop: if the future is dropped mid-flight or
the child fails before its normal cleanup, its spawned subagents are
cancelled instead of running on invisibly.
Two offline chain tests materialize an echo-inputs graph agent whose
script reports the width it sees: a parent with inputs.width gets the
parent's value back, a parent without inputs gets the child's YAML
default. Two agents.rs tests pin the exact runtime guard texts for a
config-only target and a reserved initial_prompt key.
Task T5 of plans/map-subgraph-branches-design.md §3.12.
Agent nodes gain an optional `inputs:` map of templates resolved against
the parent state and overlaid on the child graph agent's initial_state
before its executor starts. A lone {{key}} keeps the state value's JSON
type; mixed text renders to a string; initial_prompt is still owned by the
dispatcher. run_child_agent and run_active_agent_graph keep their
signatures as wrappers over the new *_with_inputs variants, and
run_agent_for_graph rejects inputs aimed at a config-only agent or
containing initial_prompt.
Design: plans/map-subgraph-branches-design.md §3.12 (C18), task T5.
PR #24 follow-up 1. The bundled-tools test called the writing
generate_bash_declarations on the repo's assets/functions/tools/*.sh,
so cargo test could rewrite tracked files. Split the build+export step
into a pure build_bash_tool so the test never touches the source tree.
A map item no longer runs a single branch node; it walks the subgraph
reachable from `branch` over `next`/`fallback` edges (branch_subgraph),
one chain per item, holding one semaphore permit for the whole chain.
Chain steps go through the shared `step()` so llm retries, fallbacks,
script `_next` and agent timeouts keep their existing semantics. A node
without `next` ends the chain (StepContext.branch_mode), a route outside
the subgraph, a fan-out, an end node or a loop past
max_loop_iterations fails the item, and every item error carries
`map node 'X': sub-branch [i] failed at node 'Y' (step n)`.
`output_key` is removed from each item's fork before the chain runs, so
the collected value must be written by the chain rather than inherited
from the parent.
Teammate identities are now per item for the whole chain: provisioned
when any flagged agent is in the subgraph, armed only before
`teammates: true` agent steps, and retired at chain exit. mark_finished
moves out of run_agent_for_graph to the two provisioning sites (frontier
task and chain exit), so timed-out and early-failing agents are also
marked finished.
Agent nodes marked teammates: true whose sub-agents run concurrently (map fan-out branches or multiple flagged agent nodes in one super-step) now form a team: each member gets a pre-provisioned inbox in a messaging-only PeerRegistry before any branch starts, a teammate roster injected into its prompt, peer resolution in agent__send_message, and a peers section in agent__list_running. Registries are scoped to a single fan-out and carry no lifecycle semantics: outputs still flow exclusively through collect_into/state reducers, finished peers reject new messages, and nothing registers in a supervisor. Also hardens the agent-recursion depth guard to work without a supervisor and adds validator warnings for zero-valued orchestration limits.
The user-scope MCP config historically lived at
<config-dir>/functions/mcp.json, a leftover from when MCP support was
part of the llm-functions tooling. It now resolves through a single
choke point with these semantics:
- Preferred location: <config-dir>/mcp.json (created there on first run)
- Historical <config-dir>/functions/mcp.json still honored when the
preferred file does not exist, so existing installs are unchanged
- If both exist, the preferred location wins
--info/.info now reports the resolved location as mcp_config_file, and
the --scope help text plus config.agent.example.yaml reference the new
default. This also removes the asymmetry with the workspace scope,
which already used .coyote/mcp.json directly.
The supervisor registry went kind-generic (TaskHandle::Agent | Job)
earlier in this branch, but the module holding the agent__* handlers
and two model-facing error strings still carried the old name:
- src/function/supervisor.rs -> src/function/agents.rs (it contains
only agent__* tool handlers, pairing with function/jobs.rs; the
kind-generic src/supervisor/ registry keeps its name)
- 'Supervisor tool failed' -> 'Agent tool failed'
- 'Unknown supervisor action' -> 'Unknown agent action'
Graph LLM nodes now own the jobs they start, on every exit path. A new
node_job_scope on RequestContext records job ids started while a node
runs: the turn-end guardrail nags only about the node's own jobs
(parallel branches no longer see each other's), and the node executor
reaps — cancels and deregisters — anything left registered when the
node exits, including error, timeout, and retry-exhaustion paths.
Cross-node job handoff is no longer possible; a crashed node takes its
in-flight jobs with it.
With inheritance gone, job__* declarations are gated on capability:
the family is only declared when at least one declared tool would pass
job__start's whitelist (shared predicate: is_backgroundable_tool). One
carve-out — while a context still owns registered jobs (job started,
tool disabled mid-session), the lifecycle verbs stay declared so a
running job can never become unreachable; job__start alone disappears.
A graph node with tools: [] now sees no job__* tools at all.
Prompt instructions, tool declarations, and graph.example.yaml updated
to the node-local semantics; +7 tests, 8 visibility pins rewritten.
job__collect's 50k-char tail cap is a safety default, but collect is
consume-once and the cap was mandatory — a model that genuinely needed
the complete output had no recourse. Add a full_result boolean that
skips the cap (tail_lines still honored; the session-wide
max_tool_result_chars limit still applies downstream), teach the
truncation banner to name the recourse, and point job__check's
output_bytes_captured at the collect decision.
Since the supervisor registry became kind-generic (agents AND jobs),
'supervisor' naming on the agent__* tool plumbing was misleading:
job__* handlers operate on the same supervisor. Rename
SUPERVISOR_FUNCTION_PREFIX -> AGENT_FUNCTION_PREFIX,
supervisor_function_declarations -> agent_function_declarations,
handle_supervisor_tool -> handle_agent_tool. No behavior change.
memory_config() only reports enabled when a global memory index or a
workspace memory store exists on disk, so asserting the handler's
'name is required' detail was environment-dependent even with the
memory pref forced on. The routing test now accepts either
memory-handler-owned message: the 'Memory tool failed' prefix alone
proves the memory__ prefix reached the memory handler.
The eval_routes_memory_prefix_to_memory_handler characterization test
inherited the host machine's memory configuration: on runners without a
memory setup, should_register_memory_tools() gates the handler off and the
error message differs. Force memory = Some(true) at ctx construction so
the test asserts the same handler path everywhere.
The spawned-agent task now pushes an agent_completed/agent_failed event
into the spawning context's notification queue before returning, so a
parent that keeps working learns mid-turn that a child finished instead
of discovering it only at the turn-end guardrail. Cancelled or
already-collected agents are suppressed by the existing drain-time
registration filter. This delivery applies regardless of whether
background jobs are enabled.
- add NotificationQueue/SystemNotification: every context owns a fresh
queue (children never inherit the parent's, avoiding first-drainer-wins
races between transcripts)
- job tasks push job_completed/job_failed events on completion, failure,
and timeout; a panic skips the push and is surfaced by the guardrail's
finished-handle enumeration and collect's JoinError mapping instead
- events for jobs already collected or cancelled are dropped at drain time
by filtering against live supervisor registration
- replace inject_escalation_notification with single-pass
merge_system_channel: pending_escalations (root-only) ordered before
system_notifications (any depth) on the last tool result of a batch;
byte-identical output when notifications are empty, proven by the
unmodified pre-merger characterization tests
Reject fast built-in file tools (fs_* / ast_grep) in job__start per the
backgroundable-tools whitelist; clean up env-snapshot temp files on every
exit of run_process_job via a drop guard; bound the output-pump awaits and
abort them on the failure path; treat signal death (no exit code) as a
failure with a teaching message; bound job__collect's post-drain join with
a SIGKILL escalation so a TERM-ignoring process cannot hang collect after
a Ctrl-C teardown; document the unguarded SIGTERM pid-reuse window; give
the injected Background Jobs prompt section a fresh line on both sides;
extract the MCP server name with strip_prefix instead of replace.
Capacity-0 audit for jobs-disabled contexts: REPL displays have no
supervisor consumers (only Ctrl-C/exit cancel_recursive at
repl/mod.rs:460,473, kind-agnostic); session save/load does not persist
supervisor state (src/config/session.rs has no supervisor references) --
nothing to test for either.
Detached tokio::process runner with a frozen JobEnvSnapshot (env-derived
bin dirs, vault-interpolated agent envs, COYOTE_TOOL_TIMEOUT resolved at
start), process_group(0) with pgid-guarded SIGTERM/SIGKILL escalation,
capture-only ring-buffer telemetry, and LLM_OUTPUT read after wait().
MCP jobs snapshot a single-entry McpRuntime holding only the validated
server and render through the same free fn as the foreground path.
job__start enforces its gates synchronously before any spawn:
jobs_enabled, the backgroundable whitelist with directionality teaching
errors, the per-request declared-names stash captured in
before_chat_completion, then capacity (lazy supervisor get-or-init in
plain sessions). job__check/list read the shared JobState cell without
consuming; job__collect blocks with the escalation early-out and applies
a tail-biased char-boundary cap plus optional tail_lines; job__cancel
kills the group with a 5s grace.
Job declarations are injected iff jobs are enabled at agent init, the
plain-session function-init sites, and the exit_agent rebuild; job__ is
carved out of enabled_tools filtering and excluded from
concrete_tool_names so REPL toggles cannot grant or revoke it.
When max_chars landed inside a multi-byte UTF-8 character of the
serialized output, s.get(..max_chars) returned None and the code fell
back to the FULL untruncated string while still prepending the
truncation marker — the "truncated" output actually grew. The cut is
now floored to the previous char boundary so the prefix is always a
valid, genuinely truncated slice.
agent__check on a finished agent delegated to agent__collect, which
returned the full (unbounded) result and consumed the handle. That
contradicted the tool's own docs and broke the check-then-collect
pattern: a second collect on the same id failed.
check now reports { status: finished } with a pointer to
agent__collect and leaves the handle registered; collect is the single
retrieval verb. The tool description and prompt table are updated to
stop promising that check returns the result.
The turn-end guardrail only counted still-running agents, so an agent
that finished before the turn ended was invisible: its uncollected
result was silently dropped. Jobs were never counted at all.
The guardrail now enumerates every registered task (running and
finished, agents and jobs) via Supervisor::list_tasks and renders a
kind-aware prompt with two sections: still-running tasks to reclaim,
and completed-but-uncollected tasks with the exact collect command.
At the force-terminate cap, finished-but-uncollected handles are
explicitly discarded with a warning naming the lost ids, so the
guardrail cannot loop forever on handles nobody will collect.
Implements T1 of plans/background-jobs-design.md (§6, R7/R8/R9):
- Supervisor.handles is now HashMap<String, TaskHandle> where
TaskHandle = Agent(AgentHandle) | Job(JobHandle); agent-facing
accessors (active_count, effective_active_count, is_finished, take,
inbox, abort_signal_for, list_agents) match only Agent variants,
preserving all existing external behavior byte-for-byte.
- New JobHandle/JobState/JobStatus/JobResult types with pgid-guarded
process-group kill discipline: Drop and cancel_all/cancel_recursive
kill the group only while state.pgid is still set (pid-reuse guard),
via libc::killpg on unix and JoinHandle::abort elsewhere.
- Per-kind job capacity: Supervisor carries max_concurrent_jobs
(builder-set, default 0); job registration rejects at capacity.
- Cross-kind teaching errors at the four agent-lookup miss sites
(agent__check/collect/cancel/send_message) when the id is a
registered job or job_-prefixed; genuinely-unknown ids keep their
existing messages.
- Supervisor init condition is now can_spawn_agents || jobs_enabled in
use_agent and both child-agent spawn paths, with agent capacity 0 in
jobs-only contexts; use_agent cancels the old supervisor recursively
before replacing it.
- max_concurrent_jobs config plumbing: global Config field, AgentConfig
override + accessor, all four AppConfig touch points including the
COYOTE_MAX_CONCURRENT_JOBS env override; shared
effective_max_concurrent_jobs/jobs_enabled predicates
(agent override -> global -> default 5; 0 disables).
- Stage dependency-free RingBuf (64 KiB default) in src/function/jobs.rs
for the upcoming job output pump.
- New sanctioned dependency: libc 0.2 under cfg(unix).
Route CallToolResult content through the render.rs content policy per
plans/mcp-resources-prompts-design.md §6 (T8): oversized text sliced at
TEXT_MAX_BYTES_CLAMP with a self-explaining truncation note, image/audio/
embedded blob content spilled (or inlined when UTF-8-clean) instead of
shipping base64 into model context, and structuredContent subject to the
same ceiling. Clamp server-controlled uri/mime metadata strings to the new
METADATA_MAX_BYTES bound in both the read and tool-result paths, sanitize
the terminal rendering of MCP dispatch errors while keeping raw text in
the tool_call_error payload, and surface resource audience annotations in
both mcp_search results and mcp_read metadata via the catalog.
Dark-Alex-17