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.
The executor integration-test module hoisted job-test paths into
module-level imports, but their only consumer is a #[cfg(unix)] test —
on Windows the imports went unused and failed -D warnings.
Graph agents could only inherit the app-wide job budget; the agent-level
header in graph.yaml now accepts max_concurrent_jobs alongside
model/temperature, flowing through AgentConfig::from_graph into the
run-wide supervisor. Deliberately graph-wide, not per-node: jobs outlive
the node that started them.
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.
- Extend the injected Background Jobs prompt guidance: system_notifications
push on completion, collect-only-when-idle wait protocol, and the graph
LLM-node collect-before-final-turn rule
- Mention the system_notifications push in the agent spawning guidance and
in the sisyphus/architect wait-protocol text (agent completions push
notifications too)
- config.example.yaml: max_concurrent_jobs (default 5, 0 = disabled)
- README: features-list entry pointing at the Background-Jobs wiki page
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.
Agents whose function pool includes job__* declarations get a Background
Jobs section teaching start/check/collect/cancel discipline and the
snapshot/no-persistence semantics. Presence of the declarations doubles
as the jobs_enabled predicate, so a context with function calling off or
max_concurrent_jobs 0 sees no job prompt text.
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).
Completes the .prompt/.temp-role split: --prompt set an ad-hoc system
role, which is what .temp-role now means everywhere. The --prompt name
is left unbound so a future one-shot MCP prompt flag can take it with
properly designed non-interactive semantics. use_prompt follows the
rename as use_temp_role.
BREAKING CHANGE: invocations using --prompt <text> must switch to
--temp-role <text>; clap rejects the old flag loudly.
Windows review findings on the bundle provenance code:
- clones now pin core.autocrlf=false and core.eol=lf so recorded sha256
values reflect repository bytes, not the machine's git config (autocrlf
on Windows previously made every text file a false conflict on update),
plus core.longpaths=true for deep bundle trees
- is_safe_relative_path additionally rejects NTFS alternate data stream
colons, reserved device names (con, nul, COM1..), and trailing dots or
spaces; such names never come from a valid checkout and previously
desynced or failed on Windows
- file ownership dedupe compares paths case-insensitively on Windows and
macOS where case variants denote one physical file (uninstalling one
bundle could previously delete another bundle's file)
- a failed git clone no longer leaks its partial tree in the temp dir,
and temp cleanup failures are logged instead of swallowed
- recording a bundle file outside the config dir (asset dir override)
now warns instead of silently producing an undeletable record
Windows review findings: reserved device names (con, nul, COM1..) and
trailing dots in server names break or desync directory creation, so
sanitize_server now escapes reserved stems, strips trailing dots, and
caps length at 64 chars. Spill writes go through a temp file + rename
so a visible file is always complete (closes a cross-process partial
read race), and eviction protection compares content-hashed file names
instead of full paths. Also drops a duplicated cfg attribute.
The non-interactive bail under test only triggers without a TTY; from a
terminal the code correctly opens the interactive selector instead, so
the test hung or failed depending on input. Same guard as the three
sibling non-interactive tests.
MCP prompts are live, server-owned catalog entries, not managed assets;
list_prompt_assets/prompt_asset_rows implied otherwise and are now
list_mcp_prompts/mcp_prompt_rows. The .list completer was also missing
the prompts kind that the usage string and unknown-kind error advertise.
The unified install parser accepts flags in any position, so the
first-argument completion list now offers all four flags instead of
only --git-host and --help.
Update the README's MCP feature entry to cover the full capability trio
(tools, resources, prompts): the capability-gated mcp_read/mcp_prompt
meta-tools, bounded results and blob spilling, and the .prompt REPL
command with staged tab-completion and .list prompts.
Per plans/mcp-resources-prompts-design.md section 10.T9.
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.
Emit an mcp_prompt_<server> declaration for servers advertising the
prompts capability, execute prompts via McpRuntime::prompt on both tool
dispatch chains, and return the flattened prompt text as the tool
result. Sanitize server-controlled prompt names, descriptions, and
argument names before terminal rendering, and attribute the .prompt
argument inquire label to its server and prompt.
Per plans/mcp-resources-prompts-design.md §5.2 (T7).
Implements plans/mcp-resources-prompts-design.md §5.1/§5.4 (T6):
- .prompt <server> <name> [key=value ...] fetches an MCP prompt and
submits the result as chat input via Input::from_str + ask(), never
through REPL line parsing; GetPromptResult messages are flattened
into one user-role block with unconditional [user]/[assistant] labels
- missing required prompt arguments are collected interactively
- .list prompts renders server/name/description/args via the unified
catalog (CatalogItem gains an arguments field), degrading per server
- staged live tab-completion: enabled+running+prompts-capable servers
(no RPC), then live prompt names, then key= argument suggestions with
(required) markers; 2s timeout per RPC, all errors degrade to silent
empty suggestions, ctx read guard dropped before blocking
- the enabled-server alias expansion is factored into a shared helper
used by both tool-scope rebuild and completion
- BREAKING: the former .prompt <text> temp-role builtin is renamed to
.temp-role <text> (behavior preserved); .prompt now belongs to MCP
prompts, and a user macro named prompt or temp-role is shadowed
Implements plans/mcp-resources-prompts-design.md section 4.3 (T4):
mcp_read_<server> declaration and handler wired to render.rs, RFC 6570
Level-1-only URI template expansion, defensive ResourceContents parsing,
per-item text paging with pattern filtering, blob spill metadata, an
overall 204800-byte multi-content ceiling, dispatch wiring on both
eval chains, and a render_text paging-stall guard.
Per-server McpServerFeatures (tools fail-open, resources/prompts
fail-closed) now drive which meta-functions are declared, with
gated_meta_function_prefixes as the single gating seam; read/prompt
declarations land together with their handlers. The server-enablement
sentinel keys on the always-emitted search name so resources-only
servers survive role filtering.
Implements plans/mcp-resources-prompts-design.md §4.4/D7 (T5).
Single content-policy module for MCP resource and tool content, per
plans/mcp-resources-prompts-design.md §4.5 (T3):
- render_text: UTF-8-boundary-safe paging with clamped max_bytes and
grep-style fancy-regex line filtering (2 lines of context, 1-based
line-number prefixes, merged hunks); offsets walk the filtered stream.
- render_blob/render_blob_at: streaming base64 decode with a 50 MiB
ceiling, UTF-8 sniff, sha256-named 0600 spill files under a sanitized
server dir with a fixed mime->ext allowlist, and best-effort
oldest-first eviction bounding the spill tree at 512 MiB.
Not yet wired to call sites; module carries #![allow(dead_code)] until
the read/prompt surfaces land.
Implements T1 of plans/mcp-resources-prompts-design.md (§4.1, §4.6):
- Replace list_tools(None) with cursor-following list_all_tools() at the
three call sites (start_server catalog build, catalog_items, describe)
so paginating servers no longer silently lose tools past page one.
- Add MCP_READ/MCP_PROMPT prefix constants (declared nowhere yet; wired
in T4/T7) plus centralized helpers MCP_META_FUNCTION_PREFIXES,
is_mcp_meta_function, and mcp_meta_function_names.
- Mechanically replace every hand-rolled 3-prefix starts_with triple
(partition in eval_tool_calls, 3 exclusion triples in
select_enabled_functions, 3 inclusion triples + per-server name
construction in select_enabled_mcp_servers) with the helpers,
preserving the existing lax starts_with matching semantics and the
mcp_invoke_* enablement sentinel (sentinel moves to search in T5).
- Behavior-neutral: dispatch chains keep their 3 arms, emission stays
at exactly 3 meta-functions per server, existing tests unmodified.
- Add unit tests: helper classification, prefix-soundness property,
lax-matching pin, ordered candidate-name construction.
The unified install parser accepts flags in any position, so the
first-argument completion list now offers all four flags instead of
only --git-host and --help.
Macro invocations (.name and .macro name) accept leading name=value
assignments before positional args: assignments set declared variables
directly so earlier variables can keep their defaults, remaining
positionals fill unassigned variables in declaration order, and the
free text after -- is never scanned for assignments. Identifier-shaped
keys that match no declared variable error with the declared list to
catch typos; non-identifier tokens containing = stay positional.
MacroVariable gains an optional description field, and tab completion
after a macro name offers name= candidates showing each variable's
description and default until the assignment prefix ends.
.install --help and .uninstall --help print a usage guide covering the
owner/repo shorthand, --git-host, --filter, --force, ref pinning, and
the bundle lifecycle; both usage error lines now point at --help. Tab
completion offers --help for both commands and --git-host on the first
.install argument, and the unified install parser accepts flags in any
argument position so completed flags work wherever they are inserted.
The empty .list bundles message now shows the REPL install form
alongside the CLI one.
The path-escape guard that uninstall applies to recorded paths now also
covers update's obsolete-file deletion through a shared check, so a
tampered store cannot turn either delete site into an arbitrary file
removal. Updates gain a working non-interactive path: --yes now applies
to --update-bundle (locally modified files, obsolete files, and modified
mcp entries are all kept; everything else refreshes), owned mcp entries
whose recorded hash still matches the local entry take the remote side
without prompting, and the non-TTY conflict bails name the flag that
actually works per surface. An update records its new commit and version
only after files and mcp entries land, so an aborted update cannot claim
content it never wrote. The store gains a version field and rejects
stores from newer builds, the corrupt-store error no longer advises the
removal that would forfeit ownership tracking, and duplicate records
tracking one source abort a rename instead of overwriting a record.
Reinstalling from a source URL reclassifies owned unmodified files as
silent refreshes just like updates. git runs with GIT_TERMINAL_PROMPT=0
and a null stdin so private or mistyped URLs fail instead of hanging.
File comparison fills buffers fully before comparing, deleting an
obsolete file prunes emptied directories, mcp.json backfill uses the
fsynced atomic writer, --list-bundles no longer triggers builtin
backfill, bundle-name completion logs store errors instead of swallowing
them and offers --yes, and REPL .uninstall rejects unknown flags.
The user-origin marker on replaced mcp.json entries is now sticky:
re-records and cross-bundle transfers only upgrade replaced to
transferred when the prior record proves bundle origin, so updating a
bundle can no longer make uninstall delete a key the user had before the
bundle replaced it. Canonical source URLs lowercase only the host, since
self-hosted forges treat repository paths as case-sensitive and
collapsing distinct repos misdirects updates and uninstalls. git clone
invocations pass '--' before the URL so a crafted source cannot be
parsed as a git flag. Lifecycle flags (--install, --install-builtins,
--update-bundle, --uninstall) and their companions now conflict
explicitly instead of first-match dispatch silently dropping actions.
--install-from returns as a hidden tombstone that errors with the
replacement instead of feeding the flag to the LLM as prompt text.
--list-bundles dispatches before config load so a pure read no longer
boots MCP servers. write_file_atomic fsyncs before the rename so a crash
cannot persist a truncated store. REPL: .uninstall accepts --yes,
.install rejects trailing tokens after a category, and .install remote
gets a migration hint. Plus polish: host validation rejects '#' and '?',
renamed_to no longer serializes null, derived names get a debug assert
against the validator, completions share DEFAULT_GIT_HOST, README
mentions skills.
Bundle names that collide with an asset category (agents, roles, skills,
macros, functions, mcp_config) are now owner-qualified at install time,
whether derived from the repo or declared by a manifest, so no bundle can
shadow a category by name. A manifest name that collides with a bundle
from a different source now prompts for confirmation interactively (a
fork or typo-squat is the likely cause); declining aborts before anything
is written, and non-interactive runs keep the deterministic
owner-qualification. Uninstall summaries now list the vault secrets the
bundle's MCP servers reference, noting they are installed by the bundle
but not removed. Also removes the dead ResolvedBundleName.migrated_from
field.
--install someuser/repo expands to https://github.com/someuser/repo;
--git-host overrides the default host and forces source interpretation
even when the value matches an installed bundle name. Two or more path
segments are accepted so nested GitLab-style groups work, and #ref
pinning applies to shorthand values. --uninstall resolves owner/repo
against recorded sources: a single match uninstalls, multiple matches
prompt an interactive selection showing each bundle's source, and
non-interactive runs bail instead of guessing.
Function docs that restated behavior already evident from names,
signatures, and code are removed; only comments carrying invariants
the code cannot express remain.
--install <git-url|name> is the single entry point for remote installs
and updates; the unified .install dispatch likewise replaces
.install remote. Flag completion for .install now applies to the
unified form.
Provenance records stored OS-native path separators, making
installed-bundles.yaml non-portable; slug derivation treated a Windows
drive letter as an scp host and swallowed the whole path into one
sanitized segment. Store paths are now always forward-slashed and
backslashes normalize before URL parsing. Test fixture repos commit a
'* -text' .gitattributes so clone-side autocrlf cannot rewrite content
assertions.
--install now takes a git URL or an installed bundle name: categories
are redirected to the new --install-builtins, installed names become
implicit updates, and source-shaped values install remotely. The old
--install-from keeps its exact behavior as a hidden deprecated alias.
The REPL's .install gains the same unified dispatch while keeping
.install <category> and .install remote <url> back-compat.
Updates re-clone a bundle's recorded source (honoring a recorded commit
pin unless a #<ref> override moves it), silently refresh files the bundle
owns that the user never modified, and fall back to the normal conflict
prompts for modified or unowned files. Files the remote no longer ships
are offered for deletion (kept by default non-interactively, staying
owned). The record is refreshed with the new commit, version, and
metadata, and stamped with an updated_at timestamp on success.
The with_macro_dirs fixture derived its temp-dir name from a wall-clock
nanosecond timestamp, so parallel tests starting in the same clock tick
shared a directory and saw each other's macro files (flaky on CI
runners with coarse tick granularity). A process id + atomic counter
makes the name unique by construction.
An invalid installed macro invoked as a top-level command fell through
to the generic unknown-command error, while .macro <name> reported the
parse/validation failure. Both paths now surface the reason.
Long agent/skill descriptions wrapped badly in the bullet-list format.
Extract a shared asset_table helper (UTF8_FULL + dynamic arrangement,
same style as the markdown renderer and .list macros) and use it for
the agents, skills, and macros listings. The skills loaded marker
keeps its color; comfy-table's custom_styling feature accounts for
ANSI sequences in column widths.
Dark-Alex-17
