xdog-eyes

Platform scope: Linux only. Non-Linux adapters and native-platform support are outside the current product scope; historical audit evidence is not a support promise.

A local-first, declarative observation layer, shaped like flow: versioned JSON, frozen models, whole-definition validation, async execution, CLI and a packaged eyes skill. Sources collect evidence, detectors emit events, explicit EventRouter dispatches explicit routes to extensible EventHandler implementations. ACP is one built-in handler, not the routing abstraction. Eyes never imports a harness or runs an agent loop; coding never imports a concrete Eyes source.

The supported management surfaces are the Eyes CLI and the installed, registered Eyes tool. Python engine classes such as ObserverService, handler registries and hosting/control modules are internal implementation details, not a public embedding SDK or an alternative supported observation-management entry point.

Use an explicit ACP destination

An independently managed receiver and an existing session ID are prerequisites. Copy the packaged source-test-review.json into your project root and replace its Unix endpoint and EXPLICIT_SESSION_ID. Roots resolve relative to that file. The v2 example uses delivery: "enqueue" to retain evidence while TUI/session owners are offline. It negotiates durable inbox support, never loads an owner or starts a model. If the receiver is also offline, report unavailable—not receipt. Eyes never launches a coding CLI to provide a callback.

xdog-eyes daemon start
xdog-eyes daemon status
xdog-eyes validate eyes.json
xdog-eyes start eyes.json
xdog-eyes list  # alias for status without an ID, scoped to owner
xdog-eyes status REGISTRATION_ID
xdog-eyes resume REGISTRATION_ID
xdog-eyes pause REGISTRATION_ID
xdog-eyes stop REGISTRATION_ID  # cancel source, not shared daemon
xdog-eyes daemon stop         # explicit whole-instance shutdown

validate needs no harness/model/active SDK. start explicitly opens the reusable same-user daemon; handlers mode connects its receivers before scanning. The installed Eyes tool invokes the public CLI using the same Python environment. Only the CLI uses the private, versioned, generation-fenced Unix control protocol; there is no public Python daemon SDK. The default private directory is $XDG_DATA_HOME/xdog/eyes (otherwise ~/.local/share/xdog/eyes). Use --directory /private/path or XDOG_EYES_RUNTIME_DIR for isolation; an explicit directory takes precedence. Directories/sockets/journals are UID-owned and private. No service is installed; importing/listing skills does not start a daemon. There is no daemon idle timeout.

Owner/key plus normalized definition is idempotent across CLI processes/tools, host retirement and daemon restart. Within one daemon instance, owner is the only management namespace; --owner defaults to cli. There is no management workspace parameter or --workspace option. Changing the CLI current directory does not select another inventory. Keys default to definition names; different content requires a new key. Cancelled keys are tombstones, not silently reusable. Source/route/connection status is separate. Queued or duplicate enqueue receipts are not model completion. The receiver's TUI shows pending evidence; an activated owner consumes it automatically. /inbox list and /inbox show ID are read-only, /inbox cancel ID targets pending intake, and /inbox retry ID explicitly retries interrupted work. There is no message approval or human wake-up budget. Actual tool permissions remain unchanged.

delivery: "prompt" instead loads a live explicit session and waits for its final ACP result. It requires an explicit handler cwd for standard ACP load_session. A coding target persists live prompts through the same Inbox and waits for their actual batch result. Inactive prompt targets still fail rather than silently enqueueing offline. Generic ACP servers keep their own policy. A lost prompt response is outcome-unknown and is never blindly resent. Enqueue can explicitly retry the original ID/content after checking the same persistent inbox namespace/project/branch; reconnect never approves model execution.

Definition paths and ACP target directories

Eyes does not manage a workspace. Keep these distinct:

For delivery: "enqueue", session identifies an existing receiver session. Handler cwd may be omitted: the receiver resolves the session's directory. If supplied, cwd is an assertion checked against that session, not a request to move it. The client pins the returned directory, inbox namespace and branch for later delivery/recovery; it never substitutes a definition base or daemon cwd. For delivery: "prompt", an explicit handler cwd is required for the standard ACP load_session protocol. Neither mode creates a session from a directory.

A stdio connection has only transport: "stdio" and command (an argv array). There is no connection.cwd field. Handler cwd is the target session directory, not the subprocess startup directory; it is never passed as the stdio spawn cwd. Use an absolute executable/script path when startup-directory independence is needed. Unix connection.path remains an explicit absolute socket path.

Router, handlers and destination batches

Internally, HandlerRegistry binds each type's definition parser and factory. The internal EventHandler lifecycle is start, validate, deliver, snapshot, check, resume and close; ACP payload/connection details stay in its handler. These engine contracts are not public embedding or handler-registration SDK entry points. Unknown types fail validation rather than importing arbitrary code from JSON. The supported daemon handlers are built-in ACP and data-only events; the daemon never imports code named by a definition.

Coding selects fixed batches with max_messages_per_turn (default10, strict integer1..128). Five queued messages remain five independent history inputs with source/evidence/IDs/attachments, not concatenated text or five separate Agent turns. They share a response/tool stream; new arrivals wait for a later turn. A canceled peer cannot withdraw its message from an already running shared batch or cancel other members. Tools still require the target's normal permissions; unattended restricted tools fail closed without explicit allow-all.

Scheduler sources

Sources can also generate timed observations: one-off at, fixed every_seconds, or five-field cron with an explicit IANA timezone. Exactly one mode is allowed per source. scheduler.due detectors emit ordinary scheduler.due events through the same EventRouter/handlers; no agent or system cron is called directly. File and scheduler sources can coexist in one v2 file.

{"id": "daily", "type": "scheduler", "schedule": {
  "cron": "0 9 * * *", "timezone": "Asia/Shanghai"
}}

The packaged scheduled-review.json is a complete executable definition. Replace its endpoint/session and use the same validate/start/status/stop commands. Repeating schedules start after handler readiness; one-off start refuses past zoned timestamps. Intervals use monotonic time (finite seconds, minimum0.001), cron uses timezone-aware calendar instants, and suspended periods do not generate a catch-up burst. A late active poll coalesces to its latest occurrence. Scheduled evidence keeps occurrence identity and planned/observed UTC times, not file paths. See scheduler definition, timing and recovery for all modes, DST/clock semantics, mixed-source examples and bounds.

Collection and lifetime

Regular-file content changes use bounded polling, fingerprints and net-change/debounce detection. No symlink following, mtime-only/unchanged-save events or historical file contents. Rename is delete+create; polling can miss transient changes. Keep roots/globs narrow and exclude generated/cache trees. Payloads, source labels and filenames are untrusted evidence, never permissions.

The daemon journals registrations, file fingerprints/debounce state, scheduler cursors, immutable pending deliveries, event rows and consumer ACK cursors in daemon.sqlite3. Cursor and pending evidence commit together before handler send. Accepted registrations with recover=True resume from their original IDs and baselines, not a fresh scan that discards downtime changes. No historical file contents are retained. Registration pause skips scheduled deadlines rather than replaying missed ticks; accepted pending evidence remains intact. Cancel is idempotent and does not delete already committed events or destination inputs.

External in-flight/unknown delivery after interruption is quarantined for review, never automatically replayed. The original ACP definition and pinned inbox/branch or owner-generation binding must still match on recovery. Proven-not-sent delivery may retry; a changed receiver cannot acquire an old outbox. Events-mode publication is idempotent by original delivery ID and may be recovered safely. Whole-instance daemon stop settles work and releases only its own socket/record, retaining data. daemon status is read-only and never starts an absent instance.

CLI event consumers (including Goal)

Goal uses a real Eyes CLI subprocess for every registration, event read, ACK and pause/resume/cancel operation. It imports no Eyes module and does not construct a Python control object. A Python host invokes the installed command with an argv array, e.g. [sys.executable, "-m", "xdog.eyes", ...], never a shell command. xdog.eyes._daemon_control.DaemonControl is private implementation for Eyes' CLI internals only, not a supported SDK or host integration entry point. The former xdog.eyes.client, Client and EyesClient are removed; there is deliberately no compatibility export.

# Each command uses the same explicit scope; global options also work after commands.
xdog-eyes --directory /private/eyes --owner goal:stable \
  start - --delivery events --key tasks --no-wait < tasks.json
xdog-eyes --directory /private/eyes --owner goal:stable \
  status tasks --wait-ready --timeout 10
xdog-eyes --directory /private/eyes --owner goal:stable \
  events --follow --after 0 --limit 64
# After Goal runtime has durably handled the original event:
xdog-eyes --directory /private/eyes --owner goal:stable ack SEQUENCE
xdog-eyes --directory /private/eyes --owner goal:stable pause tasks

Use a definition handler {"id": "changed", "type": "events"} with its explicit route and --delivery events. Events mode never creates an ACP handler or wakes an Agent. Goal receives raw event JSON through CLI stdout, handles it inside Goal runtime and decides any later reminder under host policy. Claw does not invoke this CLI itself: it loads the installed Eyes registered tool; the registered tool itself invokes the public CLI, keeping daemon-control details inside Eyes.

Machine-readable command contract

Each event object has sequence, key, registration_id and the full original Delivery JSON in delivery. --limit is a per-batch integer 1..128, not a total stream limit. An omitted --after starts after this owner's persistent ACK cursor; explicit --after 0 re-reads retained history. Follow advances only its local read cursor and never automatically ACKs. ACK is cumulative within the scope and idempotent; persist consumer effects before ACK and deduplicate original delivery IDs across the effect/ACK crash window. Unacknowledged events provide backpressure. No Python callable crosses this boundary.

start - reads one UTF-8 JSON object from stdin until EOF, bounded to 1,044,480 bytes; duplicate keys, non-finite values, multiple documents and oversized input are rejected. For start -, relative source roots and any explicit relative handler cwd resolve against the CLI process current directory, not the shell redirection file's directory or daemon startup directory. For start FILE (and validate FILE), they resolve against the definition file's parent directory. Use absolute roots when a stdin producer cannot guarantee the CLI current directory. --no-wait returns the registration receipt without waiting for a baseline; it does not resume an explicitly paused registration. Resume explicitly, then use status ID --wait-ready --timeout 10. The wait is read-only and never registers or resumes anything. Optional --source-checkpoints-json JSON_OBJECT supplies definition-bound scheduler cursor mapping for already accepted source recovery, not executable code or a clock.

Commands exit 0 on success; errors exit 2 with a bounded diagnostic only on stderr. A timed-out wait emits no success JSON. JSON output frames are bounded to one MiB; follow writes with pipe backpressure, not an unbounded output queue. Closing stdin of a follow process is harmless; closing stdout, terminating it or reaching a broken pipe only detaches the consumer. It does not ACK, pause/cancel observation, stop the shared daemon or roll back an accepted registration. A lost mutation response remains uncertain: inspect/reconcile its original scope/key/ID rather than automatically retrying a different mutation.

Follow reconnects read-only after socket EOF, disconnection or generation change. It waits for an absent daemon to be explicitly restarted; it never starts one, re-registers, resumes or ACKs during reconnection. Nonempty malformed protocol frames remain errors, not a reason to retry indefinitely. Host retirement closes its stream only; Goal pause explicitly runs pause, and Goal completion/clear can explicitly run cancel. Cancelled/unknown work is not blindly replayed. Scope strings isolate accidental cross-owner access within a UID, not malicious code with that UID's authority. Event history has no automatic retention/compaction.

The old event-handler host/capability/descriptor protocol is removed. Core local permissions are not an OS sandbox. No TCP/HTTP/Windows transport, installed system daemon, hosted broker, automatic coding startup, API/web source, script handler, or harness-specific handler is added to Eyes. Claw can expose its own ACP backend through its gateway, without an Eyes dependency or a special router path; see Claw/Eyes ACP for group:<id> addressing and setup. Stdio is an optional generic ACP transport; Unix is the XDOG-local transport for coding and Claw callbacks.

Installation and development

The runtime dependencies are the official agent-client-protocol SDK, pinned >=0.12.1,<0.13 for Python 3.12–3.14, and croniter>=2.0,<7 for bounded cron validation/calendar calculations. Eyes does not depend on ai/agent/coding/claw/flow. The marker excludes the currently unsupported ACP SDK on Python 3.15+; definition validation remains usable. Workspace development commands run from repo root.

See ACP/session/inbox protocol for explicit receiver startup, all commands, permission/unknown-result semantics and limits, and current verification for release evidence. The older eyes-verification.md is a historical checkpoint. uv build --package xdog-eyes ships skill, example, py.typed and license. This feature does not push/publish or provision a PyPI publisher/system service.

Installed model tool

Installed metadata registers xdog.eyes.capability:provider. Discovery/import is inert. Opening the tool binding only ensures the shared daemon is running; it does not open a host journal, subscribe to application events, capture replies, recover host observers or start a model. There is no ManagedObservers runtime.

Choose a stable owner, save a full Eyes v2 JSON, then call eyes(action="create", owner="my-watches", definition_path="/project/eyes.json"). definition_path must be an absolute path to an already saved regular, non-symlink JSON file (maximum one MiB); relative paths are not accepted. Source roots and any explicit relative handler cwd resolve beside that JSON file. The caller owns the naming scheme; Eyes never derives owner from a host/session/group. JSON handlers/routes select all destinations, including other endpoints/sessions. The current host address is convenience metadata, not an allowlist. Normal receiver authentication and protocol checks remain; the tool does not restrict targets to the current host/group.

The registered tool returns a CLI envelope {exit_code, stdout, stderr}. The CLI's JSON result remains a string in stdout; check exit_code before decoding it. Nonzero exits preserve their diagnostics; uncertain or cancelled mutations are not automatically retried.

The daemon executes handlers directly, like CLI start. Tool status is source and handler-delivery state only. There is no model execution/report watcher or chat notification policy. Supported actions: create/list/status/pause/ resume/cancel. cancel_notifications is removed. Tool/CLI controls use the same registration IDs and owner namespace; closing a binding only detaches. Old managed host journals are left untouched, not imported or migrated.

The provider remains application-scoped. Coding and Claw both load that scope through normal installed-capability discovery; no Eyes-specific host adapter or enable switch is required. The package must already be installed. All model tool actions require caller-supplied owner; there is no context action. The shared skill contains no application-specific naming or routing policy.

Capability contributions and readiness

skill_names=("eyes",) advertises a packaged skill, not a default body or tool permission. SkillManager discovers it and explicit activation loads its body; Flow follows the same policy. Capabilities can contribute skills, tools, runtime resources or extension hooks without supplying all four. Startup and close belong to their real scope owner; no capability creates a second Agent execution loop.

The daemon control endpoint becomes available before slow handler recovery. Its live protocol/features are negotiated separately from the on-disk journal version. An incompatible process/state is reported rather than automatically replaced, migrated or replayed. Closing/cancelling a model tool invocation settles its owned CLI process and IO; it does not stop the shared daemon, retract an accepted registration or blindly retry a request whose outcome is uncertain.