fix(cli): preserve runtime and delegation observations

README.md

@@ -57,14 +57,22 @@ Configure: `set-mode` (`build|plan|review`), `set-model`, `set-effort` (`auto|lo
 
 Work: `prompt`, `follow`, `watch`, `steer`, `queue`, `discard-queued`, `exec`, `cancel`, `runtime`
 
-`prompt --new TEXT` creates a session. `prompt SESSION_ID TEXT` continues one. `-` reads the prompt from stdin. `--delivery auto` prompts when idle, steers when the turn is steerable, otherwise queues. It waits until idle, including queued turns.
+`prompt --new TEXT` creates a session. `prompt SESSION_ID TEXT` continues one. `-` reads the prompt from stdin. `--delivery auto` prompts when idle, steers when the turn is steerable, otherwise queues. Auto delivery and idle waits require the updated backend observation metadata; unknown/unavailable status is not usable.
+
+Prompt/follow/watch wait for self-session idle, including queued turns, not for
+child workflows or detached processes. Idle is not task success. Idle results
+retain runtime observation health and separate direct delegation activity.
 
 `queue SESSION_ID -` and `steer SESSION_ID -` read stdin and return a submission
 receipt without loading history. `prompt --delivery queue|steer --until-idle false`
 also returns after submission; idle/direct prompt delivery still waits for its turn.
 
 `discard-queued` takes the `inputId` from an `input_state` event. That is the same value as `clientInputId` from the submit response.
 
+`cancel` sends a cancellation notification for the current turn. Its
+`cancellationState: "notification_sent"` confirms the local write/flush, not run
+termination or task success. It does not discard queued inputs or cancel children.
+
 `find` reports `pagesScanned`, `truncated`, `pageCapReached`, and `stopReason`
 so callers can distinguish a complete search from a result/page limit.
 
@@ -102,7 +110,11 @@ a final `type: done` result. `watch` also emits `runtime` events.
 
 Exit codes: 0 ok, 1 usage, 2 connection, 3 RPC, 4 timeout or cancel.
 
-Event types: `session`, `submitted`, `text`, `tool`, `mode`, `plan`, `input_state`, `permission`, `elicitation`, `runtime`, `update`, `done`.
+Event types: `session`, `submitted`, `text`, `tool`, `delegation`, `mode`, `plan`, `input_state`, `permission`, `elicitation`, `runtime`, `update`, `done`.
+
+`delegation` records describe workflow lifecycle separately from dispatch `tool`
+status. Compact `inspect` includes snapshot `delegations`; successful dispatch
+alone does not establish delegation success.
 
 Submission receipts retain `clientInputId`; partial-failure errors include known
 IDs and delivery state in `recovery`. Acceptance is not task completion. See

skills/qmacp/SKILL.md

@@ -28,7 +28,9 @@ keep the returned `sessionId`. Default endpoint: `ws://127.0.0.1:3030/acp/ws`.
   `--permission cancel` to fail closed. Elicitation is always cancelled.
 - JSON/NDJSON goes to stdout; connection logs go to stderr. Require exit 0 and
   JSON `ok` not false; `prompt`/`follow` also need a terminal `"type":"done"`.
-  Do not use `--pretty` on streams. Acceptance is not proof of task completion.
+  Do not use `--pretty` on streams. Acceptance and self idle are not task success.
+  `cancel` with `cancellationState: "notification_sent"` confirms only sending the
+  notification, not run termination.
 - After a timeout or uncertain submission, inspect `runtime` and history before
   retrying, steering, queuing, discarding, or cancelling. Do not claim changes
   succeeded without exit 0 and JSON `ok: true`. Retain receipts and error
@@ -52,7 +54,9 @@ qmacp --quiet runtime "$SESSION_ID"
 
 Long prompts: use stdin (`-`), not nested quoting. Make them self-contained.
 Existing-session `prompt` defaults to `--delivery auto`: idle -> prompt,
-busy and steerable -> steer, otherwise queue. It waits until idle by default.
+busy and steerable -> steer, otherwise queue. Auto delivery and idle waits require
+the updated backend observation contract; unknown/unavailable is not a usable
+phase. Waiting means self-session idle, not descendant/process completion.
 Read [live control](control.md) for delivery, waiting, cancellation, and how to
 label agent-to-agent messages.
 

skills/qmacp/callbacks.md

@@ -108,9 +108,10 @@ notify_callback() {
   steering; `steer --run-id` can target an exact active run. Do not assume
   queue-only delivery wakes an idle session. See [live control](control.md)
   for direct commands, including stdin delivery without history loading.
-- `--until-idle false` avoids waiting for all subsequent work after
-  steer/queue. It is **not** fire-and-forget for an idle session: the prompt
-  RPC still waits for that turn, bounded here by `--timeout 900`. Keep the
+- `--until-idle false` skips the self-session idle wait after steer/queue.
+  That wait never implies descendant workflow or detached process completion.
+  It is **not** fire-and-forget for an idle session: the prompt RPC still waits
+  for that turn, bounded here by `--timeout 900`. Keep the
   delivery process supervised.
 - Keep stdout NDJSON and stderr in the watcher logs; check exit status and the
   final `done` event. Submission/`done` alone does not prove the agent handled

skills/qmacp/cli.md

@@ -39,29 +39,50 @@ qmacp [--url URL] [--allow-insecure] [--pretty] [--permission allow-once|reject-
 - `prompt`, `follow`, `exec`, and opt-in `new --stream`: NDJSON.
   `watch` emits `runtime` events, then a final JSON object.
   Connection logs go to stderr.
-- Success requires exit 0 and JSON `ok` not false. `prompt`, `follow`, and
+- Command success requires exit 0 and JSON `ok` not false. `prompt`, `follow`, and
   `new --stream` also require a terminal `"type":"done"`. `exec` emits individual
   command results, not a global `done`.
 - Exit codes: 0 success, 1 usage, 2 connection, 3 RPC, 4 timeout/cancel.
 
-Stream event types: `session`, `submitted`, `text`, `tool`, `mode`, `plan`,
-`input_state`, `permission`, `elicitation`, `runtime`, `update`, `done`.
-Do not claim a cancel/discard/close/config change succeeded without exit 0 and
-JSON `ok: true`. Inspect current state after uncertain delivery rather than retrying
-blindly; see [live control](control.md).
+Stream event types: `session`, `submitted`, `text`, `tool`, `delegation`, `mode`,
+`plan`, `input_state`, `permission`, `elicitation`, `runtime`, `update`, `done`.
+`delegation` is a separate lifecycle record from `querymt/session/delegationUpdate`,
+also delivered by the exported `AcpClient::connect_with_events` channel. It retains
+backend identities, state, child/tool links, stage/activity, outcome details and
+`sourceEpoch`/`sessionSeq` when supplied, without inferred defaults. A completed
+`tool` can mean only successful dispatch while its delegation is still verifying
+or later fails. Delegation ordering is separate from self-runtime revisions;
+records are forwarded, not merged into a client-side state cache.
+
+Require exit 0 and JSON `ok: true` for command success. For `cancel`,
+`cancellationState: "notification_sent"` confirms only the local notification
+write/flush, not backend acceptance or execution termination. Inspect current
+state after uncertain delivery rather than retrying blindly; see [live control](control.md).
 
 `new --stream`/`prompt --new` emit `sessionId` before configuration. Default
 `new` still returns one object. Queue/steer prompts emit `submitted` with
 `clientInputId` and `receipt` before waiting; `done` retains them.
 
 On failure, `recovery` preserves known IDs, the failing `stage`
-(`create|configure|submit|observe`), and creation/delivery state. `not_submitted`
+(`create|configure|submit|open|observe`), and creation/delivery state. `not_submitted`
 or `rejected` differs from `unknown`; `accepted` means a receipt was received,
 not that the task completed. Inspect state before retrying an uncertain or
 accepted submission. A lost creation response cannot provide a session ID.
 
-`observationState` is `not_waited`, `idle`, or `prompt_returned`, not proof of
-task success. `text` contains only assistant text observed on this connection.
+Successful prompt `observationState` is `not_waited`, `idle`, or `prompt_returned`;
+follow reports `idle`. None proves task success. Idle results retain `runtime`,
+including backend observation health, run identity and direct delegation activity.
+Self idle does not mean descendants finished or detached OS processes exited.
+The prompt RPC `stopReason` is separate from idle evidence; it is also retained in
+recovery when observation fails after the RPC returned. It does not describe a
+later queued run. No terminal outcome is inferred from idle.
+
+Open/observation failures emit no successful `done`; recovery reports
+`observationState: "unavailable"` with the session ID and runtime evidence when
+available. A retained runtime may be the last known observation, not current
+liveness. Updated backend observation metadata is required for auto delivery and
+waits; absent/unknown/unavailable health is not a usable busy or idle state.
+`text` contains only assistant text observed on this connection.
 
 ## Permissions
 

skills/qmacp/control.md

@@ -26,10 +26,12 @@ self-contained; elicitation is always cancelled. Existing sessions accept
 - `queue`: submit input for after the current turn. Whether it starts an idle
   session depends on server support; use `auto` when that support is unknown.
 
-`--until-idle` defaults true and waits through queued turns. Pass `--timeout`
-explicitly for predictable bounds. `--until-idle false` skips waiting for idle
-following steer/queue; it is not fire-and-forget when idle delivery uses the
-prompt RPC, which still waits for that turn.
+`--until-idle` defaults true and waits for this session's own idle state, including
+queued turns. It does not wait for descendant workflows or detached OS processes,
+and idle is not task success. Pass `--timeout` explicitly for predictable bounds.
+`--until-idle false` skips waiting for idle following steer/queue; it is not
+fire-and-forget when idle delivery uses the prompt RPC, which still waits for
+that turn.
 
 Queue/steer delivery skips history loading. Direct commands return a receipt;
 `prompt --until-idle false` returns after submission when either is selected.
@@ -62,21 +64,33 @@ qmacp --quiet cancel "$SESSION_ID"
 qmacp --quiet close "$SESSION_ID"
 ```
 
-- `follow` streams session events until idle and loads cwd from the current
-  process; run it from the session workspace.
-- `watch` polls `runtime` (`"type":"runtime"`) until idle unless `--idle false`.
-  Default `follow`/`watch` timeout is 120s.
+- `follow` requires successful session open, then streams events until self idle;
+  run it from the session workspace (cwd comes from the current process).
+- `watch` polls `runtime` (`"type":"runtime"`) until self idle unless `--idle false`.
+  Default `follow`/`watch` timeout is 120s. Unknown/unavailable observation fails
+  rather than completing either command, including nonwaiting watch.
+- Auto delivery and waits require the updated backend `observation.availability`
+  contract (`live` or known `inactive`); a bare phase is not usable evidence.
+  `runtime.delegation_activity` is separate direct-workflow activity and may remain
+  live after self idle. Prompt/follow idle results retain the full `runtime`.
+  Live `delegation` events report workflow lifecycle separately from dispatch
+  `tool` events; neither changes follow's self-idle completion scope.
 - `steer --run-id` is optional; if omitted, qmacp pins `runtime.active_run_id`.
   Missing/empty run IDs fail before submission rather than steering unguarded.
 - Queued submit returns `clientInputId`. `input_state` events expose the same ID
   as `inputId`; pass that exact value to `discard-queued`.
+- `cancel` returns `cancellationState: "notification_sent"` after the local
+  WebSocket write/flush. It does not confirm backend acceptance or run termination,
+  discard queued inputs, or cancel child workflows. Inspect `runtime` separately;
+  a queued next run can keep the session busy.
 - Do not close a session unless the user asked to end it or a qmtcode test
   explicitly requires `close`.
 
 After a timeout, inspect `runtime` and session history before resending or
 changing work. Do not steer, queue, discard, or cancel just because time ran out.
 Require exit 0 and JSON `ok` not false; prompt/follow also need a terminal
-`"type":"done"`. Require JSON `ok: true` for cancel/discard/close. A submission
+`"type":"done"`. For cancel, exit 0 and JSON `ok: true` prove only notification
+sending; for discard/close they confirm the respective operation. A submission
 receipt is not proof the agent completed the task.
 
 For event-driven waiting outside the agent turn, see [callbacks](callbacks.md).

skills/qmacp/sessions.md

@@ -36,6 +36,10 @@ creation receipts, opt in to `new --stream`; see [CLI output](cli.md#output).
 
 `find` reports `pagesScanned`, `truncated`, `pageCapReached`, and `stopReason`
 (`exhausted|limit|page_cap`). A truncated empty result is not exhaustive.
+Summaries retain full backend `runtime` observation/activity metadata. Missing
+phase stays null; unknown/unavailable or unobserved status never becomes idle.
+`find --phase idle` matches only usable self-idle observations, even when managed
+dependency activity remains live. It does not test task or descendant completion.
 
 ## Inspect and configure
 
@@ -52,7 +56,12 @@ qmacp --quiet set-effort "$SESSION_ID" high    # auto|low|medium|high|max
 ```
 
 `status` is mode/model/profile/effort; `runtime` is phase/queue/steerability.
-`inspect --tools` defaults true; use `--tools false` for compact history.
+Compact `inspect` includes `delegations` from snapshot `delegationUpdates`, separate
+from dispatch `tools`, using the same records as live `delegation` events.
+`inspect --tools` defaults true; `--tools false` hides only tools, not delegations.
+Snapshot records are point observations; a later live record with the same
+`sourceEpoch` and higher `sessionSeq` is newer. Historical nonterminal rows do not
+prove live activity, and a terminal workflow record does not prove OS process exit.
 `--full` dumps the raw snapshot (large). Do not treat earlier output as current
 state; query again before acting. Require exit 0 and JSON `ok: true` for changes.
 

skills/qmacp/testing.md

@@ -27,7 +27,11 @@ requirements; do not alter unrelated sessions or user-run services.
 - Load/list features: if `caps.acp.loadSession`/`list`/`close` are false, do not keep
   calling `inspect`/`sessions`/`close`.
 
-Query `runtime` or `status` again rather than treating earlier output as current.
-Do not claim cancel/discard/close/config success without exit 0 and JSON `ok: true`.
+Query `runtime` again for execution state rather than treating earlier output as
+current; `status` is configuration only. Auto delivery and waits require updated
+backend observation metadata. Unknown/unavailable is not idle or task success.
+Require exit 0 and JSON `ok: true` for command success. Cancel's
+`cancellationState: "notification_sent"` proves only notification write/flush;
+check runtime/outcome evidence separately for termination.
 Prompt/follow checks also require a terminal `"type":"done"`, but a receipt or idle
 state alone is not semantic proof of the task result. Inspect the relevant result.

src/client.rs

@@ -585,6 +585,14 @@ async fn handle_inbound(
         event_sink.emit(event).await;
     }
 
+    if matches!(
+        method.as_str(),
+        "querymt/session/delegationUpdate" | "_querymt/session/delegationUpdate"
+    ) && let Some(event) = events::compact_delegation_update(&params)
+    {
+        event_sink.emit(event).await;
+    }
+
     if let Some(id) = value.get("id").cloned() {
         let (reply, event) = inbound_request_reply(id, &method, &params, permission)?;
         if let Some(event) = event {