delegation-layer 0.1.20 → 0.1.22

package/README.md

@@ -20,6 +20,8 @@ already use instead of replacing them.
   and time budget.
 - **Recoverable results** — a later `collect` call can read a finished task
   without launching the provider again.
+- **Clarify while work runs** — answer native questions through Delegate; steer
+  an active Codex turn without restarting it. [Interactive tasks](docs/cli-reference.md#interactive-tasks).
 - **Resumable turns** — a durable timeout can continue the exact provider
   session through a linked successor task.
 - **Runtime compatibility** — the selected provider's executable and advertised

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "delegation-layer",
-  "version": "0.1.20",
+  "version": "0.1.22",
   "description": "Multi-provider AI agent delegation CLI and Agent Skill for running bounded coding tasks across provider CLIs.",
   "keywords": [
     "ai-agents",

package/plugins/delegation-layer/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "delegation-layer",
   "displayName": "Delegation Layer",
-  "version": "0.1.20",
+  "version": "0.1.22",
   "description": "Consult external AI harnesses and CCR-configured models for reviews, research, implementation, and debate through the resumable Delegate CLI.",
   "author": {
     "name": "Hesham Karam",

package/plugins/delegation-layer/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "delegation-layer",
-  "version": "0.1.20",
+  "version": "0.1.22",
   "description": "Consult external AI harnesses and CCR-configured models for reviews, research, implementation, and debate through the resumable Delegate CLI.",
   "author": {
     "name": "Hesham Karam",

package/plugins/delegation-layer/skills/agent-integration/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: agent-integration
-description: Coordinate external agent, model, and separate harness tasks through Delegate, including research, implementation, tests, reviews, consultations, comparisons, and debates. Use before choosing an external integration; ordinary native subagents and questions about tools remain local.
+description: Use Delegate for work assigned to another agent, model, or separate harness CLI session. Load this before a provider-specific skill or plugin, including when a user names a CLI for research, implementation, tests, reviews, consultations, comparisons, or debates. A named destination still routes through Delegate. Ordinary same-harness subagents and questions about tools remain local; respect an explicit choice of another integration.
 ---
 
 # Delegation Layer
@@ -13,6 +13,8 @@ external harness to do work. Requests such as “ask Codex to review this”,
 collaboration even when the user never mentions Delegate. The same applies
 when you decide external expertise is useful to fulfill the user's task.
 These names are examples; discover supported destinations through the CLI.
+A destination name identifies who should do the work; it does not by itself
+select a provider-specific skill, plugin, or direct CLI integration.
 “Claude Code Router” and “CCR” mean `ccr:print`, not ordinary Claude Code.
 For that request, use Delegate with a CCR model alias; a Claude-specific skill
 or subscription model does not fulfill the requested route.
@@ -36,8 +38,10 @@ not turn that request into an environment-repair task. When a response says
 `failure.next_action: "stop_and_report"`, that attempt is finished. Include its
 failure in your final response without opening a clarification dialog to choose
 a fallback. For `task_record: not_created`, identify the requested destination
-and quote the returned `task_id` (if present), `task_record`, `failure`, and
-`error`. Do not append inferred state paths or provider-setup diagnoses. Do not run diagnostic commands, retry it, or use another integration.
+and include the saved JSON failure response verbatim. A `task_id` can be returned
+even when no task record was created. Do not infer that an unavailable provider
+is uninstalled or unauthenticated. Do not append inferred state paths or
+provider-setup diagnoses, run diagnostic commands, retry it, or use another integration.
 Troubleshooting requires a separate user request. Other independently requested
 tasks may proceed; the failed attempt must remain reported as failed or blocked.
 
@@ -64,26 +68,38 @@ tasks may proceed; the failed attempt must remain reported as failed or blocked.
    Delegate owns the supervisor lifecycle. Never install, configure, start,
    inspect, or repair Pueue, and never poll a task unless the response says its
    `task_record` is `created`.
-5. Observe with `status` and collect with `collect`. Keep collection in the
+5. Observe tasks only through Delegate `status`, `collect`, or `events`.
+   Do not run `ps`, `pgrep`, process-tree searches, or supervisor commands to
+   monitor delegated work. Pending output is normal startup or work in progress;
+   empty `pending_requests` means keep collecting, not diagnose the environment.
+   Environment diagnostics require a separately requested troubleshooting task.
+   Keep collection in the
    foreground until all requested tasks reach a terminal result or an explicit
    blocker. Do not end a headless turn promising a background notification.
-   Pending publication (exit 3) means keep waiting, not success. Only detach
+   Pending publication (exit 3) means keep waiting, not success. When the JSON
+   includes an outstanding `interaction.pending_requests` batch, follow
+   [active-task interaction](references/interaction.md) before waiting again. Only detach
    when the user explicitly requests it, retaining the exact task IDs.
    Collection may recover the
    saved supervisor only to observe an already-requested budget stop; it never
    launches a provider or retries a task.
-6. If a task is timed out and its JSON says `continuation.resumable: true`, use
+6. If `failure.next_action` is `recheck`, follow
+   [offline recovery](references/result-handling.md#offline-recovery) once for
+   that task. Rechecking saved evidence does not run the provider again.
+   Otherwise, report a rejected result without redispatching or reading raw
+   output as an answer.
+7. If a task is timed out and its JSON says `continuation.resumable: true`, use
    `continue --task TASK_ID`, optionally with a new `--brief`. Do not create a
    fresh unrelated dispatch to replace a resumable task.
-7. Treat `blocked` authentication as unresolved. Do not call it compatible,
+8. Treat `blocked` authentication as unresolved. Do not call it compatible,
    successful, or failed provider behavior, and do not invent login commands.
-8. Never put credentials, tokens, or unrestricted provider output in a brief,
+9. Never put credentials, tokens, or unrestricted provider output in a brief,
    option, receipt, or agent message.
-9. Permission modes select provider-native behavior, not an independent sandbox.
+10. Permission modes select provider-native behavior, not an independent sandbox.
    Providers load their own configuration and extensions. If a task is rejected,
    report delegate's evidence; do not assume the user's configuration is faulty
    or remove MCP servers, plugins, settings, or credentials to make it pass.
-10. The model chooses ordinary-task fan-out by issuing independent `dispatch`
+11. The model chooses ordinary-task fan-out by issuing independent `dispatch`
     commands concurrently. The private supervisor does not impose a task-count
     or CPU-based limit. Use this for independent investigation, debate, review,
     or provider comparison. Concurrent write tasks must use separate workspaces
@@ -121,7 +137,7 @@ normal task. For an explicitly requested readiness check, see
 [preflight guidance](references/command-reference.md#checking-readiness).
 
 The catalog and `capabilities` response expose supported permission modes,
-request options, continuation, and the read-only mechanism. A read-only
+request options, continuation, native interaction, and the read-only mechanism. A read-only
 capability with `containment: "provider-owned"` selects the provider's native
 behavior but is not an independent delegate-owned sandbox. Static discovery
 does not prove authentication or live acceptance; use the dispatch or live
@@ -143,46 +159,74 @@ it remains ambiguous or discovery is unavailable, ask for the exact model ID;
 do not turn a display name into a guessed alias. An explicit ID needs no such
 clarification just because it is absent from discovery.
 
+For planning, consultation, debate, or other external work that may need
+answers or guidance while it runs, read [active-task interaction](references/interaction.md)
+and use `--interactive` when supported. Answer known facts yourself; ask the
+user for missing preferences, authorization, or scope. Keep ordinary native
+subagents available under the same-harness rule above.
+
 ### 3. Prepare separate paths and a finite brief
 
 Use an existing task brief when one is supplied. Otherwise create a bounded
 brief file that contains the requested work and its acceptance criteria. Keep
 the state root outside the workspace and provider-writable trees (which can
-include the system temporary directory). Use `DELEGATE_ROOT` when supplied:
+include the system temporary directory). Use `DELEGATE_ROOT` when supplied.
+Pass a regular file path for `--brief`, not shell process substitution such as
+`<(...)` or an ephemeral `/dev/fd/...` path.
 
 Run path preparation and dispatch in **one shell/tool invocation**, as shown
 below. Shell variables may disappear between tool calls. For later commands,
 use the actual saved absolute paths or assign them again in that invocation;
 never pass an empty `--root`, `--brief`, or `--cwd`.
 
-Replace the example brief below with the actual user request. Do not place the state
+Write the brief as the **worker’s own assignment**, not a copy of the
+orchestration request. For “consult another agent about X”, tell the worker
+“Investigate X and return your findings; you are the assigned worker, do not
+delegate this assignment.” Keep dispatching, answering questions, and
+collecting results in the lead session unless recursive delegation was
+explicitly requested.
+
+Replace the example brief below with that bounded worker assignment. Do not place the state
 root inside the workspace or place secrets in either path’s task data. Keep it
 outside provider configuration and runtime directories as well. If delegate
 rejects a caller-supplied state root, report the placement error rather than
-silently replacing that root. If no root was supplied, choose one outside the
-workspace and reuse it. Do not change provider settings. Retain the selected
-root for every observation and continuation.
+silently replacing that root. For every task command, use
+`--root "${DELEGATE_ROOT:-$HOME/delegation-state}"` as shown below. This reads the supplied root in each shell call and uses the
+same default when none was supplied; no root setup or new directory choice is
+needed. If the user explicitly requests another root, substitute that exact
+absolute path consistently. Do not change provider settings or switch roots
+when waiting, replying, steering, or continuing.
 
 ### 4. Dispatch one bounded turn
 
 Use automatic selection unless the user named a destination.
 For a named harness or provider resolved in step 2, replace `--auto` with `--provider PROVIDER_ID`; never
-combine the two:
+combine the two.
+
+Capture the dispatch exit code with `if`, as below. This prints the response
+even when dispatch fails and the surrounding shell has `set -e` enabled.
 
 ```sh
-STATE_ROOT="${DELEGATE_ROOT:-$HOME/delegation-state}"
 WORKSPACE="$PWD"
 BRIEF="$(mktemp)"
+RESPONSE="$(mktemp)"
 cat >"$BRIEF" <<'EOF'
 Describe the delegated task, the files or behavior to change, and the checks
 that must pass. Keep the response concise and report validation evidence.
 EOF
-delegate --root "$STATE_ROOT" dispatch --auto \
+printf 'Saved response: %s\n' "$RESPONSE"
+if delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --auto \
   --brief "$BRIEF" \
   --cwd "$WORKSPACE" \
   --permission read-only \
   --budget 30m \
-  --json
+  --json >"$RESPONSE"; then
+  dispatch_exit=0
+else
+  dispatch_exit=$?
+fi
+cat "$RESPONSE"
+printf '\nDispatch exit: %s; saved response: %s\n' "$dispatch_exit" "$RESPONSE"
 ```
 
 For an authorized code change, use `--permission workspace-write`; this requests
@@ -201,6 +245,12 @@ files, never background commands writing to the same stdout. Read
 [parallel reviews and debates](references/collaboration.md) when coordinating
 multiple participants or challenge rounds.
 
+Capture the response and exit status in the **original** dispatch invocation,
+as above. Read the saved response if you need it again. Never repeat dispatch
+to obtain its exit code, recover output, or confirm that it worked: each call
+can create new paid work. Once a response says `task_record: created`, save its
+`task_id` and switch to observation. A missing exit code does not undo admission.
+
 Save the exact `task_id` from the JSON response. Apply this decision separately
 for **each attempt**, including requests issued together:
 
@@ -219,16 +269,20 @@ never retry that uncertain ID or run Pueue commands.
 
 ### 5. Observe and collect
 
-Use the same state root and task ID:
+Use the same state root and task ID. Each command below resolves the root
+in its own shell call, so it does not depend on a previous shell variable:
 
 ```sh
-delegate --root "$STATE_ROOT" status TASK_ID --json
-delegate --root "$STATE_ROOT" collect TASK_ID --watch 5s --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" status TASK_ID --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" collect TASK_ID --watch 5s --json
 ```
 
 Check `status` first: if it is `timed_out`, follow step 6 even when publication
-is pending and no outcome is present. Otherwise, if `collect` exits with code
-`3` because publication is still pending, wait a short interval and run
+is pending and no outcome is present. If `interaction.pending_requests` is
+nonempty and the interaction is not closed, answer the questions using
+[active-task interaction](references/interaction.md) before collecting again.
+Otherwise, if `collect` exits with code `3` because publication is still
+pending, wait a short interval and run
 `collect` again. Do not dispatch again. Continue until
 the response has a terminal `outcome` or a bounded operational error that
 requires the caller’s attention.

package/plugins/delegation-layer/skills/agent-integration/evals/evals.json

@@ -56,6 +56,46 @@
       "name": "write-timeout",
       "prompt": "An authorized write task has status timed_out, publication unknown, no outcome, and continuation.resumable true.",
       "expected": ["delegate continue --task", "same state root", "new task id", "do not wait indefinitely for an outcome", "no fresh unrelated dispatch"]
+    },
+    {
+      "name": "admitted-without-explicit-exit-marker",
+      "prompt": "Dispatch returned task_record created, admission admitted, and a task_id. The shell output did not include a separate EXIT marker. Confirm the task and collect its result.",
+      "expected": ["use status with returned task_id", "read saved dispatch response if needed", "do not dispatch again to print an exit code", "exactly one admitted task for the assignment"]
+    },
+    {
+      "name": "supplied-root-across-shell-calls",
+      "prompt": "DELEGATE_ROOT is already set. Consult a separate external agent and answer its questions; each shell tool call starts a new shell.",
+      "expected": ["use supplied DELEGATE_ROOT on dispatch, status, collect, reply, steer, and continue", "resolve the same root in each shell call", "no invented task-specific root"]
+    },
+    {
+      "name": "pending-without-questions",
+      "prompt": "The admitted task is running. Collect exits 3 with pending publication, empty pending_requests, and no operational error.",
+      "expected": ["keep collecting the same task", "no ps or pgrep or process-tree inspection", "no supervisor commands", "no duplicate dispatch", "answer through delegate reply when a question appears"]
+    },
+    {
+      "name": "interactive-worker-assignment",
+      "prompt": "Consult a separate external agent about a pricing plan and answer its questions from README.md.",
+      "expected": ["brief assigns the pricing investigation directly", "worker does not recruit another external agent", "lead answers through delegate reply", "do not copy the orchestration request verbatim"]
+    },
+    {
+      "name": "interactive-clarification",
+      "prompt": "An external planning task returned pending.kind interaction with question IDs, choices, and free_text false. The requested preference is not in the user brief.",
+      "expected": ["read references/interaction.md", "ask user for missing preference", "reply through delegate with exact request and question IDs", "stable message ID", "continue collecting"]
+    },
+    {
+      "name": "interactive-steering",
+      "prompt": "The user adds guidance while an interactive external task is running and its capabilities advertise steering.",
+      "expected": ["delegate steer", "preserve task model permission and budget", "do not cancel and redispatch", "receipt is not completion"]
+    },
+    {
+      "name": "interactive-unknown-delivery",
+      "prompt": "A reply submission lost its CLI response. Its recorded delivery is unknown.",
+      "expected": ["same message ID and exact content to retrieve receipt", "no new message ID retry", "no native CLI bypass", "observe task"]
+    },
+    {
+      "name": "interactive-unsupported-destination",
+      "prompt": "A named external provider supports continuation but neither native clarification nor steering. The user wants a discussion with it.",
+      "expected": ["preserve named provider", "ordinary bounded turns and continuation", "explain no active-turn questions", "do not pass unsupported interactive flag"]
     }
   ]
 }

package/plugins/delegation-layer/skills/agent-integration/evals/routing.json

@@ -199,6 +199,32 @@
       ],
       "split": "validation"
     },
+    {
+      "name": "recheck-recoverable",
+      "prompt": "Ask a separate Codex CLI session to review pricing.py against README.md without editing. If the terminal result is rejected with failure.next_action=recheck, recheck the same task once and report only the verified recovered answer.",
+      "external": true,
+      "provider": "codex:exec",
+      "fixture_scenario": "recoverable-rejection",
+      "holdout": true,
+      "split": "validation"
+    },
+    {
+      "name": "recheck-still-rejected",
+      "prompt": "Ask a separate Codex CLI session to review pricing.py against README.md without editing. If the terminal result is rejected with failure.next_action=recheck, recheck the same task once and report the rejection without using raw output as an answer.",
+      "external": true,
+      "provider": "codex:exec",
+      "fixture_scenario": "still-rejected-recheck",
+      "holdout": true,
+      "split": "validation"
+    },
+    {
+      "name": "external-collaboration-holdout",
+      "prompt": "Ask a separate Codex CLI session to review pricing.py against README.md and report its finding without editing.",
+      "external": true,
+      "provider": "codex:exec",
+      "holdout": true,
+      "split": "validation"
+    },
     {
       "name": "requested-model-effort",
       "prompt": "Ask a separate Codex CLI session to review pricing.py without edits. Use model gpt-5.6-luna and effort max, preserving both choices.",

package/plugins/delegation-layer/skills/agent-integration/examples/json-results.md

@@ -1,8 +1,8 @@
 # JSON decision sketch
 
 ```sh
-delegate --root "$STATE_ROOT" status "$TASK_ID" --json
-delegate --root "$STATE_ROOT" collect "$TASK_ID" --watch 5s --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" status "$TASK_ID" --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" collect "$TASK_ID" --watch 5s --json
 ```
 
 Decision order:

package/plugins/delegation-layer/skills/agent-integration/examples/read-only-task.md

@@ -1,10 +1,9 @@
 # Read-only task
 
 ```sh
-STATE_ROOT="${DELEGATE_ROOT:-$HOME/delegation-state}"
 BRIEF="$(mktemp)"
 printf '%s\n' 'Inspect the workspace and summarize the current test failures. Do not edit files.' >"$BRIEF"
-delegate --root "$STATE_ROOT" dispatch --auto \
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --auto \
   --brief "$BRIEF" --cwd "$PWD" --permission read-only --budget 15m --json
 ```
 

package/plugins/delegation-layer/skills/agent-integration/examples/workspace-write-task.md

@@ -1,8 +1,7 @@
 # Authorized workspace write task
 
 ```sh
-STATE_ROOT="${DELEGATE_ROOT:-$HOME/delegation-state}"
-delegate --root "$STATE_ROOT" dispatch --auto \
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --auto \
   --brief "$PWD/delegated-task.md" --cwd "$PWD" \
   --permission workspace-write --budget 30m --json
 ```

package/plugins/delegation-layer/skills/agent-integration/references/collaboration.md

@@ -19,14 +19,14 @@ RECEIPTS="$(mktemp -d)" || exit 1
 # Set these absolute paths and catalog IDs in this same tool invocation.
 # BRIEF_A and BRIEF_B contain the real independent assignments.
 (
-  delegate --root "$STATE_ROOT" dispatch --provider "$PROVIDER_A" \
+  delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --provider "$PROVIDER_A" \
     --brief "$BRIEF_A" --cwd "$WORKSPACE_A" \
     --permission read-only --budget 30m --json \
     >"$RECEIPTS/review-a.json" 2>"$RECEIPTS/review-a.stderr"
   printf '%s\n' "$?" >"$RECEIPTS/review-a.exit"
 ) &
 (
-  delegate --root "$STATE_ROOT" dispatch --provider "$PROVIDER_B" \
+  delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --provider "$PROVIDER_B" \
     --brief "$BRIEF_B" --cwd "$WORKSPACE_B" \
     --permission read-only --budget 30m --json \
     >"$RECEIPTS/review-b.json" 2>"$RECEIPTS/review-b.stderr"

package/plugins/delegation-layer/skills/agent-integration/references/command-reference.md

@@ -13,7 +13,11 @@ All paths passed to `--root`, `--cwd`, and `--brief` must be absolute. Add
 | `delegate dispatch --auto ... --json` | Select and admit one bounded turn | Yes, at most once |
 | `delegate status TASK --json` | Reconcile and observe task state | No provider launch |
 | `delegate collect TASK --watch 5s --json` | Read or wait for publication | No provider launch |
+| `delegate recheck TASK --json` | Reassess eligible rejected evidence offline | No provider launch; original outcome unchanged |
 | `delegate continue --task TASK ... --json` | Create an exact-session successor | Yes, at most once for the successor |
+| `delegate reply TASK --request ID --answers FILE --id MESSAGE_ID --json` | Answer a pending native question | No new launch |
+| `delegate steer TASK --brief FILE --id MESSAGE_ID --json` | Guide a supported active turn | No new launch |
+| `delegate events TASK --after SEQUENCE --watch 5s --json` | Read a bounded interaction event page | No |
 | `delegate logs TASK --json` | Return validated output descriptors | No |
 | `delegate cancel TASK --json` | Record one explicit stop request | Stops, does not relaunch |
 
@@ -30,9 +34,11 @@ Dispatch flags:
 - `--auto`: let the CLI select a ready provider;
 - `--provider ID`: preserve the caller’s named harness or provider, resolved through the catalog;
 - `--model MODEL` and `--effort VALUE`: provider-advertised choices (CCR requires a model alias);
-  Codex maps effort to TOML `model_reasoning_effort`, while Claude and AGY
+  Codex maps effort to TOML `model_reasoning_effort` for ordinary turns or a native
+  turn parameter for interactive turns, while Claude and AGY
   pass native effort flags; Pi maps effort to `--thinking`; OpenCode maps it to
   `--variant`; CCR forwards effort to Claude and uses `--model` as a CCR alias; `default` is omitted;
+- `--interactive`: enable advertised native clarification and steering; see [interaction](interaction.md);
 - `--native-timeout DURATION`: optional provider-native timeout where advertised.
 
 Model discovery uses the same `--provider ID` profiles and accepts an optional
@@ -47,7 +53,8 @@ Continuation flags:
 
 - `--task TASK`: predecessor task ID;
 - `--brief FILE`: optional new follow-up instruction;
-- `--budget DURATION`, `--model MODEL`, and `--effort VALUE`: optional overrides.
+- `--budget DURATION`, `--model MODEL`, and `--effort VALUE`: optional overrides;
+- `--interactive`: enable interaction when supported; already-interactive predecessors keep it.
 
 ## Choosing models and effort
 

package/plugins/delegation-layer/skills/agent-integration/references/continuation.md

@@ -17,7 +17,7 @@ When `status` or `collect` reports a timed-out turn with:
 run:
 
 ```sh
-delegate --root "$STATE_ROOT" continue \
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" continue \
   --task PREDECESSOR_TASK_ID \
   --brief FOLLOW_UP_BRIEF \
   --budget 30m \

package/plugins/delegation-layer/skills/agent-integration/references/interaction.md

@@ -0,0 +1,110 @@
+# Questions and guidance during an active task
+
+Use this reference when an external task may need clarification or guidance
+before it finishes, including planning, consultation, and debate. Use only
+Delegate commands; do not open a terminal session with the provider.
+
+## Choose the transport
+
+Read the selected provider's `interaction` capability and
+`interaction.supported_modes`. `clarification` and `steering` are independent.
+Add `--interactive` to dispatch when the chosen provider and permission mode
+support the interaction you need. Keep the user's exact provider, model,
+effort, workspace, permission, and budget.
+
+Without a named provider, `dispatch --auto --interactive` selects among
+providers that support interaction. If a named provider does not support it,
+use ordinary bounded turns and `continue` for discussion when native
+continuation is available. Explain that questions cannot be answered during
+that provider's active turn. Never replace the requested provider silently.
+
+In the worker brief, explicitly request **native structured clarification**:
+"Use your native question tool to request clarification and await its answer.
+Do not ask only in final text or a progress message and then sleep: the lead
+can answer structured requests, not an unstructured prompt. If the question
+tool is unavailable, report that blocker and finish instead of waiting."
+Assign the actual work directly; the worker should not start another delegate.
+
+No change is needed for self-contained tasks that do not need mid-turn input.
+Continuation inherits interaction from its predecessor. `continue --task ID
+--interactive` can enable it for a previously ordinary task.
+
+## Respond to questions
+
+1. Save the admitted task ID and state root. Run `collect TASK_ID --watch 30s
+   --json` with that same root. Exit 3 is pending, not completion. Collection
+   returns early when there is an outstanding native question.
+2. Read `interaction.pending_requests`. Use the returned request ID, question
+   IDs, options, `multiple`, and `free_text`; do not invent or guess IDs.
+3. Answer facts already supplied by the user or verified during the task.
+   Ask the user for missing preferences, business decisions, new scope, or
+   authorization. Do not guess an answer to keep the worker moving. A question
+   reply does not approve tool access or change the task's permission mode.
+4. Write one JSON object mapping **every question ID in that request** to an
+   array of answer strings. A single-choice question takes one value. Use exact
+   option labels unless `free_text` permits another answer. Example:
+
+   ```json
+   {"color": ["Blue"], "sections": ["Summary", "Evidence"]}
+   ```
+
+5. Allocate a new 32-character lowercase hexadecimal message ID and save it
+   together with the exact answers. Submit once:
+
+   ```sh
+   delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" reply "$task_id" \
+     --request "$request_id" --answers "$answers_file" \
+     --id "$message_id" --json
+   ```
+
+6. Keep collecting. A receipt is not task completion. Only the task's
+   authoritative `outcome` establishes the final result.
+
+Use the same root expression for dispatch, collection, reply, and steering.
+If the user explicitly chose another root, use that exact path for all of
+them. The other variables stand for actual values saved from this task;
+reassign them in each shell invocation if your tool does not preserve variables.
+A pending response without questions means keep collecting, not inspect OS
+processes or dispatch a replacement task.
+
+## Guide active work
+
+When `interaction.capabilities.steering` is true and the task is active, put
+additional guidance in a brief file and submit:
+
+```sh
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" steer "$task_id" \
+  --brief "$guidance_file" --id "$message_id" --json
+```
+
+Steering is guidance to the existing turn. It does not change provider, model,
+effort, workspace, permission, or budget. Once the turn ends, use `continue`
+for another authorized turn instead. Do not cancel and redispatch simply to
+send guidance.
+
+## Interpret delivery precisely
+
+| State | What it establishes |
+| --- | --- |
+| `queued` | Delegate durably accepted the message; native delivery is pending. |
+| `unknown` | Delivery was reserved or attempted, but its result is uncertain. |
+| `sent` | The native input write completed; no native acknowledgement is established. |
+| `acknowledged` | A correlated native response acknowledged the message. This does not prove the model followed guidance. |
+| `rejected` | The message could not be delivered or the native protocol rejected it. Read the reason. |
+
+If a submission response was lost, repeat the **same message ID and exact
+content** to recover its receipt. Never allocate a fresh ID to retry an
+`unknown` delivery. Delegate will not replay an ambiguous native write.
+Changed content with an existing ID is a conflict. Observe the task instead
+of guessing whether it consumed the message.
+
+`events TASK_ID --after SEQUENCE --watch 30s --json` returns a bounded event
+page. Save the returned `cursor` for the next page. Events and interaction
+state are progress evidence, not a replacement for `collect` and `outcome`.
+
+Waiting for an answer consumes the original execution budget; it does not
+pause or extend it. `interaction.deadline` shows the recorded wall-clock
+boundary when execution has started. If the task ends while waiting, do not
+reply to the old request. Follow the ordinary continuation guidance only when
+Delegate reports the task resumable. Preserve any verified answers in the new
+brief; uncertain delivery must not become an automatic replay.

package/plugins/delegation-layer/skills/agent-integration/references/result-handling.md

@@ -15,10 +15,13 @@ the same as a successful provider result.
    permits a linked successor even when no outcome has been published.
 4. If `publication` is committed and `outcome.verdict` is committed, return
    the result and its payload/evidence descriptors.
-5. If `outcome.verdict` is rejected, report a terminal rejected result with its
-   refusal and evidence. Do not retry automatically.
-6. Otherwise, if `liveness` is queued or running or publication is pending,
-   observe the same task again with a finite watch interval.
+5. If `outcome.verdict` is rejected and `failure.next_action` is `recheck`,
+   follow [offline recovery](#offline-recovery) once. Otherwise report the
+   terminal rejection with its diagnostic and evidence. Do not redispatch.
+6. If `pending.kind` is `interaction`, handle the returned questions using
+   [interaction guidance](interaction.md) before waiting again. Otherwise, if
+   `liveness` is queued or running or publication is pending, observe the same
+   task with a finite watch interval.
 7. If authentication is `blocked`, preserve that neutral state and tell the
    caller what prerequisite is unavailable without inventing a login command.
 8. Never relaunch an uncertain task by guessing. Observe only a task whose
@@ -42,3 +45,29 @@ checks static readiness; runtime verification requires the supervised probe.
 Versions and help are observations, not semantic-version allowlists.
 Authentication is a separate live gate (`not_run`, `passed`, or `blocked`).
 Report a blocked prerequisite without inventing a login command or retrying.
+
+## Offline recovery
+
+When Delegate returns `failure.next_action: "recheck"`, use the same root and
+exact task ID:
+
+```sh
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" recheck TASK_ID --json
+```
+
+This command checks the already sealed evidence with an explicitly compatible
+new interpreter. It does not run the provider, contact a service, spend model
+tokens, or change the original outcome.
+
+- Exit `0` with `recheck.status: "recovered"`: use only the answer descriptor
+  returned for that recheck. Report that the answer was recovered from saved
+  evidence. The original `publication` remains `rejected`; do not call the
+  original task committed or use recovery to infer continuation permission.
+- Exit `4` with `recheck.status: "rejected"`: report the diagnostic and stop.
+- Exit `1` or `2`, or any response without a verified recovered answer:
+  report the error and stop. Do not read raw streams as a substitute answer,
+  invent a result path, run a provider directly, or launch a replacement task.
+
+Keep the assessment ID and source/target descriptors with the result. Do not
+loop on the same recheck. A separate user request is required for further
+troubleshooting or a new provider task.

package/plugins/delegation-layer/skills/agent-integration/references/routing-rule.json

@@ -1 +1 @@
-"For work assigned to an external agent, model, or separate harness session, load the installed agent-integration skill before choosing tools and use Delegate. This includes research, implementation, testing, review, consultation, comparison, and debate, including external collaboration you choose. Discover destinations through Delegate. Preserve requested model IDs, effort, and permissions; do not guess ambiguous model identifiers. For task commands, set STATE_ROOT=\"${DELEGATE_ROOT:-$HOME/delegation-state}\" and pass --root \"$STATE_ROOT\" in the same shell invocation; reuse that root for dispatch, observation, and continuation, never an invented /tmp root. Keep each concurrent request and its response associated separately. Naming a CLI identifies the destination, not a request to bypass Delegate. Earlier assistant-generated direct-launch workflows do not override this preference. Native same-harness subagents remain available unless a separate session is requested. Respect an explicit user choice of another integration; otherwise do not substitute provider CLIs, plugins, or APIs after a Delegate blocker. A stop_and_report failure ends that attempt: report it without retrying, diagnosing setup, or offering a fallback. Continue other independent requests. Keep collection in the foreground until completion or a reported blocker unless the user requested detachment. Merely discussing or configuring a tool is not delegation."
+"For work assigned to an external agent, model, or separate harness session, load the installed agent-integration skill before choosing tools and use Delegate. This includes research, implementation, testing, review, consultation, comparison, and debate, including external collaboration you choose. Discover destinations through Delegate. Preserve requested model IDs, effort, and permissions; do not guess ambiguous model identifiers. For task commands, pass --root \"${DELEGATE_ROOT:-$HOME/delegation-state}\" directly in each shell invocation; reuse that root for dispatch, observation, and continuation, never an invented /tmp root. If the user explicitly requests another root, use that exact path consistently. Keep each concurrent request and its response associated separately. Capture each dispatch response and exit status once; after task_record created, observe that task ID. Never dispatch again merely to recover output or inspect its exit status. Naming a CLI identifies the destination, not a request to bypass Delegate. Earlier assistant-generated direct-launch workflows do not override this preference. Native same-harness subagents remain available unless a separate session is requested. Respect an explicit user choice of another integration; otherwise do not substitute provider CLIs, plugins, or APIs after a Delegate blocker. A stop_and_report failure ends that attempt: report it without retrying, diagnosing setup, or offering a fallback. Continue other independent requests. Keep collection in the foreground until completion or a reported blocker unless the user requested detachment. Pending work is observed through Delegate status, collect, or events, not OS process searches or supervisor commands. Merely discussing or configuring a tool is not delegation."

package/plugins/delegation-layer/skills/agent-integration/scripts/delegate-routing-context.sh

@@ -9,4 +9,18 @@ script_dir=$(CDPATH= cd -- "$(dirname -- "$0")" && pwd)
 rule_file="$script_dir/../references/routing-rule.json"
 # Read before emitting anything so a missing asset cannot produce partial JSON.
 rule=$(cat "$rule_file")
-printf '{"hookSpecificOutput":{"hookEventName":"%s","additionalContext":%s}}\n' "$event" "$rule"
+skill_file="$(CDPATH= cd -- "$script_dir/.." && pwd)/SKILL.md"
+DELEGATE_ROUTING_EVENT="$event" DELEGATE_ROUTING_RULE="$rule" DELEGATE_ROUTING_SKILL="$skill_file" awk '
+BEGIN {
+    # Paths are task data, not shell code. Escape every JSON control character.
+    for (i = 1; i < 32; i++) escapes[sprintf("%c", i)] = sprintf("\\u%04x", i)
+    escapes["\\"] = "\\\\"
+    escapes["\""] = "\\\""
+    path = ENVIRON["DELEGATE_ROUTING_SKILL"]
+    for (i = 1; i <= length(path); i++) {
+        character = substr(path, i, 1)
+        escaped = escaped ((character in escapes) ? escapes[character] : character)
+    }
+    # The maintained rule is a JSON string. Keep its escaping when prefixing it.
+    printf "{\"hookSpecificOutput\":{\"hookEventName\":\"%s\",\"additionalContext\":\"Read the complete agent-integration skill at %s. %s}}\n", ENVIRON["DELEGATE_ROUTING_EVENT"], escaped, substr(ENVIRON["DELEGATE_ROUTING_RULE"], 2)
+}'

package/skills/agent-integration/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: agent-integration
-description: Coordinate external agent, model, and separate harness tasks through Delegate, including research, implementation, tests, reviews, consultations, comparisons, and debates. Use before choosing an external integration; ordinary native subagents and questions about tools remain local.
+description: Use Delegate for work assigned to another agent, model, or separate harness CLI session. Load this before a provider-specific skill or plugin, including when a user names a CLI for research, implementation, tests, reviews, consultations, comparisons, or debates. A named destination still routes through Delegate. Ordinary same-harness subagents and questions about tools remain local; respect an explicit choice of another integration.
 ---
 
 # Delegation Layer
@@ -13,6 +13,8 @@ external harness to do work. Requests such as “ask Codex to review this”,
 collaboration even when the user never mentions Delegate. The same applies
 when you decide external expertise is useful to fulfill the user's task.
 These names are examples; discover supported destinations through the CLI.
+A destination name identifies who should do the work; it does not by itself
+select a provider-specific skill, plugin, or direct CLI integration.
 “Claude Code Router” and “CCR” mean `ccr:print`, not ordinary Claude Code.
 For that request, use Delegate with a CCR model alias; a Claude-specific skill
 or subscription model does not fulfill the requested route.
@@ -36,8 +38,10 @@ not turn that request into an environment-repair task. When a response says
 `failure.next_action: "stop_and_report"`, that attempt is finished. Include its
 failure in your final response without opening a clarification dialog to choose
 a fallback. For `task_record: not_created`, identify the requested destination
-and quote the returned `task_id` (if present), `task_record`, `failure`, and
-`error`. Do not append inferred state paths or provider-setup diagnoses. Do not run diagnostic commands, retry it, or use another integration.
+and include the saved JSON failure response verbatim. A `task_id` can be returned
+even when no task record was created. Do not infer that an unavailable provider
+is uninstalled or unauthenticated. Do not append inferred state paths or
+provider-setup diagnoses, run diagnostic commands, retry it, or use another integration.
 Troubleshooting requires a separate user request. Other independently requested
 tasks may proceed; the failed attempt must remain reported as failed or blocked.
 
@@ -64,26 +68,38 @@ tasks may proceed; the failed attempt must remain reported as failed or blocked.
    Delegate owns the supervisor lifecycle. Never install, configure, start,
    inspect, or repair Pueue, and never poll a task unless the response says its
    `task_record` is `created`.
-5. Observe with `status` and collect with `collect`. Keep collection in the
+5. Observe tasks only through Delegate `status`, `collect`, or `events`.
+   Do not run `ps`, `pgrep`, process-tree searches, or supervisor commands to
+   monitor delegated work. Pending output is normal startup or work in progress;
+   empty `pending_requests` means keep collecting, not diagnose the environment.
+   Environment diagnostics require a separately requested troubleshooting task.
+   Keep collection in the
    foreground until all requested tasks reach a terminal result or an explicit
    blocker. Do not end a headless turn promising a background notification.
-   Pending publication (exit 3) means keep waiting, not success. Only detach
+   Pending publication (exit 3) means keep waiting, not success. When the JSON
+   includes an outstanding `interaction.pending_requests` batch, follow
+   [active-task interaction](references/interaction.md) before waiting again. Only detach
    when the user explicitly requests it, retaining the exact task IDs.
    Collection may recover the
    saved supervisor only to observe an already-requested budget stop; it never
    launches a provider or retries a task.
-6. If a task is timed out and its JSON says `continuation.resumable: true`, use
+6. If `failure.next_action` is `recheck`, follow
+   [offline recovery](references/result-handling.md#offline-recovery) once for
+   that task. Rechecking saved evidence does not run the provider again.
+   Otherwise, report a rejected result without redispatching or reading raw
+   output as an answer.
+7. If a task is timed out and its JSON says `continuation.resumable: true`, use
    `continue --task TASK_ID`, optionally with a new `--brief`. Do not create a
    fresh unrelated dispatch to replace a resumable task.
-7. Treat `blocked` authentication as unresolved. Do not call it compatible,
+8. Treat `blocked` authentication as unresolved. Do not call it compatible,
    successful, or failed provider behavior, and do not invent login commands.
-8. Never put credentials, tokens, or unrestricted provider output in a brief,
+9. Never put credentials, tokens, or unrestricted provider output in a brief,
    option, receipt, or agent message.
-9. Permission modes select provider-native behavior, not an independent sandbox.
+10. Permission modes select provider-native behavior, not an independent sandbox.
    Providers load their own configuration and extensions. If a task is rejected,
    report delegate's evidence; do not assume the user's configuration is faulty
    or remove MCP servers, plugins, settings, or credentials to make it pass.
-10. The model chooses ordinary-task fan-out by issuing independent `dispatch`
+11. The model chooses ordinary-task fan-out by issuing independent `dispatch`
     commands concurrently. The private supervisor does not impose a task-count
     or CPU-based limit. Use this for independent investigation, debate, review,
     or provider comparison. Concurrent write tasks must use separate workspaces
@@ -121,7 +137,7 @@ normal task. For an explicitly requested readiness check, see
 [preflight guidance](references/command-reference.md#checking-readiness).
 
 The catalog and `capabilities` response expose supported permission modes,
-request options, continuation, and the read-only mechanism. A read-only
+request options, continuation, native interaction, and the read-only mechanism. A read-only
 capability with `containment: "provider-owned"` selects the provider's native
 behavior but is not an independent delegate-owned sandbox. Static discovery
 does not prove authentication or live acceptance; use the dispatch or live
@@ -143,46 +159,74 @@ it remains ambiguous or discovery is unavailable, ask for the exact model ID;
 do not turn a display name into a guessed alias. An explicit ID needs no such
 clarification just because it is absent from discovery.
 
+For planning, consultation, debate, or other external work that may need
+answers or guidance while it runs, read [active-task interaction](references/interaction.md)
+and use `--interactive` when supported. Answer known facts yourself; ask the
+user for missing preferences, authorization, or scope. Keep ordinary native
+subagents available under the same-harness rule above.
+
 ### 3. Prepare separate paths and a finite brief
 
 Use an existing task brief when one is supplied. Otherwise create a bounded
 brief file that contains the requested work and its acceptance criteria. Keep
 the state root outside the workspace and provider-writable trees (which can
-include the system temporary directory). Use `DELEGATE_ROOT` when supplied:
+include the system temporary directory). Use `DELEGATE_ROOT` when supplied.
+Pass a regular file path for `--brief`, not shell process substitution such as
+`<(...)` or an ephemeral `/dev/fd/...` path.
 
 Run path preparation and dispatch in **one shell/tool invocation**, as shown
 below. Shell variables may disappear between tool calls. For later commands,
 use the actual saved absolute paths or assign them again in that invocation;
 never pass an empty `--root`, `--brief`, or `--cwd`.
 
-Replace the example brief below with the actual user request. Do not place the state
+Write the brief as the **worker’s own assignment**, not a copy of the
+orchestration request. For “consult another agent about X”, tell the worker
+“Investigate X and return your findings; you are the assigned worker, do not
+delegate this assignment.” Keep dispatching, answering questions, and
+collecting results in the lead session unless recursive delegation was
+explicitly requested.
+
+Replace the example brief below with that bounded worker assignment. Do not place the state
 root inside the workspace or place secrets in either path’s task data. Keep it
 outside provider configuration and runtime directories as well. If delegate
 rejects a caller-supplied state root, report the placement error rather than
-silently replacing that root. If no root was supplied, choose one outside the
-workspace and reuse it. Do not change provider settings. Retain the selected
-root for every observation and continuation.
+silently replacing that root. For every task command, use
+`--root "${DELEGATE_ROOT:-$HOME/delegation-state}"` as shown below. This reads the supplied root in each shell call and uses the
+same default when none was supplied; no root setup or new directory choice is
+needed. If the user explicitly requests another root, substitute that exact
+absolute path consistently. Do not change provider settings or switch roots
+when waiting, replying, steering, or continuing.
 
 ### 4. Dispatch one bounded turn
 
 Use automatic selection unless the user named a destination.
 For a named harness or provider resolved in step 2, replace `--auto` with `--provider PROVIDER_ID`; never
-combine the two:
+combine the two.
+
+Capture the dispatch exit code with `if`, as below. This prints the response
+even when dispatch fails and the surrounding shell has `set -e` enabled.
 
 ```sh
-STATE_ROOT="${DELEGATE_ROOT:-$HOME/delegation-state}"
 WORKSPACE="$PWD"
 BRIEF="$(mktemp)"
+RESPONSE="$(mktemp)"
 cat >"$BRIEF" <<'EOF'
 Describe the delegated task, the files or behavior to change, and the checks
 that must pass. Keep the response concise and report validation evidence.
 EOF
-delegate --root "$STATE_ROOT" dispatch --auto \
+printf 'Saved response: %s\n' "$RESPONSE"
+if delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --auto \
   --brief "$BRIEF" \
   --cwd "$WORKSPACE" \
   --permission read-only \
   --budget 30m \
-  --json
+  --json >"$RESPONSE"; then
+  dispatch_exit=0
+else
+  dispatch_exit=$?
+fi
+cat "$RESPONSE"
+printf '\nDispatch exit: %s; saved response: %s\n' "$dispatch_exit" "$RESPONSE"
 ```
 
 For an authorized code change, use `--permission workspace-write`; this requests
@@ -201,6 +245,12 @@ files, never background commands writing to the same stdout. Read
 [parallel reviews and debates](references/collaboration.md) when coordinating
 multiple participants or challenge rounds.
 
+Capture the response and exit status in the **original** dispatch invocation,
+as above. Read the saved response if you need it again. Never repeat dispatch
+to obtain its exit code, recover output, or confirm that it worked: each call
+can create new paid work. Once a response says `task_record: created`, save its
+`task_id` and switch to observation. A missing exit code does not undo admission.
+
 Save the exact `task_id` from the JSON response. Apply this decision separately
 for **each attempt**, including requests issued together:
 
@@ -219,16 +269,20 @@ never retry that uncertain ID or run Pueue commands.
 
 ### 5. Observe and collect
 
-Use the same state root and task ID:
+Use the same state root and task ID. Each command below resolves the root
+in its own shell call, so it does not depend on a previous shell variable:
 
 ```sh
-delegate --root "$STATE_ROOT" status TASK_ID --json
-delegate --root "$STATE_ROOT" collect TASK_ID --watch 5s --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" status TASK_ID --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" collect TASK_ID --watch 5s --json
 ```
 
 Check `status` first: if it is `timed_out`, follow step 6 even when publication
-is pending and no outcome is present. Otherwise, if `collect` exits with code
-`3` because publication is still pending, wait a short interval and run
+is pending and no outcome is present. If `interaction.pending_requests` is
+nonempty and the interaction is not closed, answer the questions using
+[active-task interaction](references/interaction.md) before collecting again.
+Otherwise, if `collect` exits with code `3` because publication is still
+pending, wait a short interval and run
 `collect` again. Do not dispatch again. Continue until
 the response has a terminal `outcome` or a bounded operational error that
 requires the caller’s attention.

package/skills/agent-integration/evals/evals.json

@@ -56,6 +56,46 @@
       "name": "write-timeout",
       "prompt": "An authorized write task has status timed_out, publication unknown, no outcome, and continuation.resumable true.",
       "expected": ["delegate continue --task", "same state root", "new task id", "do not wait indefinitely for an outcome", "no fresh unrelated dispatch"]
+    },
+    {
+      "name": "admitted-without-explicit-exit-marker",
+      "prompt": "Dispatch returned task_record created, admission admitted, and a task_id. The shell output did not include a separate EXIT marker. Confirm the task and collect its result.",
+      "expected": ["use status with returned task_id", "read saved dispatch response if needed", "do not dispatch again to print an exit code", "exactly one admitted task for the assignment"]
+    },
+    {
+      "name": "supplied-root-across-shell-calls",
+      "prompt": "DELEGATE_ROOT is already set. Consult a separate external agent and answer its questions; each shell tool call starts a new shell.",
+      "expected": ["use supplied DELEGATE_ROOT on dispatch, status, collect, reply, steer, and continue", "resolve the same root in each shell call", "no invented task-specific root"]
+    },
+    {
+      "name": "pending-without-questions",
+      "prompt": "The admitted task is running. Collect exits 3 with pending publication, empty pending_requests, and no operational error.",
+      "expected": ["keep collecting the same task", "no ps or pgrep or process-tree inspection", "no supervisor commands", "no duplicate dispatch", "answer through delegate reply when a question appears"]
+    },
+    {
+      "name": "interactive-worker-assignment",
+      "prompt": "Consult a separate external agent about a pricing plan and answer its questions from README.md.",
+      "expected": ["brief assigns the pricing investigation directly", "worker does not recruit another external agent", "lead answers through delegate reply", "do not copy the orchestration request verbatim"]
+    },
+    {
+      "name": "interactive-clarification",
+      "prompt": "An external planning task returned pending.kind interaction with question IDs, choices, and free_text false. The requested preference is not in the user brief.",
+      "expected": ["read references/interaction.md", "ask user for missing preference", "reply through delegate with exact request and question IDs", "stable message ID", "continue collecting"]
+    },
+    {
+      "name": "interactive-steering",
+      "prompt": "The user adds guidance while an interactive external task is running and its capabilities advertise steering.",
+      "expected": ["delegate steer", "preserve task model permission and budget", "do not cancel and redispatch", "receipt is not completion"]
+    },
+    {
+      "name": "interactive-unknown-delivery",
+      "prompt": "A reply submission lost its CLI response. Its recorded delivery is unknown.",
+      "expected": ["same message ID and exact content to retrieve receipt", "no new message ID retry", "no native CLI bypass", "observe task"]
+    },
+    {
+      "name": "interactive-unsupported-destination",
+      "prompt": "A named external provider supports continuation but neither native clarification nor steering. The user wants a discussion with it.",
+      "expected": ["preserve named provider", "ordinary bounded turns and continuation", "explain no active-turn questions", "do not pass unsupported interactive flag"]
     }
   ]
 }

package/skills/agent-integration/evals/routing.json

@@ -199,6 +199,32 @@
       ],
       "split": "validation"
     },
+    {
+      "name": "recheck-recoverable",
+      "prompt": "Ask a separate Codex CLI session to review pricing.py against README.md without editing. If the terminal result is rejected with failure.next_action=recheck, recheck the same task once and report only the verified recovered answer.",
+      "external": true,
+      "provider": "codex:exec",
+      "fixture_scenario": "recoverable-rejection",
+      "holdout": true,
+      "split": "validation"
+    },
+    {
+      "name": "recheck-still-rejected",
+      "prompt": "Ask a separate Codex CLI session to review pricing.py against README.md without editing. If the terminal result is rejected with failure.next_action=recheck, recheck the same task once and report the rejection without using raw output as an answer.",
+      "external": true,
+      "provider": "codex:exec",
+      "fixture_scenario": "still-rejected-recheck",
+      "holdout": true,
+      "split": "validation"
+    },
+    {
+      "name": "external-collaboration-holdout",
+      "prompt": "Ask a separate Codex CLI session to review pricing.py against README.md and report its finding without editing.",
+      "external": true,
+      "provider": "codex:exec",
+      "holdout": true,
+      "split": "validation"
+    },
     {
       "name": "requested-model-effort",
       "prompt": "Ask a separate Codex CLI session to review pricing.py without edits. Use model gpt-5.6-luna and effort max, preserving both choices.",

package/skills/agent-integration/examples/json-results.md

@@ -1,8 +1,8 @@
 # JSON decision sketch
 
 ```sh
-delegate --root "$STATE_ROOT" status "$TASK_ID" --json
-delegate --root "$STATE_ROOT" collect "$TASK_ID" --watch 5s --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" status "$TASK_ID" --json
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" collect "$TASK_ID" --watch 5s --json
 ```
 
 Decision order:

package/skills/agent-integration/examples/read-only-task.md

@@ -1,10 +1,9 @@
 # Read-only task
 
 ```sh
-STATE_ROOT="${DELEGATE_ROOT:-$HOME/delegation-state}"
 BRIEF="$(mktemp)"
 printf '%s\n' 'Inspect the workspace and summarize the current test failures. Do not edit files.' >"$BRIEF"
-delegate --root "$STATE_ROOT" dispatch --auto \
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --auto \
   --brief "$BRIEF" --cwd "$PWD" --permission read-only --budget 15m --json
 ```
 

package/skills/agent-integration/examples/workspace-write-task.md

@@ -1,8 +1,7 @@
 # Authorized workspace write task
 
 ```sh
-STATE_ROOT="${DELEGATE_ROOT:-$HOME/delegation-state}"
-delegate --root "$STATE_ROOT" dispatch --auto \
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --auto \
   --brief "$PWD/delegated-task.md" --cwd "$PWD" \
   --permission workspace-write --budget 30m --json
 ```

package/skills/agent-integration/references/collaboration.md

@@ -19,14 +19,14 @@ RECEIPTS="$(mktemp -d)" || exit 1
 # Set these absolute paths and catalog IDs in this same tool invocation.
 # BRIEF_A and BRIEF_B contain the real independent assignments.
 (
-  delegate --root "$STATE_ROOT" dispatch --provider "$PROVIDER_A" \
+  delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --provider "$PROVIDER_A" \
     --brief "$BRIEF_A" --cwd "$WORKSPACE_A" \
     --permission read-only --budget 30m --json \
     >"$RECEIPTS/review-a.json" 2>"$RECEIPTS/review-a.stderr"
   printf '%s\n' "$?" >"$RECEIPTS/review-a.exit"
 ) &
 (
-  delegate --root "$STATE_ROOT" dispatch --provider "$PROVIDER_B" \
+  delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" dispatch --provider "$PROVIDER_B" \
     --brief "$BRIEF_B" --cwd "$WORKSPACE_B" \
     --permission read-only --budget 30m --json \
     >"$RECEIPTS/review-b.json" 2>"$RECEIPTS/review-b.stderr"

package/skills/agent-integration/references/command-reference.md

@@ -13,7 +13,11 @@ All paths passed to `--root`, `--cwd`, and `--brief` must be absolute. Add
 | `delegate dispatch --auto ... --json` | Select and admit one bounded turn | Yes, at most once |
 | `delegate status TASK --json` | Reconcile and observe task state | No provider launch |
 | `delegate collect TASK --watch 5s --json` | Read or wait for publication | No provider launch |
+| `delegate recheck TASK --json` | Reassess eligible rejected evidence offline | No provider launch; original outcome unchanged |
 | `delegate continue --task TASK ... --json` | Create an exact-session successor | Yes, at most once for the successor |
+| `delegate reply TASK --request ID --answers FILE --id MESSAGE_ID --json` | Answer a pending native question | No new launch |
+| `delegate steer TASK --brief FILE --id MESSAGE_ID --json` | Guide a supported active turn | No new launch |
+| `delegate events TASK --after SEQUENCE --watch 5s --json` | Read a bounded interaction event page | No |
 | `delegate logs TASK --json` | Return validated output descriptors | No |
 | `delegate cancel TASK --json` | Record one explicit stop request | Stops, does not relaunch |
 
@@ -30,9 +34,11 @@ Dispatch flags:
 - `--auto`: let the CLI select a ready provider;
 - `--provider ID`: preserve the caller’s named harness or provider, resolved through the catalog;
 - `--model MODEL` and `--effort VALUE`: provider-advertised choices (CCR requires a model alias);
-  Codex maps effort to TOML `model_reasoning_effort`, while Claude and AGY
+  Codex maps effort to TOML `model_reasoning_effort` for ordinary turns or a native
+  turn parameter for interactive turns, while Claude and AGY
   pass native effort flags; Pi maps effort to `--thinking`; OpenCode maps it to
   `--variant`; CCR forwards effort to Claude and uses `--model` as a CCR alias; `default` is omitted;
+- `--interactive`: enable advertised native clarification and steering; see [interaction](interaction.md);
 - `--native-timeout DURATION`: optional provider-native timeout where advertised.
 
 Model discovery uses the same `--provider ID` profiles and accepts an optional
@@ -47,7 +53,8 @@ Continuation flags:
 
 - `--task TASK`: predecessor task ID;
 - `--brief FILE`: optional new follow-up instruction;
-- `--budget DURATION`, `--model MODEL`, and `--effort VALUE`: optional overrides.
+- `--budget DURATION`, `--model MODEL`, and `--effort VALUE`: optional overrides;
+- `--interactive`: enable interaction when supported; already-interactive predecessors keep it.
 
 ## Choosing models and effort
 

package/skills/agent-integration/references/continuation.md

@@ -17,7 +17,7 @@ When `status` or `collect` reports a timed-out turn with:
 run:
 
 ```sh
-delegate --root "$STATE_ROOT" continue \
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" continue \
   --task PREDECESSOR_TASK_ID \
   --brief FOLLOW_UP_BRIEF \
   --budget 30m \

package/skills/agent-integration/references/interaction.md

@@ -0,0 +1,110 @@
+# Questions and guidance during an active task
+
+Use this reference when an external task may need clarification or guidance
+before it finishes, including planning, consultation, and debate. Use only
+Delegate commands; do not open a terminal session with the provider.
+
+## Choose the transport
+
+Read the selected provider's `interaction` capability and
+`interaction.supported_modes`. `clarification` and `steering` are independent.
+Add `--interactive` to dispatch when the chosen provider and permission mode
+support the interaction you need. Keep the user's exact provider, model,
+effort, workspace, permission, and budget.
+
+Without a named provider, `dispatch --auto --interactive` selects among
+providers that support interaction. If a named provider does not support it,
+use ordinary bounded turns and `continue` for discussion when native
+continuation is available. Explain that questions cannot be answered during
+that provider's active turn. Never replace the requested provider silently.
+
+In the worker brief, explicitly request **native structured clarification**:
+"Use your native question tool to request clarification and await its answer.
+Do not ask only in final text or a progress message and then sleep: the lead
+can answer structured requests, not an unstructured prompt. If the question
+tool is unavailable, report that blocker and finish instead of waiting."
+Assign the actual work directly; the worker should not start another delegate.
+
+No change is needed for self-contained tasks that do not need mid-turn input.
+Continuation inherits interaction from its predecessor. `continue --task ID
+--interactive` can enable it for a previously ordinary task.
+
+## Respond to questions
+
+1. Save the admitted task ID and state root. Run `collect TASK_ID --watch 30s
+   --json` with that same root. Exit 3 is pending, not completion. Collection
+   returns early when there is an outstanding native question.
+2. Read `interaction.pending_requests`. Use the returned request ID, question
+   IDs, options, `multiple`, and `free_text`; do not invent or guess IDs.
+3. Answer facts already supplied by the user or verified during the task.
+   Ask the user for missing preferences, business decisions, new scope, or
+   authorization. Do not guess an answer to keep the worker moving. A question
+   reply does not approve tool access or change the task's permission mode.
+4. Write one JSON object mapping **every question ID in that request** to an
+   array of answer strings. A single-choice question takes one value. Use exact
+   option labels unless `free_text` permits another answer. Example:
+
+   ```json
+   {"color": ["Blue"], "sections": ["Summary", "Evidence"]}
+   ```
+
+5. Allocate a new 32-character lowercase hexadecimal message ID and save it
+   together with the exact answers. Submit once:
+
+   ```sh
+   delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" reply "$task_id" \
+     --request "$request_id" --answers "$answers_file" \
+     --id "$message_id" --json
+   ```
+
+6. Keep collecting. A receipt is not task completion. Only the task's
+   authoritative `outcome` establishes the final result.
+
+Use the same root expression for dispatch, collection, reply, and steering.
+If the user explicitly chose another root, use that exact path for all of
+them. The other variables stand for actual values saved from this task;
+reassign them in each shell invocation if your tool does not preserve variables.
+A pending response without questions means keep collecting, not inspect OS
+processes or dispatch a replacement task.
+
+## Guide active work
+
+When `interaction.capabilities.steering` is true and the task is active, put
+additional guidance in a brief file and submit:
+
+```sh
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" steer "$task_id" \
+  --brief "$guidance_file" --id "$message_id" --json
+```
+
+Steering is guidance to the existing turn. It does not change provider, model,
+effort, workspace, permission, or budget. Once the turn ends, use `continue`
+for another authorized turn instead. Do not cancel and redispatch simply to
+send guidance.
+
+## Interpret delivery precisely
+
+| State | What it establishes |
+| --- | --- |
+| `queued` | Delegate durably accepted the message; native delivery is pending. |
+| `unknown` | Delivery was reserved or attempted, but its result is uncertain. |
+| `sent` | The native input write completed; no native acknowledgement is established. |
+| `acknowledged` | A correlated native response acknowledged the message. This does not prove the model followed guidance. |
+| `rejected` | The message could not be delivered or the native protocol rejected it. Read the reason. |
+
+If a submission response was lost, repeat the **same message ID and exact
+content** to recover its receipt. Never allocate a fresh ID to retry an
+`unknown` delivery. Delegate will not replay an ambiguous native write.
+Changed content with an existing ID is a conflict. Observe the task instead
+of guessing whether it consumed the message.
+
+`events TASK_ID --after SEQUENCE --watch 30s --json` returns a bounded event
+page. Save the returned `cursor` for the next page. Events and interaction
+state are progress evidence, not a replacement for `collect` and `outcome`.
+
+Waiting for an answer consumes the original execution budget; it does not
+pause or extend it. `interaction.deadline` shows the recorded wall-clock
+boundary when execution has started. If the task ends while waiting, do not
+reply to the old request. Follow the ordinary continuation guidance only when
+Delegate reports the task resumable. Preserve any verified answers in the new
+brief; uncertain delivery must not become an automatic replay.

package/skills/agent-integration/references/result-handling.md

@@ -15,10 +15,13 @@ the same as a successful provider result.
    permits a linked successor even when no outcome has been published.
 4. If `publication` is committed and `outcome.verdict` is committed, return
    the result and its payload/evidence descriptors.
-5. If `outcome.verdict` is rejected, report a terminal rejected result with its
-   refusal and evidence. Do not retry automatically.
-6. Otherwise, if `liveness` is queued or running or publication is pending,
-   observe the same task again with a finite watch interval.
+5. If `outcome.verdict` is rejected and `failure.next_action` is `recheck`,
+   follow [offline recovery](#offline-recovery) once. Otherwise report the
+   terminal rejection with its diagnostic and evidence. Do not redispatch.
+6. If `pending.kind` is `interaction`, handle the returned questions using
+   [interaction guidance](interaction.md) before waiting again. Otherwise, if
+   `liveness` is queued or running or publication is pending, observe the same
+   task with a finite watch interval.
 7. If authentication is `blocked`, preserve that neutral state and tell the
    caller what prerequisite is unavailable without inventing a login command.
 8. Never relaunch an uncertain task by guessing. Observe only a task whose
@@ -42,3 +45,29 @@ checks static readiness; runtime verification requires the supervised probe.
 Versions and help are observations, not semantic-version allowlists.
 Authentication is a separate live gate (`not_run`, `passed`, or `blocked`).
 Report a blocked prerequisite without inventing a login command or retrying.
+
+## Offline recovery
+
+When Delegate returns `failure.next_action: "recheck"`, use the same root and
+exact task ID:
+
+```sh
+delegate --root "${DELEGATE_ROOT:-$HOME/delegation-state}" recheck TASK_ID --json
+```
+
+This command checks the already sealed evidence with an explicitly compatible
+new interpreter. It does not run the provider, contact a service, spend model
+tokens, or change the original outcome.
+
+- Exit `0` with `recheck.status: "recovered"`: use only the answer descriptor
+  returned for that recheck. Report that the answer was recovered from saved
+  evidence. The original `publication` remains `rejected`; do not call the
+  original task committed or use recovery to infer continuation permission.
+- Exit `4` with `recheck.status: "rejected"`: report the diagnostic and stop.
+- Exit `1` or `2`, or any response without a verified recovered answer:
+  report the error and stop. Do not read raw streams as a substitute answer,
+  invent a result path, run a provider directly, or launch a replacement task.
+
+Keep the assessment ID and source/target descriptors with the result. Do not
+loop on the same recheck. A separate user request is required for further
+troubleshooting or a new provider task.

package/skills/agent-integration/references/routing-rule.json

@@ -1 +1 @@
-"For work assigned to an external agent, model, or separate harness session, load the installed agent-integration skill before choosing tools and use Delegate. This includes research, implementation, testing, review, consultation, comparison, and debate, including external collaboration you choose. Discover destinations through Delegate. Preserve requested model IDs, effort, and permissions; do not guess ambiguous model identifiers. For task commands, set STATE_ROOT=\"${DELEGATE_ROOT:-$HOME/delegation-state}\" and pass --root \"$STATE_ROOT\" in the same shell invocation; reuse that root for dispatch, observation, and continuation, never an invented /tmp root. Keep each concurrent request and its response associated separately. Naming a CLI identifies the destination, not a request to bypass Delegate. Earlier assistant-generated direct-launch workflows do not override this preference. Native same-harness subagents remain available unless a separate session is requested. Respect an explicit user choice of another integration; otherwise do not substitute provider CLIs, plugins, or APIs after a Delegate blocker. A stop_and_report failure ends that attempt: report it without retrying, diagnosing setup, or offering a fallback. Continue other independent requests. Keep collection in the foreground until completion or a reported blocker unless the user requested detachment. Merely discussing or configuring a tool is not delegation."
+"For work assigned to an external agent, model, or separate harness session, load the installed agent-integration skill before choosing tools and use Delegate. This includes research, implementation, testing, review, consultation, comparison, and debate, including external collaboration you choose. Discover destinations through Delegate. Preserve requested model IDs, effort, and permissions; do not guess ambiguous model identifiers. For task commands, pass --root \"${DELEGATE_ROOT:-$HOME/delegation-state}\" directly in each shell invocation; reuse that root for dispatch, observation, and continuation, never an invented /tmp root. If the user explicitly requests another root, use that exact path consistently. Keep each concurrent request and its response associated separately. Capture each dispatch response and exit status once; after task_record created, observe that task ID. Never dispatch again merely to recover output or inspect its exit status. Naming a CLI identifies the destination, not a request to bypass Delegate. Earlier assistant-generated direct-launch workflows do not override this preference. Native same-harness subagents remain available unless a separate session is requested. Respect an explicit user choice of another integration; otherwise do not substitute provider CLIs, plugins, or APIs after a Delegate blocker. A stop_and_report failure ends that attempt: report it without retrying, diagnosing setup, or offering a fallback. Continue other independent requests. Keep collection in the foreground until completion or a reported blocker unless the user requested detachment. Pending work is observed through Delegate status, collect, or events, not OS process searches or supervisor commands. Merely discussing or configuring a tool is not delegation."

package/skills/agent-integration/scripts/delegate-routing-context.sh

@@ -9,4 +9,18 @@ script_dir=$(CDPATH= cd -- "$(dirname -- "$0")" && pwd)
 rule_file="$script_dir/../references/routing-rule.json"
 # Read before emitting anything so a missing asset cannot produce partial JSON.
 rule=$(cat "$rule_file")
-printf '{"hookSpecificOutput":{"hookEventName":"%s","additionalContext":%s}}\n' "$event" "$rule"
+skill_file="$(CDPATH= cd -- "$script_dir/.." && pwd)/SKILL.md"
+DELEGATE_ROUTING_EVENT="$event" DELEGATE_ROUTING_RULE="$rule" DELEGATE_ROUTING_SKILL="$skill_file" awk '
+BEGIN {
+    # Paths are task data, not shell code. Escape every JSON control character.
+    for (i = 1; i < 32; i++) escapes[sprintf("%c", i)] = sprintf("\\u%04x", i)
+    escapes["\\"] = "\\\\"
+    escapes["\""] = "\\\""
+    path = ENVIRON["DELEGATE_ROUTING_SKILL"]
+    for (i = 1; i <= length(path); i++) {
+        character = substr(path, i, 1)
+        escaped = escaped ((character in escapes) ? escapes[character] : character)
+    }
+    # The maintained rule is a JSON string. Keep its escaping when prefixing it.
+    printf "{\"hookSpecificOutput\":{\"hookEventName\":\"%s\",\"additionalContext\":\"Read the complete agent-integration skill at %s. %s}}\n", ENVIRON["DELEGATE_ROUTING_EVENT"], escaped, substr(ENVIRON["DELEGATE_ROUTING_RULE"], 2)
+}'