delegation-layer 0.1.29 → 0.1.32

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "delegation-layer",
-  "version": "0.1.29",
+  "version": "0.1.32",
   "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": "0.1.29",
+  "version": "0.1.32",
   "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": "0.1.29",
+  "version": "0.1.32",
   "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

@@ -112,7 +112,13 @@ tasks may proceed; the failed attempt must remain reported as failed or blocked.
    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.
+   fresh unrelated dispatch to replace a resumable task. The one other automatic
+   route is an `interrupted` task from `status` or `collect`, and only when
+   `failure.code` is `"task_interrupted"`, `failure.next_action` is
+   `"continue_task"`, and `continuation.resumable` is `true`; see
+   [interrupted turn](references/continuation.md#interrupted-turn). A
+   cancellation, `stop_and_report`, or failed admission always takes precedence
+   over `resumable: true`.
 8. Treat `blocked` authentication as unresolved. Do not call it compatible,
    successful, or failed provider behavior, and do not invent login commands.
 9. Never put credentials, tokens, or unrestricted provider output in a brief,
@@ -321,8 +327,8 @@ 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. If `interaction.pending_requests` is
+Check `status` first: if it is `timed_out` or `interrupted`, follow step 6 even
+when publication is pending or unknown 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
@@ -336,12 +342,15 @@ is committed. A rejected outcome is terminal evidence and must be reported as
 rejected. Payload and raw output are returned as validated descriptors; do not
 copy unrestricted provider output into the control response.
 
-### 6. Continue a timed out turn
+### 6. Continue a timed out or interrupted turn
 
-When `status` or `collect` reports `timed_out`, read
-[continuation and timeout recovery](references/continuation.md). If
-`continuation.resumable` is true, use `continue --task TASK_ID` with the original
-state root and a finite budget, then observe the returned successor ID. Do not
+When `status` or `collect` reports `timed_out` or `interrupted`, read
+[continuation and timeout recovery](references/continuation.md). Unless a
+cancellation or `stop_and_report` applies, a `timed_out` task with
+`continuation.resumable: true` may use `continue --task TASK_ID`. An
+`interrupted` task may use it only under all three conditions in rule 7;
+otherwise report it. When continuing, use the original state root and a finite
+budget, then observe the returned successor ID. Do not
 wait indefinitely for pending publication or restart the task as an unrelated
 dispatch. If continuation is unavailable, report its reason. Apply the same
 `failure` and `task_record` decisions as for dispatch; never repeat an uncertain
@@ -365,8 +374,8 @@ recruit more agents unless that is explicitly part of the task.
   model/effort options or another public command.
 - Read [result handling](references/result-handling.md) when interpreting
   admission, publication, authentication, errors, or exit codes.
-- Read [continuation](references/continuation.md) for timeout recovery or an
-  exact-session follow-up.
+- Read [continuation](references/continuation.md) for timeout or interruption
+  recovery or an exact-session follow-up.
 - Read [managed workspaces](references/workspaces.md) for optional Git
   worktrees, capture, review scope, release, and forgetting.
 - Use [read-only example](examples/read-only-task.md),

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

@@ -52,8 +52,8 @@ Dispatch flags:
 - `--model MODEL` and `--effort VALUE`: provider-advertised choices (CCR requires a model alias);
   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;
+  pass native effort flags; Pi maps effort to `--thinking`; OpenCode sends model
+  and effort as acknowledged ACP session options; 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.
 
@@ -135,7 +135,9 @@ enumeration, not effort metadata. `null` means effort metadata is unknown;
 `[]` means the provider explicitly reported no choices. Discovery exits `0`
 for `available` or `partial`, `2` for `blocked` or `unavailable`, and `1` for
 `failed`. Preserve a blocked or unavailable discovery result as unresolved;
-defaults remain allowed.
+defaults remain allowed. An operational discovery failure also includes an
+optional `failure` object and a fixed `error` message; keep both fields and
+follow `failure.next_action`.
 
 ## Checking readiness
 

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

@@ -61,11 +61,14 @@ with a fresh dispatch. A created successor is only observed.
 - `unsupported` cannot be continued safely.
 
 Do not use a guessed session ID, a provider-specific resume command, or a new
-independent dispatch to recover a timeout. The CLI accepts either a validated
-terminal outcome or a durable Pueue-ended observation backed by a guardian
-lifetime-termination record with independently proven empty scope for the
-session handoff; without one of those proofs it refuses continuation. Inspect
-the predecessor again rather than bypassing that boundary.
+independent dispatch to recover a timeout. The CLI accepts a validated
+terminal outcome, a durable Pueue-ended budget stop backed by a guardian
+lifetime-termination record with independently proven empty scope, or canonical
+closed lifetime evidence of any closure kind together with a free runner lease
+for a turn with no seal or outcome. A sealed predecessor must be collected
+first. A queue-row end or a timestamp alone never qualifies; without one of those
+proofs it refuses continuation. Inspect the predecessor again rather than
+bypassing that boundary.
 
 For a managed workspace timeout, capture is a new finite auxiliary operation
 after canonical closure. The normal path proves the provider scope empty and
@@ -76,6 +79,56 @@ unavailable capture can still preserve continuation evidence; it does not
 change the provider outcome.
 
 
+## Interrupted turn
+
+A host restart can end a provider turn before it publishes anything. Delegate
+reports an interruption only when the task's stored publication state is still
+unknown, with no provider exit or outcome, its canonical lifetime closure is
+proven, and its runner lease is released. A terminated budget stop is checked
+first and reports `timed_out` instead, as does a closure whose recorded
+termination cause is the budget. Closure proves only that the provider
+scope is gone; it is not provider success, and no answer or output from that
+turn is accepted. Do not check these facts yourself or infer eligibility from
+missing files: obey the returned `status`, `failure`, `continuation`, and
+`pending` fields. For an eligible task, after real outcomes and concrete errors
+are checked, `status` and `collect` report `status: "interrupted"` with
+`publication: "unknown"`, no `outcome`, and
+`failure.code: "task_interrupted"`. `status` exits `0`; `collect` exits `1`.
+An interruption may also carry `pending.kind: "supervisor-command"`; that
+independent observation means only that a supervisor command has not
+finished, so follow the interruption's `failure` and `continuation` as usual.
+Do not infer an interruption from `publication: "unknown"` alone: a response
+with `pending.kind: "publication"` is still pending publication and carries no
+interrupted resume offer, so handle it as pending under
+[result handling](result-handling.md). Outcomes, `timed_out`, interaction,
+`lifetime-finalization`, and ordinary pending exit `3` keep their meaning.
+
+| Response | Action |
+|---|---|
+| `failure.code: "task_interrupted"`, `failure.next_action: "continue_task"`, and `continuation.resumable: true` (all three) | Run the `continue` above for that predecessor with the original root and a finite budget, then observe the successor. |
+| `interrupted` with any other combination, including `failure.next_action: "stop_and_report"` or `continuation.resumable: false` | Report the reason and stop. |
+| `timed_out` with `failure.code: "task_cancelled"` (user stop and budget stop, no outcome) | `stop_and_report`; `continuation.resumable` is `false` and no `continue_command` is returned. Report it. `collect` exit `3` here is a finite watch, not success and not a reason to keep waiting. |
+| Any other operational error | It is not an interruption; apply [result handling](result-handling.md). |
+
+`continuation.resumable: true` alone never revives an attempt that returned
+`stop_and_report`, and a failed admission keeps its `task_record` rules.
+Never continue a cancelled task automatically or start a fresh provider
+fallback. A user stop keeps `interrupted` / `task_interrupted` for a closed
+interrupted task, now with `stop_and_report` and `resumable: false`. A
+committed or rejected outcome stays authoritative and is never replaced by
+`task_cancelled`. A later, explicit user request to `continue` is a separate
+action and still validated against real closure and the recorded session.
+
+The successor uses the predecessor's observed provider session from
+`provider.ref`. Only when `provider.ref` is genuinely absent, for any eligible
+predecessor and not only an interrupted one, does `continue` fall back to the
+validated immutable session that turn was asked to resume. The response reports
+the source in `continuation.session_source`: `observed` means `provider.ref`
+supplied the identity, and `recorded_input` means the requested session supplied
+it. Neither value is evidence the provider saved any message or resumed
+successfully. A missing task reference, or an unreadable, malformed, or mismatched
+record, refuses. Never search ancestors or rerun the original task.
+
 ## Follow up after normal completion
 
 A validated terminal outcome, committed or rejected, can also support a new

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

@@ -24,7 +24,13 @@ the same as a successful provider result.
    evidence for an existing task.
 3. If `status` is `timed_out`, follow the continuation procedure before treating
    pending publication as a reason to wait. `continuation.resumable: true`
-   permits a linked successor even when no outcome has been published.
+   permits a linked successor even when no outcome has been published. An
+   `interrupted` task (`failure.code: "task_interrupted"`) follows the table in
+   [interrupted turn](continuation.md#interrupted-turn): continue only when
+   `failure.next_action` is `continue_task` and `continuation.resumable` is
+   `true`, otherwise report it. A `timed_out` task whose `failure.code` is
+   `task_cancelled` is `stop_and_report`: report it and stop, even though
+   `collect` exits `3`; that exit is not success.
 4. If `publication` is committed and `outcome.verdict` is committed, return
    the result and its payload/evidence descriptors.
 5. If `outcome.verdict` is rejected and `failure.next_action` is `recheck`,
@@ -51,7 +57,13 @@ from an inspection-worker or supervisor-observation failure.
 For `delegate models`, validate the complete JSON envelope before using any
 fact. Treat `available` and `partial` with exit `0` as observations, preserve
 `blocked` and `unavailable` with exit `2` as unresolved prerequisites, and
-preserve `failed` with exit `1` as a discovery failure. `complete` refers to
+preserve `failed` as a discovery failure. Exit `1` with `failed` is an
+operational failure; exit `2` with `failed` is an invalid or unsupported
+request, not an authentication block. Status and exit are separate
+observations. When the response
+includes the optional `failure` object and `error` message, keep both in the
+report and follow `failure.next_action`; report and stop on
+`stop_and_report`. `complete` refers to
 model enumeration. Keep `efforts: null` distinct from `efforts: []`, and do not
 apply harness-wide effort choices to every model. Discovery is advisory and
 never an admission allowlist; a provider default remains valid and a listed

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

@@ -48,7 +48,10 @@ delegate --root ROOT dispatch --isolation git-worktree --cwd ABS \
 `--base` resolves to a commit; the default is `HEAD`. Dirty and untracked
 source changes are observed but excluded from the managed checkout. The
 provider runs in the retained detached worktree, with the source-relative
-working subdirectory preserved. A request that attaches to an existing
+working subdirectory preserved. The source may itself be a managed checkout,
+such as a sibling in the same workspace container. After provisioning, the new
+workspace depends on the shared repository, not on the source checkout, so
+removing that source does not block capture, release, or forget. A request that attaches to an existing
 workspace uses its recorded checkout and source identity:
 
 ```text
@@ -117,6 +120,36 @@ facts. Normal capture covers the initial committed base through the captured
 tree, including worker commits and added untracked files that belong to the
 Git-content view. It uses a private temporary Git index and records an owned
 base pin plus tree/snapshot reference.
+
+Each provider turn also records its actual starting tree, `HEAD`, and index
+identity just before launch, so Lead edits made between turns belong to the
+next turn's start. The index identity covers every entry's mode, object, stage
+(including conflicts), path, assume-unchanged or skip-worktree flag, and
+intent-to-add (`git add -N`) state, but not stat caches or `HEAD`. A task capture's `candidate.turn` compares that start with the
+final capture: `content`, `head`, and `index` are `changed` or `unchanged`, and
+`baseline_ref` keeps the start tree. This is observed evidence, not
+enforcement: a read-only turn that commits or edits is reported, not rejected,
+and writes reverted before capture or ignored files are not observed.
+`status: "unavailable"` with `reason_code: "baseline_unavailable"` means no
+baseline exists, for example for a turn started before this evidence existed;
+treat it as unknown, never as unchanged. A capture without `candidate.turn`,
+such as one recorded before per-turn evidence existed or a release or exec
+recapture, is also unknown, never unchanged. A failed baseline refuses the
+provider launch for that task.
+
+An observed turn adds `turn-diff.patch` and `turn-name-status.z`, which diff
+that turn's start tree to the final tree with the same bounds as the
+cumulative artifacts. Use them to see which files the current turn touched;
+they exclude earlier turns and Lead edits made before its launch. Use the
+cumulative `diff.patch` and `name-status.z`, from the workspace base, to land
+or review the whole delegated result. Report a read-only turn whose `content`,
+`head`, or `index` is `changed` as a mutation, even when its turn file list is
+empty, and report an unavailable turn as unknown.
+
+`status` and `collect` for a task report that task's own capture, even after a
+later turn has used the workspace. The workspace `state` still reflects the
+resource now, so a sealed earlier capture does not make a busy, released, or
+forgotten workspace usable. `workspace status` reports the latest use.
 Optional binary patch and name-status artifacts are bounded to 8 MiB each; an
 overflow is omitted, never silently truncated. Verification diagnostics retain
 an 8 MiB prefix per stream, mark truncation explicitly, and continue draining
@@ -146,9 +179,10 @@ identity check retains the checkout and evidence for retry. `--discard` accepts
 loss of uncaptured checkout content; it does not waive native closure,
 ownership, or identity proof.
 
-Release leaves the pinned refs/blobs and capture receipts as evidence. After a
-proven release, `workspace forget ID` writes an exact inventory intent before
-removing only matching owned refs and optional artifacts by identity/CAS. It
+Release leaves the pinned refs/blobs, turn-baseline refs, and capture receipts
+as evidence. After a proven release, `workspace forget ID` writes an exact
+inventory intent before removing only matching owned refs and optional
+artifacts by identity/CAS. It
 retains capture receipts, tombstones, all provider outcomes, raw evidence, and
 payload descriptors. Readers should label an intentionally forgotten item as
 `forgotten`, not as corruption.

package/skills/agent-integration/SKILL.md

@@ -112,7 +112,13 @@ tasks may proceed; the failed attempt must remain reported as failed or blocked.
    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.
+   fresh unrelated dispatch to replace a resumable task. The one other automatic
+   route is an `interrupted` task from `status` or `collect`, and only when
+   `failure.code` is `"task_interrupted"`, `failure.next_action` is
+   `"continue_task"`, and `continuation.resumable` is `true`; see
+   [interrupted turn](references/continuation.md#interrupted-turn). A
+   cancellation, `stop_and_report`, or failed admission always takes precedence
+   over `resumable: true`.
 8. Treat `blocked` authentication as unresolved. Do not call it compatible,
    successful, or failed provider behavior, and do not invent login commands.
 9. Never put credentials, tokens, or unrestricted provider output in a brief,
@@ -321,8 +327,8 @@ 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. If `interaction.pending_requests` is
+Check `status` first: if it is `timed_out` or `interrupted`, follow step 6 even
+when publication is pending or unknown 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
@@ -336,12 +342,15 @@ is committed. A rejected outcome is terminal evidence and must be reported as
 rejected. Payload and raw output are returned as validated descriptors; do not
 copy unrestricted provider output into the control response.
 
-### 6. Continue a timed out turn
+### 6. Continue a timed out or interrupted turn
 
-When `status` or `collect` reports `timed_out`, read
-[continuation and timeout recovery](references/continuation.md). If
-`continuation.resumable` is true, use `continue --task TASK_ID` with the original
-state root and a finite budget, then observe the returned successor ID. Do not
+When `status` or `collect` reports `timed_out` or `interrupted`, read
+[continuation and timeout recovery](references/continuation.md). Unless a
+cancellation or `stop_and_report` applies, a `timed_out` task with
+`continuation.resumable: true` may use `continue --task TASK_ID`. An
+`interrupted` task may use it only under all three conditions in rule 7;
+otherwise report it. When continuing, use the original state root and a finite
+budget, then observe the returned successor ID. Do not
 wait indefinitely for pending publication or restart the task as an unrelated
 dispatch. If continuation is unavailable, report its reason. Apply the same
 `failure` and `task_record` decisions as for dispatch; never repeat an uncertain
@@ -365,8 +374,8 @@ recruit more agents unless that is explicitly part of the task.
   model/effort options or another public command.
 - Read [result handling](references/result-handling.md) when interpreting
   admission, publication, authentication, errors, or exit codes.
-- Read [continuation](references/continuation.md) for timeout recovery or an
-  exact-session follow-up.
+- Read [continuation](references/continuation.md) for timeout or interruption
+  recovery or an exact-session follow-up.
 - Read [managed workspaces](references/workspaces.md) for optional Git
   worktrees, capture, review scope, release, and forgetting.
 - Use [read-only example](examples/read-only-task.md),

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

@@ -52,8 +52,8 @@ Dispatch flags:
 - `--model MODEL` and `--effort VALUE`: provider-advertised choices (CCR requires a model alias);
   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;
+  pass native effort flags; Pi maps effort to `--thinking`; OpenCode sends model
+  and effort as acknowledged ACP session options; 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.
 
@@ -135,7 +135,9 @@ enumeration, not effort metadata. `null` means effort metadata is unknown;
 `[]` means the provider explicitly reported no choices. Discovery exits `0`
 for `available` or `partial`, `2` for `blocked` or `unavailable`, and `1` for
 `failed`. Preserve a blocked or unavailable discovery result as unresolved;
-defaults remain allowed.
+defaults remain allowed. An operational discovery failure also includes an
+optional `failure` object and a fixed `error` message; keep both fields and
+follow `failure.next_action`.
 
 ## Checking readiness
 

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

@@ -61,11 +61,14 @@ with a fresh dispatch. A created successor is only observed.
 - `unsupported` cannot be continued safely.
 
 Do not use a guessed session ID, a provider-specific resume command, or a new
-independent dispatch to recover a timeout. The CLI accepts either a validated
-terminal outcome or a durable Pueue-ended observation backed by a guardian
-lifetime-termination record with independently proven empty scope for the
-session handoff; without one of those proofs it refuses continuation. Inspect
-the predecessor again rather than bypassing that boundary.
+independent dispatch to recover a timeout. The CLI accepts a validated
+terminal outcome, a durable Pueue-ended budget stop backed by a guardian
+lifetime-termination record with independently proven empty scope, or canonical
+closed lifetime evidence of any closure kind together with a free runner lease
+for a turn with no seal or outcome. A sealed predecessor must be collected
+first. A queue-row end or a timestamp alone never qualifies; without one of those
+proofs it refuses continuation. Inspect the predecessor again rather than
+bypassing that boundary.
 
 For a managed workspace timeout, capture is a new finite auxiliary operation
 after canonical closure. The normal path proves the provider scope empty and
@@ -76,6 +79,56 @@ unavailable capture can still preserve continuation evidence; it does not
 change the provider outcome.
 
 
+## Interrupted turn
+
+A host restart can end a provider turn before it publishes anything. Delegate
+reports an interruption only when the task's stored publication state is still
+unknown, with no provider exit or outcome, its canonical lifetime closure is
+proven, and its runner lease is released. A terminated budget stop is checked
+first and reports `timed_out` instead, as does a closure whose recorded
+termination cause is the budget. Closure proves only that the provider
+scope is gone; it is not provider success, and no answer or output from that
+turn is accepted. Do not check these facts yourself or infer eligibility from
+missing files: obey the returned `status`, `failure`, `continuation`, and
+`pending` fields. For an eligible task, after real outcomes and concrete errors
+are checked, `status` and `collect` report `status: "interrupted"` with
+`publication: "unknown"`, no `outcome`, and
+`failure.code: "task_interrupted"`. `status` exits `0`; `collect` exits `1`.
+An interruption may also carry `pending.kind: "supervisor-command"`; that
+independent observation means only that a supervisor command has not
+finished, so follow the interruption's `failure` and `continuation` as usual.
+Do not infer an interruption from `publication: "unknown"` alone: a response
+with `pending.kind: "publication"` is still pending publication and carries no
+interrupted resume offer, so handle it as pending under
+[result handling](result-handling.md). Outcomes, `timed_out`, interaction,
+`lifetime-finalization`, and ordinary pending exit `3` keep their meaning.
+
+| Response | Action |
+|---|---|
+| `failure.code: "task_interrupted"`, `failure.next_action: "continue_task"`, and `continuation.resumable: true` (all three) | Run the `continue` above for that predecessor with the original root and a finite budget, then observe the successor. |
+| `interrupted` with any other combination, including `failure.next_action: "stop_and_report"` or `continuation.resumable: false` | Report the reason and stop. |
+| `timed_out` with `failure.code: "task_cancelled"` (user stop and budget stop, no outcome) | `stop_and_report`; `continuation.resumable` is `false` and no `continue_command` is returned. Report it. `collect` exit `3` here is a finite watch, not success and not a reason to keep waiting. |
+| Any other operational error | It is not an interruption; apply [result handling](result-handling.md). |
+
+`continuation.resumable: true` alone never revives an attempt that returned
+`stop_and_report`, and a failed admission keeps its `task_record` rules.
+Never continue a cancelled task automatically or start a fresh provider
+fallback. A user stop keeps `interrupted` / `task_interrupted` for a closed
+interrupted task, now with `stop_and_report` and `resumable: false`. A
+committed or rejected outcome stays authoritative and is never replaced by
+`task_cancelled`. A later, explicit user request to `continue` is a separate
+action and still validated against real closure and the recorded session.
+
+The successor uses the predecessor's observed provider session from
+`provider.ref`. Only when `provider.ref` is genuinely absent, for any eligible
+predecessor and not only an interrupted one, does `continue` fall back to the
+validated immutable session that turn was asked to resume. The response reports
+the source in `continuation.session_source`: `observed` means `provider.ref`
+supplied the identity, and `recorded_input` means the requested session supplied
+it. Neither value is evidence the provider saved any message or resumed
+successfully. A missing task reference, or an unreadable, malformed, or mismatched
+record, refuses. Never search ancestors or rerun the original task.
+
 ## Follow up after normal completion
 
 A validated terminal outcome, committed or rejected, can also support a new

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

@@ -24,7 +24,13 @@ the same as a successful provider result.
    evidence for an existing task.
 3. If `status` is `timed_out`, follow the continuation procedure before treating
    pending publication as a reason to wait. `continuation.resumable: true`
-   permits a linked successor even when no outcome has been published.
+   permits a linked successor even when no outcome has been published. An
+   `interrupted` task (`failure.code: "task_interrupted"`) follows the table in
+   [interrupted turn](continuation.md#interrupted-turn): continue only when
+   `failure.next_action` is `continue_task` and `continuation.resumable` is
+   `true`, otherwise report it. A `timed_out` task whose `failure.code` is
+   `task_cancelled` is `stop_and_report`: report it and stop, even though
+   `collect` exits `3`; that exit is not success.
 4. If `publication` is committed and `outcome.verdict` is committed, return
    the result and its payload/evidence descriptors.
 5. If `outcome.verdict` is rejected and `failure.next_action` is `recheck`,
@@ -51,7 +57,13 @@ from an inspection-worker or supervisor-observation failure.
 For `delegate models`, validate the complete JSON envelope before using any
 fact. Treat `available` and `partial` with exit `0` as observations, preserve
 `blocked` and `unavailable` with exit `2` as unresolved prerequisites, and
-preserve `failed` with exit `1` as a discovery failure. `complete` refers to
+preserve `failed` as a discovery failure. Exit `1` with `failed` is an
+operational failure; exit `2` with `failed` is an invalid or unsupported
+request, not an authentication block. Status and exit are separate
+observations. When the response
+includes the optional `failure` object and `error` message, keep both in the
+report and follow `failure.next_action`; report and stop on
+`stop_and_report`. `complete` refers to
 model enumeration. Keep `efforts: null` distinct from `efforts: []`, and do not
 apply harness-wide effort choices to every model. Discovery is advisory and
 never an admission allowlist; a provider default remains valid and a listed

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

@@ -48,7 +48,10 @@ delegate --root ROOT dispatch --isolation git-worktree --cwd ABS \
 `--base` resolves to a commit; the default is `HEAD`. Dirty and untracked
 source changes are observed but excluded from the managed checkout. The
 provider runs in the retained detached worktree, with the source-relative
-working subdirectory preserved. A request that attaches to an existing
+working subdirectory preserved. The source may itself be a managed checkout,
+such as a sibling in the same workspace container. After provisioning, the new
+workspace depends on the shared repository, not on the source checkout, so
+removing that source does not block capture, release, or forget. A request that attaches to an existing
 workspace uses its recorded checkout and source identity:
 
 ```text
@@ -117,6 +120,36 @@ facts. Normal capture covers the initial committed base through the captured
 tree, including worker commits and added untracked files that belong to the
 Git-content view. It uses a private temporary Git index and records an owned
 base pin plus tree/snapshot reference.
+
+Each provider turn also records its actual starting tree, `HEAD`, and index
+identity just before launch, so Lead edits made between turns belong to the
+next turn's start. The index identity covers every entry's mode, object, stage
+(including conflicts), path, assume-unchanged or skip-worktree flag, and
+intent-to-add (`git add -N`) state, but not stat caches or `HEAD`. A task capture's `candidate.turn` compares that start with the
+final capture: `content`, `head`, and `index` are `changed` or `unchanged`, and
+`baseline_ref` keeps the start tree. This is observed evidence, not
+enforcement: a read-only turn that commits or edits is reported, not rejected,
+and writes reverted before capture or ignored files are not observed.
+`status: "unavailable"` with `reason_code: "baseline_unavailable"` means no
+baseline exists, for example for a turn started before this evidence existed;
+treat it as unknown, never as unchanged. A capture without `candidate.turn`,
+such as one recorded before per-turn evidence existed or a release or exec
+recapture, is also unknown, never unchanged. A failed baseline refuses the
+provider launch for that task.
+
+An observed turn adds `turn-diff.patch` and `turn-name-status.z`, which diff
+that turn's start tree to the final tree with the same bounds as the
+cumulative artifacts. Use them to see which files the current turn touched;
+they exclude earlier turns and Lead edits made before its launch. Use the
+cumulative `diff.patch` and `name-status.z`, from the workspace base, to land
+or review the whole delegated result. Report a read-only turn whose `content`,
+`head`, or `index` is `changed` as a mutation, even when its turn file list is
+empty, and report an unavailable turn as unknown.
+
+`status` and `collect` for a task report that task's own capture, even after a
+later turn has used the workspace. The workspace `state` still reflects the
+resource now, so a sealed earlier capture does not make a busy, released, or
+forgotten workspace usable. `workspace status` reports the latest use.
 Optional binary patch and name-status artifacts are bounded to 8 MiB each; an
 overflow is omitted, never silently truncated. Verification diagnostics retain
 an 8 MiB prefix per stream, mark truncation explicitly, and continue draining
@@ -146,9 +179,10 @@ identity check retains the checkout and evidence for retry. `--discard` accepts
 loss of uncaptured checkout content; it does not waive native closure,
 ownership, or identity proof.
 
-Release leaves the pinned refs/blobs and capture receipts as evidence. After a
-proven release, `workspace forget ID` writes an exact inventory intent before
-removing only matching owned refs and optional artifacts by identity/CAS. It
+Release leaves the pinned refs/blobs, turn-baseline refs, and capture receipts
+as evidence. After a proven release, `workspace forget ID` writes an exact
+inventory intent before removing only matching owned refs and optional
+artifacts by identity/CAS. It
 retains capture receipts, tombstones, all provider outcomes, raw evidence, and
 payload descriptors. Readers should label an intentionally forgotten item as
 `forgotten`, not as corruption.