crewly 1.20.192 → 1.20.194

package/config/skills/agent/core/todo-add/execute.test.sh

@@ -33,6 +33,8 @@ class H(BaseHTTPRequestHandler):
             out = {'list': LIST, 'count': 1, 'tasks': [{'id': 'T1', 'title': 'Milk', 'status': 'notStarted', 'due': '2026-10-01'}]}
         elif path == '/api/microsoft-todo/tasks':
             out = {'list': LIST, 'task': {'id': 'T2', 'title': (data or {}).get('title'), 'status': 'notStarted'}}
+            if (data or {}).get('title') == 'partial':
+                out['failedSteps'] = [{'step': 'Milk', 'error': 'Forbidden: no'}]
         elif method == 'PATCH':
             out = {'list': LIST, 'task': {'id': path.rsplit('/', 1)[1], 'title': 'Milk', 'status': 'completed'}}
         elif method == 'DELETE':
@@ -69,6 +71,18 @@ run '{"list":"Groceries","title":"Eggs"}' >/dev/null
 check "json input" "$(last '.body')" '{"title":"Eggs","list":"Groceries"}'
 check "title required" "$(run_err --list Work | jq -r .error)" "--title is required"
 check "bad due" "$(run_err --title x --due tomorrow | jq -r .error)" "--due must look like 2026-10-01"
+run --title "Costco" --steps "Eggs, Milk,,Bread " >/dev/null
+check "steps: comma list, trimmed, empties dropped" "$(last '.body')" '{"title":"Costco","steps":["Eggs","Milk","Bread"]}'
+run --title "Costco" --step "Eggs, large" --step Milk >/dev/null
+check "steps: repeated --step keeps commas" "$(last '.body.steps')" '["Eggs, large","Milk"]'
+run '{"title":"Costco","steps":["Eggs","Milk"]}' >/dev/null
+check "steps: json array" "$(last '.body.steps')" '["Eggs","Milk"]'
+run '{"title":"Costco","steps":"Eggs,Milk"}' >/dev/null
+check "steps: json string" "$(last '.body.steps')" '["Eggs","Milk"]'
+run --title "Plain" --steps " , " >/dev/null
+check "steps: none left means no steps field" "$(last '.body')" '{"title":"Plain"}'
+OUT=$(run --title partial --steps "Eggs,Milk")
+check "steps: failed steps are reported" "$(printf '%s' "$OUT" | jq -c '[.success,.failedSteps]')" '[true,[{"step":"Milk","error":"Forbidden: no"}]]'
 OUT=$(run --title no-grant || true)
 check "not connected" "$(printf '%s' "$OUT" | jq -r .reason)" "not_connected"
 

package/config/skills/agent/core/todo-tasks/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: To Do Tasks
-description: List the tasks in one of the owner's Microsoft To Do lists — open tasks by default, --all to include completed ones (via the Microsoft grant held by Crewly Cloud). Read-only.
-version: 1.0.0
+description: List the tasks in one of the owner's Microsoft To Do lists — open tasks by default, --all to include completed ones, with each task's steps (via the Microsoft grant held by Crewly Cloud). Read-only.
+version: 1.1.0
 category: productivity
 skillType: claude-skill
 assignableRoles:
@@ -46,7 +46,8 @@ bash execute.sh --list Work --limit 20 # default 50, max 100
 ```
 
 `importance` appears only when not normal; `note` is plain text, cut at
-200 characters. `hasMore: true` means the list has more tasks than the
+200 characters. `steps` (checklist items, in To Do order, each with `id`,
+`title` and `checked`) appears only on tasks that have steps. `hasMore: true` means the list has more tasks than the
 limit. Use the task `id` with todo-update.
 
 ## Failures

package/config/skills/agent/core/todo-update/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: To Do Update
-description: Complete, reopen, retitle, re-date or delete a task in one of the owner's Microsoft To Do lists (via the Microsoft grant held by Crewly Cloud).
-version: 1.0.0
+description: Complete, reopen, retitle, re-date or delete a task — and add, tick, untick or remove its steps — in one of the owner's Microsoft To Do lists (via the Microsoft grant held by Crewly Cloud).
+version: 1.1.0
 category: productivity
 skillType: claude-skill
 assignableRoles:
@@ -38,8 +38,18 @@ bash execute.sh --list Work --task <id> --title "Send v4 deck" --due 2026-10-03
 bash execute.sh --list Work --task <id> --due none     # clear the due date
 bash execute.sh --list Work --task <id> --reopen
 bash execute.sh --list Work --task <id> --delete       # cannot be undone
+bash execute.sh --list Groceries --task <id> --add-steps "Butter,Ham"
+bash execute.sh --list Groceries --task <id> --check-steps "Eggs,Milk"    # tick steps
+bash execute.sh --list Groceries --task <id> --uncheck-steps Eggs         # untick
+bash execute.sh --list Groceries --task <id> --remove-steps Bread         # delete a step
 ```
 
+Steps are named by title (case-insensitive) or by step id from todo-tasks.
+`--add-step` / `--check-step` / `--uncheck-step` / `--remove-step` take one
+step each and can be repeated, for a step whose title contains a comma.
+Adding steps changes the existing task; it never creates a new one. A step
+name that matches nothing (or two steps) fails before anything is changed.
+
 `--task` is the id from todo-tasks. `--list` is the list the task is in
 (name or id; the default list when omitted). Changes can be combined in one
 call; `--delete` ignores the others.
@@ -48,6 +58,7 @@ call; `--delete` ignores the others.
 
 ```json
 {"success":true,"list":"Groceries","task":{"id":"AAMk…","title":"Milk","status":"completed","completedAt":"2026-09-23"}}
+{"success":true,"list":"Groceries","task":{"id":"AAMk…","title":"Costco","status":"notStarted","steps":[{"id":"…","title":"Eggs","checked":true},{"id":"…","title":"Butter","checked":false}]}}
 {"success":true,"deleted":true,"list":"Work","taskId":"AAMk…"}
 ```
 

package/config/skills/agent/core/todo-update/execute.sh

@@ -10,6 +10,7 @@
 #   bash execute.sh --list Work --task <taskId> --due none          # clear the due date
 #   bash execute.sh --list Work --task <taskId> --reopen
 #   bash execute.sh --list Work --task <taskId> --delete
+#   bash execute.sh --list Groceries --task <taskId> --add-steps "Butter,Ham" --check-steps "Eggs"
 #   bash execute.sh '{"list":"Work","task":"AAMk…","complete":true}'
 # =============================================================================
 set -euo pipefail
@@ -24,6 +25,7 @@ Usage:
   bash execute.sh --list Work --task <taskId> --due none          # clear the due date
   bash execute.sh --list Work --task <taskId> --reopen
   bash execute.sh --list Work --task <taskId> --delete
+  bash execute.sh --list Groceries --task <taskId> --add-steps "Butter,Ham" --check-steps "Eggs"
   bash execute.sh '{"list":"Work","task":"AAMk…","complete":true}'
 
 Options:
@@ -36,6 +38,12 @@ Options:
   --note         Replace the note
   --importance   low | normal | high
   --delete       Delete the task (cannot be undone)
+  --add-steps    Comma-separated steps to append to the task
+  --check-steps  Comma-separated steps to tick, by title (case-insensitive) or step id
+  --uncheck-steps  Comma-separated steps to untick, by title or id
+  --remove-steps Comma-separated steps to delete, by title or id
+  --add-step, --check-step, --uncheck-step, --remove-step
+                 One step each; repeat for more (use when a step contains a comma)
   --help | -h    Show this help
 EOF_USAGE
 }
@@ -48,6 +56,17 @@ fail_from() {
 
 uri() { jq -rn --arg v "$1" '$v|@uri'; }
 
+# Append step titles / ids to a JSON array: $1 = array, $2 = text, $3 =
+# "split" to split the text on commas. Trimmed; empty ones are dropped.
+add_steps() {
+  jq -c --arg v "$2" --arg mode "${3:-}" '. + (if $mode == "split" then ($v | split(",")) else [$v] end | map(gsub("^\\s+|\\s+$"; "")) | map(select(length > 0)))' <<<"$1"
+}
+
+# Step list from the JSON input: $1 = key; an array or a comma string.
+input_steps() {
+  printf '%s' "$INPUT" | jq -c --arg k "$1" '.[$k] // [] | if type == "string" then split(",") else map(tostring) end | map(gsub("^\\s+|\\s+$"; "")) | map(select(length > 0))'
+}
+
 # api_call may print a one-line warning to stderr (no CREWLY_SESSION_NAME);
 # the backend answer is always the last line. On failure print the mapped
 # failure JSON (fail_from) and return 1.
@@ -63,6 +82,7 @@ if [[ $# -gt 0 && ${1:0:1} == '{' ]]; then
   shift || true
 fi
 TASK=""; LIST=""; COMPLETE=""; TITLE=""; DUE=""; NOTE=""; IMPORTANCE=""; DELETE=""
+ADD_STEPS='[]'; CHECK_STEPS='[]'; UNCHECK_STEPS='[]'; REMOVE_STEPS='[]'
 while [[ $# -gt 0 ]]; do
   case "$1" in
     --task)       [ $# -ge 2 ] || error_exit "--task requires a value";       TASK="$2";       shift 2 ;;
@@ -74,6 +94,14 @@ while [[ $# -gt 0 ]]; do
     --note)       [ $# -ge 2 ] || error_exit "--note requires a value";       NOTE="$2";       shift 2 ;;
     --importance) [ $# -ge 2 ] || error_exit "--importance requires a value"; IMPORTANCE="$2"; shift 2 ;;
     --delete)     DELETE="1"; shift ;;
+    --add-steps)     [ $# -ge 2 ] || error_exit "--add-steps requires a value";     ADD_STEPS=$(add_steps "$ADD_STEPS" "$2" split);         shift 2 ;;
+    --add-step)      [ $# -ge 2 ] || error_exit "--add-step requires a value";      ADD_STEPS=$(add_steps "$ADD_STEPS" "$2");               shift 2 ;;
+    --check-steps)   [ $# -ge 2 ] || error_exit "--check-steps requires a value";   CHECK_STEPS=$(add_steps "$CHECK_STEPS" "$2" split);     shift 2 ;;
+    --check-step)    [ $# -ge 2 ] || error_exit "--check-step requires a value";    CHECK_STEPS=$(add_steps "$CHECK_STEPS" "$2");           shift 2 ;;
+    --uncheck-steps) [ $# -ge 2 ] || error_exit "--uncheck-steps requires a value"; UNCHECK_STEPS=$(add_steps "$UNCHECK_STEPS" "$2" split); shift 2 ;;
+    --uncheck-step)  [ $# -ge 2 ] || error_exit "--uncheck-step requires a value";  UNCHECK_STEPS=$(add_steps "$UNCHECK_STEPS" "$2");       shift 2 ;;
+    --remove-steps)  [ $# -ge 2 ] || error_exit "--remove-steps requires a value";  REMOVE_STEPS=$(add_steps "$REMOVE_STEPS" "$2" split);   shift 2 ;;
+    --remove-step)   [ $# -ge 2 ] || error_exit "--remove-step requires a value";   REMOVE_STEPS=$(add_steps "$REMOVE_STEPS" "$2");         shift 2 ;;
     --help|-h)    print_usage; exit 0 ;;
     *) error_exit "Unknown option: $1" ;;
   esac
@@ -88,6 +116,10 @@ if [ -n "$INPUT_JSON" ]; then
   [ -z "$NOTE" ]       && NOTE=$(printf '%s' "$INPUT" | jq -r '.note // empty')
   [ -z "$IMPORTANCE" ] && IMPORTANCE=$(printf '%s' "$INPUT" | jq -r '.importance // empty')
   [ -z "$DELETE" ]     && DELETE=$(printf '%s' "$INPUT" | jq -r 'if .delete == true then "1" else empty end')
+  [ "$ADD_STEPS" = '[]' ]     && ADD_STEPS=$(input_steps addSteps)
+  [ "$CHECK_STEPS" = '[]' ]   && CHECK_STEPS=$(input_steps checkSteps)
+  [ "$UNCHECK_STEPS" = '[]' ] && UNCHECK_STEPS=$(input_steps uncheckSteps)
+  [ "$REMOVE_STEPS" = '[]' ]  && REMOVE_STEPS=$(input_steps removeSteps)
 fi
 [ -n "$TASK" ] || error_exit "--task is required (ids come from todo-tasks)"
 if [ -n "$DELETE" ]; then
@@ -100,16 +132,22 @@ fi
 if [ -n "$DUE" ] && [ "$DUE" != "none" ] && ! [[ "$DUE" =~ ^[0-9]{4}-[0-9]{2}-[0-9]{2}$ ]]; then
   error_exit "--due must look like 2026-10-01, or none to clear it"
 fi
-if [ -z "$COMPLETE" ] && [ -z "$TITLE" ] && [ -z "$DUE" ] && [ -z "$NOTE" ] && [ -z "$IMPORTANCE" ]; then
-  error_exit "nothing to change: give --complete, --reopen, --title, --due, --note, --importance or --delete"
+STEP_CHANGES="${ADD_STEPS}${CHECK_STEPS}${UNCHECK_STEPS}${REMOVE_STEPS}"
+if [ -z "$COMPLETE" ] && [ -z "$TITLE" ] && [ -z "$DUE" ] && [ -z "$NOTE" ] && [ -z "$IMPORTANCE" ] && [ "$STEP_CHANGES" = '[][][][]' ]; then
+  error_exit "nothing to change: give --complete, --reopen, --title, --due, --note, --importance, a step change or --delete"
 fi
 BODY=$(jq -cn --arg list "$LIST" --arg complete "$COMPLETE" --arg title "$TITLE" --arg due "$DUE" --arg note "$NOTE" --arg imp "$IMPORTANCE" \
+  --argjson add "$ADD_STEPS" --argjson check "$CHECK_STEPS" --argjson uncheck "$UNCHECK_STEPS" --argjson remove "$REMOVE_STEPS" \
   '{}
       + (if $list != "" then {list: $list} else {} end)
       + (if $complete != "" then {complete: ($complete == "true")} else {} end)
       + (if $title != "" then {title: $title} else {} end)
       + (if $due == "none" then {due: null} elif $due != "" then {due: $due} else {} end)
       + (if $note != "" then {note: $note} else {} end)
-      + (if $imp != "" then {importance: $imp} else {} end)')
+      + (if $imp != "" then {importance: $imp} else {} end)
+      + (if ($add | length) > 0 then {addSteps: $add} else {} end)
+      + (if ($check | length) > 0 then {checkSteps: $check} else {} end)
+      + (if ($uncheck | length) > 0 then {uncheckSteps: $uncheck} else {} end)
+      + (if ($remove | length) > 0 then {removeSteps: $remove} else {} end)')
 RESPONSE=$(call PATCH "/microsoft-todo/tasks/$(uri "$TASK")" "$BODY") || { printf '%s\n' "$RESPONSE"; exit 1; }
 printf '%s' "$RESPONSE" | jq -c '{success: true, list: .data.list.name, task: .data.task}'

package/config/skills/agent/core/todo-update/execute.test.sh

@@ -75,6 +75,13 @@ check "json input" "$(last '[.path,.body]')" '["/api/microsoft-todo/tasks/T7",{"
 check "task required" "$(run_err --complete | jq -r .error)" "--task is required (ids come from todo-tasks)"
 check "nothing to change" "$(run_err --task T1 | jq -r .error | cut -c1-17)" "nothing to change"
 check "bad due" "$(run_err --task T1 --due 1/10 | jq -r .error)" "--due must look like 2026-10-01, or none to clear it"
+run --list Groceries --task T1 --add-steps "Butter, Ham" --check-steps Eggs --uncheck-step "Milk, oat" --remove-steps Bread >/dev/null
+check "steps: body" "$(last '[.method,.path,.body]')" '["PATCH","/api/microsoft-todo/tasks/T1",{"list":"Groceries","addSteps":["Butter","Ham"],"checkSteps":["Eggs"],"uncheckSteps":["Milk, oat"],"removeSteps":["Bread"]}]'
+run --task T1 --add-step Butter >/dev/null
+check "steps: a step change alone is a change" "$(last '.body')" '{"addSteps":["Butter"]}'
+run '{"task":"T1","checkSteps":["Eggs"],"removeSteps":"Bread,Ham"}' >/dev/null
+check "steps: json input" "$(last '.body')" '{"checkSteps":["Eggs"],"removeSteps":["Bread","Ham"]}'
+check "steps: blank is still nothing to change" "$(run_err --task T1 --add-steps ' , ' | jq -r .error | cut -c1-17)" "nothing to change"
 OUT=$(run --list no-grant --task T1 --complete || true)
 check "not connected" "$(printf '%s' "$OUT" | jq -r .reason)" "not_connected"
 

package/config/skills/marketplace-registry.ts

@@ -118,10 +118,10 @@ export const SKILL_FILE_EXCLUDES: ReadonlyArray<RegExp> = [/\.test\./, /^mock-/,
  * never installed.
  *
  * Deliberately FLAT: files in subdirectories (e.g. remotion-video/templates/)
- * are not listed. The shipped CLI writes each listed file without creating
- * parent directories, so a nested path would fail the whole install on every
- * CLI released so far. List them only after the CLI creates parent dirs AND
- * the older CLIs are no longer in use.
+ * are not listed. CLIs released before #800 write each listed file without
+ * creating parent directories, so a nested path fails the whole install on
+ * them. The installers create parent dirs since #800; list nested files only
+ * once the older CLIs are no longer in use.
  *
  * @param skillDir - Absolute skill directory
  * @returns File names relative to the skill directory

package/config/skills/orchestrator/complete-task/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: Complete Task
-description: Mark a task as complete in the task management system.
-version: 1.0.0
+description: Mark a WorkItem complete in the V3 task pool, with evidence.
+version: 1.1.0
 category: management
 skillType: claude-skill
 assignableRoles:
@@ -24,18 +24,25 @@ execution:
 
 # Complete Task
 
-Mark a task as complete in the task management system.
+Mark a WorkItem complete via `POST /api/task-pool/complete/:id`.
 
 ## Usage
 
 ```bash
-bash config/skills/orchestrator/complete-task/execute.sh '{"taskId":"task-123","result":"success"}'
+bash config/skills/orchestrator/complete-task/execute.sh '{"workItemId":"wi-123","summary":"Shipped the report","evidence":[{"type":"artifact","path":"/abs/path/report.md"},{"type":"command","command":"npm test","exitCode":0}]}'
 ```
 
 ## Parameters
 
-Pass the full JSON body as expected by `POST /api/task-management/complete`.
+| Parameter | Required | Description |
+|-----------|----------|-------------|
+| `workItemId` (alias `taskId`) | Yes | The WorkItem to complete |
+| `summary` (alias `result`) | Yes | What was produced |
+| `evidence` | Yes* | Evidence array (#873): `{"type":"artifact","path":…}` (must exist, or an https URL), `{"type":"command","command":…,"exitCode":0,"outputTail":…}`, or `{"type":"blocked","step":…,"reason":…}` (records the item as blocked, not done). *Accepted without it this release with a `warning`; required from the next |
+| `output` | No | Structured output stored on the WorkItem before completion |
+| `agentId` / `sessionName` | No | Who completed it (defaults to `crewly-orc`) |
 
 ## Output
 
-JSON confirmation of task completion.
+The server's JSON response. A `warning` from the server (completion without
+evidence) is repeated on stderr.

package/config/skills/orchestrator/complete-task/execute.sh

@@ -6,7 +6,15 @@
 # on the WorkItem (replaces `<taskId>.output.json`) before completion.
 #
 # Input shape:
-#   { "workItemId": "abc-123", "summary": "...", "output"?: { ... } }
+#   { "workItemId": "abc-123", "summary": "...", "output"?: { ... },
+#     "evidence"?: [ {"type":"artifact","path":"..."} |
+#                    {"type":"command","command":"...","exitCode":0,"outputTail":"..."} |
+#                    {"type":"blocked","step":"...","reason":"..."} ] }
+#
+# `evidence` (#873) is sent as result.evidence. The server checks it: missing
+# artifacts and non-zero exit codes are refused, a `blocked` entry records the
+# item as blocked instead of done, and no evidence at all is accepted this
+# release with a `warning` (refused from the next).
 #
 # Backwards-compat: a `taskId` field is accepted as alias for `workItemId`.
 
@@ -26,6 +34,11 @@ OUTPUT=$(printf '%s' "$INPUT" | jq -c '.output // empty')
 # work — pass that session name instead).
 AGENT_ID=$(printf '%s' "$INPUT" | jq -r '.agentId // .sessionName // "crewly-orc"')
 
+EVIDENCE=$(printf '%s' "$INPUT" | jq -c '.evidence // empty')
+if [ -n "$EVIDENCE" ] && [ "$(printf '%s' "$EVIDENCE" | jq -r 'type')" != "array" ]; then
+  error_exit "evidence must be a JSON array, e.g. [{\"type\":\"artifact\",\"path\":\"/abs/file\"}]"
+fi
+
 require_param "workItemId" "$WORK_ITEM_ID"
 
 # Persist structured output on the WorkItem before completing, so that
@@ -47,6 +60,12 @@ fi
 COMPLETE_BODY=$(jq -n \
   --arg agentId "$AGENT_ID" \
   --arg summary "$SUMMARY" \
-  '{agentId: $agentId, result: {summary: $summary}}')
+  --argjson evidence "${EVIDENCE:-null}" \
+  '{agentId: $agentId, result: ({summary: $summary} + (if $evidence != null then {evidence: $evidence} else {} end))}')
 
-api_call POST "/task-pool/complete/${WORK_ITEM_ID}" "$COMPLETE_BODY"
+COMPLETE_RESPONSE=$(api_call POST "/task-pool/complete/${WORK_ITEM_ID}" "$COMPLETE_BODY")
+printf '%s\n' "$COMPLETE_RESPONSE"
+EVIDENCE_WARNING=$(printf '%s' "$COMPLETE_RESPONSE" | jq -r '.warning // empty' 2>/dev/null || true)
+if [ -n "$EVIDENCE_WARNING" ]; then
+  jq -n --arg w "$EVIDENCE_WARNING" '{warning: $w}' >&2
+fi

package/config/skills/orchestrator/delegate-task/execute.sh

@@ -255,13 +255,13 @@ if [ "$USE_STRUCTURED" = "true" ] && [ -n "$TITLE" ]; then
 
   [ -n "$CONTEXT" ] && TASK_MESSAGE="${TASK_MESSAGE}\n\nAdditional context: ${CONTEXT}"
   [ -n "$DEADLINE_HINT" ] && TASK_MESSAGE="${TASK_MESSAGE}\n\n**Deadline hint**: ${DEADLINE_HINT}"
-  TASK_MESSAGE="${TASK_MESSAGE}\n\n---\nWhen done, report back with report-status, passing the workItemId from your [CREWLY-DISPATCH] notice (or from get-my-tasks) so the right WorkItem is closed: bash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"workItemId\":\"<your WorkItem id>\",\"status\":\"done\",\"summary\":\"<brief summary>\",\"projectPath\":\"${PROJECT_PATH}\"}'"
+  TASK_MESSAGE="${TASK_MESSAGE}\n\n---\nWhen done, report back with report-status, passing the workItemId from your [CREWLY-DISPATCH] notice (or from get-my-tasks) so the right WorkItem is closed: bash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"workItemId\":\"<your WorkItem id>\",\"status\":\"done\",\"summary\":\"<brief summary>\",\"evidence\":[{\"type\":\"artifact\",\"path\":\"<file you produced>\"},{\"type\":\"command\",\"command\":\"<check you ran>\",\"exitCode\":0}],\"projectPath\":\"${PROJECT_PATH}\"}' — done needs evidence (files that exist, checks with exit code 0); if a step failed, send [{\"type\":\"blocked\",\"step\":\"<step>\",\"reason\":\"<why>\"}] instead, never done."
   TASK_MESSAGE="${TASK_MESSAGE}\n\nBefore reporting done, persist key findings using: bash ${CREWLY_ROOT}/config/skills/agent/core/remember/execute.sh '{\"agentId\":\"${TO}\",\"content\":\"<key findings>\",\"category\":\"pattern\",\"scope\":\"project\",\"projectPath\":\"${PROJECT_PATH}\"}'"
 else
   # Legacy free-text format (backwards compatible)
   TASK_MESSAGE="New task from orchestrator (priority: ${PRIORITY}):\n\n${TASK}"
   [ -n "$CONTEXT" ] && TASK_MESSAGE="${TASK_MESSAGE}\n\nContext: ${CONTEXT}"
-  TASK_MESSAGE="${TASK_MESSAGE}\n\nWhen done, report back with report-status, passing the workItemId from your [CREWLY-DISPATCH] notice (or from get-my-tasks) so the right WorkItem is closed: bash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"workItemId\":\"<your WorkItem id>\",\"status\":\"done\",\"summary\":\"<brief summary>\",\"projectPath\":\"${PROJECT_PATH}\"}'"
+  TASK_MESSAGE="${TASK_MESSAGE}\n\nWhen done, report back with report-status, passing the workItemId from your [CREWLY-DISPATCH] notice (or from get-my-tasks) so the right WorkItem is closed: bash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"workItemId\":\"<your WorkItem id>\",\"status\":\"done\",\"summary\":\"<brief summary>\",\"evidence\":[{\"type\":\"artifact\",\"path\":\"<file you produced>\"},{\"type\":\"command\",\"command\":\"<check you ran>\",\"exitCode\":0}],\"projectPath\":\"${PROJECT_PATH}\"}' — done needs evidence (files that exist, checks with exit code 0); if a step failed, send [{\"type\":\"blocked\",\"step\":\"<step>\",\"reason\":\"<why>\"}] instead, never done."
   TASK_MESSAGE="${TASK_MESSAGE}\n\nBefore reporting done, persist key findings using: bash ${CREWLY_ROOT}/config/skills/agent/core/remember/execute.sh '{\"agentId\":\"${TO}\",\"content\":\"<key findings>\",\"category\":\"pattern\",\"scope\":\"project\",\"projectPath\":\"${PROJECT_PATH}\"}'"
 fi
 

package/config/skills/orchestrator/harness-login/SKILL.md

@@ -38,8 +38,15 @@ account ("帮我重新登陆 claude code", "换个账号", "send me the login li
 ```bash
 bash config/skills/orchestrator/harness-login/execute.sh --harness claude
 bash config/skills/orchestrator/harness-login/execute.sh --harness codex --switch-account
+bash config/skills/orchestrator/harness-login/execute.sh --harness claude --account work
 ```
 
+`--account <name>` signs in another of the owner's **own** Claude Code accounts
+("login claude work"): it gets its own config dir and login, and agents move to
+it when an earlier runtime in their fallback order runs out (chain entry
+`claude-code@work`). Nothing is restarted. Add only Claude Code accounts that
+the owner owns; never use it for another person's account.
+
 ## What happens
 
 1. Crewly starts the harness's own login command in a terminal it keeps alive
@@ -75,7 +82,7 @@ in one line to finish it in Setup).
 - `antigravity` answers `no_link_login` (it uses a Gemini API key, entered in
   Setup); Gemini CLI is enterprise-only. Pass the message on in one line.
 - The owner can also just write 「重新登录 claude」 / 「换个账号登录 codex」 /
-  `relogin claude` in your DM — Crewly handles that without you.
+  `relogin claude` / `login claude@work` in your DM — Crewly handles that without you.
 
 ## Errors
 
@@ -86,3 +93,4 @@ in one line to finish it in Setup).
 | `owner_request_unverifiable` | The owner's chat history could not be read |
 | `unknown_harness` | Use claude, codex or antigravity |
 | `no_link_login` | The harness has no link login (Antigravity, Gemini) |
+| `invalid_account` | Account names are 1–32 lower-case letters, digits, `-` or `_` |

package/config/skills/orchestrator/harness-login/execute.sh

@@ -19,6 +19,7 @@
 # Usage:
 #   bash execute.sh --harness claude
 #   bash execute.sh --harness codex --switch-account
+#   bash execute.sh --harness claude --account work   # another of the owner's Claude accounts
 #   bash execute.sh '{"harness":"claude","switchAccount":true}'
 # =============================================================================
 set -euo pipefail
@@ -28,22 +29,25 @@ source "${SCRIPT_DIR}/../_common/lib.sh"
 print_usage() {
   cat <<'EOF_USAGE'
 Usage:
-  bash execute.sh --harness <claude|codex|antigravity> [--switch-account]
-  bash execute.sh '{"harness":"claude","switchAccount":false}'
+  bash execute.sh --harness <claude|codex|antigravity> [--switch-account] [--account <name>]
+  bash execute.sh '{"harness":"claude","switchAccount":false,"account":"work"}'
 
 Options:
   --harness         claude (Claude Code), codex (Codex CLI) or antigravity (agy) — required
   --switch-account  The owner wants a different account (changes the wording of the DM)
+  --account         Sign in another of the owner's own Claude Code accounts (used as a
+                    runtime fallback, chain entry claude-code@<name>) instead of the default login
   --help | -h       Show this help
 EOF_USAGE
 }
 
-INPUT_JSON=""; HARNESS=""; SWITCH=false
+INPUT_JSON=""; HARNESS=""; SWITCH=false; ACCOUNT=""
 if [[ $# -gt 0 && ${1:0:1} == '{' ]]; then INPUT_JSON="$1"; shift || true; fi
 while [[ $# -gt 0 ]]; do
   case "$1" in
     --harness)        [ $# -ge 2 ] || error_exit "--harness requires a value"; HARNESS="$2"; shift 2 ;;
     --switch-account) SWITCH=true; shift ;;
+    --account)        [ $# -ge 2 ] || error_exit "--account requires a value"; ACCOUNT="$2"; shift 2 ;;
     --help|-h)        print_usage; exit 0 ;;
     *) error_exit "Unknown option: $1" ;;
   esac
@@ -52,6 +56,7 @@ if [ -n "$INPUT_JSON" ]; then
   INPUT=$(read_json_input "$INPUT_JSON")
   [ -z "$HARNESS" ] && HARNESS=$(printf '%s' "$INPUT" | jq -r '.harness // .harnessId // empty')
   [ "$(printf '%s' "$INPUT" | jq -r '.switchAccount // false')" = "true" ] && SWITCH=true
+  [ -z "$ACCOUNT" ] && ACCOUNT=$(printf '%s' "$INPUT" | jq -r '.account // empty')
 fi
 require_param "harness (--harness)" "$HARNESS"
 
@@ -59,7 +64,10 @@ require_param "harness (--harness)" "$HARNESS"
 HARNESS_ID=$(printf '%s' "$HARNESS" | tr '[:upper:]' '[:lower:]' | sed -E 's/[[:space:]_]+/-/g')
 [[ "$HARNESS_ID" =~ ^[a-z0-9-]+$ ]] || error_exit "Invalid harness: $HARNESS (use claude, codex or antigravity)"
 
-BODY=$(jq -cn --argjson sw "$SWITCH" '{switchAccount: $sw}')
+ACCOUNT=$(printf '%s' "$ACCOUNT" | tr '[:upper:]' '[:lower:]')
+[ -z "$ACCOUNT" ] || [[ "$ACCOUNT" =~ ^[a-z0-9][a-z0-9_-]{0,31}$ ]] || error_exit "Invalid account: $ACCOUNT (lower-case letters, digits, - or _)"
+
+BODY=$(jq -cn --argjson sw "$SWITCH" --arg acct "$ACCOUNT" '{switchAccount: $sw} + (if $acct != "" then {account: $acct} else {} end)')
 
 ERR_FILE=$(mktemp); trap 'rm -f "$ERR_FILE"' EXIT
 RESPONSE=$(api_call POST "/harness/${HARNESS_ID}/owner-login" "$BODY" 2>"$ERR_FILE") || {

package/config/skills/orchestrator/harness-login/execute.test.sh

@@ -54,6 +54,12 @@ check "display name → id, switch flag" "$(jq -c '{path, body}' "$STUB_LOG")" '
 OUT=$(run '{"harness":"claude","switchAccount":true}')
 check "json input" "$(jq -c .body "$STUB_LOG")" '{"switchAccount":true}'
 
+OUT=$(run --harness claude --account Work)
+check "account (lower-cased)" "$(jq -c .body "$STUB_LOG")" '{"switchAccount":false,"account":"work"}'
+
+OUT=$(run '{"harness":"claude","account":"b"}')
+check "json account" "$(jq -c .body "$STUB_LOG")" '{"switchAccount":false,"account":"b"}'
+
 OUT=$(run --harness codex); RC=$?
 check "no owner request: exit 1" "$RC" "1"
 check "no owner request: reason" "$(printf '%s' "$OUT" | jq -r .reason)" "owner_request_not_found"
@@ -67,6 +73,7 @@ check "other agents refused" "$(printf '%s' "$OUT" | jq -r .reason)" "orchestrat
 check "missing harness" "$(run_err)" "Missing required parameter: harness (--harness)"
 check "path injection refused" "$(run_err --harness '../orc')" "Invalid harness: ../orc (use claude, codex or antigravity)"
 check "unknown option" "$(run_err --yes)" "Unknown option: --yes"
+check "invalid account" "$(run_err --harness claude --account '../x')" "Invalid account: ../x (lower-case letters, digits, - or _)"
 
 echo "=== Results: $PASS passed, $FAIL failed ==="
 [ "$FAIL" -eq 0 ]

package/config/skills/registry.json

@@ -1,6 +1,6 @@
 {
   "schemaVersion": 2,
-  "lastUpdated": "2026-09-29T01:47:07.718Z",
+  "lastUpdated": "2026-10-02T23:58:51.779Z",
   "cdnBaseUrl": "https://raw.githubusercontent.com/stevehuang0115/crewly/main",
   "source": "github",
   "items": [
@@ -1412,7 +1412,7 @@
       "name": "SEO Ops",
       "description": "Search Console-driven SEO operations for any site: query patterns (low-CTR top-3, near-miss 4-20, rising queries, keyword cannibalization), per-URL report cards, pre-publish SEO/AEO checks, a gated programmatic-page queue, and a live-diff gate that stops agents from removing things from live pages. Use when auditing organic search, deciding what to rewrite, or before changing a live page. For writing posts use seo-blog-writer instead.",
       "author": "Crewly",
-      "version": "1.0.0",
+      "version": "1.1.0",
       "category": "development",
       "tags": [
         "seo",
@@ -1425,12 +1425,12 @@
       "downloads": 0,
       "rating": 0,
       "createdAt": "2026-09-29T01:43:29.904Z",
-      "updatedAt": "2026-09-29T01:47:07.718Z",
+      "updatedAt": "2026-10-02T23:58:51.779Z",
       "source": "config/skills/agent/marketplace/seo-ops",
       "assets": {
         "archive": "config/skills/agent/marketplace/seo-ops",
         "checksum": "",
-        "sizeBytes": 95495
+        "sizeBytes": 122232
       },
       "metadata": {
         "skillType": "claude-skill",
@@ -1447,7 +1447,8 @@
           "keyword cannibalization",
           "pre-publish seo check",
           "programmatic seo pages",
-          "live page diff"
+          "live page diff",
+          "experiment metric"
         ],
         "files": [
           "execute.sh",

package/config/skills/team-leader/delegate-task/execute.sh

@@ -170,7 +170,7 @@ if [ -n "$CONTEXT" ]; then
 fi
 
 # Build a structured task message from Team Leader
-TASK_MESSAGE="New task from Team Leader (priority: ${PRIORITY}):\n\n**[REQUIRED] When done, you MUST:** (1) Output a text summary of your work, findings, and any issues. (2) Call report-status, passing the workItemId from your [CREWLY-DISPATCH] notice (or from get-my-tasks) so the right WorkItem is closed:\nbash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"workItemId\":\"<your WorkItem id>\",\"status\":\"done\",\"summary\":\"<brief summary>\",\"projectPath\":\"${PROJECT_PATH}\"}'\n\n---\n\n${TASK}"
+TASK_MESSAGE="New task from Team Leader (priority: ${PRIORITY}):\n\n**[REQUIRED] When done, you MUST:** (1) Output a text summary of your work, findings, and any issues. (2) Call report-status, passing the workItemId from your [CREWLY-DISPATCH] notice (or from get-my-tasks) so the right WorkItem is closed:\nbash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"workItemId\":\"<your WorkItem id>\",\"status\":\"done\",\"summary\":\"<brief summary>\",\"evidence\":[{\"type\":\"artifact\",\"path\":\"<file you produced>\"},{\"type\":\"command\",\"command\":\"<check you ran>\",\"exitCode\":0}],\"projectPath\":\"${PROJECT_PATH}\"}' — done needs evidence (files that exist, checks with exit code 0); if a step failed, send [{\"type\":\"blocked\",\"step\":\"<step>\",\"reason\":\"<why>\"}] instead, never done.\n\n---\n\n${TASK}"
 [ -n "$CONTEXT" ] && TASK_MESSAGE="${TASK_MESSAGE}\n\nContext: ${CONTEXT}"
 
 # V3-only producer (spec 2026-05-06-task-management-v1-deprecation.md):

package/config/skills/team-leader/verify-output/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: Verify Output
 description: "Execute the team's verification pipeline against a completed task. Reads verificationPipeline configuration from the team template and runs each verification step sequentially."
-version: 1.0.0
+version: 1.1.0
 category: quality
 skillType: claude-skill
 assignableRoles:
@@ -33,6 +33,18 @@ Supports two modes:
 1. **Explicit checks**: Pass `checks[]` array directly with specific verification steps
 2. **Template pipeline**: Pass `templateId` to automatically load verification steps from the team template's `verificationPipeline`
 
+## Evidence first (#873)
+
+The skill reads the worker's **evidence block** (`output.evidence` on the
+WorkItem: artifacts that exist, commands with exit codes, blocked steps) before
+running any check, and returns it as the first field of the verdict.
+
+- `evidenceWarning` is set when the item was marked done **without evidence**
+  (accepted only while the server is in evidence warn mode) — do not verify it
+  on its summary alone: check the work directly or send it back asking for
+  evidence. It is also set when the worker reported `blocked` steps.
+- The warning, when present, leads `feedback`.
+
 ## When to Use
 
 - When a worker reports a task as `done`
@@ -118,6 +130,11 @@ The template's `passPolicy` controls the overall pass/fail determination:
 
 ```json
 {
+  "evidence": [
+    { "type": "artifact", "path": "src/auth.ts" },
+    { "type": "command", "command": "npm test", "exitCode": 0, "outputTail": "42 passed" }
+  ],
+  "evidenceWarning": null,
   "passed": false,
   "score": 67,
   "feedback": "Verification failed: 1/3 checks failed (policy: critical_only).",

package/config/skills/team-leader/verify-output/execute.sh

@@ -36,6 +36,25 @@ WI_DATA=$(api_call GET "/task-pool/items/${WORK_ITEM_ID}" 2>/dev/null || echo '{
 WI_BRIEF=$(echo "$WI_DATA" | jq -r '.data.briefMarkdown // empty' 2>/dev/null || true)
 WI_OUTPUT_RAW=$(echo "$WI_DATA" | jq -c '.data.output // empty' 2>/dev/null || true)
 
+# Evidence first (#873). The worker's completion carries an evidence block
+# (artifacts that exist, commands with exit codes, or blocked steps), stored
+# at output.evidence. Read it before anything else, and flag an item that was
+# marked done WITHOUT evidence — accepted only because the server is in warn
+# mode this release — so it is not verified on its summary alone.
+WI_FOUND=$(echo "$WI_DATA" | jq -r 'if (.data // null) != null then "yes" else "no" end' 2>/dev/null || echo "no")
+WI_EVIDENCE=$(echo "$WI_DATA" | jq -c 'if (.data.output.evidence | type) == "array" then .data.output.evidence else [] end' 2>/dev/null || echo '[]')
+EVIDENCE_WARNING=""
+if [ "$WI_FOUND" = "yes" ]; then
+  if [ "$(printf '%s' "$WI_EVIDENCE" | jq 'length')" = "0" ]; then
+    EVIDENCE_WARNING="No completion evidence: this WorkItem was marked done without artifacts or commands (accepted only in evidence warn mode). Do not verify on the summary alone — check the work directly, or send it back asking for evidence."
+  else
+    BLOCKED_STEPS=$(printf '%s' "$WI_EVIDENCE" | jq -r '[.[] | select(.type == "blocked") | "\(.step): \(.reason)"] | join("; ")')
+    if [ -n "$BLOCKED_STEPS" ]; then
+      EVIDENCE_WARNING="The worker reported blocked steps (the item is blocked, not done): ${BLOCKED_STEPS}"
+    fi
+  fi
+fi
+
 # Compose `TASK_OUTPUT` for content-scan checks. Prefer worker-supplied
 # structured output, falling back to the brief if no output was set.
 if [ -n "$WI_OUTPUT_RAW" ] && [ "$WI_OUTPUT_RAW" != "null" ] && [ "$WI_OUTPUT_RAW" != "" ]; then
@@ -316,6 +335,11 @@ else
   FEEDBACK="Verification failed: ${TOTAL_FAILED}/${TOTAL_CHECKS} checks failed (policy: ${PASS_POLICY})."
 fi
 
+# The evidence finding leads the feedback.
+if [ -n "$EVIDENCE_WARNING" ]; then
+  FEEDBACK="${EVIDENCE_WARNING} ${FEEDBACK}"
+fi
+
 # Check if any results require manual TL review
 REQUIRES_REVIEW=$(echo "$RESULTS" | jq '[.[] | select(.requiresReview == true)] | length')
 if [ "$REQUIRES_REVIEW" -gt 0 ]; then
@@ -323,6 +347,8 @@ if [ "$REQUIRES_REVIEW" -gt 0 ]; then
 fi
 
 jq -n \
+  --argjson evidence "$WI_EVIDENCE" \
+  --arg evidenceWarning "$EVIDENCE_WARNING" \
   --arg passed "$PASSED" \
   --arg score "$SCORE" \
   --arg feedback "$FEEDBACK" \
@@ -333,6 +359,8 @@ jq -n \
   --arg workerId "$WORKER_ID" \
   --arg templateId "$TEMPLATE_ID" \
   '{
+    evidence: $evidence,
+    evidenceWarning: (if $evidenceWarning != "" then $evidenceWarning else null end),
     passed: ($passed == "true"),
     score: ($score | tonumber),
     feedback: $feedback,