crewly 1.12.1 → 1.13.1

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

@@ -82,30 +82,49 @@ require_param "task (--task)" "$TASK"
 # entitled to push back per the Brief Reception Protocol when these are
 # absent. Spec: .crewly/specs/2026-05-03-agent-improvement-p0-execution.md
 # §"Fix P0-3".
+#
+# Scans the WHOLE delegated brief, not just `--task`. The Request Contract
+# can legitimately live in `--task`, in `--context`, or be split across the
+# two, and delegators routinely put the long-form G/O/E in `--context` while
+# `--task` carries the one-line ask. Scanning only `--task` warned
+# "Request Contract incomplete" at briefs whose contract was fully present,
+# just in the other field. That false positive is corrosive: this warning is
+# the signal workers use to decide whether to push back under the Brief
+# Reception Protocol, so crying wolf trains them to ignore a real one.
+#
+# The check is NOT weakened — a marker genuinely absent from BOTH fields
+# still warns, and the missing-field list is still per-marker.
+#
+# $1 = task, $2 = context (either may be empty).
 warn_missing_request_contract() {
-  local task="$1"
+  # Join with a newline rather than concatenating: the patterns below anchor
+  # on `^`, which grep evaluates per line, and a seam newline also prevents a
+  # task ending in "go" plus a context starting "al:" from fabricating a
+  # spurious "goal:" match across the boundary.
+  local brief
+  brief="$(printf '%s\n%s' "${1:-}" "${2:-}")"
   local missing=()
   # Goal: the standalone word "Goal" (with optional ** markdown). Also
   # accept "Objective" as a synonym some upstream callers use.
-  if ! echo "$task" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(goal|objective)(:|s?\b)'; then
+  if ! echo "$brief" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(goal|objective)(:|s?\b)'; then
     missing+=("Goal")
   fi
   # Outcome: "Outcome" or "Expected Outcome".
-  if ! echo "$task" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(expected )?outcome(:|s?\b)'; then
+  if ! echo "$brief" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(expected )?outcome(:|s?\b)'; then
     missing+=("Outcome")
   fi
   # Eval: "Eval", "Eval Criteria", "Acceptance Criteria", "Evaluation Criteria".
-  if ! echo "$task" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(eval|evaluation criteria|acceptance criteria)(:|s?\b)'; then
+  if ! echo "$brief" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(eval|evaluation criteria|acceptance criteria)(:|s?\b)'; then
     missing+=("Eval")
   fi
   if [ ${#missing[@]} -gt 0 ]; then
     local list
     list=$(IFS=, ; echo "${missing[*]}")
-    echo "{\"warning\":\"Request Contract incomplete: brief is missing markers for: ${list}. Per P0-3 spec, every delegated subtask MUST include Goal + Expected Outcome + Eval Criteria. Workers may push back via the Brief Reception Protocol. Source: .crewly/specs/2026-05-03-agent-improvement-p0-execution.md §Fix P0-3.\"}" >&2
+    echo "{\"warning\":\"Request Contract incomplete: brief is missing markers for: ${list}. Neither the task nor the context field carries them. Per P0-3 spec, every delegated subtask MUST include Goal + Expected Outcome + Eval Criteria. Workers may push back via the Brief Reception Protocol. Source: .crewly/specs/2026-05-03-agent-improvement-p0-execution.md §Fix P0-3.\"}" >&2
   fi
 }
 
-warn_missing_request_contract "$TASK"
+warn_missing_request_contract "$TASK" "$CONTEXT"
 
 # Resolve Crewly root from this script path:
 # config/skills/team-leader/delegate-task/execute.sh -> project root
@@ -144,13 +163,80 @@ 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:\nbash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"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>\",\"projectPath\":\"${PROJECT_PATH}\"}'\n\n---\n\n${TASK}"
 [ -n "$CONTEXT" ] && TASK_MESSAGE="${TASK_MESSAGE}\n\nContext: ${CONTEXT}"
 
+# V3-only producer (spec 2026-05-06-task-management-v1-deprecation.md):
+# We no longer write a `.crewly/tasks/delegated/*.md` file via the legacy
+# `/task-management/create` endpoint. Instead the long-form brief travels
+# on the WorkItem itself (`briefMarkdown` field) and the WI is the sole
+# durable record of this delegation.
+#
+# ORDERING (WI 65578471): the WorkItem is created BEFORE delivery is
+# attempted, and that order is load-bearing. Previously the pool-add ran
+# only after delivery succeeded, so a delivery failure exited the script
+# with NO RECORD ANYWHERE — not in the pool, not targeted, nothing for a
+# reconciler or a human to find. Every other orphan class in this system
+# at least left a WorkItem behind; this one left the intent with no system
+# of record at all.
+#
+# On a delivery failure the WI is deliberately left `queued` with its
+# `target` set, plus an explicit `[UNDELIVERED]` note (see below). That
+# state is chosen, not incidental:
+#   - `queued` + `target` is recoverable — the wake gate admits a worker
+#     for a queued WI targeting them, so the task can still be delivered.
+#   - `queued` is NOT `running`, so it never masquerades as in-flight work.
+#   - the note makes an undelivered item distinguishable from a normally
+#     queued one, so a silent nothing is not traded for a silent something.
+TASK_ID=""
+
+case "$PRIORITY" in
+  critical|urgent) WI_PRIORITY="critical" ;;
+  high)            WI_PRIORITY="high" ;;
+  low)             WI_PRIORITY="low" ;;
+  *)               WI_PRIORITY="medium" ;;
+esac
+
+WI_TITLE="$(echo "$TASK" | head -c 200)"
+
+POOL_BODY=$(jq -n \
+  --arg type "delegate" \
+  --arg owner "team_lead" \
+  --arg target "$TO" \
+  --arg title "$WI_TITLE" \
+  --arg description "$TASK_MESSAGE" \
+  --arg briefMarkdown "$TASK" \
+  --arg priority "$WI_PRIORITY" \
+  --arg projectPath "${PROJECT_PATH:-}" \
+  '{type: $type, owner: $owner, target: $target, title: $title, description: $description, briefMarkdown: $briefMarkdown, metadata: ({priority: $priority} + (if $projectPath != "" then {projectPath: $projectPath} else {} end))}')
+
+POOL_RESULT=$(api_call POST "/task-pool/add" "$POOL_BODY" 2>/dev/null || echo '{"success":false}')
+POOL_OK=$(echo "$POOL_RESULT" | jq -r '.success // "false"' 2>/dev/null)
+TASK_ID=$(echo "$POOL_RESULT" | jq -r '.data.id // .workItemId // empty' 2>/dev/null || true)
+
+if [ "$POOL_OK" != "true" ]; then
+  # Unchanged from before the reorder: warn and still attempt delivery.
+  # Aborting here instead would be a behavioural change beyond this WI's
+  # scope, so it is deliberately NOT made. The residual gap (record
+  # creation fails, task delivered anyway) is reported, not silently
+  # fixed — it is the mirror of the bug fixed here and wants its own call.
+  echo "{\"warning\":\"Failed to create WorkItem in TaskPool — delivering anyway, so this task will have no durable record\",\"details\":$(echo "$POOL_RESULT" | jq -c . 2>/dev/null || echo '{}')}" >&2
+fi
+
 # Deliver the task message with fallback strategy:
 # 1. Try normal delivery (waitForReady)
 # 2. If fails, try force delivery
 # 3. If still fails, auto-start the worker then retry
+# waitTimeout 15000: the INITIAL attempt against a worker expected to be
+# already running, so it fails fast and lets the fallback ladder below take
+# over. The retry after auto-start uses 30000 — see the note there; the two
+# values differ deliberately and must not be normalised to one number.
+#
+# ORIGIN NOT ESTABLISHED (WI dae79289): 15000 matches no delivery-related
+# constant. Seven constants share the value, none of them about terminal
+# delivery, so there is nothing here that this literal can honestly be said
+# to mirror. Left exactly as-is rather than attached to a plausible-looking
+# constant that does not actually govern it.
 BODY=$(jq -n --arg message "$TASK_MESSAGE" '{message: $message, waitForReady: true, waitTimeout: 15000}')
 
 DELIVER_OK=true
@@ -183,6 +269,16 @@ if [ "$DELIVER_OK" = "false" ]; then
           echo '{"info":"Worker '"$TO"' start triggered — retrying delivery in 10s..."}' >&2
           sleep 10
           # Retry delivery after agent boots
+          # waitTimeout 30000, double the initial 15000 above. This retry
+          # follows an auto-start, so the worker is booting from cold and
+          # legitimately needs longer to reach a prompt than one that was
+          # expected to be up already. The asymmetry is the point: it is not
+          # drift between two copies of the same call.
+          #
+          # ORIGIN NOT ESTABLISHED (WI dae79289): the value coincides with
+          # EVENT_DELIVERY_CONSTANTS.TOTAL_DELIVERY_TIMEOUT, but coincidence
+          # of value is not evidence of relationship — eleven constants share
+          # 30000. Not claimed as a mirror.
           RETRY_BODY=$(jq -n --arg message "$TASK_MESSAGE" '{message: $message, waitForReady: true, waitTimeout: 30000}')
           api_call POST "/terminal/${TO}/deliver" "$RETRY_BODY" && STARTED=true || {
             # Final fallback: force deliver
@@ -193,46 +289,27 @@ if [ "$DELIVER_OK" = "false" ]; then
     fi
 
     if [ "$STARTED" = "false" ]; then
-      echo '{"error":"Failed to deliver task to '"$TO"'. Worker is offline and could not be auto-started. Ensure the worker is a team member with a valid session.","to":"'"$TO"'","teamId":"'"$TEAM_ID"'"}'
+      # WI 65578471 — delivery failed, but the WorkItem already exists
+      # (created above, deliberately, before delivery was attempted). Mark
+      # it explicitly so an undelivered item is never mistaken for a
+      # normally queued one, then fail loudly. The WI stays `queued` with
+      # its `target`, which is the recoverable state: the worker can still
+      # be woken for it and re-delivery can be retried.
+      #
+      # Best-effort: a failure to annotate must not swallow the delivery
+      # error, which is the more important signal.
+      if [ -n "$TASK_ID" ]; then
+        NOTE_AUTHOR="${CREWLY_SESSION_NAME:-${FROM_SESSION:-team-leader}}"
+        NOTE_BODY=$(jq -n --arg author "$NOTE_AUTHOR" --arg note "[UNDELIVERED] Delivery to ${TO} failed — worker offline and could not be auto-started. WorkItem left queued and targeted for re-delivery." \
+          '{author: $author, note: $note}')
+        api_call POST "/task-pool/items/${TASK_ID}/notes" "$NOTE_BODY" >/dev/null 2>&1 || true
+      fi
+      echo '{"error":"Failed to deliver task to '"$TO"'. Worker is offline and could not be auto-started. Ensure the worker is a team member with a valid session.","to":"'"$TO"'","teamId":"'"$TEAM_ID"'","workItemId":"'"$TASK_ID"'","workItemState":"queued+targeted, marked [UNDELIVERED] — re-delivery can be retried"}'
       exit 1
     fi
   }
 fi
 
-# V3-only producer (spec 2026-05-06-task-management-v1-deprecation.md):
-# We no longer write a `.crewly/tasks/delegated/*.md` file via the legacy
-# `/task-management/create` endpoint. Instead the long-form brief travels
-# on the WorkItem itself (`briefMarkdown` field) and the WI is the sole
-# durable record of this delegation.
-TASK_ID=""
-
-case "$PRIORITY" in
-  critical|urgent) WI_PRIORITY="critical" ;;
-  high)            WI_PRIORITY="high" ;;
-  low)             WI_PRIORITY="low" ;;
-  *)               WI_PRIORITY="medium" ;;
-esac
-
-WI_TITLE="$(echo "$TASK" | head -c 200)"
-
-POOL_BODY=$(jq -n \
-  --arg type "delegate" \
-  --arg owner "team_lead" \
-  --arg target "$TO" \
-  --arg title "$WI_TITLE" \
-  --arg description "$TASK_MESSAGE" \
-  --arg briefMarkdown "$TASK" \
-  --arg priority "$WI_PRIORITY" \
-  --arg projectPath "${PROJECT_PATH:-}" \
-  '{type: $type, owner: $owner, target: $target, title: $title, description: $description, briefMarkdown: $briefMarkdown, priority: $priority} + (if $projectPath != "" then {projectPath: $projectPath} else {} end)')
-
-POOL_RESULT=$(api_call POST "/task-pool/add" "$POOL_BODY" 2>/dev/null || echo '{"success":false}')
-POOL_OK=$(echo "$POOL_RESULT" | jq -r '.success // "false"' 2>/dev/null)
-TASK_ID=$(echo "$POOL_RESULT" | jq -r '.data.id // .workItemId // empty' 2>/dev/null || true)
-
-if [ "$POOL_OK" != "true" ]; then
-  echo "{\"warning\":\"Failed to create WorkItem in TaskPool — task delivered to terminal but no durable record\",\"details\":$(echo "$POOL_RESULT" | jq -c . 2>/dev/null || echo '{}')}" >&2
-fi
 
 # Set up idle event subscription for TL monitoring
 MONITOR_IDLE=$(printf '%s' "$INPUT" | jq -r 'if .monitor.idleEvent == null then true else .monitor.idleEvent end')
@@ -251,11 +328,23 @@ if [ -z "$RESOLVED_SESSION" ]; then
 fi
 
 if [ "$MONITOR_IDLE" = "true" ] && [ -n "$RESOLVED_SESSION" ]; then
+  # ttlMinutes is deliberately OMITTED so the server default applies.
+  #
+  # This used to hardcode `ttlMinutes: 120`, silently capping every delegation
+  # subscription at 2h against a system default of 8h
+  # (EVENT_BUS_CONSTANTS.DEFAULT_SUBSCRIPTION_TTL_MINUTES = 480, applied by
+  # event-bus.service.ts via `input.ttlMinutes ?? DEFAULT`). A subscription
+  # that expires 6h early stops the TL ever being woken for a long-running
+  # delegation — the same failure shape as a lease lapsing under an agent that
+  # is simply heads-down.
+  #
+  # Omitting beats re-hardcoding 480: the constant stays the single source of
+  # truth and the skill cannot drift from it again.
   SUB_BODY=$(jq -n \
     --arg eventType "agent:idle" \
     --arg sessionName "$TO" \
     --arg subscriber "$RESOLVED_SESSION" \
-    '{eventType: $eventType, filter: {sessionName: $sessionName}, subscriberSession: $subscriber, oneShot: true, ttlMinutes: 120}')
+    '{eventType: $eventType, filter: {sessionName: $sessionName}, subscriberSession: $subscriber, oneShot: true}')
   SUB_RESULT=$(api_call POST "/events/subscribe" "$SUB_BODY" 2>/dev/null || true)
   SUB_ID=$(echo "$SUB_RESULT" | jq -r '.data.id // empty' 2>/dev/null || true)
   if [ -n "$SUB_ID" ]; then

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

@@ -1,12 +1,34 @@
 #!/bin/bash
 # Tests for TL delegate-task execute.sh
 # Covers: CREWLY_SESSION_NAME resolution, fromSession fallback,
-#         graceful monitoring skip, delegation success
+#         graceful monitoring skip, delegation success, and the Request
+#         Contract checker (which scans BOTH `task` and `context`).
 set -euo pipefail
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 PASS=0
 FAIL=0
 
+# ---------------------------------------------------------------------------
+# Extract the REAL warn_missing_request_contract out of execute.sh.
+#
+# This file used to keep a hand-copied duplicate of that function inside the
+# mock, with a comment asking future editors to keep it "byte-equivalent".
+# It did not stay byte-equivalent, and worse, nothing could detect that:
+# changing the function in execute.sh left all 16 assertions green, because
+# they were exercising the copy. Pulling the function straight out of the
+# shipped script means production edits are felt here immediately, and makes
+# the mutation test in the commit message meaningful.
+# ---------------------------------------------------------------------------
+REAL_FN_FILE="$(mktemp)"
+export REAL_FN_FILE
+sed -n '/^warn_missing_request_contract() {$/,/^}$/p' "$SCRIPT_DIR/execute.sh" > "$REAL_FN_FILE"
+if ! grep -q 'Request Contract incomplete' "$REAL_FN_FILE"; then
+  echo "FATAL: could not extract warn_missing_request_contract() from execute.sh" >&2
+  exit 1
+fi
+# shellcheck source=/dev/null
+source "$REAL_FN_FILE"
+
 assert_succeeds() {
   local desc="$1"
   shift
@@ -74,27 +96,11 @@ shift
 # Load the real shared lib first
 source "${REAL_SKILL_DIR}/../_common/lib.sh"
 
-# Request Contract warning helper (mirror of the real script's
-# warn_missing_request_contract — must stay byte-equivalent so the test
-# exercises actual production logic, not a stub).
-warn_missing_request_contract() {
-  local task="$1"
-  local missing=()
-  if ! echo "$task" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(goal|objective)(:|s?\b)'; then
-    missing+=("Goal")
-  fi
-  if ! echo "$task" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(expected )?outcome(:|s?\b)'; then
-    missing+=("Outcome")
-  fi
-  if ! echo "$task" | grep -qiE '(^|[^a-zA-Z])(\*\*)?(eval|evaluation criteria|acceptance criteria)(:|s?\b)'; then
-    missing+=("Eval")
-  fi
-  if [ ${#missing[@]} -gt 0 ]; then
-    local list
-    list=$(IFS=, ; echo "${missing[*]}")
-    echo "{\"warning\":\"Request Contract incomplete: brief is missing markers for: ${list}. Per P0-3 spec, every delegated subtask MUST include Goal + Expected Outcome + Eval Criteria. Workers may push back via the Brief Reception Protocol. Source: .crewly/specs/2026-05-03-agent-improvement-p0-execution.md §Fix P0-3.\"}" >&2
-  fi
-}
+# Load the REAL warn_missing_request_contract, extracted from execute.sh by
+# the harness above. Do NOT re-declare it here — a hand-copy silently
+# decouples these scenarios from the code that actually ships.
+# shellcheck source=/dev/null
+source "$REAL_FN_FILE"
 
 # Track API calls for verification
 API_CALLS_LOG=$(mktemp)
@@ -158,9 +164,14 @@ FROM_SESSION=$(echo "$INPUT" | jq -r '.fromSession // empty')
 require_param "to" "$TO"
 require_param "task" "$TASK"
 
-# Request Contract warning (P0-3): emit non-fatal warning when brief is
-# missing Goal / Expected Outcome / Eval markers. Mirror of the real script.
-warn_missing_request_contract "$TASK"
+# Request Contract warning (P0-3): emit the non-fatal warning when the brief
+# is missing Goal / Expected Outcome / Eval markers.
+#
+# Run the PRODUCTION call line verbatim, with this mock's $TASK / $CONTEXT
+# bound. Hardcoding the arguments here would let a regression at the real
+# call site — dropping $CONTEXT, say — slip past every scenario below, which
+# is exactly how the earlier hand-copied version of this mock went blind.
+eval "$(grep -E '^warn_missing_request_contract "' "${REAL_SKILL_DIR}/execute.sh")"
 
 # Validate hierarchy
 if [ -n "$TEAM_ID" ] && [ -n "$TL_MEMBER_ID" ]; then
@@ -192,7 +203,7 @@ fi
 
 TASK_MESSAGE="New task from Team Leader (priority: ${PRIORITY}):\n\n${TASK}"
 [ -n "$CONTEXT" ] && TASK_MESSAGE="${TASK_MESSAGE}\n\nContext: ${CONTEXT}"
-TASK_MESSAGE="${TASK_MESSAGE}\n\nWhen done, report back using: bash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"status\":\"done\",\"summary\":\"<brief summary>\"}'"
+TASK_MESSAGE="${TASK_MESSAGE}\n\nWhen done, report back with report-status, passing the workItemId from your [CREWLY-DISPATCH] notice: bash ${CREWLY_ROOT}/config/skills/agent/core/report-status/execute.sh '{\"sessionName\":\"${TO}\",\"workItemId\":\"<your WorkItem id>\",\"status\":\"done\",\"summary\":\"<brief summary>\"}'"
 
 BODY=$(jq -n --arg message "$TASK_MESSAGE" '{message: $message, waitForReady: true, waitTimeout: 15000}')
 api_call POST "/terminal/${TO}/deliver" "$BODY"
@@ -231,7 +242,7 @@ if [ "$MONITOR_IDLE" = "true" ] && [ -n "$RESOLVED_SESSION" ]; then
     --arg eventType "agent:idle" \
     --arg sessionName "$TO" \
     --arg subscriber "$RESOLVED_SESSION" \
-    '{eventType: $eventType, filter: {sessionName: $sessionName}, subscriberSession: $subscriber, oneShot: true, ttlMinutes: 120}')
+    '{eventType: $eventType, filter: {sessionName: $sessionName}, subscriberSession: $subscriber, oneShot: true}')
   SUB_RESULT=$(api_call POST "/events/subscribe" "$SUB_BODY" 2>/dev/null || true)
   SUB_ID=$(echo "$SUB_RESULT" | jq -r '.data.id // empty' 2>/dev/null || true)
   if [ -n "$SUB_ID" ]; then
@@ -510,7 +521,149 @@ else
   PASS=$((PASS + 1))
 fi
 
+# -----------------------------------------------------------------------
+# Scenario 15: Contract lives ENTIRELY in `context` — must NOT warn
+#
+# This is the reported bug. A delegator whose `task` is the one-line ask and
+# whose full Goal / Expected Outcome / Eval Criteria sit in `context` was
+# told the contract was "incomplete" even though every marker was present.
+# -----------------------------------------------------------------------
+echo "Scenario 15: full contract in context field only — no warning"
+CTX_FULL="**Goal:** Ship the importer. **Expected Outcome:** CSVs land in prod. **Eval Criteria:** all tests green."
+RAW_OUTPUT=$(env CREWLY_SESSION_NAME="sam-tl" bash "$MOCK_SCRIPT" "$SKILL_DIR" \
+  "{\"to\":\"crewly-product-leo-dev\",\"task\":\"build the importer\",\"context\":\"$CTX_FULL\"}" 2>&1)
+if echo "$RAW_OUTPUT" | grep -q "Request Contract incomplete"; then
+  echo "  FAIL: contract present in context was still reported incomplete: $RAW_OUTPUT"
+  FAIL=$((FAIL + 1))
+else
+  echo "  PASS: contract in context suppresses the warning"
+  PASS=$((PASS + 1))
+fi
+
+# -----------------------------------------------------------------------
+# Scenario 16: Contract SPLIT across task and context — must NOT warn
+# -----------------------------------------------------------------------
+echo "Scenario 16: contract split across task and context — no warning"
+RAW_OUTPUT=$(env CREWLY_SESSION_NAME="sam-tl" bash "$MOCK_SCRIPT" "$SKILL_DIR" \
+  '{"to":"crewly-product-leo-dev","task":"**Goal:** ship the importer","context":"**Expected Outcome:** CSVs land in prod. **Eval Criteria:** all tests green."}' 2>&1)
+if echo "$RAW_OUTPUT" | grep -q "Request Contract incomplete"; then
+  echo "  FAIL: split contract was reported incomplete: $RAW_OUTPUT"
+  FAIL=$((FAIL + 1))
+else
+  echo "  PASS: split contract suppresses the warning"
+  PASS=$((PASS + 1))
+fi
+
+# -----------------------------------------------------------------------
+# Scenario 17: Check is NOT weakened — a field absent from BOTH still warns
+# -----------------------------------------------------------------------
+echo "Scenario 17: field missing from both fields still warns (check not weakened)"
+RAW_OUTPUT=$(env CREWLY_SESSION_NAME="sam-tl" bash "$MOCK_SCRIPT" "$SKILL_DIR" \
+  '{"to":"crewly-product-leo-dev","task":"**Goal:** ship it","context":"**Expected Outcome:** it ships."}' 2>&1)
+if echo "$RAW_OUTPUT" | grep -q "Request Contract incomplete" && \
+   echo "$RAW_OUTPUT" | grep -q "Eval"; then
+  echo "  PASS: still warns for the one marker absent from both fields"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL: expected an Eval-missing warning, got: $RAW_OUTPUT"
+  FAIL=$((FAIL + 1))
+fi
+
+# -----------------------------------------------------------------------
+# Scenario 18: Nothing anywhere — all three still named
+# -----------------------------------------------------------------------
+echo "Scenario 18: empty contract in both fields names Goal+Outcome+Eval"
+RAW_OUTPUT=$(env CREWLY_SESSION_NAME="sam-tl" bash "$MOCK_SCRIPT" "$SKILL_DIR" \
+  '{"to":"crewly-product-leo-dev","task":"just do the thing","context":"here is some background prose"}' 2>&1)
+if echo "$RAW_OUTPUT" | grep -q "Request Contract incomplete" && \
+   echo "$RAW_OUTPUT" | grep -q "Goal" && \
+   echo "$RAW_OUTPUT" | grep -q "Outcome" && \
+   echo "$RAW_OUTPUT" | grep -q "Eval"; then
+  echo "  PASS: all three markers reported missing"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL: expected all three named, got: $RAW_OUTPUT"
+  FAIL=$((FAIL + 1))
+fi
+
+# -----------------------------------------------------------------------
+# Scenarios 19-22: direct unit tests against the REAL extracted function.
+#
+# These call warn_missing_request_contract() itself (sourced from
+# execute.sh at the top of this file), so they pin the two-argument
+# signature and the scan surface with no mock in between.
+# -----------------------------------------------------------------------
+echo "Scenario 19: function scans arg 2 (context) for markers"
+FN_OUT=$(warn_missing_request_contract "bare ask" \
+  "Goal: x. Expected Outcome: y. Eval Criteria: z." 2>&1 || true)
+if [ -z "$FN_OUT" ]; then
+  echo "  PASS: no warning when all markers are in the context argument"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL: expected silence, got: $FN_OUT"
+  FAIL=$((FAIL + 1))
+fi
+
+echo "Scenario 20: function still scans arg 1 (task) on its own"
+FN_OUT=$(warn_missing_request_contract \
+  "Goal: x. Expected Outcome: y. Eval Criteria: z." "" 2>&1 || true)
+if [ -z "$FN_OUT" ]; then
+  echo "  PASS: no warning when all markers are in the task argument"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL: expected silence, got: $FN_OUT"
+  FAIL=$((FAIL + 1))
+fi
+
+echo "Scenario 21: function warns when a marker is in neither argument"
+FN_OUT=$(warn_missing_request_contract "Goal: x." "Expected Outcome: y." 2>&1 || true)
+if echo "$FN_OUT" | grep -q "Request Contract incomplete" && echo "$FN_OUT" | grep -q "Eval"; then
+  echo "  PASS: names Eval as missing from both arguments"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL: expected Eval-missing warning, got: $FN_OUT"
+  FAIL=$((FAIL + 1))
+fi
+
+echo "Scenario 22: markers are not fabricated across the task/context seam"
+# "go" + "al: ..." must not join into "goal:" — the newline join prevents it.
+FN_OUT=$(warn_missing_request_contract "please go" \
+  "al: nothing here. Expected Outcome: y. Eval Criteria: z." 2>&1 || true)
+if echo "$FN_OUT" | grep -q "Goal"; then
+  echo "  PASS: seam does not fabricate a Goal marker"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL: 'go'+'al:' across the seam was accepted as a Goal marker: $FN_OUT"
+  FAIL=$((FAIL + 1))
+fi
+
+# -----------------------------------------------------------------------
+# Scenario 23: source-level guard on the production call site.
+#
+# The unit tests above cannot see which arguments execute.sh actually
+# passes. This asserts the call site itself, so reverting it to the
+# task-only form fails here even though the function stays correct.
+# -----------------------------------------------------------------------
+echo "Scenario 23: execute.sh calls the checker with BOTH task and context"
+SRC=$(cat "$SKILL_DIR/execute.sh")
+if printf '%s' "$SRC" | grep -q -- 'warn_missing_request_contract "$TASK" "$CONTEXT"'; then
+  echo "  PASS: call site passes \$TASK and \$CONTEXT"
+  PASS=$((PASS + 1))
+else
+  echo "  FAIL: execute.sh does not call warn_missing_request_contract with both fields"
+  FAIL=$((FAIL + 1))
+fi
+
+if printf '%s' "$SRC" | grep -qE '^warn_missing_request_contract "\$TASK"$'; then
+  echo "  FAIL: execute.sh still has a task-only call to the contract checker"
+  FAIL=$((FAIL + 1))
+else
+  echo "  PASS: no task-only call to the contract checker remains"
+  PASS=$((PASS + 1))
+fi
+
 # Cleanup
+rm -f "$REAL_FN_FILE"
 rm -rf "$MOCK_DIR"
 
 echo ""

package/config/vitest-exclude.constants.ts

@@ -0,0 +1,56 @@
+/**
+ * Test-collection exclusion patterns for root-level vitest runs.
+ *
+ * The repository root has no vitest project of its own — the real vitest
+ * suites live in `frontend/` and `packages/chat-ui/`, each with their own
+ * config. A bare `npx vitest` at the root therefore falls back to vitest's
+ * built-in defaults, whose exclude list covers only `node_modules`, `dist`,
+ * `cypress` and a handful of dotfolders (`.idea`, `.git`, `.cache`,
+ * `.output`, `.temp`).
+ *
+ * `.claude/` is NOT in that list. It is agent scratch space and holds full
+ * git worktree copies of this repository, so a root-level run collects
+ * thousands of stale duplicate test files — files that belong to other
+ * agents' in-flight branches and are frequently written in jest syntax.
+ * Collecting them produces a wall of import errors that has nothing to do
+ * with the working tree under review.
+ *
+ * These patterns are kept here, rather than inline in `vitest.config.ts`,
+ * so the exclusion is a named, testable contract: `vitest-exclude.constants.test.ts`
+ * fails if the pattern is changed or the config stops applying it.
+ *
+ * @see vitest.config.ts — the sole consumer, which spreads these on top of
+ *      vitest's `defaultExclude` (never replacing it).
+ */
+
+/**
+ * Glob matching everything inside the `.claude/` agent scratch directory, at
+ * any depth. Covers `.claude/worktrees/<branch>/**` (full repo copies) and
+ * `.claude/agents/**`.
+ *
+ * Deliberately matches `**\/.claude/**` rather than `.claude/**` so that a
+ * nested checkout cannot smuggle the directory back into collection.
+ */
+export const AGENT_SCRATCH_EXCLUDE_PATTERN = '**/.claude/**';
+
+/**
+ * Exclusion globs applied to root-level vitest runs, on top of vitest's own
+ * `defaultExclude`.
+ *
+ * IMPORTANT: consumers must SPREAD vitest's `defaultExclude` alongside these.
+ * Assigning `test.exclude` replaces the built-in list outright, which would
+ * silently re-enable collection of `node_modules` and `dist`.
+ *
+ * @example
+ * ```typescript
+ * import { defineConfig, defaultExclude } from 'vitest/config';
+ * import { ROOT_VITEST_EXCLUDE_PATTERNS } from './config/vitest-exclude.constants.js';
+ *
+ * export default defineConfig({
+ *   test: { exclude: [...defaultExclude, ...ROOT_VITEST_EXCLUDE_PATTERNS] },
+ * });
+ * ```
+ */
+export const ROOT_VITEST_EXCLUDE_PATTERNS: readonly string[] = [
+  AGENT_SCRATCH_EXCLUDE_PATTERN,
+];

package/dist/backend/backend/src/controllers/chat/chat.controller.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chat.controller.d.ts","sourceRoot":"","sources":["../../../../../../backend/src/controllers/chat/chat.controller.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,OAAO,EAAE,QAAQ,EAAE,YAAY,EAAE,MAAM,SAAS,CAAC;AAiB/D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mDAAmD,CAAC;AAoB7F;;;;;GAKG;AACH,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,mBAAmB,GAAG,IAAI,CAEzE;AAED;;;;;GAKG;AACH,wBAAgB,2BAA2B,CACzC,OAAO,EAAE,OAAO,yDAAyD,EAAE,wBAAwB,GAClG,IAAI,CAEN;AAMD;;;;;;;;GAQG;AACH,wBAAsB,WAAW,CAC/B,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAwIf;AAED;;;;;;;GAOG;AACH,wBAAsB,WAAW,CAC/B,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CA8Cf;AAED;;;;;;;GAOG;AACH,wBAAsB,UAAU,CAC9B,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAsBf;AAMD;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAsB,aAAa,CACjC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAgKf;AAMD;;;;;;;GAOG;AACH,wBAAsB,gBAAgB,CACpC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAyBf;AAED;;;;;;;;GAQG;AACH,wBAAsB,sBAAsB,CAC1C,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAwBf;AAED;;;;;;;GAOG;AACH,wBAAsB,eAAe,CACnC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAsBf;AAED;;;;;;;GAOG;AACH,wBAAsB,kBAAkB,CACtC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAcf;AAED;;;;;;;GAOG;AACH,wBAAsB,kBAAkB,CACtC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CA8Bf;AAED;;;;;;;GAOG;AACH,wBAAsB,mBAAmB,CACvC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAqBf;AAED;;;;;;;GAOG;AACH,wBAAsB,qBAAqB,CACzC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAqBf;AAED;;;;;;;GAOG;AACH,wBAAsB,kBAAkB,CACtC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAcf;AAED;;;;;;;GAOG;AACH,wBAAsB,iBAAiB,CACrC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAcf;AAMD;;;;;;;GAOG;AACH,wBAAsB,aAAa,CACjC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAYf;AAMD;;;;;;;GAOG;AACH,wBAAgB,mBAAmB,CACjC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,IAAI,CAwBN;AAED;;;;;;;GAOG;AACH,wBAAgB,oBAAoB,CAClC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,IAAI,CAmBN;AAED;;;;;;;GAOG;AACH,wBAAgB,oBAAoB,CAClC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,IAAI,CA4BN;AAMD;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,mBAAmB,CACvC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAiBf"}
+{"version":3,"file":"chat.controller.d.ts","sourceRoot":"","sources":["../../../../../../backend/src/controllers/chat/chat.controller.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,OAAO,KAAK,EAAE,OAAO,EAAE,QAAQ,EAAE,YAAY,EAAE,MAAM,SAAS,CAAC;AAiB/D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,mDAAmD,CAAC;AAoB7F;;;;;GAKG;AACH,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,mBAAmB,GAAG,IAAI,CAEzE;AAED;;;;;GAKG;AACH,wBAAgB,2BAA2B,CACzC,OAAO,EAAE,OAAO,yDAAyD,EAAE,wBAAwB,GAClG,IAAI,CAEN;AAMD;;;;;;;;GAQG;AACH,wBAAsB,WAAW,CAC/B,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAwIf;AAED;;;;;;;GAOG;AACH,wBAAsB,WAAW,CAC/B,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CA8Cf;AAED;;;;;;;GAOG;AACH,wBAAsB,UAAU,CAC9B,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAsBf;AAMD;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAsB,aAAa,CACjC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAqLf;AAMD;;;;;;;GAOG;AACH,wBAAsB,gBAAgB,CACpC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAyBf;AAED;;;;;;;;GAQG;AACH,wBAAsB,sBAAsB,CAC1C,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAwBf;AAED;;;;;;;GAOG;AACH,wBAAsB,eAAe,CACnC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAsBf;AAED;;;;;;;GAOG;AACH,wBAAsB,kBAAkB,CACtC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAcf;AAED;;;;;;;GAOG;AACH,wBAAsB,kBAAkB,CACtC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CA8Bf;AAED;;;;;;;GAOG;AACH,wBAAsB,mBAAmB,CACvC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAqBf;AAED;;;;;;;GAOG;AACH,wBAAsB,qBAAqB,CACzC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAqBf;AAED;;;;;;;GAOG;AACH,wBAAsB,kBAAkB,CACtC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAcf;AAED;;;;;;;GAOG;AACH,wBAAsB,iBAAiB,CACrC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAcf;AAMD;;;;;;;GAOG;AACH,wBAAsB,aAAa,CACjC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAYf;AAMD;;;;;;;GAOG;AACH,wBAAgB,mBAAmB,CACjC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,IAAI,CAwBN;AAED;;;;;;;GAOG;AACH,wBAAgB,oBAAoB,CAClC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,IAAI,CAmBN;AAED;;;;;;;GAOG;AACH,wBAAgB,oBAAoB,CAClC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,IAAI,CA4BN;AAMD;;;;;;;;;;;;;;GAcG;AACH,wBAAsB,mBAAmB,CACvC,GAAG,EAAE,OAAO,EACZ,GAAG,EAAE,QAAQ,EACb,IAAI,EAAE,YAAY,GACjB,OAAO,CAAC,IAAI,CAAC,CAiBf"}

package/dist/backend/backend/src/controllers/chat/chat.controller.js

@@ -313,7 +313,13 @@ export async function agentResponse(req, res, next) {
             return;
         }
         const chatService = getChatService();
-        // Resolve conversation: use provided ID or get/create the current one
+        // Resolve conversation: use provided ID or get/create the current one.
+        //
+        // Whether the caller NAMED the conversation matters downstream: the
+        // fallback below picks the globally-current conversation, which is a fine
+        // default for storing a message but is NOT evidence that this message
+        // belongs to that thread (issue #731).
+        const conversationIdWasExplicit = Boolean(conversationId);
         let resolvedConversationId = conversationId;
         if (!resolvedConversationId) {
             const current = await chatService.getCurrentConversation();
@@ -363,13 +369,29 @@ export async function agentResponse(req, res, next) {
             // within the cadence, the enforcer will nudge it with
             // [DELIVER_REQUIRED]. The service no-ops on non-slack conversations
             // and on non-delivery markers (e.g. [IN_PROGRESS]).
+            //
+            // Only when the caller NAMED the thread (issue #731): a cron-driven
+            // daily task reports completion with no conversationId, so the fallback
+            // above hands back whatever thread happens to be globally current —
+            // typically some long-resolved thread. Tracking that as a pending
+            // delivery made the watchdog demand a deliverable in an unrelated thread
+            // every single day, and because markPendingDelivery re-arms the reminder
+            // counter, the nudges never aged out.
             try {
-                const { OrcDeliveryEnforcerService } = await import('../../services/orc/orc-delivery-enforcer.service.js');
-                OrcDeliveryEnforcerService.getInstance()?.markPendingDelivery({
-                    conversationId: resolvedConversationId,
-                    agentSender: senderName,
-                    text: content,
-                });
+                if (conversationIdWasExplicit) {
+                    const { OrcDeliveryEnforcerService } = await import('../../services/orc/orc-delivery-enforcer.service.js');
+                    OrcDeliveryEnforcerService.getInstance()?.markPendingDelivery({
+                        conversationId: resolvedConversationId,
+                        agentSender: senderName,
+                        text: content,
+                    });
+                }
+                else {
+                    logger.debug('Skipping delivery tracking — conversation was inferred, not named', {
+                        senderName,
+                        conversationId: resolvedConversationId,
+                    });
+                }
             }
             catch (enforcerErr) {
                 logger.warn('OrcDeliveryEnforcer.markPendingDelivery threw (swallowed)', {