ci: integrate new tasks cli features in ai triage workflow (#20007)

Integrates new experimental tasks CLI commands into traiage workflow

Sample run:
https://github.com/coder/coder/actions/runs/18095432003/job/51485262065

.github/workflows/traiage.yaml

@@ -84,8 +84,6 @@ jobs:
           RUN_ID: ${{ github.run_id }}
           TEMPLATE_PARAMETERS: ${{ secrets.TRAIAGE_TEMPLATE_PARAMETERS }}
           TEMPLATE_PRESET: ${{ inputs.template_preset }}
-          # TODO: replace with coder exp task command
-          APP_SLUG: ccw
         run: |
           # Fetch issue description using `gh` CLI
           issue_description=$(gh issue view "${ISSUE_URL}")
@@ -105,7 +103,7 @@ jobs:
           export TASK_NAME="${PREFIX}-${CONTEXT_KEY}-${RUN_ID}"
           echo "Creating task: $TASK_NAME"
           ./scripts/traiage.sh create
-          ./scripts/traiage.sh wait
+          coder exp task status "$TASK_NAME" --watch
           echo "TASK_NAME=${TASK_NAME}" >> "${GITHUB_OUTPUT}"
           echo "TASK_NAME=${TASK_NAME}" >> "${GITHUB_ENV}"
 
@@ -122,7 +120,6 @@ jobs:
         id: generate-summary
         env:
           ARCHIVE_URL: ${{ steps.create-archive.outputs.archive_url }}
-          APP_SLUG: ccw
           BUCKET_PREFIX: "gs://coder-traiage-outputs/traiage"
           CONTEXT_KEY: ${{ steps.extract-context.outputs.context_key }}
           GITHUB_TOKEN: ${{ github.token }}

scripts/traiage.sh

@@ -30,8 +30,8 @@ create() {
 	set -e
 
 	if [[ "${TASK_NAME}" == $(jq -r '.name' <<<"${task_json}") ]]; then
-		# TODO(Cian): Send PROMPT to the agent in the existing workspace.
-		echo "Task \"${TASK_NAME}\" already exists."
+		echo "Task \"${TASK_NAME}\" already exists. Sending prompt to existing task."
+		prompt
 		exit 0
 	fi
 
@@ -67,193 +67,55 @@ ssh_config() {
 }
 
 prompt() {
-	requiredenvs CODER_URL CODER_SESSION_TOKEN TASK_NAME APP_SLUG PROMPT
+	requiredenvs CODER_URL CODER_SESSION_TOKEN TASK_NAME PROMPT
 
-	wait_agentapi_stable
+	${CODER_BIN} \
+		--url "${CODER_URL}" \
+		--token "${CODER_SESSION_TOKEN}" \
+		exp tasks status "${TASK_NAME}" \
+		--watch >/dev/null
 
-	username=$(curl \
-		--fail \
-		--header "Coder-Session-Token: ${CODER_SESSION_TOKEN}" \
-		--location \
-		--show-error \
-		--silent \
-		"${CODER_URL}/api/v2/users/me" | jq -r '.username')
+	${CODER_BIN} \
+		--url "${CODER_URL}" \
+		--token "${CODER_SESSION_TOKEN}" \
+		exp tasks send "${TASK_NAME}" \
+		--stdin \
+		<<<"${PROMPT}"
 
-	payload="{
-		\"content\": \"${PROMPT}\",
-		\"type\": \"user\"
-	}"
+	${CODER_BIN} \
+		--url "${CODER_URL}" \
+		--token "${CODER_SESSION_TOKEN}" \
+		exp tasks status "${TASK_NAME}" \
+		--watch >/dev/null
 
-	response=$(curl \
-		--data-raw "${payload}" \
-		--fail \
-		--header "Content-Type: application/json" \
-		--header "Coder-Session-Token: ${CODER_SESSION_TOKEN}" \
-		--location \
-		--request POST \
-		--show-error \
-		--silent \
-		"${CODER_URL}/@${username}/${TASK_NAME}/apps/${APP_SLUG}/message" | jq -r '.ok')
-	if [[ "${response}" != "true" ]]; then
-		echo "Failed to send prompt"
-		exit 1
-	fi
-
-	# Wait for agentapi to process the response and return the last agent message
-	wait_agentapi_stable
-
-	CODER_USERNAME="${username}" last_message
+	last_message
 }
 
 last_message() {
-	requiredenvs CODER_URL CODER_SESSION_TOKEN CODER_USERNAME TASK_NAME APP_SLUG PROMPT
-	last_msg=$(curl \
-		--fail \
-		--header "Content-Type: application/json" \
-		--header "Coder-Session-Token: ${CODER_SESSION_TOKEN}" \
-		--location \
-		--show-error \
-		--silent \
-		"${CODER_URL}/@${CODER_USERNAME}/${TASK_NAME}/apps/${APP_SLUG}/messages" |
-		jq -r 'last(.messages[] | select(.role=="agent") | [.])')
+	requiredenvs CODER_URL CODER_SESSION_TOKEN TASK_NAME PROMPT
 
+	last_msg_json=$(
+		${CODER_BIN} \
+			--url "${CODER_URL}" \
+			--token "${CODER_SESSION_TOKEN}" \
+			exp tasks logs "${TASK_NAME}" \
+			--output json
+	)
+	last_output_msg=$(jq -r 'last(.[] | select(.type=="output")) | .content' <<<"${last_msg_json}")
+	# HACK: agentapi currently doesn't split multiple messages, so you can end up with tool
+	# call responses in the output.
+	last_msg=$(tac <<<"${last_output_msg}" | sed '/^● /q' | tr -d '●' | tac)
 	echo "${last_msg}"
 }
 
 wait_agentapi_stable() {
-	requiredenvs CODER_URL CODER_SESSION_TOKEN TASK_NAME APP_SLUG
-	username=$(curl \
-		--fail \
-		--header "Coder-Session-Token: ${CODER_SESSION_TOKEN}" \
-		--location \
-		--show-error \
-		--silent \
-		"${CODER_URL}/api/v2/users/me" | jq -r '.username')
+	requiredenvs CODER_URL CODER_SESSION_TOKEN TASK_NAME
 
-	for attempt in {1..120}; do
-		# First wait for task to start
-		set +o pipefail
-		task_status=$("${CODER_BIN}" \
-			--url "${CODER_URL}" \
-			--token "${CODER_SESSION_TOKEN}" \
-			exp tasks status "${TASK_NAME}" \
-			--output json |
-			jq -r '.status')
-		set -o pipefail
-		echo "Task status is ${task_status}"
-		if [[ "${task_status}" == "running" ]]; then
-			echo "Task is running"
-			break
-		fi
-		echo "Waiting for task status to be running (attempt ${attempt}/120)"
-		sleep 5
-	done
-
-	for attempt in {1..120}; do
-		# Workspace agent must be healthy
-		set +o pipefail
-		healthy=$("${CODER_BIN}" \
-			--url "${CODER_URL}" \
-			--token "${CODER_SESSION_TOKEN}" \
-			exp tasks status "${TASK_NAME}" \
-			--output json |
-			jq -r '.workspace_agent_health.healthy')
-		set -o pipefail
-		if [[ "${healthy}" == "true" ]]; then
-			echo "Workspace agent is healthy"
-			break
-		fi
-		echo "Workspace agent not yet healthy (attempt ${attempt}/120)"
-		sleep 5
-	done
-
-	for attempt in {1..120}; do
-		# AgentAPI application should not be 404'ing
-		agentapi_app_status_code=$(curl \
-			--header "Content-Type: application/json" \
-			--header "Coder-Session-Token: ${CODER_SESSION_TOKEN}" \
-			--location \
-			--request GET \
-			--show-error \
-			--silent \
-			--output /dev/null \
-			--write-out '%{http_code}' \
-			"${CODER_URL}/@${username}/${TASK_NAME}/apps/${APP_SLUG}/status")
-		echo "Workspace app ${APP_SLUG} returned ${agentapi_app_status_code}"
-		if [[ "${agentapi_app_status_code}" == "200" ]]; then
-			echo "AgentAPI is running"
-			break
-		fi
-		echo "AgentAPI not yet running (attempt ${attempt}/120)"
-		sleep 5
-	done
-
-	for attempt in {1..120}; do
-		# AgentAPI must be stable
-		set +o pipefail
-		agentapi_status=$(curl \
-			--header "Content-Type: application/json" \
-			--header "Coder-Session-Token: ${CODER_SESSION_TOKEN}" \
-			--location \
-			--request GET \
-			--show-error \
-			--silent \
-			"${CODER_URL}/@${username}/${TASK_NAME}/apps/${APP_SLUG}/status" | jq -r '.status')
-		set -o pipefail
-		if [[ "${agentapi_status}" == "stable" ]]; then
-			echo "AgentAPI stable"
-			break
-		fi
-		echo "Waiting for AgentAPI to report stable status (attempt ${attempt}/120)"
-		sleep 5
-	done
-
-	# At this point the task is running, the workspace agent is healthy,
-	# the AgentAPI app is running and the AgentAPI reports stable status.
-	# Now we wait for the Task Agent to report 'idle', 'complete', or 'failure'.
-	for attempt in {1..120}; do
-		task_json=$(${CODER_BIN} \
-			--url "${CODER_URL}" \
-			--token "${CODER_SESSION_TOKEN}" \
-			exp tasks status "${TASK_NAME}" \
-			--output json)
-		set +o pipefail
-		current_state=$(jq -r '.current_state.state' <<<"${task_json}")
-		current_message=$(jq -r '.current_state.message' <<<"${task_json}")
-		set -o pipefail
-
-		# The current_state or current_message may be null if the Task Agent
-		# has not yet reported its status.
-		if [[ -z "${current_state}" ]] || [[ -z "${current_message}" ]]; then
-			echo "Waiting for Task Agent to report state (attempt ${attempt}/120)"
-			sleep 5
-			continue
-		fi
-		break
-	done
-
-	# Finally, wait for the Task Agent to report 'idle', 'complete', or 'failure'. This is unbounded, as we don't know how long the agent will take.
-	while true; do
-		task_json=$(${CODER_BIN} \
-			--url "${CODER_URL}" \
-			--token "${CODER_SESSION_TOKEN}" \
-			exp tasks status "${TASK_NAME}" \
-			--output json)
-		set +o pipefail
-
-		current_state=$(jq -r '.current_state.state' <<<"${task_json}")
-		current_message=$(jq -r '.current_state.message' <<<"${task_json}")
-		set -o pipefail
-		if [[ "${current_state}" == "working" ]]; then
-			echo "Task Agent is ${current_state} and reports: \"${current_message}\""
-			echo "Waiting for Task Agent to report idle state (attempt ${attempt}/120)"
-			sleep 5
-			continue
-		else
-			echo "Task Agent is ${current_state} and reports: \"${current_message}\""
-			break
-		fi
-	done
+	${CODER_BIN} \
+		--url "${CODER_URL}" \
+		--token "${CODER_SESSION_TOKEN}" \
+		exp tasks status "${TASK_NAME}" \
+		--watch
 }
 
 archive() {
@@ -287,7 +149,7 @@ archive() {
 }
 
 summary() {
-	requiredenvs CODER_URL CODER_SESSION_TOKEN TASK_NAME APP_SLUG
+	requiredenvs CODER_URL CODER_SESSION_TOKEN TASK_NAME
 	ssh_config
 
 	# We want the heredoc to be expanded locally and not remotely.