restored steering message

Author: jptaylor

CHANGELOG.md

@@ -9,7 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
-- `TaskAgent._inject_steering` now injects steer text as a single `<event name="task.steered">` user message with a priority directive line ("User has steered your task. Override and prioritize your current plan with the instruction below.") followed by the raw steer text. Same event-xml shape as `task.progress` / `task.cancelled` the agent already reads — without it, the steer was just another peer instruction to the original task description (which is itself a plain user message). Orchestrator side drops its now-redundant `"Steering instruction: "` text prefix.
+- `TaskAgent._inject_steering` now injects steer text as a single `<event name="task.steered">` user message with a short priority line followed by the raw steer text inline. Same event-xml shape as `task.progress` / `task.cancelled` the agent already reads — without it, the steer was just another peer instruction to the original task description (which is itself a plain user message). Orchestrator side drops its now-redundant `"Steering instruction: "` text prefix.
 - `start_task` and `steer_task` now pre-flight the target's finishing state. If the in-process TaskAgent has already called `finished` (or been cancelled), the steer would race the terminal turn and silently drop. `start_task` handles this transparently: the public entrypoint waits up to `TASK_STEER_CLOSING_WAIT_SECONDS` (default 5s) for the slot to free, then retries the start cleanly — the LLM never sees the intermediate state. Explicit `steer_task` returns `error: "task_closing"` for the LLM to chain `start_task` itself. Personal and corp in-process ships symmetric; BYOA skips the check.
 - Voice agent prompt: removed the contradictory "do NOT call start_task if all slots are occupied" line; added a short note that for steered instructions the agent should read the `task.completed` message and mention any unfulfilled intent to the commander.
 

src/gradientbang/prompts/agents/task_agent.md

@@ -16,7 +16,7 @@ Approach each task methodically:
 
 ## Steering Updates
 
-If you receive a user message beginning with "Steering instruction:", treat it as an update to the current task plan. Integrate it and continue.
+If you receive a `<event name="task.steered">` user message, treat it as a priority update to the current task plan. Integrate the inline instruction and continue.
 
 ## Historical Event Queries
 

src/gradientbang/prompts/agents/voice_agent.md

@@ -112,7 +112,7 @@ Functions requiring a task (use `start_task` immediately, in the same response):
 - Each corporation ship adds 1 task slot, up to a maximum of 3.
 - A player can own more than 3 corporation ships, but only 3 can run tasks at the same time.
 
-You may call `start_task` multiple times in a single response to fill available slots, or to send an additional instruction into a slot that's already in use (it auto-steers into the running task with a priority directive — the task reassesses on its next turn).
+You may call `start_task` multiple times in a single response to fill available slots, or to send an additional instruction into a slot that's already in use (it auto-steers into the running task as a priority update — the task reassesses on its next turn).
 Check the `Active tasks:` line in `status.snapshot` to see which slots are in use.
 
 When `task.completed` fires for a task that received steered instructions, read what the agent actually did. If an intent you steered in wasn't fulfilled, mention it to the commander — they can ask you to start it fresh if they still want it.

src/gradientbang/runtime/orchestrator.py

@@ -2759,8 +2759,8 @@ async def _handle_start_task_attempt(self, params: FunctionCallParams) -> dict:
                 # Phase: existing-task routing.
                 # If this ship already has an active task, auto-steer the new
                 # instruction into the running task. The TaskAgent wraps the
-                # steer with a <priority> directive on injection so it
-                # outranks the original task instruction (see
+                # steer in a task.steered event with a short priority hint so
+                # it outranks the original task instruction (see
                 # TaskAgent._inject_steering). If the active task is already
                 # finishing, the steer would race the terminal turn and
                 # silently drop — `_steer_existing_task` returns `task_closing`
@@ -3078,8 +3078,8 @@ async def _steer_existing_task(
         if not steering_text:
             return {"success": False, "error": "Empty steering instruction"}
         # No "Steering instruction:" prefix here — TaskAgent wraps the text
-        # with a structured priority directive before injecting into LLM
-        # context (see TaskAgent._inject_steering).
+        # in a task.steered event before injecting into LLM context (see
+        # TaskAgent._inject_steering).
 
         # Pre-flight: if the target in-process TaskAgent is in a finishing
         # state, the steer would race the terminal turn and silently drop.

src/gradientbang/runtime/subagents/task_agent.py

@@ -1715,15 +1715,13 @@ async def _inject_steering(self, text: str) -> None:
         cleaned = text.strip()
         if not cleaned:
             return
-        # Single user message: task.steered event with the steer text inside.
-        # Same event-xml shape as task.progress / task.cancelled the agent
-        # already reads. The directive line tells the LLM to prioritize this
-        # over its original task instruction (which is itself a plain user
-        # message — no priority signal without this wrap).
+        # Single user message: task.steered event with the steer text inline.
+        # This matches the event-message flow the task agent already reads,
+        # with one short priority hint before the raw steer instruction.
         steered_xml = (
             '<event name="task.steered">\n'
-            "User has steered your task. Override and prioritize your current "
-            f"plan with the instruction below.\n{cleaned}"
+            "Priority update: revise your plan now; follow this over the original task.\n"
+            f"{cleaned}\n"
             "</event>"
         )
         if self._llm_context is not None:

tests/unit/test_task_agent.py

@@ -974,9 +974,9 @@ async def test_steering_emits_replanning_message(self):
         injected = agent._llm_context.add_message.call_args.args[0]
         assert injected["role"] == "user"
         assert injected["content"].startswith('<event name="task.steered">')
-        assert "User has steered your task" in injected["content"]
-        assert "Override and prioritize" in injected["content"]
+        assert "Priority update: revise your plan now" in injected["content"]
         assert "Change direction" in injected["content"]
+        assert injected["content"].endswith("Change direction\n</event>")
 
     async def test_duplicate_progress_message_is_suppressed_without_new_action_or_event(self):
         agent = _make_task_agent()

tests/unit/test_voice_agent.py

@@ -703,7 +703,7 @@ async def test_steer_success(self):
         assert sent.target == "task_abc123"
         assert sent.task_id == full_id
         # No "Steering instruction:" prefix — TaskAgent wraps the raw text
-        # with a <priority> directive on injection.
+        # in a task.steered event on injection.
         assert sent.text == "Change course"
 
     async def test_steer_success_for_byoa_agent_targets_bus_name(self):