Add turn_end hook

agent-schema.json

@@ -620,6 +620,13 @@
             "$ref": "#/definitions/HookDefinition"
           }
         },
+        "turn_end": {
+          "type": "array",
+          "description": "Hooks that run once per agent turn when the turn finishes — the symmetric counterpart of turn_start. Fires no matter why the turn ended: a normal stop, an error, a hook-driven shutdown, the loop detector, or context cancellation. The reason is reported via the hook input's reason field ('normal', 'continue', 'steered', 'error', 'canceled', 'hook_blocked', 'loop_detected'). Observational; output is ignored.",
+          "items": {
+            "$ref": "#/definitions/HookDefinition"
+          }
+        },
         "before_llm_call": {
           "type": "array",
           "description": "Hooks that run just before each model call (after turn_start has assembled the messages). Use for observability, cost guardrails, or auditing without contributing system messages — turn_start is the right event for the latter.",

docs/configuration/hooks/index.md

@@ -43,6 +43,7 @@ docker-agent dispatches the following hook events:
 | `session_start`             | When a session begins or resumes                                                    | No         |
 | `user_prompt_submit`        | Once per user message, after submission and before the model runs                   | Yes        |
 | `turn_start`                | At the start of every agent turn (each model call)                                  | No         |
+| `turn_end`                  | At the end of every agent turn — fires no matter why the turn ended                 | No         |
 | `before_llm_call`           | Just before every model call (after `turn_start`)                                   | Yes        |
 | `after_llm_call`            | After every successful model call, before the response is recorded                  | No         |
 | `session_end`               | When a session terminates                                                           | No         |
@@ -225,6 +226,7 @@ In addition to the common fields, each event ships its own payload:
 | `session_start`             | `source` — one of `startup`, `resume`, `clear`, `compact`                                                     |
 | `user_prompt_submit`        | `prompt` — the text the user just submitted                                                                   |
 | `turn_start`                | _none_ (just the common fields)                                                                               |
+| `turn_end`                  | `agent_name`, `reason` — one of `normal`, `continue`, `steered`, `error`, `canceled`, `hook_blocked`, `loop_detected` |
 | `before_llm_call`           | _none_                                                                                                        |
 | `after_llm_call`            | `agent_name`, `stop_response`, `last_user_message`                                                            |
 | `session_end`               | `reason` — one of `clear`, `logout`, `prompt_input_exit`, `other`                                             |
@@ -495,6 +497,24 @@ Use `on_error` and `on_max_iterations` instead of `notification` when you want a
 
 `turn_start` fires at the start of every agent turn (each model call). Anything you contribute via `additional_context` (or plain stdout) is appended as a **transient** system message for that turn only — it is *not* persisted to the session. Use it for fast-moving signals like the date, current git state, or per-turn prompt files. The built-in hooks `add_date`, `add_prompt_files`, `add_git_status`, and `add_git_diff` all target this event.
 
+### Turn-End: per-turn finalizer
+
+`turn_end` is the symmetric counterpart of `turn_start`. It fires once per turn when the iteration finishes — no matter why. The runtime guarantees the dispatch on every exit path (a normal stop, an error, a hook-driven shutdown, the loop detector, even context cancellation), and it uses `context.WithoutCancel` internally so handlers run to completion on Ctrl+C.
+
+The `reason` field classifies the exit:
+
+| `reason`        | When |
+| --------------- | ---- |
+| `normal`        | Model finished cleanly with no follow-up |
+| `continue`      | More iterations to come (e.g. tool calls, follow-up message) |
+| `steered`       | Drained steered messages prompted a re-entry |
+| `error`         | Model call failed (`handleStreamError` exited the loop) |
+| `canceled`      | Context was cancelled (e.g. Ctrl+C) |
+| `hook_blocked`  | `before_llm_call` or `post_tool_use` denied the call |
+| `loop_detected` | The consecutive-tool-call loop detector terminated the turn |
+
+`turn_end` is observational — the result is ignored. Use it to time turns, accumulate per-turn metrics (token usage, tool counts), or notify external observability pipelines symmetrically with `turn_start`.
+
 ### Before/After-LLM-Call: budget guards and model auditing
 
 `before_llm_call` fires immediately before every model call (after `turn_start` has assembled the messages). It cannot contribute context — use `turn_start` for that — but it can **stop the run** by returning `decision: block` (or exit code 2). The built-in `max_iterations` hook implements a hard cap on top of this event.

examples/hooks.yaml

@@ -16,6 +16,7 @@
 #   session_start     - one-time setup; AdditionalContext PERSISTS in the session
 #   user_prompt_submit- runs once per user message, before the first LLM call
 #   turn_start        - per-turn context; AdditionalContext is TRANSIENT
+#   turn_end          - per-turn finalizer; fires no matter why the turn ended
 #   before_llm_call   - just before each model call (observability, guardrails)
 #   after_llm_call    - just after a successful model call
 #   session_end       - cleanup when the session terminates
@@ -62,6 +63,7 @@
 #   /tmp/agent-session.log         (session_start, session_end)
 #   /tmp/agent-prompts.log         (user_prompt_submit)
 #   /tmp/agent-llm-calls.log       (before_llm_call, after_llm_call)
+#   /tmp/agent-turns.log           (turn_end)
 #   /tmp/agent-tool-results.log    (post_tool_use)
 #   /tmp/agent-permissions.log     (permission_request)
 #   /tmp/agent-compactions.log     (pre_compact)
@@ -225,6 +227,35 @@ agents:
             - GUIDELINES.md
             - PROJECT.md
 
+      # ====================================================================
+      # TURN-END - runs ONCE per turn after the iteration finishes — the
+      # symmetric counterpart of turn_start. Fires no matter why the turn
+      # ended: a normal stop, an error, a hook-driven shutdown, the loop
+      # detector, or context cancellation. The reason is reported via
+      # the .reason field:
+      #
+      #   normal           - model finished cleanly, no follow-up
+      #   continue         - more iterations to come (e.g. tool calls)
+      #   steered          - drained steered messages prompted a re-entry
+      #   error            - model call failed (handleStreamError)
+      #   canceled         - context cancellation (Ctrl+C, parent ctx done)
+      #   hook_blocked     - before_llm_call or post_tool_use signalled stop
+      #   loop_detected    - degenerate consecutive-tool-call loop
+      #
+      # Observational; the result is ignored. Use it to time turns,
+      # accumulate per-turn metrics (token usage, tool counts), or notify
+      # external observability pipelines.
+      # ====================================================================
+      turn_end:
+        - type: command
+          timeout: 5
+          command: |
+            INPUT=$(cat)
+            SESSION_ID=$(echo "$INPUT" | jq -r '.session_id // "unknown"')
+            AGENT=$(echo "$INPUT" | jq -r '.agent_name // "unknown"')
+            REASON=$(echo "$INPUT" | jq -r '.reason // "unknown"')
+            echo "[$(date)] [←] $SESSION_ID $AGENT turn ended (reason=$REASON)" >> /tmp/agent-turns.log
+
       # ====================================================================
       # BEFORE-LLM-CALL - fires just before every model invocation, AFTER
       # turn_start has assembled the messages slice. Use for observability

pkg/config/latest/types.go

@@ -1709,6 +1709,15 @@ type HooksConfig struct {
 	// turn instead of bloating the message history on every resume.
 	TurnStart []HookDefinition `json:"turn_start,omitempty" yaml:"turn_start,omitempty"`
 
+	// TurnEnd hooks run once per agent turn when the turn finishes —
+	// the symmetric counterpart of TurnStart. Fires no matter why the
+	// turn ended: a normal stop, an error, a hook-driven shutdown, the
+	// loop detector, or context cancellation. The reason is reported
+	// in the hook input's reason field ("normal", "continue",
+	// "steered", "error", "canceled", "hook_blocked",
+	// "loop_detected"). Observational; output is ignored.
+	TurnEnd []HookDefinition `json:"turn_end,omitempty" yaml:"turn_end,omitempty"`
+
 	// BeforeLLMCall hooks run just before each model call (after
 	// turn_start). Use this for observability, cost guardrails, or
 	// auditing without contributing system messages — turn_start is the
@@ -1815,6 +1824,7 @@ func (h *HooksConfig) IsEmpty() bool {
 		len(h.SessionStart) == 0 &&
 		len(h.UserPromptSubmit) == 0 &&
 		len(h.TurnStart) == 0 &&
+		len(h.TurnEnd) == 0 &&
 		len(h.BeforeLLMCall) == 0 &&
 		len(h.AfterLLMCall) == 0 &&
 		len(h.SessionEnd) == 0 &&
@@ -1971,6 +1981,13 @@ func (h *HooksConfig) validate() error {
 		}
 	}
 
+	// Validate TurnEnd hooks
+	for i, hook := range h.TurnEnd {
+		if err := hook.validate("turn_end", i); err != nil {
+			return err
+		}
+	}
+
 	// Validate BeforeLLMCall hooks
 	for i, hook := range h.BeforeLLMCall {
 		if err := hook.validate("before_llm_call", i); err != nil {

pkg/hooks/dispatch_test.go

@@ -25,6 +25,7 @@ var onlyHooks = map[EventType]*Config{
 	EventPostToolUse:      {PostToolUse: matcherWildcard},
 	EventSessionStart:     {SessionStart: trueHook},
 	EventTurnStart:        {TurnStart: trueHook},
+	EventTurnEnd:          {TurnEnd: trueHook},
 	EventBeforeLLMCall:    {BeforeLLMCall: trueHook},
 	EventAfterLLMCall:     {AfterLLMCall: trueHook},
 	EventSessionEnd:       {SessionEnd: trueHook},

pkg/hooks/executor.go

@@ -79,6 +79,7 @@ func compileEvents(c *Config) map[EventType][]matcher {
 		EventSessionStart:           flat(c.SessionStart),
 		EventUserPromptSubmit:       flat(c.UserPromptSubmit),
 		EventTurnStart:              flat(c.TurnStart),
+		EventTurnEnd:                flat(c.TurnEnd),
 		EventBeforeLLMCall:          flat(c.BeforeLLMCall),
 		EventAfterLLMCall:           flat(c.AfterLLMCall),
 		EventSessionEnd:             flat(c.SessionEnd),

pkg/hooks/hooks_test.go

@@ -124,6 +124,13 @@ func TestConfigIsEmpty(t *testing.T) {
 			},
 			expected: false,
 		},
+		{
+			name: "with turn_end",
+			config: Config{
+				TurnEnd: []Hook{{Type: HookTypeCommand}},
+			},
+			expected: false,
+		},
 	}
 
 	for _, tt := range tests {
@@ -706,7 +713,7 @@ func TestPlainStdoutBecomesAdditionalContext(t *testing.T) {
 	observationalEvents := []EventType{
 		EventBeforeLLMCall, EventAfterLLMCall, EventOnError,
 		EventOnMaxIterations, EventNotification, EventOnUserInput, EventSessionEnd,
-		EventBeforeCompaction, EventAfterCompaction,
+		EventBeforeCompaction, EventAfterCompaction, EventTurnEnd,
 	}
 
 	for _, ev := range contextEvents {
@@ -750,6 +757,8 @@ func configWithFlatHook(ev EventType, h Hook) *Config {
 		cfg.SessionStart = []Hook{h}
 	case EventTurnStart:
 		cfg.TurnStart = []Hook{h}
+	case EventTurnEnd:
+		cfg.TurnEnd = []Hook{h}
 	case EventBeforeLLMCall:
 		cfg.BeforeLLMCall = []Hook{h}
 	case EventAfterLLMCall:

pkg/hooks/types.go

@@ -51,6 +51,15 @@ const (
 	// EventTurnStart fires at the start of every agent turn (each model
 	// call). AdditionalContext is injected transiently and never persisted.
 	EventTurnStart EventType = "turn_start"
+	// EventTurnEnd fires once per agent turn (each model call) when the
+	// turn finishes — symmetric to [EventTurnStart]. It runs no matter
+	// why the turn ended: a normal stop, an error, a hook-driven
+	// shutdown, the loop detector, or context cancellation. The reason
+	// is reported in [Input.Reason] using one of the turnEndReason*
+	// constants in the runtime package ("normal", "continue",
+	// "steered", "error", "canceled", "hook_blocked",
+	// "loop_detected"). Observational; output is ignored.
+	EventTurnEnd EventType = "turn_end"
 	// EventBeforeLLMCall fires immediately before each model call.
 	// Returning decision="block" (or continue=false / exit code 2)
 	// stops the run loop before the model is invoked — useful for hard
@@ -209,6 +218,8 @@ type Input struct {
 	// PreCompact specific: "manual", "auto", "overflow", "tool_overflow".
 	Source string `json:"source,omitempty"`
 	// SessionEnd specific: "clear", "logout", "prompt_input_exit", "other".
+	// TurnEnd specific: "normal", "continue", "steered", "error",
+	// "canceled", "hook_blocked", "loop_detected".
 	Reason string `json:"reason,omitempty"`
 	// Stop / AfterLLMCall / SubagentStop: the model's final response content.
 	StopResponse string `json:"stop_response,omitempty"`

pkg/runtime/hooks.go

@@ -132,6 +132,48 @@ func (r *LocalRuntime) executeTurnStartHooks(ctx context.Context, sess *session.
 	}, events))
 }
 
+// Reason values reported in [hooks.Input.Reason] when [hooks.EventTurnEnd]
+// fires. The runtime guarantees that turn_end runs once per turn that
+// fired turn_start, no matter how the turn exited; the reason classifies
+// which exit path the runtime took.
+const (
+	// turnEndReasonNormal — the model finished the turn cleanly and the
+	// run loop is about to break out (no further iterations).
+	turnEndReasonNormal = "normal"
+	// turnEndReasonContinue — the turn finished cleanly and the loop is
+	// about to start a new iteration (e.g. after tool calls, or after a
+	// stop with a queued follow-up).
+	turnEndReasonContinue = "continue"
+	// turnEndReasonSteered — the turn finished and was followed by
+	// drained steered messages, prompting a new iteration.
+	turnEndReasonSteered = "steered"
+	// turnEndReasonError — the model call failed and the runtime is
+	// shutting down the run (handleStreamError returned non-retry).
+	turnEndReasonError = "error"
+	// turnEndReasonCanceled — the turn ended because the stream context
+	// was cancelled (e.g. user Ctrl+C). Includes deferred firing on
+	// any return path while ctx is done.
+	turnEndReasonCanceled = "canceled"
+	// turnEndReasonHookBlocked — a hook (before_llm_call or
+	// post_tool_use) signalled run termination via a deny verdict.
+	turnEndReasonHookBlocked = "hook_blocked"
+	// turnEndReasonLoopDetected — the consecutive-tool-call loop
+	// detector terminated the turn.
+	turnEndReasonLoopDetected = "loop_detected"
+)
+
+// executeTurnEndHooks fires turn_end once per turn — symmetric to
+// turn_start. Observational; the result is discarded. Reason is one
+// of the turnEndReason* constants above and is reported via
+// [hooks.Input.Reason] so handlers can branch on the exit path.
+func (r *LocalRuntime) executeTurnEndHooks(ctx context.Context, sess *session.Session, a *agent.Agent, reason string, events chan Event) {
+	r.dispatchHook(ctx, a, hooks.EventTurnEnd, &hooks.Input{
+		SessionID: sess.ID,
+		AgentName: a.Name(),
+		Reason:    reason,
+	}, events)
+}
+
 // contextMessages converts a context-providing hook's AdditionalContext
 // into a one-element transient system-message slice ready to thread
 // through [session.Session.GetMessages]. Returns nil for empty results