loop: replace plan file re-injection with structured TaskState re-injection

pkg/agent/core.go

@@ -364,9 +364,6 @@ func NewAgentCore(cfg AgentCoreConfig) Core {
 	}
 	if cfg.SessionID != "" {
 		os.MkdirAll(filepath.Join(ollieTmpDir(), cfg.SessionID), 0700) //nolint:errcheck
-		if mount := os.Getenv("OLLIE"); mount != "" {
-			run.PlanFile = filepath.Join(mount, "s", cfg.SessionID, "plan")
-		}
 	}
 
 	log := cfg.Log

pkg/agent/loop.go

@@ -1,12 +1,10 @@
 package agent
 
 import (
-	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
-	"os"
 	"strings"
 	"sync"
 	"time"
@@ -31,8 +29,7 @@ const (
 )
 
 // planReinjectInterval is the number of tool-call rounds between periodic
-// plan re-injection. Every N steps the current contents of the plan file
-// are surfaced back into the conversation to keep the model on track.
+// task state re-injection into the conversation.
 const planReinjectInterval = 10
 
 type toolExecutor func(ctx context.Context, name string, args json.RawMessage) (string, error)
@@ -56,27 +53,10 @@ type agentConfig struct {
 	// When reached, a soft nudge is injected and the loop exits cleanly.
 	// 0 means unlimited.
 	MaxSteps int
-	// PlanFile is the path to the session plan file (e.g. s/<id>/plan).
-	// When set, the loop re-surfaces its contents every planReinjectInterval
-	// tool-call rounds. Empty string disables plan re-injection.
-	PlanFile string
 	// IncrToolCallCount increments and returns the session-wide tool call counter.
 	IncrToolCallCount func() int64
 }
 
-// readPlanFile returns the contents of the plan file, or empty string if
-// the file does not exist, is empty, or cannot be read.
-func readPlanFile(path string) string {
-	if path == "" {
-		return ""
-	}
-	b, err := os.ReadFile(path)
-	if err != nil || len(bytes.TrimSpace(b)) == 0 {
-		return ""
-	}
-	return string(b)
-}
-
 func run(ctx context.Context, cfg agentConfig, state state) error {
 	var step int
 	var consecutiveErrors int // rounds where every tool call returned an error
@@ -97,7 +77,7 @@ func run(ctx context.Context, cfg agentConfig, state state) error {
 			history = append([]backend.Message{{Role: "system", Content: cfg.preamble}}, history...)
 		}
 		if ts := state.taskState(); ts != nil {
-			if msg := ts.render(); msg != "" {
+			if msg := ts.render(); msg != "" && len(history) > 0 {
 				// Insert after system prompt, before conversation.
 				history = append([]backend.Message{history[0], {Role: "user", Content: msg}}, history[1:]...)
 			}
@@ -493,18 +473,16 @@ func run(ctx context.Context, cfg agentConfig, state state) error {
 			break
 		}
 
-		// Periodically re-surface the plan file so it stays visible as tool
-		// results accumulate in the context. Only fires when PlanFile is set
-		// and the file has content.
-		if cfg.PlanFile != "" && step > 0 && step%planReinjectInterval == 0 {
-			if plan := readPlanFile(cfg.PlanFile); plan != "" {
-				state.update(backend.Message{
-					Role: "user",
-					Content: fmt.Sprintf(
-						"[system: your current plan:\n\n%s\n\nKeep this plan current — mark completed steps and revise if your approach has changed. Continue.]",
-						plan,
-					),
-				}, nil)
+		// Periodic task state re-injection: every N tool rounds, surface the
+		// structured task state back into the conversation to keep the model on track.
+		if step > 0 && step%planReinjectInterval == 0 {
+			if ts := state.taskState(); ts != nil {
+				if msg := ts.render(); msg != "" {
+					state.update(backend.Message{
+						Role:    "user",
+						Content: "[system: review your current task state and continue. Update it if your approach has changed.]\n\n" + msg,
+					}, nil)
+				}
 			}
 		}
 

pkg/agent/loop_maxsteps_test.go

@@ -4,8 +4,6 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
-	"os"
-	"path/filepath"
 	"strings"
 	"testing"
 
@@ -132,17 +130,11 @@ func TestMaxStepsExactBoundary(t *testing.T) {
 	}
 }
 
-// TestPlanReinjection verifies that the plan file contents are re-surfaced
+// TestTaskStateReinjection verifies that the loop re-injects the task state
 // into the conversation every planReinjectInterval tool rounds.
-func TestPlanReinjection(t *testing.T) {
-	// Write a plan file to a temp location.
-	planPath := filepath.Join(t.TempDir(), "plan")
-	os.WriteFile(planPath, []byte("- [ ] step one\n- [ ] step two\n"), 0644)
-
+func TestTaskStateReinjection(t *testing.T) {
 	// We need planReinjectInterval+1 tool rounds so the re-injection fires
 	// at step == planReinjectInterval (0-indexed, checked after increment).
-	// The check is: step > 0 && step%planReinjectInterval == 0, and step is
-	// incremented at the end of the loop, so it fires after the 10th round.
 	n := planReinjectInterval + 1
 	var responses []mockResponse
 	for i := range n {
@@ -155,27 +147,38 @@ func TestPlanReinjection(t *testing.T) {
 
 	mb := &mockBackend{respond: sequentialStream(responses)}
 	cfg := agentConfig{
-		Backend:  mb,
-		Tools:    []backend.Tool{{Name: "tool"}},
-		PlanFile: planPath,
+		Backend: mb,
+		Tools:   []backend.Tool{{Name: "tool"}},
 		Exec: func(_ context.Context, _ string, _ json.RawMessage) (string, error) {
 			return "ok", nil
 		},
 	}
 
-	s := newState()
+	s := &taskStateState{
+		simpleState: simpleState{},
+		ts:          &TaskState{Objective: "test objective", PlanStep: "step one"},
+	}
 	if err := run(context.Background(), cfg, s); err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
 
-	var planSeen bool
+	var seen bool
 	for _, m := range s.history() {
-		if m.Role == "user" && strings.Contains(m.Content, "step one") && strings.Contains(m.Content, "your current plan") {
-			planSeen = true
+		if m.Role == "user" && strings.Contains(m.Content, "test objective") && strings.Contains(m.Content, "task state") {
+			seen = true
 			break
 		}
 	}
-	if !planSeen {
-		t.Error("expected plan re-injection message in conversation history")
+	if !seen {
+		t.Error("expected task state re-injection message in conversation history")
 	}
 }
+
+// taskStateState wraps simpleState with a non-nil TaskState.
+type taskStateState struct {
+	simpleState
+	ts *TaskState
+}
+
+func (s *taskStateState) taskState() *TaskState       { return s.ts }
+func (s *taskStateState) updateTaskState(ts TaskState) { s.ts = &ts }