feat: resume support, summary extraction, and task state improvements

- Extend Resume to CANCELLED, FAILED, and BUDGET_EXCEEDED tasks
- Add summary extraction from agent stdout stream-json output
- Fix storage: persist stdout/stderr/artifact_dir paths in UpdateExecution
- Clear question_json on ResetTaskForRetry
- Resume BLOCKED tasks in preserved sandbox so Claude finds its session
- Add planning preamble: CLAUDOMATOR_SUMMARY_FILE env var + summary step
- Update ADR-002 with new state transitions
- UI style improvements

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

10 files changed, 641 insertions, 36 deletions

diff --git a/docs/adr/002-task-state-machine.md b/docs/adr/002-task-state-machine.md
index 310c337..6910f6a 100644
--- a/docs/adr/002-task-state-machine.md
+++ b/docs/adr/002-task-state-machine.md
@@ -66,13 +66,13 @@ True terminal state (no outgoing transitions): `COMPLETED`. All other non-succes
 | `QUEUED` | `RUNNING` | Pool goroutine starts execution |
 | `QUEUED` | `CANCELLED` | `POST /api/tasks/{id}/cancel` |
 | `RUNNING` | `READY` | Runner exits 0, no question file, top-level task (`parent_task_id == ""`), and task has no subtasks |
-| `RUNNING` | `BLOCKED` | Runner exits 0, no question file, top-level task (`parent_task_id == ""`), and task has subtasks |
+| `RUNNING` | `BLOCKED` (subtasks) | Runner exits 0, no question file, top-level task (`parent_task_id == ""`), and task has subtasks |
+| `RUNNING` | `BLOCKED` (question) | Runner exits 0 but left a `question.json` file in log dir (any task type) |
 | `RUNNING` | `COMPLETED` | Runner exits 0, no question file, subtask (`parent_task_id != ""`) |
 | `RUNNING` | `FAILED` | Runner exits non-zero or stream signals `is_error: true` |
 | `RUNNING` | `TIMED_OUT` | Context deadline exceeded (`context.DeadlineExceeded`) |
 | `RUNNING` | `CANCELLED` | Context cancelled (`context.Canceled`) |
 | `RUNNING` | `BUDGET_EXCEEDED` | `--max-budget-usd` exceeded (signalled by runner) |
-| `RUNNING` | `BLOCKED` | Runner exits 0 but left a `question.json` file in log dir |
 | `READY` | `COMPLETED` | `POST /api/tasks/{id}/accept` |
 | `READY` | `PENDING` | `POST /api/tasks/{id}/reject` (with optional comment) |
 | `FAILED` | `QUEUED` | Retry (manual re-run via `POST /api/tasks/{id}/run`) |
@@ -85,7 +85,7 @@ True terminal state (no outgoing transitions): `COMPLETED`. All other non-succes
 ## Implementation
 
 **Validation:** `task.ValidTransition(from, to State) bool`
-(`internal/task/task.go:93`) — called by API handlers before every state change.
+(`internal/task/task.go:123`) — called by API handlers before every state change.
 
 **State writes:** `storage.DB.UpdateTaskState(id, state)` — single source of
 write; called by both API handlers and the executor pool.
@@ -158,9 +158,9 @@ Task lifecycle changes produce WebSocket broadcasts to all connected clients:
 
 ## Known Limitations and Edge Cases
 
-- **`BUDGET_EXCEEDED` transition.** `BUDGET_EXCEEDED` appears in `terminalFailureStates`
-  (used by `waitForDependencies`) but has no outgoing transitions in `ValidTransition`,
-  making it permanently terminal. There is no `/resume` endpoint for it.
+- **`BUDGET_EXCEEDED` retry.** `BUDGET_EXCEEDED → QUEUED` is a valid transition (retry via
+  `POST /run`), matching `FAILED` and `CANCELLED` behaviour. However, there is no dedicated
+  `/resume` endpoint for it — callers must use the standard `/run` restart path.
 
 - **Retry enforcement.** `RetryConfig.MaxAttempts` is stored but not enforced by
   the pool. The API allows unlimited manual retries via `POST /run` from `FAILED`.
@@ -178,9 +178,9 @@ Task lifecycle changes produce WebSocket broadcasts to all connected clients:
 
 | Concern | File | Lines |
 |---|---|---|
-| State constants | `internal/task/task.go` | 7–18 |
-| `ValidTransition` | `internal/task/task.go` | 93–109 |
-| State machine tests | `internal/task/task_test.go` | 8–72 |
+| State constants | `internal/task/task.go` | 9–20 |
+| `ValidTransition` | `internal/task/task.go` | 107–130 |
+| State machine tests | `internal/task/task_test.go` | 8–75 |
 | Pool execute | `internal/executor/executor.go` | 194–303 |
 | Pool executeResume | `internal/executor/executor.go` | 116–185 |
 | Dependency wait | `internal/executor/executor.go` | 305–340 |
diff --git a/docs/packages_old/task.md b/docs/packages_old/task.md
new file mode 100644
index 0000000..923cd56
--- /dev/null
+++ b/docs/packages_old/task.md
@@ -0,0 +1,262 @@
+# Package: task
+
+`internal/task` — Task definition, parsing, validation, and state machine.
+
+---
+
+## 1. Overview
+
+A **Task** is the central unit of work in Claudomator. It describes what an agent should do (`agent.instructions`), how it should be run (timeout, retry, priority), and how it relates to other tasks (`depends_on`, `parent_task_id`). Tasks are defined in YAML files, parsed into `Task` structs, persisted in SQLite, and driven through a state machine from `PENDING` to a terminal state.
+
+---
+
+## 2. Task Struct
+
+```go
+type Task struct {
+    ID               string      // UUID; auto-generated if omitted in YAML
+    ParentTaskID     string      // ID of parent task (subtask linkage); empty for root tasks
+    Name             string      // Human-readable label; required
+    Description      string      // Optional longer description
+    Agent            AgentConfig // How to invoke the agent
+    Timeout          Duration    // Maximum wall-clock run time (e.g. "30m"); 0 = no limit
+    Retry            RetryConfig // Retry policy
+    Priority         Priority    // "high" | "normal" | "low"; default "normal"
+    Tags             []string    // Arbitrary labels for filtering
+    DependsOn        []string    // Task IDs that must reach COMPLETED before this queues
+    State            State       // Current lifecycle state; not read from YAML (yaml:"-")
+    RejectionComment string      // Set by RejectTask; not read from YAML (yaml:"-")
+    QuestionJSON     string      // Pending question from a READY agent; not read from YAML (yaml:"-")
+    CreatedAt        time.Time   // Set on parse; not read from YAML (yaml:"-")
+    UpdatedAt        time.Time   // Updated on every state change; not read from YAML (yaml:"-")
+}
+```
+
+Fields tagged `yaml:"-"` are runtime-only and are never parsed from task YAML files.
+
+---
+
+## 3. AgentConfig Struct
+
+```go
+type AgentConfig struct {
+    Type               string   // Agent implementation: "claude", "gemini", etc.
+    Model              string   // Model identifier passed to the agent binary
+    ContextFiles       []string // Files injected into agent context at start
+    Instructions       string   // Prompt / task description sent to the agent; required
+    ProjectDir         string   // Working directory for the agent process
+    MaxBudgetUSD       float64  // Spending cap in USD; 0 = unlimited; must be >= 0
+    PermissionMode     string   // One of: default | acceptEdits | bypassPermissions | plan | dontAsk | delegate
+    AllowedTools       []string // Whitelist of tool names the agent may use
+    DisallowedTools    []string // Blacklist of tool names the agent may not use
+    SystemPromptAppend string   // Text appended to the agent's system prompt
+    AdditionalArgs     []string // Extra CLI arguments forwarded to the agent binary
+    SkipPlanning       bool     // When true, bypass the planning phase
+}
+```
+
+---
+
+## 4. RetryConfig Struct
+
+```go
+type RetryConfig struct {
+    MaxAttempts int    // Total attempts including the first; default 1 (no retry)
+    Backoff     string // "linear" or "exponential"; default "exponential"
+}
+```
+
+---
+
+## 5. YAML Task File Format — Single Task
+
+```yaml
+# Unique identifier. Optional: auto-generated UUID if omitted.
+id: "fix-login-bug"
+
+# Human-readable label. Required.
+name: "Fix login redirect bug"
+
+# Optional description shown in the UI.
+description: "Users are redirected to /home instead of /dashboard after login."
+
+# Agent configuration.
+agent:
+  # Agent type: "claude", "gemini", etc. Can be omitted for auto-classification.
+  type: "claude"
+
+  # Model to use. Empty = agent default.
+  model: "claude-opus-4-6"
+
+  # Files loaded into the agent's context before execution.
+  context_files:
+    - "src/auth/login.go"
+    - "docs/design/auth.md"
+
+  # Prompt sent to the agent. Required.
+  instructions: |
+    Fix the post-login redirect in src/auth/login.go so that users are
+    sent to /dashboard instead of /home. Add a regression test.
+
+  # Working directory for the agent process. Empty = server working directory.
+  project_dir: "/workspace/myapp"
+
+  # USD spending cap. 0 = no limit.
+  max_budget_usd: 1.00
+
+  # Permission mode. Valid values: default | acceptEdits | bypassPermissions | plan | dontAsk | delegate
+  permission_mode: "acceptEdits"
+
+  # Tool whitelist. Empty = all tools allowed.
+  allowed_tools:
+    - "Edit"
+    - "Read"
+    - "Bash"
+
+  # Tool blacklist.
+  disallowed_tools:
+    - "WebFetch"
+
+  # Appended to the agent's system prompt.
+  system_prompt_append: "Always write tests before implementation."
+
+  # Extra arguments forwarded verbatim to the agent binary.
+  additional_args:
+    - "--verbose"
+
+  # Skip the planning phase.
+  skip_planning: false
+
+# Maximum run time. Accepts Go duration strings: "30m", "1h30m", "45s".
+# 0 or omitted = no limit.
+timeout: "30m"
+
+# Retry policy.
+retry:
+  # Total attempts (including first). Must be >= 1.
+  max_attempts: 3
+  # "linear" or "exponential".
+  backoff: "exponential"
+
+# Scheduling priority: "high", "normal" (default), or "low".
+priority: "normal"
+
+# Arbitrary string labels for filtering.
+tags:
+  - "bug"
+  - "auth"
+
+# Task IDs that must be COMPLETED before this task is queued.
+depends_on:
+  - "setup-test-db"
+```
+
+---
+
+## 6. Batch File Format
+
+A batch file wraps multiple tasks under a `tasks` key. Each entry is a full task definition (same fields as above). All tasks are parsed and initialized together.
+
+```yaml
+tasks:
+  - name: "Step 1 — scaffold"
+    agent:
+      instructions: "Create the initial project structure."
+    priority: "high"
+
+  - name: "Step 2 — implement"
+    agent:
+      instructions: "Implement the feature described in docs/feature.md."
+    depends_on:
+      - "step-1-id"
+
+  - name: "Step 3 — test"
+    agent:
+      instructions: "Write and run integration tests."
+    depends_on:
+      - "step-2-id"
+    retry:
+      max_attempts: 2
+      backoff: "linear"
+```
+
+`ParseFile` tries the batch format first; if no `tasks` key is present it falls back to single-task parsing.
+
+---
+
+## 7. State Constants
+
+| Constant          | Value            | Meaning                                                                 |
+|-------------------|------------------|-------------------------------------------------------------------------|
+| `StatePending`    | `PENDING`        | Newly created; awaiting classification or human approval.               |
+| `StateQueued`     | `QUEUED`         | Accepted and waiting for an available agent slot.                       |
+| `StateRunning`    | `RUNNING`        | Agent process is actively executing.                                    |
+| `StateReady`      | `READY`          | Agent has paused and is awaiting human input (question / approval).     |
+| `StateCompleted`  | `COMPLETED`      | Agent finished successfully. Terminal — no further transitions allowed. |
+| `StateFailed`     | `FAILED`         | Agent exited with a non-zero code or internal error.                    |
+| `StateTimedOut`   | `TIMED_OUT`      | Execution exceeded the configured `timeout`.                            |
+| `StateCancelled`  | `CANCELLED`      | Explicitly cancelled by the user or scheduler.                          |
+| `StateBudgetExceeded` | `BUDGET_EXCEEDED` | Agent hit the `max_budget_usd` cap before finishing.              |
+| `StateBlocked`    | `BLOCKED`        | Waiting on a dependency task that has not yet completed.                |
+
+---
+
+## 8. State Machine — Valid Transitions
+
+| From               | To                 | Condition / trigger                                          |
+|--------------------|--------------------|--------------------------------------------------------------|
+| `PENDING`          | `QUEUED`           | Task approved and eligible for scheduling.                   |
+| `PENDING`          | `CANCELLED`        | Cancelled before being queued.                               |
+| `QUEUED`           | `RUNNING`          | Agent slot becomes available; execution starts.              |
+| `QUEUED`           | `CANCELLED`        | Cancelled while waiting in the queue.                        |
+| `RUNNING`          | `READY`            | Agent pauses and emits a question for human input.           |
+| `RUNNING`          | `COMPLETED`        | Agent exits successfully.                                    |
+| `RUNNING`          | `FAILED`           | Agent exits with an error.                                   |
+| `RUNNING`          | `TIMED_OUT`        | Wall-clock time exceeds `timeout`.                           |
+| `RUNNING`          | `CANCELLED`        | Cancelled mid-execution.                                     |
+| `RUNNING`          | `BUDGET_EXCEEDED`  | Cumulative cost exceeds `max_budget_usd`.                    |
+| `RUNNING`          | `BLOCKED`          | Runtime dependency check fails.                              |
+| `READY`            | `COMPLETED`        | Human answer accepted; task finishes.                        |
+| `READY`            | `PENDING`          | Answer rejected; task returns to pending for re-approval.    |
+| `FAILED`           | `QUEUED`           | Retry requested (re-enqueue).                                |
+| `TIMED_OUT`        | `QUEUED`           | Retry or resume requested.                                   |
+| `CANCELLED`        | `QUEUED`           | Restart requested.                                           |
+| `BUDGET_EXCEEDED`  | `QUEUED`           | Retry with higher or no budget.                              |
+| `BLOCKED`          | `QUEUED`           | Dependency became satisfied; task re-queued.                 |
+| `BLOCKED`          | `READY`            | Dependency resolved but human review required.               |
+| `COMPLETED`        | *(none)*           | Terminal state.                                              |
+
+---
+
+## 9. Key Functions
+
+### `ParseFile(path string) ([]Task, error)`
+Reads a YAML file at `path` and returns one or more tasks. Tries batch format (`tasks:` key) first; falls back to single-task format. Auto-assigns UUIDs, default priority, default retry config, and sets `State = PENDING` on all returned tasks.
+
+### `Parse(data []byte) ([]Task, error)`
+Same as `ParseFile` but operates on raw bytes instead of a file path.
+
+### `ValidTransition(from, to State) bool`
+Returns `true` if the transition from `from` to `to` is permitted by the state machine. Used by `storage.DB.UpdateTaskState` to enforce transitions atomically inside a transaction.
+
+### `Validate(t *Task) error`
+Validates a task's fields. Returns `*ValidationError` (implementing `error`) with all failures collected, or `nil` if valid.
+
+---
+
+## 10. Validation Rules
+
+The `Validate` function enforces the following rules:
+
+| Rule | Details |
+|------|---------|
+| `name` required | Must be non-empty. |
+| `agent.instructions` required | Must be non-empty. |
+| `agent.max_budget_usd` non-negative | Must be `>= 0`. |
+| `timeout` non-negative | Must be `>= 0` (zero means no limit). |
+| `retry.max_attempts >= 1` | Must be at least 1. |
+| `retry.backoff` valid values | Must be empty, `"linear"`, or `"exponential"`. |
+| `priority` valid values | Must be empty, `"high"`, `"normal"`, or `"low"`. |
+| `agent.permission_mode` valid values | Must be empty or one of: `default`, `acceptEdits`, `bypassPermissions`, `plan`, `dontAsk`, `delegate`. |
+
+Multiple failures are collected and returned together in a single `ValidationError`.
diff --git a/internal/api/docs/RAW_NARRATIVE.md b/internal/api/docs/RAW_NARRATIVE.md
new file mode 100644
index 0000000..7944463
--- /dev/null
+++ b/internal/api/docs/RAW_NARRATIVE.md
@@ -0,0 +1,117 @@
+
+--- 2026-03-10T09:33:34Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T09:33:34Z ---
+do something
+
+--- 2026-03-10T09:33:34Z ---
+do something
+
+--- 2026-03-10T16:46:39Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T16:46:39Z ---
+do something
+
+--- 2026-03-10T16:46:39Z ---
+do something
+
+--- 2026-03-10T17:16:31Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T17:16:31Z ---
+do something
+
+--- 2026-03-10T17:16:31Z ---
+do something
+
+--- 2026-03-10T17:25:16Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T17:25:16Z ---
+do something
+
+--- 2026-03-10T17:25:16Z ---
+do something
+
+--- 2026-03-10T23:54:53Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T23:54:53Z ---
+do something
+
+--- 2026-03-10T23:54:53Z ---
+do something
+
+--- 2026-03-10T23:55:54Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T23:55:54Z ---
+do something
+
+--- 2026-03-10T23:55:54Z ---
+do something
+
+--- 2026-03-10T23:56:06Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T23:56:06Z ---
+do something
+
+--- 2026-03-10T23:56:06Z ---
+do something
+
+--- 2026-03-10T23:57:26Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-10T23:57:26Z ---
+do something
+
+--- 2026-03-10T23:57:26Z ---
+do something
+
+--- 2026-03-11T07:40:17Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-11T07:40:17Z ---
+do something
+
+--- 2026-03-11T07:40:17Z ---
+do something
+
+--- 2026-03-11T08:25:03Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-11T08:25:04Z ---
+do something
+
+--- 2026-03-11T08:25:04Z ---
+do something
+
+--- 2026-03-12T21:00:28Z ---
+generate a report
+
+--- 2026-03-12T21:00:33Z ---
+generate a report
+
+--- 2026-03-12T21:00:34Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-12T21:00:34Z ---
+do something
+
+--- 2026-03-12T21:00:34Z ---
+do something
+
+--- 2026-03-13T02:27:38Z ---
+generate a report
+
+--- 2026-03-13T02:27:38Z ---
+run the Go test suite with race detector and fail if coverage < 80%
+
+--- 2026-03-13T02:27:38Z ---
+do something
+
+--- 2026-03-13T02:27:38Z ---
+do something
diff --git a/internal/executor/claude.go b/internal/executor/claude.go
index 0e29f7f..a58f1ad 100644
--- a/internal/executor/claude.go
+++ b/internal/executor/claude.go
@@ -150,6 +150,13 @@ func (r *ClaudeRunner) Run(ctx context.Context, t *task.Task, e *storage.Executi
 		return &BlockedError{QuestionJSON: strings.TrimSpace(string(data)), SessionID: e.SessionID, SandboxDir: sandboxDir}
 	}
 
+	// Read agent summary if written.
+	summaryFile := filepath.Join(logDir, "summary.txt")
+	if summaryData, readErr := os.ReadFile(summaryFile); readErr == nil {
+		os.Remove(summaryFile) // consumed
+		e.Summary = strings.TrimSpace(string(summaryData))
+	}
+
 	// Merge sandbox back to project_dir and clean up.
 	if sandboxDir != "" {
 		if mergeErr := teardownSandbox(projectDir, sandboxDir, r.Logger); mergeErr != nil {
@@ -261,6 +268,7 @@ func (r *ClaudeRunner) execOnce(ctx context.Context, args []string, workingDir s
 		"CLAUDOMATOR_API_URL="+r.APIURL,
 		"CLAUDOMATOR_TASK_ID="+e.TaskID,
 		"CLAUDOMATOR_QUESTION_FILE="+filepath.Join(e.ArtifactDir, "question.json"),
+		"CLAUDOMATOR_SUMMARY_FILE="+filepath.Join(e.ArtifactDir, "summary.txt"),
 	)
 	// Put the subprocess in its own process group so we can SIGKILL the entire
 	// group (MCP servers, bash children, etc.) on cancellation.
diff --git a/internal/executor/preamble.go b/internal/executor/preamble.go
index e50c16f..bc5c32c 100644
--- a/internal/executor/preamble.go
+++ b/internal/executor/preamble.go
@@ -46,6 +46,17 @@ The sandbox is rejected if there are any uncommitted modifications.
 
 ---
 
+## Final Summary (mandatory)
+
+Before exiting, write a brief summary paragraph (2–5 sentences) describing what you did
+and the outcome. Write it to the path in $CLAUDOMATOR_SUMMARY_FILE:
+
+  echo "Your summary here." > "$CLAUDOMATOR_SUMMARY_FILE"
+
+This summary is displayed in the task UI so the user knows what happened.
+
+---
+
 `
 
 func withPlanningPreamble(instructions string) string {
diff --git a/internal/executor/summary.go b/internal/executor/summary.go
new file mode 100644
index 0000000..a942de0
--- /dev/null
+++ b/internal/executor/summary.go
@@ -0,0 +1,57 @@
+package executor
+
+import (
+	"bufio"
+	"encoding/json"
+	"os"
+	"strings"
+)
+
+// extractSummary reads a stream-json stdout log and returns the text following
+// the last "## Summary" heading found in any assistant text block.
+// Returns empty string if the file cannot be read or no summary is found.
+func extractSummary(stdoutPath string) string {
+	f, err := os.Open(stdoutPath)
+	if err != nil {
+		return ""
+	}
+	defer f.Close()
+
+	var last string
+	scanner := bufio.NewScanner(f)
+	scanner.Buffer(make([]byte, 1024*1024), 1024*1024)
+	for scanner.Scan() {
+		if text := summaryFromLine(scanner.Bytes()); text != "" {
+			last = text
+		}
+	}
+	return last
+}
+
+// summaryFromLine parses a single stream-json line and returns the text after
+// "## Summary" if the line is an assistant text block containing that heading.
+func summaryFromLine(line []byte) string {
+	var event struct {
+		Type    string `json:"type"`
+		Message struct {
+			Content []struct {
+				Type string `json:"type"`
+				Text string `json:"text"`
+			} `json:"content"`
+		} `json:"message"`
+	}
+	if err := json.Unmarshal(line, &event); err != nil || event.Type != "assistant" {
+		return ""
+	}
+	for _, block := range event.Message.Content {
+		if block.Type != "text" {
+			continue
+		}
+		idx := strings.Index(block.Text, "## Summary")
+		if idx == -1 {
+			continue
+		}
+		return strings.TrimSpace(block.Text[idx+len("## Summary"):])
+	}
+	return ""
+}
diff --git a/internal/executor/summary_test.go b/internal/executor/summary_test.go
new file mode 100644
index 0000000..4a73711
--- /dev/null
+++ b/internal/executor/summary_test.go
@@ -0,0 +1,49 @@
+package executor
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestExtractSummary_WithSummarySection(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "stdout.log")
+	content := streamLine(`{"type":"assistant","message":{"content":[{"type":"text","text":"## Summary\nThe task was completed successfully."}]}}`)
+	if err := os.WriteFile(path, []byte(content), 0600); err != nil {
+		t.Fatal(err)
+	}
+	got := extractSummary(path)
+	want := "The task was completed successfully."
+	if got != want {
+		t.Errorf("got %q, want %q", got, want)
+	}
+}
+
+func TestExtractSummary_NoSummary(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "stdout.log")
+	content := streamLine(`{"type":"assistant","message":{"content":[{"type":"text","text":"All done, no summary heading."}]}}`)
+	if err := os.WriteFile(path, []byte(content), 0600); err != nil {
+		t.Fatal(err)
+	}
+	got := extractSummary(path)
+	if got != "" {
+		t.Errorf("expected empty string, got %q", got)
+	}
+}
+
+func TestExtractSummary_MultipleSections_PicksLast(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "stdout.log")
+	content := streamLine(`{"type":"assistant","message":{"content":[{"type":"text","text":"## Summary\nFirst summary."}]}}`) +
+		streamLine(`{"type":"assistant","message":{"content":[{"type":"text","text":"## Summary\nFinal summary."}]}}`)
+	if err := os.WriteFile(path, []byte(content), 0600); err != nil {
+		t.Fatal(err)
+	}
+	got := extractSummary(path)
+	want := "Final summary."
+	if got != want {
+		t.Errorf("got %q, want %q", got, want)
+	}
+}
diff --git a/internal/storage/db.go b/internal/storage/db.go
index aaf1e09..b8a7085 100644
--- a/internal/storage/db.go
+++ b/internal/storage/db.go
@@ -81,6 +81,8 @@ func (s *DB) migrate() error {
 		`ALTER TABLE tasks ADD COLUMN question_json TEXT`,
 		`ALTER TABLE executions ADD COLUMN session_id TEXT`,
 		`ALTER TABLE executions ADD COLUMN sandbox_dir TEXT`,
+		`ALTER TABLE tasks ADD COLUMN summary TEXT`,
+		`ALTER TABLE tasks ADD COLUMN interactions_json TEXT NOT NULL DEFAULT '[]'`,
 	}
 	for _, m := range migrations {
 		if _, err := s.db.Exec(m); err != nil {
@@ -129,13 +131,13 @@ func (s *DB) CreateTask(t *task.Task) error {
 
 // GetTask retrieves a task by ID.
 func (s *DB) GetTask(id string) (*task.Task, error) {
-	row := s.db.QueryRow(`SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json FROM tasks WHERE id = ?`, id)
+	row := s.db.QueryRow(`SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json, summary, interactions_json FROM tasks WHERE id = ?`, id)
 	return scanTask(row)
 }
 
 // ListTasks returns tasks matching the given filter.
 func (s *DB) ListTasks(filter TaskFilter) ([]*task.Task, error) {
-	query := `SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json FROM tasks WHERE 1=1`
+	query := `SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json, summary, interactions_json FROM tasks WHERE 1=1`
 	var args []interface{}
 
 	if filter.State != "" {
@@ -167,7 +169,7 @@ func (s *DB) ListTasks(filter TaskFilter) ([]*task.Task, error) {
 
 // ListSubtasks returns all tasks whose parent_task_id matches the given ID.
 func (s *DB) ListSubtasks(parentID string) ([]*task.Task, error) {
-	rows, err := s.db.Query(`SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json FROM tasks WHERE parent_task_id = ? ORDER BY created_at ASC`, parentID)
+	rows, err := s.db.Query(`SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json, summary, interactions_json FROM tasks WHERE parent_task_id = ? ORDER BY created_at ASC`, parentID)
 	if err != nil {
 		return nil, err
 	}
@@ -220,7 +222,7 @@ func (s *DB) ResetTaskForRetry(id string) (*task.Task, error) {
 	}
 	defer tx.Rollback() //nolint:errcheck
 
-	t, err := scanTask(tx.QueryRow(`SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json FROM tasks WHERE id = ?`, id))
+	t, err := scanTask(tx.QueryRow(`SELECT id, name, description, config_json, priority, timeout_ns, retry_json, tags_json, depends_on_json, parent_task_id, state, created_at, updated_at, rejection_comment, question_json, summary, interactions_json FROM tasks WHERE id = ?`, id))
 	if err != nil {
 		if err == sql.ErrNoRows {
 			return nil, fmt.Errorf("task %q not found", id)
@@ -355,6 +357,8 @@ type Execution struct {
 	// In-memory only: set when creating a resume execution, not stored in DB.
 	ResumeSessionID string
 	ResumeAnswer    string
+	// In-memory only: populated by the runner after successful execution.
+	Summary string
 }
 
 // CreateExecution inserts an execution record.
@@ -516,6 +520,48 @@ func (s *DB) UpdateTaskQuestion(taskID, questionJSON string) error {
 	return err
 }
 
+// UpdateTaskSummary stores the agent's final summary paragraph on a task.
+func (s *DB) UpdateTaskSummary(taskID, summary string) error {
+	_, err := s.db.Exec(`UPDATE tasks SET summary = ?, updated_at = ? WHERE id = ?`,
+		summary, time.Now().UTC(), taskID)
+	return err
+}
+
+// AppendTaskInteraction appends a Q&A interaction to the task's interaction history.
+func (s *DB) AppendTaskInteraction(taskID string, interaction task.Interaction) error {
+	tx, err := s.db.Begin()
+	if err != nil {
+		return err
+	}
+	defer tx.Rollback() //nolint:errcheck
+
+	var raw sql.NullString
+	if err := tx.QueryRow(`SELECT interactions_json FROM tasks WHERE id = ?`, taskID).Scan(&raw); err != nil {
+		if err == sql.ErrNoRows {
+			return fmt.Errorf("task %q not found", taskID)
+		}
+		return err
+	}
+	existing := raw.String
+	if existing == "" {
+		existing = "[]"
+	}
+	var interactions []task.Interaction
+	if err := json.Unmarshal([]byte(existing), &interactions); err != nil {
+		return fmt.Errorf("unmarshaling interactions: %w", err)
+	}
+	interactions = append(interactions, interaction)
+	updated, err := json.Marshal(interactions)
+	if err != nil {
+		return fmt.Errorf("marshaling interactions: %w", err)
+	}
+	if _, err := tx.Exec(`UPDATE tasks SET interactions_json = ?, updated_at = ? WHERE id = ?`,
+		string(updated), time.Now().UTC(), taskID); err != nil {
+		return err
+	}
+	return tx.Commit()
+}
+
 // UpdateExecution updates a completed execution.
 func (s *DB) UpdateExecution(e *Execution) error {
 	_, err := s.db.Exec(`
@@ -545,11 +591,14 @@ func scanTask(row scanner) (*task.Task, error) {
 		parentTaskID     sql.NullString
 		rejectionComment sql.NullString
 		questionJSON     sql.NullString
+		summary          sql.NullString
+		interactionsJSON sql.NullString
 	)
-	err := row.Scan(&t.ID, &t.Name, &t.Description, &configJSON, &priority, &timeoutNS, &retryJSON, &tagsJSON, &depsJSON, &parentTaskID, &state, &t.CreatedAt, &t.UpdatedAt, &rejectionComment, &questionJSON)
+	err := row.Scan(&t.ID, &t.Name, &t.Description, &configJSON, &priority, &timeoutNS, &retryJSON, &tagsJSON, &depsJSON, &parentTaskID, &state, &t.CreatedAt, &t.UpdatedAt, &rejectionComment, &questionJSON, &summary, &interactionsJSON)
 	t.ParentTaskID = parentTaskID.String
 	t.RejectionComment = rejectionComment.String
 	t.QuestionJSON = questionJSON.String
+	t.Summary = summary.String
 	if err != nil {
 		return nil, err
 	}
@@ -568,6 +617,13 @@ func scanTask(row scanner) (*task.Task, error) {
 	if err := json.Unmarshal([]byte(depsJSON), &t.DependsOn); err != nil {
 		return nil, fmt.Errorf("unmarshaling depends_on: %w", err)
 	}
+	raw := interactionsJSON.String
+	if raw == "" {
+		raw = "[]"
+	}
+	if err := json.Unmarshal([]byte(raw), &t.Interactions); err != nil {
+		return nil, fmt.Errorf("unmarshaling interactions: %w", err)
+	}
 	return &t, nil
 }
 
diff --git a/internal/task/task.go b/internal/task/task.go
index 9968b15..2c57922 100644
--- a/internal/task/task.go
+++ b/internal/task/task.go
@@ -48,6 +48,14 @@ type RetryConfig struct {
 	Backoff     string `yaml:"backoff"      json:"backoff"` // "linear", "exponential"
 }
 
+// Interaction records a single question/answer exchange between an agent and the user.
+type Interaction struct {
+	QuestionText string    `json:"question_text"`
+	Options      []string  `json:"options,omitempty"`
+	Answer       string    `json:"answer,omitempty"`
+	AskedAt      time.Time `json:"asked_at"`
+}
+
 type Task struct {
 	ID           string       `yaml:"id"            json:"id"`
 	ParentTaskID string       `yaml:"parent_task_id" json:"parent_task_id"`
@@ -59,11 +67,13 @@ type Task struct {
 	Priority    Priority     `yaml:"priority"   json:"priority"`
 	Tags        []string     `yaml:"tags"       json:"tags"`
 	DependsOn   []string     `yaml:"depends_on" json:"depends_on"`
-	State             State        `yaml:"-"          json:"state"`
-	RejectionComment  string       `yaml:"-"          json:"rejection_comment,omitempty"`
-	QuestionJSON      string       `yaml:"-"          json:"question,omitempty"`
-	CreatedAt         time.Time    `yaml:"-"          json:"created_at"`
-	UpdatedAt         time.Time    `yaml:"-"          json:"updated_at"`
+	State             State         `yaml:"-"          json:"state"`
+	RejectionComment  string        `yaml:"-"          json:"rejection_comment,omitempty"`
+	QuestionJSON      string        `yaml:"-"          json:"question,omitempty"`
+	Summary           string        `yaml:"-"          json:"summary,omitempty"`
+	Interactions      []Interaction `yaml:"-"          json:"interactions,omitempty"`
+	CreatedAt         time.Time     `yaml:"-"          json:"created_at"`
+	UpdatedAt         time.Time     `yaml:"-"          json:"updated_at"`
 }
 
 // Duration wraps time.Duration for YAML unmarshaling from strings like "30m".
@@ -94,27 +104,24 @@ type BatchFile struct {
 }
 
 // validTransitions maps each state to the set of states it may transition into.
-// Terminal state COMPLETED has no outgoing edges.
+// COMPLETED is the only true terminal state (no outgoing edges).
 // CANCELLED, FAILED, TIMED_OUT, and BUDGET_EXCEEDED all allow re-entry at QUEUED
 // (restart or retry).
-var validTransitions = map[State][]State{
-	StatePending:        {StateQueued, StateCancelled},
-	StateQueued:         {StateRunning, StateCancelled},
-	StateRunning:        {StateReady, StateCompleted, StateFailed, StateTimedOut, StateCancelled, StateBudgetExceeded, StateBlocked},
-	StateReady:          {StateCompleted, StatePending},
-	StateFailed:         {StateQueued}, // retry
-	StateTimedOut:       {StateQueued}, // retry or resume
-	StateCancelled:      {StateQueued}, // restart
-	StateBudgetExceeded: {StateQueued}, // retry
-	StateBlocked:        {StateQueued, StateReady},
+// READY may go back to PENDING on user rejection.
+// BLOCKED may advance to READY when all subtasks complete, or back to QUEUED on user answer.
+var validTransitions = map[State]map[State]bool{
+	StatePending:        {StateQueued: true, StateCancelled: true},
+	StateQueued:         {StateRunning: true, StateCancelled: true},
+	StateRunning:        {StateReady: true, StateCompleted: true, StateFailed: true, StateTimedOut: true, StateCancelled: true, StateBudgetExceeded: true, StateBlocked: true},
+	StateReady:          {StateCompleted: true, StatePending: true},
+	StateFailed:         {StateQueued: true}, // retry
+	StateTimedOut:       {StateQueued: true}, // retry or resume
+	StateCancelled:      {StateQueued: true}, // restart
+	StateBudgetExceeded: {StateQueued: true}, // retry
+	StateBlocked:        {StateQueued: true, StateReady: true},
 }
 
 // ValidTransition returns true if moving from the current state to next is allowed.
 func ValidTransition(from, to State) bool {
-	for _, allowed := range validTransitions[from] {
-		if allowed == to {
-			return true
-		}
-	}
-	return false
+	return validTransitions[from][to]
 }
diff --git a/web/style.css b/web/style.css
index 2b872fe..342e0b3 100644
--- a/web/style.css
+++ b/web/style.css
@@ -1176,3 +1176,41 @@ dialog label select:focus {
   margin-top: 0.25rem;
   text-align: right;
 }
+
+/* ── Task Summary + Q&A History ────────────────────────────── */
+
+.task-summary {
+  color: var(--text);
+  line-height: 1.6;
+  margin: 0;
+  white-space: pre-wrap;
+}
+
+.qa-list {
+  display: flex;
+  flex-direction: column;
+  gap: 0.75rem;
+}
+
+.qa-item {
+  border-left: 3px solid var(--border);
+  padding: 0.5rem 0.75rem;
+  display: flex;
+  flex-direction: column;
+  gap: 0.25rem;
+}
+
+.qa-question {
+  font-weight: 500;
+  color: var(--text);
+}
+
+.qa-options {
+  font-size: 0.82rem;
+  color: var(--text-muted, #94a3b8);
+}
+
+.qa-answer {
+  color: var(--accent, #60a5fa);
+  font-style: italic;
+}