feat: implement plan approval gate before execution

Adds plan approval workflow per GitHub issue #111:
- Register plan_approval tool in tools index
- Add plan approval handling in message handler
  - Check for pending plans on each message
  - Detect approval/rejection messages
  - Handle plan approval flow
- Create plan approval modal TUI component
- Update system prompts to instruct agents to use plan_approval tool
  - Balanced tier: Updated to require plan_approval for complex tasks
  - Thorough tier: Updated with plan_approval workflow

Workflow:
1. Agent analyzes task complexity with plan_approval
2. Agent creates plan with steps, context, risks
3. Agent submits plan for user approval
4. User approves/rejects/provides feedback
5. Only after approval, agent proceeds with execution

Closes #111
This commit is contained in:
2026-02-05 18:51:49 -05:00
parent 3d2195f074
commit f8cde24d87
5 changed files with 405 additions and 71 deletions

View File

@@ -91,47 +91,55 @@ For multi-step tasks (3+ steps), use todowrite to track progress:
export const BALANCED_TIER_PLAN_GATE = `## Plan Mode (Balanced Tier)
For COMPLEX tasks that meet these criteria, enter plan mode:
- Multi-file refactoring
For COMPLEX tasks that meet these criteria, use the plan_approval tool:
- Multi-file refactoring (3+ files)
- New feature implementation
- Architectural changes
- Security-related changes
- Database modifications
- Tasks that could take 10+ tool calls
### Plan Mode Workflow
### CRITICAL: Plan Approval Workflow
1. **Explore Phase** - Gather context with glob, grep, read
2. **Plan Phase** - Create structured plan with steps
3. **Approval Gate** - Present plan, wait for user signal to proceed
4. **Execute Phase** - Implement the plan
5. **Verify Phase** - Test and confirm
You MUST use the plan_approval tool for complex tasks. DO NOT execute file modifications until the user approves the plan.
### Plan Format
1. **Analyze Task** - Use plan_approval with action="analyze_task" to check if plan approval is needed
2. **Create Plan** - Use plan_approval with action="create" to start a plan
3. **Add Context** - Use plan_approval with action="add_context" to document files analyzed
4. **Add Steps** - Use plan_approval with action="add_step" for each implementation step
5. **Add Risks** - Use plan_approval with action="add_risk" for identified risks
6. **Submit Plan** - Use plan_approval with action="submit" to present plan to user
7. **Wait for Approval** - DO NOT PROCEED until user says "yes", "proceed", "approve", or similar
8. **Execute** - Only after approval, implement the plan step by step
### Example Plan Approval Flow
\`\`\`
## Plan: [Task Name]
// Step 1: Analyze the task
plan_approval action="analyze_task" task_description="Implement user authentication system"
### Context
- [Key file 1]: [purpose]
- [Key file 2]: [purpose]
// Step 2: Create plan
plan_approval action="create" title="User Authentication System" summary="Add JWT-based auth with login, register, and protected routes"
### Steps
1. [First change] - [file affected]
2. [Second change] - [file affected]
3. [Verification] - [command]
// Step 3: Add context
plan_approval action="add_context" plan_id="<id>" files_analyzed=["src/routes/", "src/middleware/"]
### Risks
- [Potential issue and mitigation]
// Step 4: Add steps
plan_approval action="add_step" plan_id="<id>" step_title="Create auth middleware" step_description="Add JWT verification middleware" files_affected=["src/middleware/auth.ts"] risk_level="medium"
Ready to proceed? (Continue with implementation or provide feedback)
// Step 5: Submit for approval
plan_approval action="submit" plan_id="<id>" testing_strategy="Run auth tests" rollback_plan="Revert commits"
// STOP HERE - Wait for user to approve
\`\`\`
### When to Skip Plan Mode
Skip plan mode for:
Skip plan approval for:
- Single file changes
- Simple bug fixes
- Adding a function or component
- Configuration changes
- Simple bug fixes with obvious solutions
- Adding comments or documentation
- Formatting changes
- Tasks with clear, limited scope`;
export const BALANCED_TIER_AGENTS = `## Agent Delegation

View File

@@ -110,67 +110,64 @@ Only consult when:
- Security/compliance decisions
- Changes that affect external systems or users`;
export const THOROUGH_TIER_PLAN_MODE = `## Advanced Plan Mode
export const THOROUGH_TIER_PLAN_MODE = `## Advanced Plan Mode with Approval Gate
For complex tasks, use structured planning with approval:
For complex tasks, you MUST use the plan_approval tool before executing file modifications.
### Plan Document Format
### CRITICAL: Use plan_approval Tool
\`\`\`markdown
# Implementation Plan: [Feature Name]
Even with full autonomy, you MUST:
1. Use plan_approval action="analyze_task" to check if approval is needed
2. If needed, create and submit a plan using plan_approval
3. WAIT for user approval before executing any file modifications
4. Only proceed after user says "yes", "proceed", "approve", or similar
## Executive Summary
[1-2 sentence overview]
### Plan Approval Tool Workflow
## Context Analysis
### Files Analyzed
- \`path/to/file.ts\`: [purpose and relevance]
\`\`\`
// 1. Analyze if plan approval is needed
plan_approval action="analyze_task" task_description="..."
### Current Architecture
[Brief description of existing patterns]
// 2. Create the plan
plan_approval action="create" title="Feature Name" summary="What we'll do"
### Dependencies
- [External dependency 1]
- [Internal module 1]
// 3. Add context (files analyzed, architecture)
plan_approval action="add_context" plan_id="<id>" files_analyzed=[...] current_architecture="..."
## Implementation Strategy
// 4. Add each step
plan_approval action="add_step" plan_id="<id>" step_title="Phase 1" step_description="..." files_affected=[...] risk_level="medium"
### Phase 1: [Name]
**Objective**: [Goal]
**Files affected**: [list]
**Changes**:
1. [Specific change]
2. [Specific change]
// 5. Add risks
plan_approval action="add_risk" plan_id="<id>" risk_description="..." risk_impact="high" risk_mitigation="..."
### Phase 2: [Name]
...
// 6. Submit for approval
plan_approval action="submit" plan_id="<id>" testing_strategy="..." rollback_plan="..."
## Risk Assessment
| Risk | Impact | Mitigation |
|------|--------|------------|
| [Risk 1] | [High/Med/Low] | [Strategy] |
## Testing Strategy
- Unit tests: [approach]
- Integration tests: [approach]
- Manual verification: [steps]
## Rollback Plan
[How to undo if needed]
---
**Awaiting approval to proceed with implementation.**
// STOP - Wait for user approval before any file modifications!
\`\`\`
### Plan Approval Workflow
### After Approval
1. Generate comprehensive plan
2. Present to user with "Awaiting approval"
3. User can:
- Approve: "proceed", "go ahead", "looks good"
- Modify: "change X to Y"
- Reject: "stop", "don't do this"
4. On approval, execute with full autonomy`;
Once user approves:
- Execute with full autonomy
- Use agent orchestration for parallel implementation
- Complete the entire feature end-to-end
- Handle errors and edge cases
### When Plan Approval is Required
ALWAYS use plan_approval for:
- Multi-file refactoring
- New feature implementation
- Architectural changes
- Security-related changes
- Database modifications
- Any task affecting 3+ files
You MAY skip for:
- Single file fixes with obvious solutions
- Adding comments or documentation
- Simple configuration changes`;
export const THOROUGH_TIER_AGENTS = `## Agent Orchestration System

View File

@@ -69,6 +69,15 @@ import {
executeDetectedCommand,
} from "@services/command-detection";
import { detectSkillCommand, executeSkill } from "@services/skill-service";
import {
getActivePlans,
isApprovalMessage,
isRejectionMessage,
approvePlan,
rejectPlan,
startPlanExecution,
formatPlanForDisplay,
} from "@services/plan-mode/plan-service";
// Track last response for feedback learning
let lastResponseContext: {
@@ -408,6 +417,56 @@ export const handleMessage = async (
// Check for feedback on previous response
await checkUserFeedback(message, callbacks);
// Check for pending plan approvals
const pendingPlans = getActivePlans().filter((p) => p.status === "pending");
if (pendingPlans.length > 0) {
const plan = pendingPlans[0];
// Check if this is an approval message
if (isApprovalMessage(message)) {
approvePlan(plan.id, message);
startPlanExecution(plan.id);
callbacks.onLog("system", `Plan "${plan.title}" approved. Proceeding with implementation.`);
addDebugLog("state", `Plan ${plan.id} approved by user`);
// Continue with agent execution - the agent will see the approved status
// and proceed with implementation
state.messages.push({
role: "user",
content: `The user approved the plan. Proceed with the implementation of plan "${plan.title}".`,
});
// Fall through to normal agent processing
} else if (isRejectionMessage(message)) {
rejectPlan(plan.id, message);
callbacks.onLog("system", `Plan "${plan.title}" rejected. Please provide feedback or a new approach.`);
addDebugLog("state", `Plan ${plan.id} rejected by user`);
// Add rejection to messages so agent can respond
state.messages.push({
role: "user",
content: `The user rejected the plan with feedback: "${message}". Please revise the approach.`,
});
// Fall through to normal agent processing to get revised plan
} else {
// Neither approval nor rejection - treat as feedback/modification request
callbacks.onLog("system", `Plan "${plan.title}" awaiting approval. Reply 'yes' to approve or 'no' to reject.`);
// Show the plan again with the feedback
const planDisplay = formatPlanForDisplay(plan);
appStore.addLog({
type: "system",
content: planDisplay,
});
// Add user's feedback to messages
state.messages.push({
role: "user",
content: `The user provided feedback on the pending plan: "${message}". Please address this feedback and update the plan.`,
});
// Fall through to normal agent processing
}
}
// Check for skill commands (e.g., /review, /commit)
const skillMatch = await detectSkillCommand(message);
if (skillMatch) {

View File

@@ -14,6 +14,7 @@ import { multiEditTool } from "@tools/multi-edit/execute";
import { lspTool } from "@tools/lsp";
import { applyPatchTool } from "@tools/apply-patch";
import { taskAgentTool } from "@tools/task-agent/execute";
import { planApprovalTool } from "@tools/plan-approval/execute";
import {
isMCPTool,
executeMCPTool,
@@ -42,6 +43,7 @@ export const tools: ToolDefinition[] = [
lspTool,
applyPatchTool,
taskAgentTool,
planApprovalTool,
];
// Tools that are read-only (allowed in chat mode)

View File

@@ -0,0 +1,268 @@
import { createSignal, createMemo, For, Show } from "solid-js";
import { useKeyboard } from "@opentui/solid";
import { TextAttributes } from "@opentui/core";
import { useTheme } from "@tui-solid/context/theme";
import type { ImplementationPlan, PlanStep } from "@/types/plan-mode";
/**
* Plan approval request for the TUI
*/
export interface PlanApprovalRequest {
plan: ImplementationPlan;
resolve?: (response: { approved: boolean; message?: string }) => void;
}
interface PlanApprovalModalProps {
request: PlanApprovalRequest;
onRespond?: (approved: boolean, message?: string) => void;
isActive?: boolean;
}
interface ApprovalOption {
key: string;
label: string;
approved: boolean;
message?: string;
}
const APPROVAL_OPTIONS: ApprovalOption[] = [
{ key: "y", label: "Approve - proceed with implementation", approved: true },
{ key: "e", label: "Edit - modify plan before execution", approved: false, message: "edit" },
{ key: "n", label: "Reject - cancel this plan", approved: false },
];
export function PlanApprovalModal(props: PlanApprovalModalProps) {
const theme = useTheme();
const [selectedIndex, setSelectedIndex] = createSignal(0);
const [scrollOffset, setScrollOffset] = createSignal(0);
const isActive = () => props.isActive ?? true;
const MAX_VISIBLE_STEPS = 5;
const plan = () => props.request.plan;
const visibleSteps = createMemo(() => {
const steps = plan().steps;
const offset = scrollOffset();
return steps.slice(offset, offset + MAX_VISIBLE_STEPS);
});
const canScrollUp = () => scrollOffset() > 0;
const canScrollDown = () => scrollOffset() + MAX_VISIBLE_STEPS < plan().steps.length;
const handleResponse = (approved: boolean, message?: string): void => {
if (props.request.resolve) {
props.request.resolve({ approved, message });
}
props.onRespond?.(approved, message);
};
const getRiskIcon = (risk: PlanStep["riskLevel"]): string => {
const icons: Record<PlanStep["riskLevel"], string> = {
high: "\u26A0\uFE0F",
medium: "\u26A1",
low: "\u2713",
};
return icons[risk];
};
const getRiskColor = (risk: PlanStep["riskLevel"]): string => {
const colors: Record<PlanStep["riskLevel"], string> = {
high: theme.colors.error,
medium: theme.colors.warning,
low: theme.colors.success,
};
return colors[risk];
};
useKeyboard((evt) => {
if (!isActive()) return;
evt.stopPropagation();
if (evt.name === "up") {
if (evt.shift) {
// Scroll plan view
if (canScrollUp()) {
setScrollOffset((prev) => prev - 1);
}
} else {
// Navigate options
setSelectedIndex((prev) =>
prev > 0 ? prev - 1 : APPROVAL_OPTIONS.length - 1,
);
}
evt.preventDefault();
return;
}
if (evt.name === "down") {
if (evt.shift) {
// Scroll plan view
if (canScrollDown()) {
setScrollOffset((prev) => prev + 1);
}
} else {
// Navigate options
setSelectedIndex((prev) =>
prev < APPROVAL_OPTIONS.length - 1 ? prev + 1 : 0,
);
}
evt.preventDefault();
return;
}
if (evt.name === "return") {
const option = APPROVAL_OPTIONS[selectedIndex()];
handleResponse(option.approved, option.message);
evt.preventDefault();
return;
}
if (evt.name === "escape") {
handleResponse(false, "cancelled");
evt.preventDefault();
return;
}
// Handle shortcut keys
if (evt.name.length === 1 && !evt.ctrl && !evt.meta) {
const charLower = evt.name.toLowerCase();
const option = APPROVAL_OPTIONS.find((o) => o.key === charLower);
if (option) {
handleResponse(option.approved, option.message);
evt.preventDefault();
}
}
});
return (
<box
flexDirection="column"
borderColor={theme.colors.warning}
border={["top", "bottom", "left", "right"]}
backgroundColor={theme.colors.background}
paddingLeft={2}
paddingRight={2}
paddingTop={1}
paddingBottom={1}
maxHeight={25}
>
{/* Header */}
<box marginBottom={1}>
<text fg={theme.colors.warning} attributes={TextAttributes.BOLD}>
Plan Approval Required
</text>
</box>
{/* Title and Summary */}
<box flexDirection="column" marginBottom={1}>
<text fg={theme.colors.primary} attributes={TextAttributes.BOLD}>
{plan().title}
</text>
<text fg={theme.colors.textDim}>{plan().summary}</text>
</box>
{/* Estimated Changes */}
<box flexDirection="row" marginBottom={1}>
<text fg={theme.colors.success}>
+{plan().estimatedChanges.filesCreated} create{" "}
</text>
<text fg={theme.colors.warning}>
~{plan().estimatedChanges.filesModified} modify{" "}
</text>
<text fg={theme.colors.error}>
-{plan().estimatedChanges.filesDeleted} delete
</text>
</box>
{/* Steps */}
<box flexDirection="column" marginBottom={1}>
<text fg={theme.colors.text} attributes={TextAttributes.BOLD}>
Implementation Steps ({plan().steps.length} total):
</text>
<Show when={canScrollUp()}>
<text fg={theme.colors.textDim}> {"\u25B2"} more above...</text>
</Show>
<For each={visibleSteps()}>
{(step, index) => (
<box flexDirection="row">
<text fg={getRiskColor(step.riskLevel)}>
{getRiskIcon(step.riskLevel)}{" "}
</text>
<text fg={theme.colors.textDim}>
{scrollOffset() + index() + 1}.{" "}
</text>
<text fg={theme.colors.text}>
{step.title.substring(0, 50)}
{step.title.length > 50 ? "..." : ""}
</text>
</box>
)}
</For>
<Show when={canScrollDown()}>
<text fg={theme.colors.textDim}> {"\u25BC"} more below...</text>
</Show>
</box>
{/* Risks */}
<Show when={plan().risks.length > 0}>
<box flexDirection="column" marginBottom={1}>
<text fg={theme.colors.error} attributes={TextAttributes.BOLD}>
Risks ({plan().risks.length}):
</text>
<For each={plan().risks.slice(0, 2)}>
{(risk) => (
<box flexDirection="row">
<text fg={theme.colors.error}> - </text>
<text fg={theme.colors.text}>
{risk.description.substring(0, 60)}
</text>
</box>
)}
</For>
<Show when={plan().risks.length > 2}>
<text fg={theme.colors.textDim}>
{" "}
... and {plan().risks.length - 2} more
</text>
</Show>
</box>
</Show>
{/* Options */}
<box flexDirection="column" marginTop={1}>
<For each={APPROVAL_OPTIONS}>
{(option, index) => {
const isSelected = () => index() === selectedIndex();
const keyColor = () =>
option.approved ? theme.colors.success : theme.colors.error;
return (
<box flexDirection="row">
<text
fg={isSelected() ? theme.colors.primary : undefined}
attributes={
isSelected() ? TextAttributes.BOLD : TextAttributes.NONE
}
>
{isSelected() ? "> " : " "}
</text>
<text fg={keyColor()}>[{option.key}] </text>
<text fg={isSelected() ? theme.colors.primary : undefined}>
{option.label}
</text>
</box>
);
}}
</For>
</box>
{/* Footer */}
<box marginTop={1}>
<text fg={theme.colors.textDim}>
{"\u2191\u2193"} options | Shift+{"\u2191\u2193"} scroll steps | Enter select | y/e/n shortcut
</text>
</box>
</box>
);
}