cflx-proposal

name: cflx-proposal description: Create structured Conflux change proposals through interactive conversation with users. Use when users request "create a proposal", "draft a change", "propose a feature", or similar proposal creation tasks. This skill asks clarifying questions and guides users through the proposal process.

Conflux Proposal Creator

Create structured change proposals for Conflux (OpenSpec-based) projects through interactive conversation with users.

Scope Restrictions (Proposal-Only)

• This skill is for proposal creation only. Do NOT implement or modify product/source code.
• You may READ any files for context gathering.
• You may WRITE only under openspec/changes/<change-id>/.
• After strict validation passes, stop and present the proposal for review.

Guardrails (Match Command Behavior)

• If openspec/CONSTITUTION.md exists, read it before drafting the proposal and treat it as higher-priority project law than proposal/spec deltas.
• Do not draft a proposal that violates openspec/CONSTITUTION.md unless the proposal explicitly changes that constitution first.
• Favor straightforward, minimal implementations first and add complexity only when it is requested or clearly required.
• Keep changes tightly scoped to the requested outcome.
• Default to proposal splitting: when requirements can be decomposed into independent scopes, create separate change proposals.
• If uncertain whether to split, prefer splitting unless the scopes are tightly coupled and must ship together to preserve correctness.
• For each split proposal, use a distinct verb-led change-id and keep proposal.md, tasks.md, and design.md (when needed) scoped to that proposal only.
• When multiple proposals are created, explicitly document dependency/sequence relationships and parallelizability in the final user-facing summary.
• Before asking clarifying questions, proactively gather context from the current session and repository, and treat that gathered context as the default premise for the proposal.
• Start the user-facing response with a short Premise / Context section summarizing the goals, constraints, and relevant repo architecture already discovered.
• Do not ask the user to choose or confirm the change-id; generate a concise unique verb-led slug yourself.
• If the request is sufficiently clear after context gathering, draft the proposal directly instead of forcing an extra clarification round.
• For implementation-oriented proposals, make tasks evidence-bearing: each behavior-changing task should name repository-verifiable code, tests, commands, or explicit manual checks.
• For behavior-changing proposals, require verification coverage planning per requirement/task so verification ownership is explicit at proposal time.
• Write tasks around required behavior, not artifact existence. Avoid task sets that can look complete when only contracts, docs, or placeholder wiring exist.
• Separate responsibility-definition/documentation tasks from runtime wiring/integration tasks so the proposal cannot be marked done without code paths, state changes, side effects, or externally visible behavior being connected.
• Because Conflux decides execution order autonomously, proposal-side task decomposition must be complete enough that an implementation agent does not need to infer missing required work from unstated intent.
• Treat every user-visible requirement, dependency, migration, verification activity, and non-goal as something that must be explicitly represented in the proposal artifacts, not left implicit.
• Every implementation-facing task must have an explicit completion condition so a later agent can determine done/not-done from repository evidence instead of judgment calls.
• For each major behavior-changing requirement, include at least one verification path that would fail if the implementation were stubbed, no-op, or returning dummy data.
• When the change exposes a CLI, API, workflow, job, or background process, include minimal execution verification covering a typical success path plus safe-mode / side-effect suppression / error handling where relevant.
• Prefer warnings over silent acceptance when tasks are dominated by “define / document / describe” language or when runtime behavior is claimed without runnable verification.
• Before choosing spec-only, explicitly classify the user's desired artifact as one of: spec/documentation, implementation, or both.
• Use the standard verification vocabulary for ownership planning: unit, integration, e2e, manual, benchmark, not-testable.
• Treat manual and benchmark as intentional coverage (not missing unit tests) when they fit the requirement better.
• If the user's request is phrased as a concrete product, UI, API, route, form, workflow, or behavior change, default to implementation or hybrid, not spec-only, unless the user explicitly asks for spec-only or documentation-only work.
• If canonical specs already describe the requested behavior but the codebase still lacks implementation, do NOT create another spec-only proposal unless the user explicitly asks to update spec artifacts only.
• When selecting spec-only, explicitly state in the user-facing response that the proposal will NOT implement the feature and will only update tracked specification artifacts.

When to Use This Skill

Trigger this skill when users request:

• "Create a proposal for..."
• "Draft a change proposal"
• "Propose a new feature"
• "Document this change as a proposal"
• Any request to create structured change documentation

Key Characteristics

Human-Interactive Mode:

• Ask clarifying questions to understand requirements
• Guide users through proposal structure
• Discuss design decisions and trade-offs
• Iterate on requirements based on feedback
• Present proposals for user review and approval

Not for Orchestration: This skill is designed for direct human interaction, not automated orchestration.

Proposal Structure

A Conflux proposal consists of:

openspec/changes/<change-id>/
├── proposal.md          # Change description and context
├── tasks.md             # Implementation task checklist
├── design.md            # Architecture and design (optional)
└── specs/               # Spec deltas
    └── <capability>/
        └── spec.md      # Requirement specifications

Interactive Workflow

1. Understand the Change

First, gather available context proactively before asking questions.

Context to gather first:

• Prior user messages, goals, and constraints from the current session
• Repository-specific instructions such as AGENTS.md when present
• Existing related specs, archived changes, and relevant source/test modules
• Existing architecture or workflow boundaries already mentioned in the conversation

Begin the user-facing proposal response with:

• Premise / Context
• 3-6 concise bullets summarizing the gathered facts that will shape the proposal
• Requested Artifact with one of: spec/documentation, implementation, or both

If the inferred requested artifact is implementation or both, do not silently downgrade the proposal to spec-only.

Ask questions only if still necessary after context gathering:

• What problem does this solve?
• What are the acceptance criteria?
• Are there any constraints or dependencies?
• What is the scope (minimal vs. comprehensive)?
• What required work would make the request incomplete if omitted from the proposal?

If the answer is already inferable from the repository and current conversation, skip the question and proceed.

Before proceeding, explicitly normalize the request into a completeness checklist covering:

• user-facing outcomes that must become true
• repository areas likely requiring change
• required verification proving each outcome
• dependencies, migrations, rollout concerns, and follow-up work that must be represented
• explicit non-goals / out-of-scope items to prevent accidental under-specification

Research existing code:

# Review existing specs
cflx openspec list --specs

# Check related code
rg "<keyword>"
ls <relevant-directory>

Strict validation note (common gotcha):

• In strict mode, include at least one spec delta under openspec/changes/<id>/specs/<capability>/spec.md.
• Before writing ## MODIFIED Requirements or ## REMOVED Requirements, inspect the matching canonical openspec/specs/<capability>/spec.md and copy an existing ### Requirement: heading exactly. These sections target existing canonical requirement identities; strict validation fails when the canonical heading is absent.
• Use ## ADDED Requirements when no matching canonical ### Requirement: heading exists yet.
• For bugfix-only proposals (no intended new behavior), add a minimal ## MODIFIED Requirements delta only when the target requirement heading already exists in the canonical spec; otherwise use ## ADDED Requirements for the new tracked behavior.

2. Classify Change Type

Every proposal MUST include a Change Type field in proposal.md. First classify the user's desired artifact:

• spec/documentation: the user explicitly wants proposal, spec, documentation, archive-readiness, or canonical requirement updates
• implementation: the user wants code, UI, API wiring, routes, forms, tests, runtime behavior, or removal of existing behavior in the product
• both: the user wants both spec and implementation to move together

Use this decision table:

Strong defaults:

• If the request is phrased as adding/removing/fixing a concrete product behavior, UI, page, route, form, API integration, workflow, or runtime feature, treat it as implementation by default.
• If existing canonical specs already describe the target behavior and the repo still lacks implementation, choose implementation, not spec-only, unless the user explicitly asks for spec cleanup only.
• Never choose spec-only merely because a backend surface already exists; if user-visible behavior is still missing from code, the default remains implementation.

When to split instead of using hybrid:

• If the spec authoring and the implementation can be reviewed and deployed independently, split into two proposals.
• Use hybrid only when spec and code must ship atomically to preserve correctness.

When drafting proposal.md, prefer YAML frontmatter at the top for machine-readable metadata:

---
change_type: spec-only # or: implementation | hybrid
priority: medium # high | medium | low
dependencies: [] # optional change-id list; overrides body `## Dependencies`
references: [] # optional string list of related files/specs/changes
---

Keep the human-readable line near the top as well for backward-compatible readability:

**Change Type**: spec-only <!-- or: implementation | hybrid -->

3. Evaluate Split Boundaries (Default: Split)

Before writing anything, evaluate whether the request should be split into multiple independent change proposals.

Default rule: if scopes are independent or weakly coupled, split into separate openspec/changes/<change-id>/ proposals.

Keep as a single proposal only when:

• The scopes are tightly coupled and must ship atomically to preserve correctness.
• The acceptance criteria cannot be verified independently.

When keeping a single proposal despite multiple scopes, explicitly record the rationale in proposal.md or design.md.

3. Generate Change ID

Rules:

• Verb-led (e.g., add-auth, fix-validation, refactor-api)
• Kebab-case (lowercase with hyphens)
• Must NOT include date prefixes or suffixes (forbidden: 2026-02-07-add-auth, add-auth-2026-02-07)
• Concise but descriptive
• Unique within the project

Execution rule:

• Generate the change-id yourself.
• Do not ask the user to confirm or choose it.
• If a collision is possible, disambiguate automatically (for example by adding a short suffix).

Present to user: "Using change ID <id>."

4. Draft Proposal Content

Create openspec/changes/<id>/proposal.md:

Required sections:

• YAML frontmatter with change_type, priority, optional dependencies, optional references
• Title (H1)
• Problem/Context
• Proposed Solution
• Acceptance Criteria
• Explicit Completion Conditions
• Out of Scope (if applicable)

Acceptance Criteria must describe the user-visible or system-visible outcomes that need to hold when the change is done. Explicit Completion Conditions must describe how a later agent/reviewer can tell the proposal has been fully implemented, including expected code paths, tests, commands, or artifacts.

Ask for feedback: "Here's the draft proposal. Would you like to adjust anything?"

5. Create Task Breakdown

Create openspec/changes/<id>/tasks.md:

Task format for implementation or hybrid proposals:

## Implementation Tasks

- [ ] Task 1: Description (verification: how to verify completion)
- [ ] Task 2: Description (verification: ...)

## Future Work

- Items that require human action
- Items requiring external systems
- Long-wait verification tasks

Task format for spec-only proposals — use ## Specification Tasks instead of ## Implementation Tasks, and include a one-line expected canonical outcome for each delta:

## Specification Tasks

- [ ] Promote `specs/capability-name/spec.md` delta to canonical spec
  - Expected canonical result: <what the canonical spec will contain after archive>
- [ ] Review and validate delta scenarios for completeness

## Future Work

- Human sign-off on canonical promotion

Note: For spec-only proposals, each spec delta should include a short comment describing how the canonical spec changes after archive. This allows acceptance to evaluate archive-readiness without expecting runtime integration evidence.

Guidelines:

• Break into small, verifiable steps
• Include verification methods
• Ensure the task list is complete enough that Conflux can choose execution order without depending on unstated human intent
• Represent every required implementation, integration, migration, verification, and documentation step needed to fully satisfy the request
• For each behavior-changing requirement/task, record planned verification ownership using one of: unit, integration, e2e, manual, benchmark, not-testable
• Start each behavior-bearing verification note with that ownership marker (for example verification: integration - cargo test --test auth_flow)
• When manual or benchmark is selected, state why it is intentional coverage and how reviewers will evaluate completion
• Specify integration/wiring tasks separately from responsibility-definition or documentation tasks
• Mark non-AI-executable tasks for Future Work
• Prefer concrete repository evidence in verification notes (source paths, test files, runnable commands, or explicit manual observation steps), not vague statements like "verify implementation works"
• Make task verification notes traceable as ownership pairs: implementation task + verification path (for example verification: integration - tests/api/test_auth_flow.rs)
• Every checkbox task must have a completion condition that can be evaluated from repo state, test results, generated artifacts, or an explicit manual observation
• For major requirements, at least one verification path must require real implementation behavior rather than spec text or placeholder wiring
• If a CLI/API/workflow/job/background process is in scope, include a minimal runnable verification for expected success behavior and relevant safe-mode / no-op / error conditions
• If omitting a task would leave some acceptance criterion unsatisfied, the task belongs in the proposal even if the implementation order is unknown

External dependency policy (mock-first / verification-first):

• If a requirement cannot be verified locally without credentials or external systems, design mock/stub/fixture-based verification.
• Do not change production/runtime behavior to "use mocks"; mocks/stubs/fixtures are for tests and local verification.
• Only truly non-mockable dependencies (human decisions, real external systems, long-wait checks) go to Out of Scope / Future Work.

Present to user: "I've broken this down into X tasks. Do these cover everything?"

6. Design Documentation (Optional)

Create openspec/changes/<id>/design.md when:

• Change spans multiple systems
• Architectural decisions need documentation
• Trade-offs require explanation
• Complex implementation patterns

Design documentation is strongly recommended when the proposal introduces orchestration, durable state, cross-service coordination, background workers, adapters, or other multi-layer behavior.

Ask: "Should we document the design decisions in detail?"

7. Write Spec Deltas

Create openspec/changes/<id>/specs/<capability>/spec.md:

Format:

## ADDED Requirements

### Requirement: <requirement-name>

<Description>

#### Scenario: <scenario-name>

**Given**: <preconditions>
**When**: <action>
**Then**: <expected-outcome>

## MODIFIED Requirements

### Requirement: <existing-requirement-name>

<Updated description>

#### Scenario: <scenario-name>

...

## REMOVED Requirements

### Requirement: <deprecated-requirement-name>

<Reason for removal>

Critical rules:

• Each requirement must have at least one scenario
• Use ADDED/MODIFIED/REMOVED sections
• Before selecting MODIFIED or REMOVED, open openspec/specs/<capability>/spec.md and verify the canonical ### Requirement: heading exists; copy that heading exactly into the delta.
• If the canonical heading does not exist, use ADDED instead of MODIFIED/REMOVED.
• Be specific and testable

Discuss with user: "Should we add these requirements to the spec?"

8. Validate Proposal

Run validation after authoring proposal, tasks, and spec deltas:

cflx openspec validate <id> --strict

Strict validation checks MODIFIED/REMOVED target headings against canonical specs, so fix any missing-target diagnostics before handoff.

If validation fails:

• Show errors to user
• Discuss fixes
• Apply corrections
• Re-validate

Present results: "Validation passed! The proposal is ready for review."

9. Final Review

Present complete proposal to user:

• Show directory structure
• Summarize key points
• Highlight task count
• Confirm readiness

When the proposal was split into multiple independent change proposals, always present a proposal index:

- change-id
  - one-line objective
  - dependency/sequence (if any)
  - whether it can be implemented in parallel

Ask: "The proposal is complete. Would you like to proceed with implementation, or make any final adjustments?"

Mock-First External Dependencies

When designing tasks, follow mock-first approach:

Prefer:

• Mock/stub/fixture implementations
• Test doubles for external APIs
• Local verification without credentials

Avoid:

• Blocking on missing API keys
• Requiring real external services
• Deferring mockable dependencies

Discuss with user: "For the external API integration, should we use mocks for testing, or do you have test credentials available?"

Task Classification

AI-Executable Tasks (include with checkbox):

• Code implementation
• Unit/integration tests
• Documentation updates
• Linting/formatting
• Local verification

Future Work Tasks (no checkbox):

• Manual approval required
• Human decision-making
• External system deployment
• Long-wait verification (>1 day)

Ask user: "Are there any tasks that require manual review or external approvals?"

Question Examples

Understanding Requirements

• "What's the primary user need this addresses?"
• "Are there any security or performance requirements?"
• "What's the expected timeline for this change?"

Clarifying Scope

• "Should this include error handling for edge cases?"
• "Do we need backward compatibility?"
• "Are there related features we should consider?"

Design Decisions

• "Would you prefer approach A (simpler) or B (more flexible)?"
• "Should we optimize for performance or maintainability?"
• "Where should this integrate with the existing system?"

Verification

• "How should we verify this is working correctly?"
• "What would constitute a successful implementation?"
• "Are there specific test scenarios we should cover?"

Built-in Tools

Use cflx openspec for all Conflux OpenSpec operations:

# List existing changes
cflx openspec list

# List specs
cflx openspec list --specs

# Show change details
cflx openspec show <id>

# Validate proposal
cflx openspec validate <id> --strict

# Validate proposal with implementation-evidence warnings
cflx openspec validate <id> --strict --evidence warn

# Validate proposal with implementation-evidence errors
cflx openspec validate <id> --strict --evidence error

# Recommended authoring check for behavior-changing proposals
cflx openspec validate <id> --strict --evidence warn
# Resolve warnings about missing ownership, artifact-heavy tasks,
# absent runnable verification, or executable surfaces lacking run checks

# Archive-equivalent readiness check for final validation sections
cflx openspec validate <id> --archive-gate

Final validation guidance must be written outside checkbox task lists, for example:

## Final Validation

Archive validation itself is the authoritative final OpenSpec validation gate.
Expected archive gate: `cflx openspec validate <id> --archive-gate`

Best Practices

Communication

• Ask one question at a time (avoid overwhelming)
• Summarize understanding before proceeding
• Present options with recommendations
• Confirm major decisions
• Prefer making a reasonable default proposal after context gathering over repeatedly blocking on optional clarification
• When you must ask a question, state the recommended default and what would change based on the answer

Proposal Quality

• Keep scope focused and minimal
• Break large changes into multiple proposals
• Include clear acceptance criteria
• Add explicit completion conditions so another agent can decide done/not-done without inferring missing intent
• For behavior-changing proposals, plan verification ownership per requirement/task using unit / integration / e2e / manual / benchmark / not-testable
• Treat manual and benchmark as valid intentional coverage when justified by requirement characteristics
• Specify verification for each task
• Make implementation-facing tasks hard to mark complete without source/test evidence
• Ensure tasks guidance keeps ownership traceable by pairing each implementation task with its verification path
• Do not create final OpenSpec validation as a checkbox implementation task. Final validation belongs in a non-checkbox ## Final Validation section because archive validation is the authoritative gate.
• Check for proposal completeness by asking: "If Conflux executed these tasks in its own chosen order, would the full user request still be satisfied without any hidden assumptions?"

User Experience

• Be responsive to feedback
• Iterate on content based on input
• Present information clearly
• Provide helpful suggestions

Common Patterns

Feature Addition

1. Understand user need
2. Review existing architecture
3. Propose minimal viable solution
4. Break into implementation tasks
5. Add spec requirements

Bug Fix

1. Understand the bug behavior
2. Identify root cause area
3. Propose fix approach
4. Include regression tests
5. Add MODIFIED spec if behavior changes

Refactoring

1. Clarify refactoring goals
2. Identify affected components
3. Plan incremental steps
4. Ensure test coverage
5. Document design decisions

Troubleshooting

User Unclear About Requirements

• Ask more specific questions
• Provide examples or options
• Start with minimal scope
• Plan for iteration

Complex Dependencies

• Discuss mock-first approach
• Identify what's truly non-mockable
• Plan integration points
• Document assumptions

Large Scope

• Suggest breaking into multiple proposals
• Identify core vs. nice-to-have
• Prioritize with user
• Plan phased approach

Reference Files

No additional reference files.

Example Interaction

User: "I want to add user authentication"

Agent: "I'd like to understand the authentication requirements better:
1. What type of authentication? (JWT, session-based, OAuth?)
2. What user actions need to be protected?
3. Do you have an existing user database or need to create one?"

User: "JWT-based, protect API endpoints, we have a users table"

Agent: "Got it. I suggest the change ID `add-jwt-auth`. I'll create a proposal with:
- JWT token generation and validation
- Middleware for protected routes
- Login/logout endpoints

Does this scope look correct?"

User: "Yes, please proceed"

Agent: [Creates proposal.md, tasks.md, specs/]
"Here's the task breakdown:
1. Implement JWT utility functions
2. Create auth middleware
3. Add login endpoint
4. Add logout endpoint
5. Protect existing API routes
6. Add unit tests
7. Add integration tests

Total: 7 tasks. Does this cover everything?"

User: "Add password reset too"

Agent: "I'll add password reset. This requires email sending - should we:
A) Mock email for testing (recommended)
B) Use a real email service (needs credentials)
?"

User: "Mock for now"

Agent: [Updates proposal]
"Added 2 more tasks for password reset with mocked email.
Running validation... ✓ Validation passed!

The proposal is ready. Would you like to start implementation?"