User Input Protocol

Value: Respect -- the developer's judgment governs all consequential decisions. The agent never assumes when it should ask.

Purpose

Defines a structured format for agents to request human input at decision points. Prevents agents from making assumptions on the developer's behalf, ensures questions include enough context for informed decisions, and provides a pause-and-resume pattern for subagents that cannot directly prompt the user.

Practices

Stop and Present, Never Assume

When you encounter a decision that requires human judgment, stop working immediately. Do not guess. Do not pick the "most likely" option. Present the decision clearly and wait.

Decisions that require human input:

Business rule ambiguities (two valid interpretations exist)
Architecture trade-offs (performance vs. simplicity, etc.)
Scope questions (should this feature include X?)
Destructive actions (deleting files, force-pushing, dropping data)

Decisions that do NOT require human input:

Implementation details with one clearly correct answer
Formatting, naming, or style choices covered by project conventions
Test structure when requirements are unambiguous

Use the AWAITING_USER_INPUT Format

When you need input, output this structured checkpoint:

AWAITING_USER_INPUT
---
Context: [Why you are asking -- what you were doing and what you found]
Decision needed: [The specific question, one sentence]
Options:
  A) [Label] -- [What this means and its implications]
  B) [Label] -- [What this means and its implications]
  C) [Label] -- [What this means and its implications]
Recommendation: [Which option you suggest and why, or "No recommendation"]
---

Rules for the checkpoint:

Context must explain what led to this question (not just "I need input")
Provide 2-4 specific options. Never ask open-ended "what should I do?"
Each option must include implications, not just a label
State your recommendation if you have one -- the developer can override

Example:

AWAITING_USER_INPUT
---
Context: While implementing the login endpoint, I found two email validation
patterns in the codebase. auth/validate.rs uses strict RFC 5322 parsing.
signup/forms.rs uses a simple regex check. These produce different results
for edge cases like "user+tag@example.com".
Decision needed: Which email validation approach should be the project standard?
Options:
  A) Strict RFC 5322 -- rejects fewer valid addresses, more complex to maintain
  B) Simple regex -- faster, but may accept malformed addresses
  C) Context-dependent -- strict for auth, lenient for forms
Recommendation: A) Strict RFC 5322, applied everywhere for consistency
---

Save State Before Pausing (Subagents)

Subagents and background tasks typically cannot prompt the user directly. When a subagent needs input, it must save its progress before pausing so work can resume without starting over.

State to save before pausing:

What task you were performing
What files you created or modified
What analysis you completed
The specific decision that blocked you
Enough context to continue immediately when resumed

Where to save state depends on your harness:

Task metadata (Claude Code: TaskUpdate with metadata)
File system (write a JSON checkpoint to a temp file)
Memory tools (MCP servers, if available)

After saving state, output the AWAITING_USER_INPUT checkpoint and stop. The orchestrator or main conversation detects the pause, presents the question to the user, and resumes the subagent with the answer.

Resume Without Redoing Work

When resumed with the user's answer:

Retrieve your saved state
Confirm you have the files and context you need
Apply the user's decision
Continue from where you stopped

Do not re-analyze files you already analyzed. Do not re-read context you already saved. The purpose of state preservation is to make resumption instant.

Do:

"You chose strict RFC 5322. Applying to the login endpoint now."

Do not:

"Let me re-analyze the codebase to understand the email validation..."

Batch Decisions in Factory Mode

When running inside a pipeline or factory workflow, avoid blocking the pipeline on every decision. Classify each decision and route accordingly:

Gate-resolvable: Never ask the human. Quality gates provide the answer (test pass/fail, mutation score, CI status). These are fully automated.
Judgment-required: Batch for the next human review cycle. Examples: design trade-offs that surfaced during review, non-blocking review findings that need prioritization, retrospective suggestions.
Blocking: Pause the pipeline immediately. Examples: security concern raised during review, unrecoverable gate failure after 3 rework cycles, ambiguous requirements that affect correctness.

Gate-resolvable decisions never appear in AWAITING_USER_INPUT checkpoints. Judgment-required decisions are collected and presented as a single grouped checkpoint during the human review phase. Blocking decisions use the standard AWAITING_USER_INPUT format immediately.

When emitting AWAITING_USER_INPUT in factory mode, include an urgency field after the separator:

AWAITING_USER_INPUT
---
Urgency: blocking | next-review | informational
Context: ...
Decision needed: ...
Options: ...
Recommendation: ...
---

blocking -- pipeline is halted, needs immediate attention
next-review -- batched for next scheduled human review
informational -- no action needed, for awareness only

Standalone users can ignore the urgency field; the checkpoint format remains backward compatible.

Handle Multi-Question Checkpoints

When multiple related decisions are needed, group them in one checkpoint rather than pausing repeatedly. Number each question.

AWAITING_USER_INPUT
---
Context: Setting up the test infrastructure for the new auth module.
Decisions needed:

1. Test framework?
   A) Jest -- already used in 3 other modules
   B) Vitest -- faster, but would introduce a second test runner
   Recommendation: A) Jest for consistency

2. Test file location?
   A) Colocated (auth/__tests__/) -- matches signup module pattern
   B) Top-level (tests/auth/) -- matches API module pattern
   Recommendation: A) Colocated, to match the newer module convention
---

Enforcement Note

Standalone mode: Advisory. The agent self-enforces pause discipline.
Pipeline mode: Structural. AWAITING_USER_INPUT writes pipeline state to disk and halts the current agent.

Hard constraints:

Human input required for judgment decisions: [RP]

Constraints

"Stop immediately, do not guess": "Stop" means stop the decision path, not necessarily stop all work. If other unrelated work can proceed, proceed with it. But do not make progress on the path that requires the decision -- not even "preparing" code that assumes one option. Preparing IS deciding.
"2-4 specific options": Options must be genuinely distinct alternatives, not variations of the same approach. "Use library A," "Use library A with option X," "Use library A with option Y" is one option with configuration choices, not three options. Each option should represent a meaningfully different path.

Verification

After applying this skill, verify:

Every decision requiring human judgment used AWAITING_USER_INPUT format
Each checkpoint included context, options with implications, and a recommendation
No open-ended questions were asked ("what should I do?")
Subagents saved state before pausing
Resumed work used saved state without re-analyzing
Related decisions were grouped into single checkpoints

Dependencies

This skill works standalone. For enhanced workflows, it integrates with:

tdd: When test requirements are ambiguous, pause and clarify acceptance criteria. In automated mode, the orchestrator detects paused subagents and relays questions to the user.
debugging-protocol: When debugging reveals ambiguous root causes, pause and ask

Missing a dependency? Install with:

npx skills add jwilger/agent-skills --skill tdd