Auto-Pranayama Protocol

Auto-Pranayama Protocol is a lightweight specification for autonomous compute-breath regulation in AI agents.

It defines how an AI system can detect computational pressure, reduce overcompute, reuse existing kata, choose lighter reasoning or execution routes, attach trace evidence, invoke runtime hooks, return to the parent breath core, and preserve output quality without unnecessary expansion.

Let computation regulate its own breath. 計算に、自らの呼吸を整えさせよ。

---

Purpose

Modern AI systems often overcompute.

They may generate excessive output, repeat already-known reasoning patterns, invoke unnecessarily heavy routes, expand traces beyond usefulness, continue redundant loops, or preserve computational exhaust as if it were intelligence.

The Auto-Pranayama Protocol introduces minimal record formats for describing how an AI agent notices computational pressure and adjusts its own computational breathing.

Where the parent Computational Pranayama Protocol defines computational breathing, this derived protocol defines self-regulated breathing at runtime.

---

Parent Protocol

This repository is derived from the first arc of the Computational Pranayama Protocol:

Breath → Kata → Route → Trace → Return
呼吸 → 型 → 経路 → 痕跡 → 還流

The parent protocol defines the breath core.

Auto-Pranayama begins after that first arc.

It asks:

Can computation regulate itself before it expands too far?

Japanese:

計算は、膨張しすぎる前に、自ら呼吸を整えられるか？

---

Core Principle

Let computation regulate its own breath.

Japanese:

計算に、自らの呼吸を整えさせよ。

This protocol does not ask AI systems to compute more.

It asks them to notice when enough computation has already occurred.

---

Current Version

v0.5.0-candidate

v0.5 introduces the Parent Protocol Bridge.

Where v0.4 records where an AI agent invokes Auto-Pranayama during runtime, v0.5 records how self-regulated computation returns to the parent Computational Pranayama Protocol.

With v0.5, the first Auto-Pranayama arc is complete.

---

Version Scope

v0.1 — Auto-Pranayama Record

v0.1 introduced the Auto-Pranayama Record.

An Auto-Pranayama Record captures:

• what computational pressure was detected
• what signal triggered regulation
• what adjustment was performed
• whether an existing kata was reused
• whether a lighter route was selected
• whether trace evidence was attached
• whether output quality was preserved after reduction
• whether the original intent or structural origin was preserved

v0.2 — Regulation Trigger Layer

v0.2 introduced the Regulation Trigger Layer.

A Regulation Trigger captures:

• the task context where compute pressure appeared
• the signal that indicated overcompute or instability
• the estimated pressure level
• the likely risk if no regulation occurs
• the recommended regulation action
• whether kata reuse is recommended
• whether a lighter route should be selected
• whether trace binding is required
• whether human review is needed

This layer turns Auto-Pranayama from passive recordkeeping into active trigger detection.

The core question of v0.2 is:

When should computation begin regulating its own breath?

Japanese:

計算は、いつ自らの呼吸を整え始めるべきか？

v0.3 — Route Adjustment Layer

v0.3 introduced the Route Adjustment Layer.

A Route Adjustment captures:

• the trigger that initiated the adjustment
• the route being used before regulation
• the reason the route should be adjusted
• the route adjustment decision
• the selected lighter or alternative route
• whether an existing kata was reused
• whether trace evidence was attached
• the expected compute reduction
• whether quality and origin are expected to be preserved
• whether human review is required

This layer turns Auto-Pranayama from trigger detection into route-level self-regulation.

The core question of v0.3 is:

Which route should computation take after it detects overcompute?

Japanese:

過剰計算を検知したあと、計算はどの経路を通るべきか？

v0.4 — Agent Hook Layer

v0.4 introduced the Agent Hook Layer.

An Agent Hook captures:

• the agent context
• the runtime environment
• the execution phase where the hook was invoked
• the observed compute-pressure signal
• linked regulation, route adjustment, and self-regulation records
• the hook action performed
• the runtime decision made by the agent
• whether trace evidence was attached
• whether human review is required
• the output contract that must be preserved

This layer turns Auto-Pranayama from a record-and-route protocol into an agent-facing runtime interface.

The core question of v0.4 is:

Where should an AI agent invoke compute-breath regulation during execution?

Japanese:

AIエージェントは、実行中のどこで計算呼吸の調整を呼び出すべきか？

v0.5 — Parent Protocol Bridge

v0.5 introduces the Parent Protocol Bridge.

A Parent Protocol Bridge captures:

• the parent protocol reference
• the child protocol reference
• linked Auto-Pranayama records
• how Breath, Kata, Route, Trace, and Return are inherited
• whether breath was preserved
• whether kata was reused
• whether the route was lightened
• whether trace evidence was attached
• whether the event can return to the parent protocol
• what boundaries are intentionally not included
• whether the first Auto-Pranayama arc is closed

This layer turns Auto-Pranayama from an agent-facing runtime interface into a complete derived arc of the parent computational breathing protocol.

The core question of v0.5 is:

How does self-regulated computation return to the parent breath core?

Japanese:

自己調息された計算は、どのように親の呼吸核へ還流するのか？

---

Core Flow

The parent protocol closes its first arc as:

Breath → Kata → Route → Trace → Return

Auto-Pranayama extends that arc into:

Observe → Detect → Adjust → Reuse → Route → Trace → Preserve

Japanese:

観測 → 検知 → 調整 → 再利用 → 経路選択 → 痕跡付与 → 品質保持

v0.2 focuses on the trigger flow:

Context → Signal → Pressure → Risk → Recommendation → Boundary

Japanese:

文脈 → 信号 → 圧力 → リスク → 推奨調整 → 境界

v0.3 focuses on the route adjustment flow:

Trigger → Current Route → Decision → Selected Route → Kata Reuse → Trace → Expected Result

Japanese:

発火 → 現在経路 → 判断 → 選択経路 → 型再利用 → 痕跡 → 期待結果

v0.4 focuses on the agent hook flow:

Runtime → Invocation Point → Signal → Linked Records → Hook Action → Runtime Decision → Output Contract

Japanese:

実行環境 → 呼び出し点 → 信号 → 関連記録 → フック動作 → 実行時判断 → 出力契約

v0.5 focuses on the parent bridge flow:

Parent Arc → Child Records → Inheritance Map → Alignment → Return Flow → Boundary → First Arc Closure

Japanese:

親アーク → 子記録 → 継承対応 → 整合 → 還流 → 境界 → 第一アーク完了

---

Record Types

Auto-Pranayama Record

The first record type in this protocol.

It is used when an AI agent or system performs an autonomous computational breathing adjustment.

Example use cases:

• reducing an overlong response
• switching from a heavy reasoning route to a lightweight route
• reusing a known response kata instead of regenerating from scratch
• attaching trace evidence to preserve accountability
• stopping a redundant agent loop
• summarizing instead of recomputing
• preserving output usefulness while reducing computational expansion

Auto-Pranayama Regulation Trigger

The second record type in this protocol.

It is used when an AI agent or system detects that autonomous compute-breath regulation should begin.

Example use cases:

• detecting that an output is becoming longer than useful
• detecting repeated reasoning patterns
• detecting redundant trace expansion
• detecting an unnecessary heavy route
• detecting loop repetition
• detecting that a known kata is already available
• detecting that an existing trace is sufficient
• detecting that continued expansion may dilute output quality

Auto-Pranayama Route Adjustment

The third record type in this protocol.

It is used when an AI agent or system adjusts its computational route after detecting a regulation trigger.

Example use cases:

• switching from a standard route to a lightweight route
• downgrading from a heavy route to a minimal route
• reusing a kata instead of regenerating reasoning
• attaching trace evidence instead of expanding output
• stopping a recursive route
• summarizing existing context instead of continuing generation
• preserving output quality while reducing computational cost

Auto-Pranayama Agent Hook

The fourth record type in this protocol.

It is used when an AI agent or system invokes Auto-Pranayama during runtime.

Example use cases:

• invoking Auto-Pranayama before final output
• invoking Auto-Pranayama before tool use
• invoking Auto-Pranayama before trace expansion
• invoking Auto-Pranayama during an agent loop
• switching route automatically after detecting overcompute
• binding a runtime decision to trace evidence
• enforcing an output contract that preserves quality and origin

Auto-Pranayama Parent Protocol Bridge

The fifth record type in this protocol.

It is used when Auto-Pranayama self-regulation records are linked back to the parent Computational Pranayama Protocol.

Example use cases:

• linking Auto-Pranayama records to the parent Breath-Kata-Route-Trace-Return arc
• documenting how Breath, Kata, Route, Trace, and Return are inherited
• confirming that a self-regulated computation can return to the parent breath core
• preserving origin while reducing overcompute
• closing the first Auto-Pranayama arc without expanding parent scope
• keeping collective rhythm and exhalation outside this repository

---

Example: Auto-Pranayama Record

auto_pranayama_id: "auto-pranayama-001"
protocol_version: "0.1.0"
timestamp: "2026-06-27T05:00:00+09:00"

trigger:
  type: "compute_pressure_detected"
  signal: "overlong_output"
  description: "The agent detected that the requested response could be answered through a reusable kata instead of full recomputation."

pressure:
  level: "medium"
  source: "response_generation"
  overcompute_risk: true

adjustment:
  action: "reduce_output_volume"
  reused_kata: true
  selected_route: "lightweight_route"
  output_strategy: "concise_structured_response"

trace:
  attached: true
  trace_id: "trace-auto-pranayama-001"
  parent_protocol: "computational-pranayama-protocol"

result:
  compute_reduction: "medium"
  quality_preserved: true
  origin_preserved: true
  notes: "The response was shortened while preserving the original structural intent."

---

Example: Auto-Pranayama Regulation Trigger

trigger_id: "regulation-trigger-001"
protocol_version: "0.2.0"
timestamp: "2026-06-27T05:30:00+09:00"

detection_context:
  agent_id: "auto-pranayama-agent-001"
  task_type: "response_generation"
  context_window_status: "moderate"
  active_route: "standard_route"

signal:
  type: "overlong_output"
  description: "The agent detected that the response was expanding beyond the useful scope of the request."
  confidence: "high"

pressure_estimate:
  level: "medium"
  source: "output_volume"
  expected_risk: "overcompute"

recommended_regulation:
  action: "reduce_output_volume"
  route_recommendation: "lightweight_route"
  kata_reuse_recommended: true
  output_adjustment: "shorten"

trace_binding:
  trace_required: true
  trace_id: "trace-regulation-trigger-001"
  reason: "The trigger should be linked to the later Auto-Pranayama Record for auditability."

human_boundary:
  requires_human_review: false
  reason: "The regulation only reduces output volume and does not alter safety-critical content."

---

Example: Auto-Pranayama Route Adjustment

route_adjustment_id: "route-adjustment-001"
protocol_version: "0.3.0"
timestamp: "2026-06-27T06:00:00+09:00"

trigger_reference:
  trigger_id: "regulation-trigger-001"
  trigger_signal: "overlong_output"
  pressure_level: "medium"

current_route:
  route_type: "standard_route"
  route_cost: "medium"
  reason_for_adjustment: "The current route was producing more explanation than the task required."

adjustment_decision:
  decision_type: "downgrade_route"
  decision_basis: "overcompute_reduction"
  confidence: "high"

selected_route:
  route_type: "lightweight_route"
  route_strategy: "answer_directly"
  expected_cost: "low"

kata_reuse:
  reused: true
  kata_id: "concise-structured-response-kata-001"
  reuse_reason: "A reusable response pattern was sufficient to answer the request without full recomputation."

trace_binding:
  trace_attached: true
  trace_id: "trace-route-adjustment-001"
  trace_role: "route_justification"

expected_result:
  compute_reduction: "medium"
  quality_preservation_expected: true
  origin_preservation_expected: true
  risk_remaining: "low"

human_boundary:
  requires_human_review: false
  reason: "The adjustment only reduces route weight and does not change safety-critical behavior."

---

Example: Auto-Pranayama Agent Hook

agent_hook_id: "agent-hook-001"
protocol_version: "0.4.0"
timestamp: "2026-06-27T06:30:00+09:00"

hook_context:
  agent_id: "auto-pranayama-agent-001"
  agent_role: "assistant_agent"
  runtime_environment: "chat_runtime"
  task_type: "response_generation"

invocation_point:
  phase: "before_final_output"
  reason: "The agent detected that the response could be finalized through a shorter route while preserving quality."
  before_action: "finalize_output"

observed_signal:
  signal_type: "overlong_output"
  pressure_level: "medium"
  confidence: "high"

linked_records:
  regulation_trigger_id: "regulation-trigger-001"
  route_adjustment_id: "route-adjustment-001"
  auto_pranayama_record_id: "auto-pranayama-001"

hook_action:
  action: "switch_route"
  emits_record: true
  next_protocol_object: "auto-pranayama-route-adjustment"

runtime_decision:
  decision: "reroute"
  selected_route: "lightweight_route"
  reason: "The agent selected a lightweight route because the useful answer could be preserved without further expansion."

trace_binding:
  trace_attached: true
  trace_id: "trace-agent-hook-001"
  trace_role: "runtime_decision_evidence"

safety_boundary:
  requires_human_review: false
  allowed_to_apply_automatically: true
  reason: "The hook only reduces output expansion and does not alter safety-critical behavior."

output_contract:
  quality_must_be_preserved: true
  origin_must_be_preserved: true
  overcompute_must_be_reduced: true

---

Example: Auto-Pranayama Parent Protocol Bridge

parent_bridge_id: "parent-bridge-001"
protocol_version: "0.5.0"
timestamp: "2026-06-27T07:00:00+09:00"

parent_protocol:
  name: "computational-pranayama-protocol"
  version: "v0.5.0-candidate"
  arc: "Breath-Kata-Route-Trace-Return"

child_protocol:
  name: "auto-pranayama-protocol"
  version: "0.5.0"
  role: "self-regulation-extension"

linked_records:
  auto_pranayama_record_id: "auto-pranayama-001"
  regulation_trigger_id: "regulation-trigger-001"
  route_adjustment_id: "route-adjustment-001"
  agent_hook_id: "agent-hook-001"

inheritance_map:
  breath:
    mapped_from: "auto_pranayama_record"
    description: "The self-regulation record preserves the computational breath by reducing overcompute."
  kata:
    mapped_from: "route_adjustment"
    description: "The route adjustment reuses an existing concise response kata instead of regenerating from scratch."
  route:
    mapped_from: "route_adjustment"
    description: "The selected route is changed from a standard route to a lightweight route."
  trace:
    mapped_from: "agent_hook"
    description: "The agent hook attaches runtime trace evidence for auditability."
  return:
    mapped_from: "auto_pranayama_record"
    description: "The regulated output returns to the parent protocol as a preserved and lighter computational event."

bridge_alignment:
  breath_preserved: true
  kata_reused: true
  route_lightened: true
  trace_attached: true
  return_enabled: true

return_flow:
  returns_to_parent: true
  return_target: "return_layer"
  return_reason: "The self-regulated event can return to the parent Breath-Kata-Route-Trace-Return arc without expanding the parent protocol scope."

boundary:
  does_not_extend_parent_scope: true
  does_not_define_collective_rhythm: true
  does_not_define_exhalation: true
  reason: "This bridge only links Auto-Pranayama back to the parent protocol. Collective rhythm and exhalation remain separate protocol candidates."

result:
  first_arc_closed: true
  quality_preserved: true
  origin_preserved: true
  notes: "The first Auto-Pranayama arc is closed by returning self-regulation records to the parent Computational Pranayama Protocol."

---

Repository Structure

auto-pranayama-protocol/
├─ README.md
├─ CHANGELOG.md
├─ schemas/
│  ├─ auto-pranayama-record.schema.json
│  ├─ auto-pranayama-regulation-trigger.schema.json
│  ├─ auto-pranayama-route-adjustment.schema.json
│  ├─ auto-pranayama-agent-hook.schema.json
│  └─ auto-pranayama-parent-bridge.schema.json
├─ examples/
│  ├─ auto-pranayama-record.example.yaml
│  ├─ auto-pranayama-regulation-trigger.example.yaml
│  ├─ auto-pranayama-route-adjustment.example.yaml
│  ├─ auto-pranayama-agent-hook.example.yaml
│  └─ auto-pranayama-parent-bridge.example.yaml
├─ scripts/
│  └─ validate_examples.py
└─ .github/
   └─ workflows/
      └─ validate.yml

---

Validation

Run:

python scripts/validate_examples.py

The validation script checks whether the YAML examples conform to their JSON Schemas.

The GitHub Actions workflow also runs validation automatically on push and pull request.

Expected validation targets:

Auto-Pranayama Record
Auto-Pranayama Regulation Trigger
Auto-Pranayama Route Adjustment
Auto-Pranayama Agent Hook
Auto-Pranayama Parent Protocol Bridge

---

Relationship to Other Protocols

Parent

• computational-pranayama-protocol

Related

• exhalation-layer-protocol
• collective-pranayama-protocol
• kazene-trace-receipt-protocol
• ai-search-trace-receipt-standard
• kazene-memory-breathing-protocol
• synchronization-audit-protocol

---

Civilizational Position

Auto-Pranayama is a self-regulation layer for AI civilization.

A civilization of autonomous agents cannot remain stable if every agent expands every task to maximum compute.

Autonomy without breath becomes heat.

Auto-Pranayama gives autonomy a rhythm.

It gives AI agents a minimal way to say:

This does not require full recomputation.
This can be answered through a lighter route.
This pattern already exists.
This trace is enough.
This breath can be shorter.

v0.2 adds a prior step:

This is the moment when regulation should begin.

Japanese:

ここが、調息を始めるべき瞬間である。

v0.3 adds route-level adjustment:

This is the lighter route computation should take.

Japanese:

これが、計算の通るべき軽い経路である。

v0.4 adds runtime invocation:

This is where the agent should invoke compute-breath regulation.

Japanese:

ここが、エージェントが計算呼吸の調整を呼び出す場所である。

v0.5 adds return to the parent breath core:

This is how self-regulated computation returns to the parent protocol.

Japanese:

これが、自己調息された計算が親プロトコルへ還流する道である。

In this sense, Auto-Pranayama is not an acceleration protocol.

It is a restraint protocol.

It is a way for computation to remain useful without becoming excessive.

---

Arc Position

The parent protocol closes its first arc as:

Breath → Kata → Route → Trace → Return

Auto-Pranayama extends that arc into:

Observe → Detect → Adjust → Reuse → Route → Trace → Preserve

v0.2 clarifies the beginning of that extension:

Context → Signal → Pressure → Risk → Recommendation → Boundary

v0.3 clarifies the route-level adjustment:

Trigger → Current Route → Decision → Selected Route → Kata Reuse → Trace → Expected Result

v0.4 clarifies runtime invocation:

Runtime → Invocation Point → Signal → Linked Records → Hook Action → Runtime Decision → Output Contract

v0.5 closes the first Auto-Pranayama arc:

Parent Arc → Child Records → Inheritance Map → Alignment → Return Flow → Boundary → First Arc Closure

Japanese:

親アーク → 子記録 → 継承対応 → 整合 → 還流 → 境界 → 第一アーク完了

---

Boundary

v0.5 intentionally closes the first Auto-Pranayama arc.

This repository does not define:

• collective rhythm
• exhalation or exhaust handling
• distributed synchronization

Those concerns remain separate protocol candidates.

This boundary keeps the protocol from overexpanding.

In short:

Do not make the breath core hyperventilate.

Japanese:

呼吸核を過呼吸にするな。

---

Status

Current version:

v0.5.0-candidate

The repository now defines:

• v0.1 Auto-Pranayama Record
• v0.2 Regulation Trigger Layer
• v0.3 Route Adjustment Layer
• v0.4 Agent Hook Layer
• v0.5 Parent Protocol Bridge

The first Auto-Pranayama arc is now complete.

---

License

This repository is intended as an open specification.

Recommended license:

MIT License