本文へ移動
cccskills
無料GitHub で公開

agent-conversation-protocols

Designs and audits closed, ordered, replay-safe message protocols between already admitted agent bodies. Use when a multi-agent exchange needs explicit epochs, audiences, per-sender sequencing, gathers, terminal fences, acknowledgements, and deterministic replay. NOT for selecting or spawning agents, granting identity or authority, choosing a runtime topology, general human facilitation, wire transport implementation, or deciding whether evidence is true.

インストール方法を見る

含まれるファイル(15)

  • SKILL.md9.1 KB
  • CHANGELOG.md1.1 KB
  • diagrams/01_terminal-fence.md427 B
  • diagrams/02_replay-order.md517 B
  • diagrams/INDEX.md240 B
  • examples/valid-trace.json3.4 KB
  • references/envelope-and-ordering.md917 B
  • references/fipa-a2a-mcp-boundary.md1.9 KB
  • references/gathers-and-terminals.md984 B
  • references/INDEX.md526 B
  • references/pattern-selection-boundary.md1.9 KB
  • schemas/conversation-trace-v2.schema.json4.9 KB
  • scripts/test-bundle.mjs2.3 KB
  • scripts/validate-conversation-trace.mjs8.9 KB
  • tests/activation.md793 B

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Agent Conversation Protocols

Source and implementation boundary

The local trace contract borrows vocabulary from external interaction protocols only where the referenced source establishes it. It is not a wire-conformance claim, and trace identifiers are neither business-operation identifiers nor identity or effect authority. See references/fipa-a2a-mcp-boundary.md.

Design a finite protocol over identities and authority supplied by other systems. A valid trace says which messages were accepted and why. It never creates a principal, body, lease, capability, truth verdict, or permission to cause an effect.

Use this skill when

  • already admitted bodies need a request/response, critique, handoff, or gather protocol;
  • duplicate, stale, reordered, missing, or replayed messages must fail deterministically;
  • a fixed participant set must converge to a closed terminal;
  • a transcript must replay to the same semantic state;
  • a terminal fence and explicit acknowledgement set are required.

Do not use it for

  • choosing, spawning, admitting, resurrecting, or paying agents;
  • deciding star, mesh, tree, or runtime placement;
  • ordinary human discourse or conflict mediation;
  • serialization, queues, sockets, or transport deployment;
  • adjudicating claim truth or granting action authority.

Protocol design procedure

flowchart TD
  A[Receive admitted participant snapshot] --> B[Bind protocol, conversation, epoch, scope, and bounds]
  B --> C[Declare closed message types and fixed gather membership]
  C --> D[Accept only current audience-bound envelopes]
  D --> E{Sequence state}
  E -->|duplicate| F[Record duplicate; no reducer transition]
  E -->|gap or reorder| G[Block sender stream pending exact sequence]
  E -->|next exact sequence| H[Verify digest, expiry, generation, and scope]
  H --> I[Apply deterministic reducer]
  I --> J{Terminal predicate?}
  J -->|No| D
  J -->|Yes| K[Append terminal fence]
  K --> L[Collect fixed acknowledgement set]
  L --> M[Close with truthEffect NONE and authorityEffect NONE]
  1. Bind the snapshot. Record exact participant principals, body generations, audience labels, protocol epoch, scope, deadline, and message ceiling. This skill does not admit them.
  2. Close the language. Enumerate message kinds and payload types. Unknown kinds fail closed.
  3. Sequence per sender. Sequence numbers start at 1 and are contiguous inside one epoch and body generation. A duplicate is idempotent only when its full envelope digest matches; otherwise it is equivocation.
  4. Filter before reduce. Reject stale epoch/generation, wrong scope/audience, invalid verification receipt, expired message, and unresolved causation before changing semantic state.
  5. Freeze gathers. Membership and reducer are fixed when the gather opens. Silence, timeout, majority, and acknowledgement do not imply assent, truth, or authority.
  6. Fence termination. Exactly one terminal fence closes semantic mutation. Required participants acknowledge the fence; an acknowledgement means receipt only.
  7. Replay semantically. Replaying the accepted envelopes in canonical order must reproduce the state digest and terminal record.

Non-authority invariants

  • Messages transport assertions; they do not mint identity, body continuity, admission, leases, capabilities, or permissions.
  • ACK proves only protocol receipt by the named participant generation.
  • GATHER_RESULT is a deterministic reduction of declared inputs, not a truth verdict.
  • A protocol terminal does not prove an external effect stopped, succeeded, failed, or settled.
  • Every terminal has truthEffect: NONE and authorityEffect: NONE.

Ordering and replay rules

ConditionRequired disposition
exact next sender sequencevalidate, then reduce
byte-identical duplicaterecord duplicate, no second reduction
same identity/sequence with different digestE_EQUIVOCATION
sequence gap or later message firstE_SEQUENCE_GAP
stale epoch or body generationE_STALE_EPOCH / E_STALE_GENERATION
wrong audience or scopeE_AUDIENCE_SCOPE
unknown causation parentE_CAUSATION_GAP
message after terminal fenceE_POST_TERMINAL_MESSAGE
terminal missing required acknowledgementsvalid BLOCKED, never complete

Gather rules

  • membership is a sorted, nonempty list of participant refs and cannot change.
  • ALL waits for one accepted contribution from every member.
  • QUORUM requires an explicit integer threshold and preserves every dissenting contribution.
  • FIRST_SUCCESS is permitted only for inert candidate production and must use a deterministic tie-break; it cannot decide truth or authority.
  • Timeout yields an explicit partial/blocked result. It never fabricates missing input.
  • Reducers consume typed message IDs and return a digest-bound semantic value plus unresolved members.

Terminals

COMPLETED, BLOCKED, TIMED_OUT, CANCELLED, and DISSENT_RECORDED are the only terminals. COMPLETED means only that the protocol's declared semantic predicate was met. Any external lifecycle or effect claim requires independent evidence.

Anti-patterns

Ack means done

Wrong: treat delivery acknowledgement as task completion or stopped process.

Right: receipt, semantic closure, lifecycle state, and external effect state remain separate.

Majority mints truth

Wrong: collapse a quorum result into fact or permission.

Right: preserve inputs and dissent; pass the candidate to an independent evaluator or authority.

Dynamic gather membership

Wrong: add a friendly voter after seeing early results.

Right: freeze membership and epoch before contributions.

Replay by timestamp

Wrong: trust wall-clock order across senders.

Right: verify per-sender sequence and explicit causation, then use a canonical tie-break only for concurrent events.

Validator coverage boundary

The bundled validator checks the local JSON shape, closed enums, participant and gather joins, supplied message ordering, causation references, acknowledgement sets, and digest format. It does not verify cryptographic receipts, payload bytes, reducer output, delivery, storage, identity, canonical concurrency order, runtime liveness, or external effects. A valid stateDigest is format-checked input, not a recomputed semantic result. See the terminal-fence sequence and the supplied-sequence boundary.

Output contract

Emit a JSON trace conforming to schemas/conversation-trace-v2.schema.json. Validate it with:

node skills/agent-conversation-protocols/scripts/validate-conversation-trace.mjs \
  skills/agent-conversation-protocols/examples/valid-trace.json

node skills/agent-conversation-protocols/scripts/test-bundle.mjs

Load only when needed

Truth posture

This skill statically validates supplied traces. It does not prove transport delivery, durable storage, independent identity, runtime liveness, containment, or external effects.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Expert in 2000s-era music visualization (Milkdrop, AVS, Geiss) and modern WebGL implementations. Specializes in Butterchurn integration, Web Audio API AnalyserNode FFT data, GLSL shaders for audio-reactive visuals, and psychedelic generative art. Activate on "Milkdrop", "music visualization", "WebGL visualizer", "Butterchurn", "audio reactive", "FFT visualization", "spectrum analyzer". NOT for simple bar charts/waveforms (use basic canvas), video editing, or non-audio visuals.

日本語の概要は準備中です。原文の説明を表示しています。

curiositech/port-daddy22026年10月8日 更新

Expert legal research agent for finding and scraping expungement data state by state. Knows authoritative sources, URL patterns, Firecrawl configuration, and 2026 legal landscape.

日本語の概要は準備中です。原文の説明を表示しています。

curiositech/port-daddy22026年10月8日 更新

Expert in 3D computer vision labeling tools, workflows, and AI-assisted annotation for LiDAR, point clouds, and sensor fusion. Covers SAM4D/Point-SAM, human-in-the-loop architectures, and vertical-specific training strategies. Activate on '3D labeling', 'point cloud annotation', 'LiDAR labeling', 'SAM 3D', 'SAM4D', 'sensor fusion annotation', '3D bounding box', 'semantic segmentation point cloud'. NOT for 2D image labeling (use clip-aware-embeddings), general ML training (use ml-engineer), video annotation without 3D (use computer-vision-pipeline), or VLM prompt engineering (use prompt-engineer).

日本語の概要は準備中です。原文の説明を表示しています。

curiositech/port-daddy22026年10月8日 更新

Apply crisis decision-making research to agent routing, uncertainty triage, and coordination failure analysis in time-pressured systems. Use when diagnosing handoff failures, analytical paralysis, or expert judgment under incomplete information. NOT for routine coding, simple CRUD design, or static single-agent tasks with complete information.

日本語の概要は準備中です。原文の説明を表示しています。

curiositech/port-daddy22026年10月8日 更新

Use for insight, reframing, contradiction, impasse, and anomaly-driven problem solving when execution effort no longer helps. NOT for routine optimization, error correction, or well-specified tasks with known solution paths.

日本語の概要は準備中です。原文の説明を表示しています。

curiositech/port-daddy22026年10月8日 更新

Apply cognitive task analysis to expert work that depends on perceptual cues, branching judgment, and recurring monitoring loops. Use when decomposing expert capability into agent structure, simulation design, or validation interviews. NOT for ordinary step-by-step SOP capture or simple pipelines with no tacit cue layer.

日本語の概要は準備中です。原文の説明を表示しています。

curiositech/port-daddy22026年10月8日 更新

curiositech のスキルをすべて見る

このスキルの問題を報告する