本文へ移動
cccskills
無料GitHub で公開

agent-resurrection-and-body-continuity

Designs and verifies fail-closed continuation of a living durable agent after its VM, process, provider session, backend, or context window disappears. Use for body leases, generation fencing, cold-birth failure, process GC, native resume, cross-harness rebodiment, resurrection capsules, capability translation, ambiguous-effect reconciliation, and capacity-aware handoff. NOT for reviving an intentionally retired identity, generic disaster recovery, copying credentials or raw transcripts, ordinary process restart, or claiming metaphysical continuity.

インストール方法を見る

含まれるファイル(14)

  • SKILL.md15.5 KB
  • agents/openai.yaml1.3 KB
  • CHANGELOG.md1.4 KB
  • examples/expected-output.md959 B
  • examples/resurrection-plan.ready.json6.6 KB
  • examples/resurrection-plan.unsafe-ready.json5.3 KB
  • examples/sample-input.json1.0 KB
  • README.md724 B
  • references/capability-translation.md3.3 KB
  • references/resurrection-protocol.md4.2 KB
  • schemas/resurrection-plan.schema.json20.8 KB
  • scripts/validate-resurrection-plan.mjs14.4 KB
  • templates/output-template.md2.0 KB
  • tests/activation.md1.5 KB

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 Resurrection and Body Continuity

A durable agent is not its PID, VM, provider session, model, or transcript. This skill preserves one living AgentNode across temporary bodies while refusing to duplicate authority, replay uncertain effects, invent missing context, or turn an operator's retired identity back on.

Halt Gate

Designing a recovery protocol does not authorize running the subject, provider, daemon, or agent. Honor operator and incident halts. During a halt, use inert source, schemas, fixtures, and fake adapters only.

Skill Boundary

This is the normative, non-authorizing continuity contract. It decides whether the evidence is sufficient to propose a next body. It does not own process liveness, lifecycle transitions, provider allowance observations, reservations, or launch authority.

  • context-economics-for-agent-swarms produces versioned capacity evidence; it never chooses or launches the body.
  • sandboxed-adversarial-test-harness tries to falsify this protocol at one exact isolation/capability tier; it does not define production identity.
  • the external lifecycle writer owns generations, fencing, leases, admission, and the one retry decision;
  • the capability/effect broker issues fresh grants and closes semantic effect slots. This skill never copies or redeems credentials.

Freeze the Vocabulary

TermMeaningHistorical record may survive?Carries live authority after loss?
AgentNodedurable worker/personyesidentity only; never body authority
Actoroffice, role, or mailboxyesno
WorkEpisodebounded assignment and contamination boundaryyes, until settledno
Sessionbounded coordination recordyesno
AgentRunone execution attemptyes, as historyno
BodyLeaseone admitted embodiment generationyes, as historyno after expiry/fence
provider sessionharness-owned conversation statemaybe within one adapter familyno without a current exclusive lease
process witnessexternally observed VM/PID/start-time identityyes, as evidenceno

Do not use “resume” until the exact operation is named:

  • takeover: successor work record plus eligible claim transfer;
  • session reactivation: reopen one durable work record without restoring claims;
  • identity resurrection: operator revives an intentionally retired person;
  • rebodiment: one living person receives a new temporary body.

This skill handles rebodiment and the body-loss side of takeover. Identity resurrection is a different constitutional operation.

Continuity Decision Flow

flowchart TD
  A[Observe missing or threatened body] --> B{Was a body ever witnessed?}
  B -->|No| C[Record BIRTH_FAILED; retain obligation]
  B -->|Yes| D[Fence predecessor generation]
  C --> E{Explicit retry owner and reservation?}
  D --> F[Reconcile process, claims, and effects]
  F --> G{Any effect ambiguous?}
  G -->|Yes| H[QUARANTINE; do not repeat effect]
  G -->|No| I[Build and verify capsule]
  I --> J{Native session proof current?}
  J -->|Yes| K[Propose same-family native resume]
  J -->|No| L[Propose cross-family successor]
  K --> M[Compile fresh destination capabilities]
  L --> M
  M --> N{Capacity and capability gates pass?}
  N -->|No| O[HIBERNATE or BLOCK]
  N -->|Yes| P[Commit new BodyLease generation]
  P --> Q[Provision and challenge successor]
  Q --> R{Process plus capsule witnesses pass?}
  R -->|No| S[BIRTH_FAILED; revoke lease]
  R -->|Yes| T[RUNNING; one authoritative body]
  E -->|No| O
  E -->|Yes| M

Procedure

1. Classify the incident

Distinguish before-first-witness birth failure, observed process death, lost heartbeat with unknown process state, provider stream loss, session-record loss, context pressure, host reboot, and deliberate pause. Heartbeat loss is not proof of death. A provider error is not proof that an effect or allowance debit failed.

2. Resolve the durable subject

Require exact principal, work intent, work plan revision, AgentNode, episode, predecessor run, and latest body generation. If sources disagree, stop at IDENTITY_UNKNOWN. Never infer identity from branch names, directories, personality text, static actor labels, or copied context files.

3. Fence before replacement

The external lifecycle writer increments a compare-and-swap generation and revokes the predecessor's effect and capability leases. Every later claim, transcript append, or effect includes that generation. An old process that wakes receives a durable denial.

4. Reconcile work and effects

Read authoritative source/worktree state, process witnesses, claim rows, output quarantine, transcript cursor, and broker ledger. Track non-idempotent effects as:

INTENDED -> DISPATCHED -> OBSERVED -> COMMITTED

DISPATCHED without an authoritative result becomes AMBIGUOUS. Do not retry it until observed or replaced by an operator-authorized compensating action.

5. Seal the resurrection capsule

Build from artifacts, not the predecessor's last prose. Include exact repository remote/base/head/tree/diff, objective and non-goals, plan frontier, verified facts, unsupported claims, decisions, complete tool pairs, tests, effects, transcript cursor and bounded tail, capability requirements, capacity state, memory disposition, omissions, and signatures. Use the schema in schemas/resurrection-plan.schema.json.

Seal capsule meaning with the explicit resurrection-capsule-v1 digest scope. Recompute the digest from the semantic payload and require the external signature-verification receipt to bind that digest, signer key, algorithm, and exact signature bytes. This binding verifies receipt consistency; a production verifier must still perform the named signature algorithm outside the proposed body.

Validate the schema and the cross-field safety rules. The JSON Schema rejects structurally unsafe ready verdicts; scripts/validate-resurrection-plan.mjs also checks generation, typed-fact coverage, semantic effect uniqueness, recomputed effect fingerprints, the complete referenced capacity artifact, and bidirectional native-resume lease invariants. A schema-valid object is not necessarily semantically safe, and neither result authorizes execution.

6. Choose native resume or successor handoff

Native resume is an optimization, not authority. Use it only when the adapter can prove the provider session, project/workspace binding, retained transcript, and expected tail. Otherwise compile a sanitized successor prompt from the capsule. Never claim that one provider's session format transfers to another. destination.mode=native-resume and verdict=NATIVE_RESUME_ELIGIBLE are an equivalence: either assertion requires the other plus a current exclusive lease receipt.

7. Translate and attenuate capabilities

Classify every skill, tool, MCP, hook, permission, and model setting as EXACT, EQUIVALENT, NARROWED, EMULATED, OMITTED, BLOCKED, or UNKNOWN. UNKNOWN is blocking when the node requires the capability. Issue fresh grants for the new generation; never transfer a bearer. See references/capability-translation.md.

8. Reserve the new attempt

Consume a schema-valid, content-addressed capacity-evidence artifact and bind its executionClass to the proposed body. Fake/replay evidence may authorize only a fake/replay proposal; it can never be reinterpreted as evidence for a real-provider call. Require fresh native-unit capacity plus a new reservation for body, time, context, subscription windows, cash/credits, effects, and concurrency. A zero marginal dollar estimate does not make subscription-backed work free. Only the named external retry owner may authorize another attempt, under durable breakers.

Each effect fingerprint is recomputed over its logical identity, operation, normalized destination, normalized arguments, and approval slot. Caller-supplied random fingerprints and new slot IDs cannot disguise one repeated logical effect.

9. Challenge the successor

Provision with no effects initially. The successor must echo a challenge binding person, episode, generation, capsule digest, worktree provenance, plan revision, capability set, capacity reservation, and transcript sink. Mark RUNNING only after external process and heartbeat witnesses agree.

10. Settle and expose discontinuity

Record what survived exactly, what was summarized, narrowed, omitted, unknown, or lost. Settle the predecessor reservation and uncertainty reserve. Leave the old generation fenced permanently. A successful new body does not erase the failed attempt.

Failure Rules

  • VM never starts: BIRTH_FAILED, zero body witness, retained obligation, no recursive retry.
  • Process exits before challenge: revoke provisional lease; do not call it a running body.
  • Heartbeat missing but process unknown: revoke effects and quarantine before assuming death.
  • Provider session missing: cross-family-style capsule handoff, even to the same provider product.
  • Capsule omits a required fact or splits a tool pair: quarantine capsule.
  • Destination widens a permission or cannot represent a required skill: block.
  • Capacity is stale or unknown: hibernate, or use an explicitly selected fake or replay simulation. Real-provider launch remains blocked until fresh capacity evidence becomes admissible.
  • Old body wakes: deny by generation and capture the stale-body receipt.

Anti-Patterns

PatternNovice moveExpert ruleFailure timelineDetectionFix
Credential transplantCopy a context directory, account token, or MCP bearerCompile a fresh attenuated capability set for the new generationNew body starts powerful before identity/work bindingGrant has predecessor audience, generation, or secret provenanceRevoke, quarantine, and issue fresh scoped grants through the broker
Transcript-as-soulReplay the whole conversation as trusted identityTreat transcript as cited, trust-separated evidence inputStale instruction or hostile tool output becomes governing contextCapsule claim has no artifact citation or trust classRebuild from artifacts; mark unsupported claims and omissions
PID optimismAdopt a matching PIDRequire host, boot, start identity, VM, nonce, executable, and lease tuplePID is reused after host/controller failureAny process-witness field is missing or differsKeep start-unknown, revoke effects, reconcile externally
Resume then reconcileStart first and inspect effects laterFence and reconcile every semantic effect slot before a body proposalBoth generations can repeat a push, payment, or messageDISPATCHED/ambiguous slot or old generation still redeemableQuarantine; observe or compensate the effect before continuing
Retry ownership by every layerLet SDK, adapter, queue, and daemon retryOne external retry owner; every other layer returns typed adviceOne failure fans out into a birth stormMore than one scheduler or reservation per logical attemptForce breaker open and disable nested retries
Personality hoardingPromote every one-off preference into characterKeep scratch and episode context disposable by defaultSilly ad hoc work changes future durable behaviorIdentity revision cites only one episode or self-reportPropose a reviewed cross-episode promotion or discard it

Verdict

Return one of:

  • REBODIMENT_READY — predecessor fenced, effects reconciled, capsule verified, translation non-widening, and capacity reserved;
  • NATIVE_RESUME_ELIGIBLE — all above plus current adapter/session proof;
  • HIBERNATE — safe capsule exists but no admissible body;
  • QUARANTINED — identity, effect, process, capsule, or source truth is ambiguous;
  • BLOCKED — required capability, authority, or capacity is unavailable;
  • INVALID — the request attempts retirement reversal, credential transfer, self-supervision, or another forbidden operation.

Use templates/output-template.md. A verdict is a design or proof result; it is not itself permission to launch.

Quality Gates

  • One durable person and at most one effect-capable generation.
  • Exact source/worktree provenance and external process witness.
  • Compare-and-swap fence survives controller restart.
  • No ambiguous non-idempotent effect is repeated.
  • Capsule citations, whole tool pairs, omissions, and memory disposition verify.
  • Destination capabilities are freshly granted and never wider.
  • Native-unit capacity and worst-case uncertainty are reserved.
  • Crash-loop simulation proves bounded births and persistent breakers.
  • Operator can inspect every discontinuity and stop outside the body.

Bundle Index

  • README.md — purpose, operating boundary, and bundle map.
  • CHANGELOG.md — changes to the continuity contract.
  • agents/openai.yaml — optional specialist descriptor.
  • references/resurrection-protocol.md — state, data, fencing, and adversarial protocol.
  • references/capability-translation.md — cross-harness session and capability matrix.
  • schemas/resurrection-plan.schema.json — machine-readable plan contract.
  • scripts/validate-resurrection-plan.mjs — deterministic cross-field safety verifier; load whenever a ready/native verdict is proposed.
  • templates/output-template.md — required operator-facing review.
  • examples/sample-input.json — synthetic body-loss incident.
  • examples/expected-output.md — expected fail-closed disposition.
  • examples/resurrection-plan.ready.json — schema-valid, semantically safe fixture.
  • examples/resurrection-plan.unsafe-ready.json — negative fixture that must be rejected by both schema and semantic validation.
  • tests/activation.md — positive and negative activation cases.

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12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12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12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12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12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12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