本文へ移動
cccskills
無料GitHub で公開

drydock-program-architecture

Designs, reconciles, and reviews an end-to-end Drydock program for safely staging agent runtimes: immutable worktree provenance, disposable VM containment, typed effect brokerage, global body accounting, resurrection, subscription-capacity economics, context compaction, operator control, evidence, and promotion. Use when several Drydock contracts must compose into one architecture, implementation hypertree, diagram set, or release plan. NOT for launching Port Daddy, running an untrusted subject, replacing focused containment/resurrection/capacity audits, or treating a design packet as runtime authority.

インストール方法を見る

含まれるファイル(25)

  • SKILL.md19.3 KB
  • agents/openai.yaml545 B
  • CHANGELOG.md1.5 KB
  • examples/drydock-resurrection-hypertree.json33.7 KB
  • examples/expected-output.md1.0 KB
  • examples/hypertree-execution.review-loop.json19.6 KB
  • examples/INDEX.md487 B
  • README.md957 B
  • references/agent-lifecycle-and-operator-control.md49.1 KB
  • references/architecture-decisions.md8.1 KB
  • references/controlled-agent-simulation.md107.0 KB
  • references/delivery-and-proof.md7.9 KB
  • references/diagram-atlas.md27.8 KB
  • references/hypertree-execution-observatory.md20.6 KB
  • references/INDEX.md1.0 KB
  • references/knowledge-map.md6.1 KB
  • references/resurrection-capacity-and-context-control.md79.5 KB
  • schemas/drydock-resurrection-hypertree.schema.json6.3 KB
  • schemas/hypertree-execution.schema.json22.8 KB
  • scripts/audit-drydock-program-skill.mjs6.0 KB
  • scripts/INDEX.md568 B
  • scripts/validate-drydock-resurrection-hypertree.mjs5.5 KB
  • scripts/validate-hypertree-execution.mjs22.7 KB
  • templates/architecture-packet.md700 B
  • tests/activation.md1.2 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。

Drydock Program Architecture

Drydock is the proving ground between “we wrote a safety mechanism” and “this agent runtime may touch a machine, network, credential, repository, provider, or budget.” Use this skill to join the focused contracts without erasing their separate authorities.

Halt Gate

A design, schema, fixture, or diagram is not a run lease. If the operator has halted Port Daddy or any subject runtime, remain at static tier T0:

  • inspect source and immutable evidence;
  • edit documents, schemas, validators, fixtures, and inert UI artifacts;
  • run only ordinary source validators that cannot start the subject;
  • do not invoke the halted CLI, daemon, apps, hooks, agents, providers, or MCP;
  • label dynamic gates BLOCKED or NOT_PROVISIONED, never PASS.

Stop immediately if the requested work requires a dynamic observation that the halt forbids. Preserve the missing proof as a named gate.

What This Skill Owns

This skill owns composition. It answers:

  1. Which component owns each authority?
  2. Which focused contract is normative for each decision?
  3. How do identity, bodies, capacity, context, effects, evidence, and operator control meet without circular trust?
  4. In what order may the implementation be built and promoted?
  5. Which diagrams and receipts let another engineer audit the answer?

It does not duplicate or overrule the focused skills. Load them at the decision boundaries below.

DecisionNormative skillRequired artifact
Attempt admission, reservation, ambiguity, settlement handoffconserved-capacity-admission-and-settlementConserved attempt record with terminal holds
Authority-filtered context and backend translationtrust-typed-context-compilerTrust-typed Context IR and continuation capsule
Deterministic fake-system evaluation and replaytrial-basin-deterministic-systems-evaluationSealed evaluation envelope and typed result vector
Host/guest containment, brokered I/O, hostile testssandboxed-adversarial-test-harnessDrydock review and tier verdict
Missing, dead, resumed, or cross-backend bodyagent-resurrection-and-body-continuitySchema-valid resurrection plan
Subscription allowance, context pressure, compactioncontext-economics-for-agent-swarmsSchema-valid capacity evidence
Bonds, bounties, slashing, custody, or paymentcryptoeconomic-protocol-securityOpen threat assessment with legal and evidence holds
Operator-visible proofagent-visual-evidence-manifestEvidence manifest with zoom paths

Stable Vocabulary

  • AgentNode: durable actor identity and obligations. Not a provider session.
  • Run: one admitted attempt against one bounded work node.
  • Body generation: replaceable VM/process/backend embodiment of an AgentNode.
  • Witness: host-observed process/VM identity, including generation and fence.
  • Capsule: verified plan, evidence, context, effects, and capability inputs for continuation. It is not a credential bundle.
  • CapacityVector: native provider allowance windows plus confidence and reset horizons. It is not a fabricated dollar total.
  • Effect ticket: one attenuated, typed, lease-bound operation. It is not a raw secret or general network socket.
  • Receipt: append-only claim from a named witness class. A signature proves authorship, not the truth of the signed claim.
  • Promotion: a new authority grant for one exact digest and tier. It is not a property inherited from a similar build.

Architecture Invariants

Every packet must preserve all of these:

  1. The canonical checkout is a clean, read-only projection of live origin/main; work occurs only in fresh linked worktrees or guest-local copies derived from an exact commit.
  2. At most one body generation can hold effect authority for an AgentNode.
  3. One external admission writer reserves global and scoped capacity before any asynchronous launch.
  4. The subject cannot reach host mounts, ambient sockets, credentials, network, providers, or GitHub except through enumerated typed channels.
  5. Unknown body, effect, capacity, provenance, or teardown state fails closed.
  6. Dollars, subscription native units, compute, context, and concurrency remain separate conserved ledgers; a zero marginal price never means zero scarcity.
  7. A successor gets newly compiled capabilities and a verified capsule, never a copied credential, hidden context directory, or inherited PID identity.
  8. Every dynamic claim names its external witness, exact artifact digests, and replay or inspection path.
  9. Operator Stop, fence, and spend denial remain outside the guest and provider session.
  10. No single agent, daemon, UI, or guest both requests, authorizes, executes, and certifies the same consequential transition.

Composition Procedure

1. Freeze the exact proposition

Write one falsifiable sentence naming subject digests, adversary, tier, bounded effects, native capacity, lifecycle outcome, and external witnesses. Replace “safe,” “contained,” and “resumable” with observable limits.

Record the current operator halt, current repository commit, target platform, and every fact that is unavailable. Use the state labels SHIPPED, PARTIAL, TARGET, DEFERRED, REJECTED, and UNKNOWN; never blur planned joins into implemented joins.

2. Build the authority ledger

For every component, list:

  • authority it may hold;
  • authority it must never hold;
  • durable records it writes;
  • external facts it observes;
  • process and release boundary;
  • failure mode and fail-closed terminal state.

At minimum include the operator surface, command verifier, controller, lifecycle ledger, capacity broker, effect broker, cognition compiler, VM adapter, reaper, guest driver, subject, body, and evidence projection. Load references/architecture-decisions.md when assigning components or languages.

3. Seal provenance before execution design

Define exact identities for source commit/tree, worktree derivation, guest image, controller build, test bundle, scenario, policy, price/capacity observations, tools, skills, prompts, and output quarantine. The host controller must reject:

  • canonical/main checkout paths;
  • dirty or wrong-remote sources;
  • mutable branch names without resolved commit/tree identities;
  • unsealed setup scripts, package hooks, test configs, or transforms;
  • output that attempts to overwrite input or bypass review worktrees.

4. Compose the focused contracts

Run the contracts in this order:

  1. Admission and conserved capacity: prove there is one bounded reservation for this attempt or deny launch. Unknown provider allowance stays scarce.
  2. Resurrection: establish AgentNode, predecessor generation, fence, capsule, and allowed successor shape.
  3. Context: compile authority-filtered evidence, obligations, omissions, and backend translations without admitting a successor.
  4. Containment: prove the exact body can execute only at the approved tier.
  5. Evaluation: seal schedule, faults, witnesses, oracles, and negative controls before interpreting a deterministic fake or replay.
  6. Mechanism security: when settlement is enabled, analyze custody, liquidity, collusion, Sybil, oracle, ordering, dispute, and externality risk. A no-settlement profile skips economic effects; it does not make usage free.

If any focused contract returns UNKNOWN, BLOCKED, QUARANTINED, INCOMPLETE, or UNCERTAIN, the integrated packet cannot say ready. Preserve the original verdict and explain which downstream nodes remain unreachable.

5. Design cognition without making it authority

Partition cognition into narrow compilers:

  • Intent shaper: turns operator language into a proposed WorkIntent.
  • Architect: grows a versioned hypertree and keeps fog nodes non-executable.
  • Context steward: compacts evidence and prepares a cited capsule.
  • Promptwright: translates trusted and untrusted context for one backend.
  • Capability compiler: emits EXACT, EQUIVALENT, NARROWED, OMITTED, or BLOCKED mappings for tools, MCPs, skills, hooks, prompts, and permissions.

None of these may launch, reserve, expose credentials, or declare its own output accepted. Ad hoc work defaults ephemeral and isolated; durability is an explicit promotion based on recurring obligation, continuity value, or operator choice.

6. Account globally before launch

The lifecycle ledger must reserve before birth and reconcile before reopening after any controller crash. Include global, repo, project, provider, ancestry, identity, and operator scopes. Initial automatic birth retries and recursive child depth are zero.

Every body has a generation, lease, fencing token, backend session handle, VM/process witness, heartbeat deadline, and teardown state. A PID alone proves nothing. Breakers survive process restart and calendar reset; only a separate authorized reset transition may reopen them.

7. Preserve economic and context truth

Model each scarce thing in its native unit. For subscriptions capture observed remaining percentage/units, window, reset time, auth mode, shared bucket, freshness, parser provenance, confidence, reserve, burn forecast, and unresolved attempt hold. Keep a separate cash/credit ledger.

Before an atomic action, predict whether the action plus checkpoint tail fits. If not, stop at a safe boundary, seal and verify a capsule, then compact, hibernate, or rebody. Never split a tool request from its result or summarize an unreconciled side effect as complete.

8. Draw the proof, not decoration

Load references/diagram-atlas.md and select every view needed by the risk:

  • context/trust topology for boundary claims;
  • sequence diagrams for admission, cold birth, stop, and resurrection races;
  • state diagrams for lifecycle, effects, and context;
  • ER diagram for identity and receipt cardinality;
  • flowcharts for provenance, capacity, and capability decisions;
  • implementation DAG for delivery dependencies;
  • operator journey for human control and zoom paths.

Each diagram must state what it proves and what it does not. Diagrams never substitute for schemas, tests, or runtime evidence.

9. Compile an executable hypertree

Start from examples/drydock-resurrection-hypertree.json and validate it with:

node scripts/validate-drydock-resurrection-hypertree.mjs examples/drydock-resurrection-hypertree.json

Every executable node needs immutable inputs, owner role, worker lifetime, skills, outputs, dependencies, acceptance gates, and failure terminal. Fog nodes may gather evidence or questions but may not launch. Every launcher must depend transitively on global spawn breakers and capacity admission.

Keep four topologies explicit instead of forcing every concern into the DAG:

  • a static DAG decides which nodes are eligible;
  • a bounded workflow routes review, rework, escalation, and approval;
  • manager-driven rounds assign workers and independent reviewers;
  • an append-only event stream feeds deterministic operator projections.

Load references/hypertree-execution-observatory.md when specifying node input/output contracts, reviewer tiers, rework bounds, or shared HTML, Swift, and Rust execution views. Validate the inert review-loop fixture with:

node scripts/validate-hypertree-execution.mjs examples/hypertree-execution.review-loop.json

10. Produce operator and reviewer surfaces

Operator views are projections over typed records, never parallel authority. Every aggregate must zoom within two actions to its session, body witness, worktree diff, command/test output, claim, transcript, capacity observation, or receipt. Show honest STARTING, STALE, OFFLINE, UNKNOWN, FENCED, and QUARANTINED states.

Provide one calm edge signal for meaningful new activity, then a stable live state; honor reduced motion. Expose start-one-worker, inspect, pause, stop, cancel, resume, switch backend, capacity, permissions, MCP/skill fitting, artifact, PR, and mobile-join journeys as scripted real-product states.

11. Define the proof ladder and release boundary

Use references/delivery-and-proof.md. Advance only from static contracts to fake deterministic guest, replay guest, separately approved provider canary, fixture worker, crew, and federation. A higher tier requires a new grant. Hosted CI may prove source behavior; it cannot prove local hypervisor, credential custody, provider billing, or operator pixel truth without the matching witness.

Anti-Patterns

Who is fooledTemptationCorrection
Novice“A clean path check protects main.”Make main absent from the guest and reject its inode/path identity outside the subject.
Novice“PID 418 is the same agent.”Resolve AgentNode, body generation, platform handle, and process start witness.
Expert“The ledger caps spend.”Prove provider-side custody or label only protocol authority bounded.
Expert“We can expose 139 tools but hide most schemas.”Compile a phase-scoped capability pack; schema visibility is not authorization.
Timeline“Retry after 500 ms” fixes crash storms.Persist reservation, ancestry, breaker, and reconciliation state before adding backoff.
Timeline“The next model can read the transcript.”Fence the predecessor, reconcile effects, verify a capsule, and recompile capabilities first.
Product“Green CI means Drydock is ready.”Name the exact witness each claim still lacks and keep the tier blocked.

Output Contract

Use templates/architecture-packet.md. A complete packet contains:

  • exact proposition, state labels, and halt status;
  • authority/component ledger and trust topology;
  • immutable provenance manifest and canonical-checkout exclusion proof;
  • capacity evidence and reservation logic;
  • resurrection plan and body/accounting model;
  • context/capability translation report;
  • implementation hypertree and critical path;
  • selected diagram atlas views with proof limits;
  • scenario/fault matrix, witness-labeled receipts, and residual risks;
  • operator journeys and two-action zoom paths;
  • tier verdict and next permitted action.

Quality Gates

Before calling the packet complete:

  1. Validate the skill bundle, self-containment, and Mermaid structure; run node scripts/audit-drydock-program-skill.mjs for the integrated contract.
  2. Validate hypertree JSON against schemas/drydock-resurrection-hypertree.schema.json and the semantic validator.
  3. Validate the execution fixture against schemas/hypertree-execution.schema.json and scripts/validate-hypertree-execution.mjs; require exact plan binding, distinct producer/reviewer/manager identities, complete outputs, bounded rework, and identical HTML/Swift/Rust projections.
  4. Run positive and negative activation cases in tests/activation.md.
  5. Prove every local source link in references/knowledge-map.md exists.
  6. Confirm no canonical checkout or raw credential appears in an executable input.
  7. Confirm every launch node is gated by admission, capacity, and persistent breakers.
  8. Confirm every PASS claim names an exact tier, digest, and external witness.
  9. Have a skeptical reviewer argue safety, liveness, economics, usability, and evidence independently.
  10. Run the focused skill mutation suites before trusting their examples; a composition packet cannot upgrade a failed subordinate validator.

Bundle Index

Load only what the current decision requires:

  • references/INDEX.md — progressive-disclosure routing for the full knowledge corpus.
  • references/knowledge-map.md — corpus map, authority, and exact loading triggers.
  • references/architecture-decisions.md — component/process/language placement.
  • references/diagram-atlas.md — comprehensive decision and proof diagrams.
  • references/delivery-and-proof.md — staged implementation and promotion gates.
  • references/hypertree-execution-observatory.md — typed execution, review ladder, bounded rework, shared projections, and observatory delivery slices.
  • examples/drydock-resurrection-hypertree.json — 30-node machine plan.
  • examples/hypertree-execution.review-loop.json — inert bad-to-good review-loop trace.
  • examples/INDEX.md — example and executable-plan routing.
  • schemas/drydock-resurrection-hypertree.schema.json — structural contract.
  • schemas/hypertree-execution.schema.json — closed execution/event/review contract.
  • scripts/validate-drydock-resurrection-hypertree.mjs — semantic DAG/digest checks.
  • scripts/validate-hypertree-execution.mjs — semantic event, identity, review, rework, and projection checks.
  • scripts/audit-drydock-program-skill.mjs — bundle, links, diagrams, routing, and moved-source audit.
  • scripts/INDEX.md — validator loading and invocation map.
  • templates/architecture-packet.md — final deliverable shape.
  • examples/expected-output.md — concise worked packet.
  • tests/activation.md — positive and negative skill-routing cases.
  • agents/openai.yaml — portable agent-card metadata and default prompt.

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