跳到正文
cccskills
免费在 GitHub 公开

dag-context-bridger

Builds scoped, provenance-preserving context packets between DAG nodes. NOT for execution or aggregation.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(2)

  • SKILL.md5.2 KB
  • references/provenance-context-and-loss-audit.md5.8 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

DAG context bridger

Build a context packet for one downstream task, preserving what it is, where it came from, what authority permits, and what was omitted. Compression saves tokens; it cannot make a source true, authorize disclosure, or repair a missing prerequisite.

1. Bound the packet

State the downstream question, artifact/version requirements, recipient authority, retention/redaction policy, token or byte budget, and non-negotiable obligations such as counterexamples, caveats, approval constraints, and exact quotations. Assign each fragment a stable ID, source URI or artifact digest, time/version, producer activity, responsible agent, disclosure class, and derivation relation. PROV-O supplies useful entity/activity/agent and derivation vocabulary; it does not prove truth or authorization.

flowchart LR
  I[Scoped request and recipient authority] --> F[Filter disclosure and retention]
  F --> S[Select provenance-bound fragments]
  S --> O[Check obligation coverage]
  O --> C[Compress with source links]
  C --> P[Packet, omissions, and loss audit]

2. Retrieve only eligible evidence

Filter authority before ranking. For a hybrid search, retain lexical and dense lists only when dense vectors have compatible spaceId; fuse ranks with RRF only after that gate. RRF's sum 1/(k + rank) is a rank vote, not relevance probability. Include dependency and provenance rankings when they illuminate input lineage, but make their role visible.

flowchart TB
  A[Authorized corpus] --> L[Lexical retrieval]
  A --> D[Dense retrieval with compatible spaceId]
  A --> G[Dependency and provenance candidates]
  L --> R[RRF or recorded selection rule]
  D --> R
  G --> R
  R --> V[Obligation and contradiction coverage]
  V --> K[Context packet with source IDs]

3. Choose a packet mode and preserve evidence through compression

Choose full-forward, selective plus summary, output-only, progressive loading, or hierarchical summaries according to recipient need and authority. Use the manifest procedure in the reference to calculate the recipient tokenizer budget plus response reserve; fixed token thresholds are not portable. Cache keys include artifact/version, recipient scope, model/tokenizer, policy revision, and summary method. Expiry, contradiction, or policy change invalidates the cache and holds work that depends on the invalidated packet.

Use extractive snippets for exact claims, constraints, numbers, and caveats that cannot safely be paraphrased; use generated summaries only with links to their source fragments. A budget is task-specific: do not apply fixed token cutoffs or a generic relevance threshold. Deduplicate only after comparing provenance and version: equal text can carry different authority or time meaning. Record each omitted fragment and whether it was excluded for authority, budget, irrelevance, incompatibility, or uncertainty.

Worked positive case

A reviewer needs a deployment recommendation. The packet includes the signed approval record, the tested artifact digest, a source caveat, and an extract of the failed negative test. A duplicate prose summary is omitted with its source IDs. The output says the packet is decision support, not an approval itself.

Worked negative case

Two fragments say the same thing, but one is from an expired policy revision. Do not deduplicate them as interchangeable. Retain the current fragment, report the stale one and its version, and avoid silently presenting the claim as timeless.

4. Audit and hand off

Before sending, check every stated obligation against included fragment IDs. Return the recipient/scope, source and derivation records, filters, compatible vector-space evidence, summary method, omissions, unresolved contradictions, and a request for additional evidence where coverage fails. Read provenance context and loss audit. Cite W3C PROV-O for provenance concepts and Cormack et al., 2009 for rank fusion. Neither source guarantees safe disclosure, correctness, or adequate context.

Scope and handoff checks

This is a completed-node, single-DAG information handoff. It does not spawn agents, combine final results, alter topology, implement real-time streaming, or establish durable cross-run memory. Route those needs to the executor, result aggregator, graph builder or storage owner. Before release, verify recipient scope, required-obligation coverage, measured packet budget, source version/lineage, omission reasons, contradiction status and recipient readback. The reference gives cache/cycle recovery and a concrete packet worksheet.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

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.

中文简介正在准备中。当前显示原文说明。

curiositech/port-daddy32026年10月19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.

中文简介正在准备中。当前显示原文说明。

curiositech/port-daddy32026年10月19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).

中文简介正在准备中。当前显示原文说明。

curiositech/port-daddy32026年10月19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.

中文简介正在准备中。当前显示原文说明。

curiositech/port-daddy32026年10月19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.

中文简介正在准备中。当前显示原文说明。

curiositech/port-daddy32026年10月19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.

中文简介正在准备中。当前显示原文说明。

curiositech/port-daddy32026年10月19日 更新

查看 curiositech 的全部技能

报告此技能的问题