跳到正文
cccskills
免费在 GitHub 公开

history-recall

Use when a user requests the reasoning behind one matched session by query or handle. Not for peer transfer: use history-sync. Not for store registration: use history-source-registry.

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.md4.3 KB
  • agents/openai.yaml153 B

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

History session brief

Contract

FieldBound contract
TriggerUser explicitly requests the reasoning behind one matched session by query or opaque handle.
AuthorityRead-only. No file, VCS, credential, paid, published, deployed, or remote mutation. Reads the local session index and transcripts only.
Side effectNone.
DoneOne capped typed record reports problem category, decisions, outcome, touched-file handles, lifecycle, spawn edges, and message-count buckets plus an opaque handle; no prose digest or raw text.

Inputs

  • query (required unless handle is supplied): an exact token, error string, function name, flag, or file path: matches strongest; otherwise the question in the user's own words, ranked as a fallback when nothing matches exactly.
  • handle (optional): an opaque session handle returned by a prior brief; when supplied, open that session directly instead of searching.
  • harness (optional): filter the search to one harness.
  • byte_budget (required): the maximum output size in bytes. Defaults to 4096 when the caller does not supply one. The record is trimmed to fit; the cap is paid out of the budget, not added to it.
  • trust_policy (required): path to the trust policy file that governs which imported sessions are withheld or demoted. Defaults to $HISTORY_DIR/trust-policy.json when present.
  • handle_registry (required): path to the handle registry that maps opaque handles to session files. Defaults to $HISTORY_DIR/handle-registry.jsonl when present.

Procedure

  1. Resolve the target session. If handle is supplied, open the session it names directly. Otherwise search the local index with query, trying exact tokens first and falling back to ranked matches when nothing matches exactly; take the single best-matching session.
  2. Upgrade the matched hit to the whole session. A hit carries only the messages that matched the query, so the decision, usually worded nothing like the query: is not among them. Read the whole session before extracting any field.
  3. Apply the trust policy. Keep withheld imported sessions out of the record and do not leak a session the policy excludes. Attach lifecycle state to imported promoted notes and demote rejected ones, so a session a reader rejected does not surface as authoritative.
  4. Extract the typed fields from the whole session: problem category (the user's opening problem statement, noise-filtered), decisions (assistant conclusions worded as choices), outcome (whether the session reports backing out or finishing), touched-file handles (the few files the session worked on most), lifecycle (accepted, rejected, superseded, or stale for imported notes), spawn edges (parent session, agent name, and kind where the harness records the edge, never inferred), and message-count buckets (user and assistant turn counts).
  5. Cap the record to the byte budget. When the record is trimmed, set a capped flag and keep the opaque handle so the caller can re-open the whole session; the cap is paid out of the budget, not added to it. Never emit raw transcript text or a prose digest.
  6. Return one typed record plus the opaque handle.

Failure and recovery

  • No match: return an empty record with the handle absent and a tier of none when no query matched any session, or relevance when matches existed but were filtered by the trust policy; do not fabricate a session.
  • Index unreadable or rebuild required: return a blocked result naming the index state; do not mutate the index.
  • Policy withheld every match: return an empty record with a policy_withheld count; do not bypass the policy.
  • Partial result: the record is atomic; a field that cannot be extracted is omitted, never guessed.
  • Non-mutation: no file, index, or transcript is written; a failed run leaves the store unchanged.

Output

One typed record: handle, harness, project, tier, capped, problem_category, decisions, outcome, touched_files, lifecycle, spawn_edges, message_counts. No prose digest and no raw text. The opaque handle re-opens the whole session on a later call.

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

Use when explaining System V AMD64, ARM AAPCS, RISC-V psABI, stack frames, variadic calls, or FFI register rules. Not for the Rust FFI binding layer: use rust-ffi.

中文简介正在准备中。当前显示原文说明。

OutlineDriven/odin-claude-plugin382026年9月29日 更新

Use when configuring ADC sampling time, DMA-driven ADC, calibration, or DAC channel setup on bare-metal MCUs. Not for the DMA stream itself: use dma-baremetal.

中文简介正在准备中。当前显示原文说明。

OutlineDriven/odin-claude-plugin382026年9月29日 更新

af-xdp

免费

Use when creating AF_XDP sockets, configuring UMEM and XSK rings, writing an XDP redirect program, or choosing copy versus zero-copy mode. Not for full kernel bypass: use dpdk.

中文简介正在准备中。当前显示原文说明。

OutlineDriven/odin-claude-plugin382026年9月29日 更新

Use when a completed session needs an agent-environment retrospective. Not for an engineering retrospective from telemetry: use engineering-retrospective.

中文简介正在准备中。当前显示原文说明。

OutlineDriven/odin-claude-plugin382026年9月29日 更新

Use when a redacted, trimmed agent transcript must be appended to a GitHub PR or issue body, with human approval and preview. Not for automated or model-initiated insertion.

中文简介正在准备中。当前显示原文说明。

OutlineDriven/odin-claude-plugin382026年9月29日 更新

agents-md

免费

Use when a repo needs agent setup, AGENTS.md added or made lean, CLAUDE.md audited, or agent instructions scored or pruned. Not for remote, credential, publish, deploy, or irreversible changes.

中文简介正在准备中。当前显示原文说明。

OutlineDriven/odin-claude-plugin382026年9月29日 更新

查看 OutlineDriven 的全部技能

报告此技能的问题