trace-to-skill-inducer

星标69

分支9

更新时间2026年7月19日 02:22

Use when you have captured session evidence — session-retro / session- observatory-live traces in .planning/patterns/, tool logs, correction records — and want to induce a reusable skill from it. Segments the traces into candidate skill units (an LLM judgment, not a deterministic parse) and decomposes each candidate into a four-part structured spec: workflow structure, execution semantics, and runtime attachments (verification, safety, rollback, state). It emits a spec object, NOT a finished SKILL.md, and hands that spec to skill-forge. It sits between skill-integration (upstream frequency detector) and skill-forge (downstream author). Backed by Agent-Trace-to-Skill Induction (arxiv 2606.06893v1). Triggers on inducing a skill from captured traces, turning a repeated pattern into a skill spec, and preparing evidence for skill-forge.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

Tibsfox

Tibsfox/gsd-skill-creator

打开 GitHub 仓库查看创作者相关仓库

下载

Trace-to-Skill Inducer

Turn captured interaction traces into a structured skill spec the skill-forge loop can author from. Segment the traces into candidate skill units, decompose each candidate into workflow structure + execution semantics + runtime attachments, scrub sensitive data, and hand the spec downstream. This is the induction step of the skill lifecycle on this system: it converts raw session evidence into a design contract, and stops there.

Why

skill-integration frequency-detects that a tool sequence recurs, but a raw recurrence count is not a skill — it has no declared preconditions, verification, rollback, or state model, so authoring straight from it produces under-specified skills that pass validate and then misbehave in skill-counterfactual-audit. The failure this prevents is scope collision: if induction emits a finished SKILL.md, it overlaps skill-forge and two authors fight over the same file. Draw the boundary so induction feeds authoring — spec out, not skill out.

Data classes touched

Session traces from .planning/patterns/ are project-internal. A trace can incidentally capture a credential value (a token echoed into a tool arg) or Fox Companies IP / a MEMORY.md "never surface" record (private origins, Center Camp trust rules). Boundary rule: the induced spec may reference such a value by name (e.g. RH_POSTGRES_URL, Fox-IP:<slug>) but must never embed the secret value itself. A spec carrying a live credential or a never-surface record is fail-closed: do not emit it — escalate to the security-hygiene gate.

How

Gather evidence. Read the trace set for the target pattern from .planning/patterns/ (session-retro / session-observatory-live JSONL). Only proceed on a pattern skill-integration already flagged, or one you can confirm recurs in ≥ 3 distinct sessions. Fewer than 3 → skip (§When to skip).
Segment into candidate skill units. A candidate is a goal-directed span with a stable entry precondition and a stable exit postcondition. This is an LLM judgment — do not treat tool-sequence equality as the segment boundary; two traces reaching the same goal via different tool order are one candidate (§Robustness rule).
Decompose each candidate into the four-part spec (this is the induction payload, not a SKILL.md):
- Workflow structure — ordered steps, branch points, loop/iteration.
- Execution semantics — tools invoked, arg schema, side effects, and which steps touch shared repo state (git, worktrees, refinery-merge queue).
- Runtime attachments — the verification check that proves the step worked, the safety gate (ProcessContext/LoaderContext chokepoints, PreToolUse commit hook), the rollback action, and any state the skill must persist (Grove content-addressed store / MEMORY.md).
Scrub. Apply the §Data-classes boundary rule — replace any credential or never-surface value with a named reference before the spec leaves this skill.
Emit the spec, hand to skill-forge. Output the structured spec object and route it to skill-forge; do not scaffold or write SKILL.md frontmatter here.
Low-confidence segmentation → defer. If step 2 cannot draw a stable boundary (candidate spans overlap, or entry/exit conditions are unclear), emit no spec and hand the raw evidence to skill-forge's HITL / a human, rather than guessing a unit.

Robustness rule

Judge candidates by effect, not surface phrasing. Cluster traces by the goal they achieve and the pre/post-conditions they satisfy, not by identical tool calls or wording. A candidate that recurs only because the same literal command string appears is a weaker unit than one whose outcome recurs.

Confidence / failure model

Segmentation wraps an LLM judgment — it is semi-decidable, not a deterministic check, so it can over- or under-segment. This skill reduces the chance of authoring an under-specified skill; it does not guarantee a correct unit. Fail-closed default: on any uncertainty about a candidate that touches shared repo state, sensitive memory, or self-modification, escalate (to skill-forge HITL / mayor-coordinator) rather than silently emit a spec. The refinery-merge queue never auto-resolves conflicts; induction inherits that posture — never auto-emit past an unresolved boundary.

When to skip

The pattern recurs in fewer than 3 sessions — collect more traces first.
skill-integration has not surfaced it and you cannot confirm frequency — it may be a one-off, not a skill.
A finished SKILL.md already exists for this behaviour — route to skill-causal-curation (keep/repair/retire) instead of re-inducing.
The only available trace is a single session with no repetition — there is no reusable unit to induce.

Integration

skill-integration (upstream) — its frequency detection is the trigger; this skill consumes what it flags.
session-retro / session-observatory-live (evidence source) — write the .planning/patterns/ traces this skill segments.
skill-forge (downstream) — receives the structured spec and does the authoring/validate/critique/ship; this skill never writes the SKILL.md.
security-hygiene — the scrub + never-surface escalation runs under it.

同仓库更多 Skills

同仓库

goal-ambiguity-gate

Tibsfox/gsd-skill-creator

Use the moment a request is underspecified and you are about to act on an ASSUMED goal. Enumerate the candidate goals the request could mean, project the action or artifact each one implies, and check whether they diverge: if two or more plausible goals lead to materially different actions or artifacts, ask exactly ONE targeted clarifying question; otherwise proceed on the most-likely goal and state the assumption in one line. Clarification is an evidence-producing action, not a delay — but one question, never an interrogation. Distinct from gsd-spec-phase (GSD-phase-bound, emits a heavy SPEC.md) and intent-router (fetch strategy, not goal disambiguation). Backed by agent goal-state inference (arxiv 2606.16813v1). Triggers on acting under an assumed goal when the request admits more than one materially different reading.

2026-07-1969

memory-hubness-gate

Tibsfox/gsd-skill-creator

Use at write time to vector/embedding memory — Grove content-addressed insertion, chroma/pgvector upserts, memory-consolidation promoting session traces to MEMORY.md, or embedding externally-ingested content. Scores each candidate record against a fixed panel of sentinel queries and quarantines any record that would become the nearest neighbor of too many unrelated queries — a hub — whether from adversarial poisoning or accidental over-generality. This is the memory-record-side sibling of skill-injection-guardian (file-side) and the write-side complement of memory-use-warrant (read-side). Quarantine, never silently drop; a human reviews. Backed by the admission-time hubness gate (arxiv 2606.19692v1). Triggers on inserting into vector memory, consolidating memory, and embedding stranger content.

2026-07-1969

memory-use-warrant

Tibsfox/gsd-skill-creator

Run this appropriateness check the moment you are about to integrate a retrieved long-term memory — a Grove content-addressed hit, a chroma/pgvector neighbour, a memory-consolidation digest, or a MEMORY.md line — into a response, especially anything touching private origins, Fox Companies IP, credentials, or Center Camp / consent-governed content. It answers a question intent-router never asks: not WHAT to fetch or HOW, but WHETHER a correctly-retrieved item should reach output. Relevance is not appropriateness — a perfect similarity match can still be a boundary violation. Default is FAIL-CLOSED: if in-context authorization is uncertain, the memory may inform behaviour but must not be surfaced. Backed by RBI-Eval (arxiv 2606.06055v1). Triggers on surfacing recalled sensitive memory into a response.

2026-07-1969

operational-anchor-preservation

Tibsfox/gsd-skill-creator

Use before you shorten, compress, or rewrite the BODY of an existing SKILL.md or agent .md — the moment you are about to cut text to reduce length. First inventory the file's operational anchors (exact code / command / API snippets, fail-closed workflow guards, and rule / threshold statements), preserve every one, then judge the rewrite by expected downstream task cost (exploration, debugging, recovery tokens) rather than by the resulting line count. A shorter skill that strips an anchor makes the agent MORE expensive per task, so length is never the objective and rewrite is not compression. Backed by Preserving Operational Anchors When Compressing Agent Instructions (arxiv 2606.09421v2). Triggers on shortening or refactoring a skill or agent body.

2026-07-1969

semantic-merge-adjudicator

Tibsfox/gsd-skill-creator

Use when parallel agents in a convoy or across git worktrees hit a write conflict on shared state — a git tree, a config, or a shared document — and you must decide whether the conflict is semantically real. Classifies each conflicting hunk as independent (keep both), redundant (pick one), or a genuine clash (escalate), instead of blind locks, abort-and-retry, or escalate-everything. Sits ABOVE refinery-merge and NEVER blind auto-merges: it wraps an LLM judgment, defaults to escalate on any uncertainty, and unconditionally escalates any hunk touching .planning/, config, credentials, or safety-critical paths. Backed by CoAgent (arxiv 2606.15376v1). Triggers on write conflicts between parallel agents on shared repo state.

2026-07-1969

skill-coverage-gate

Tibsfox/gsd-skill-creator

Use when reviewing, shipping, or regression-testing a SKILL.md and you hold one or more captured trajectories where that skill was actually loaded. Translates the skill's instructions into behavior constraints, then reports which were EXERCISED (coverage) versus untested, and flags any that were in-scope but VIOLATED (compliance) — a cheap single-trajectory test-adequacy gate. It surfaces untested instructions rather than auto-blocking, and escalates fail-closed only when a violation touches shared repo state, never-surface memory, or self-modification. Complements skill-counterfactual-audit (which measures EFFECT) and skill-causal-curation (which measures contribution); this one asks ADEQUACY. Backed by the agent-skill-coverage metric (arxiv 2606.20659v2). Triggers on skill review, ship gating, and regression-testing a skill against its trajectories.

2026-07-1969

name	trace-to-skill-inducer
description	Use when you have captured session evidence — session-retro / session- observatory-live traces in .planning/patterns/, tool logs, correction records — and want to induce a reusable skill from it. Segments the traces into candidate skill units (an LLM judgment, not a deterministic parse) and decomposes each candidate into a four-part structured spec: workflow structure, execution semantics, and runtime attachments (verification, safety, rollback, state). It emits a spec object, NOT a finished SKILL.md, and hands that spec to skill-forge. It sits between skill-integration (upstream frequency detector) and skill-forge (downstream author). Backed by Agent-Trace-to-Skill Induction (arxiv 2606.06893v1). Triggers on inducing a skill from captured traces, turning a repeated pattern into a skill spec, and preparing evidence for skill-forge.
description-frequency	on-demand
user-invocable	true
version	1.0.0
format	"2025-10-02T00:00:00.000Z"
triggers	["induce a skill from these captured session traces","turn this repeated pattern in .planning/patterns/ into a skill spec","prepare a structured spec from trace evidence for skill-forge"]
updated	"2026-07-18T00:00:00.000Z"
status	ACTIVE
source	arxiv 2606.06893v1 (Agent-Trace-to-Skill Induction)

Trace-to-Skill Inducer

Why

Data classes touched

How

Gather evidence. Read the trace set for the target pattern from .planning/patterns/ (session-retro / session-observatory-live JSONL). Only proceed on a pattern skill-integration already flagged, or one you can confirm recurs in ≥ 3 distinct sessions. Fewer than 3 → skip (§When to skip).
Segment into candidate skill units. A candidate is a goal-directed span with a stable entry precondition and a stable exit postcondition. This is an LLM judgment — do not treat tool-sequence equality as the segment boundary; two traces reaching the same goal via different tool order are one candidate (§Robustness rule).
Decompose each candidate into the four-part spec (this is the induction payload, not a SKILL.md):
- Workflow structure — ordered steps, branch points, loop/iteration.
- Execution semantics — tools invoked, arg schema, side effects, and which steps touch shared repo state (git, worktrees, refinery-merge queue).
- Runtime attachments — the verification check that proves the step worked, the safety gate (ProcessContext/LoaderContext chokepoints, PreToolUse commit hook), the rollback action, and any state the skill must persist (Grove content-addressed store / MEMORY.md).
Scrub. Apply the §Data-classes boundary rule — replace any credential or never-surface value with a named reference before the spec leaves this skill.
Emit the spec, hand to skill-forge. Output the structured spec object and route it to skill-forge; do not scaffold or write SKILL.md frontmatter here.
Low-confidence segmentation → defer. If step 2 cannot draw a stable boundary (candidate spans overlap, or entry/exit conditions are unclear), emit no spec and hand the raw evidence to skill-forge's HITL / a human, rather than guessing a unit.

Robustness rule

Confidence / failure model

When to skip

The pattern recurs in fewer than 3 sessions — collect more traces first.
skill-integration has not surfaced it and you cannot confirm frequency — it may be a one-off, not a skill.
A finished SKILL.md already exists for this behaviour — route to skill-causal-curation (keep/repair/retire) instead of re-inducing.
The only available trace is a single session with no repetition — there is no reusable unit to induce.

Integration

skill-integration (upstream) — its frequency detection is the trigger; this skill consumes what it flags.
session-retro / session-observatory-live (evidence source) — write the .planning/patterns/ traces this skill segments.
skill-forge (downstream) — receives the structured spec and does the authoring/validate/critique/ship; this skill never writes the SKILL.md.
security-hygiene — the scrub + never-surface escalation runs under it.