菜单
Infer which product domains a brownfield codebase evidences by cross-walking scan-index.json signals (framework idioms, dependencies, config, readme, CI) against Garura's canonical domain KB. Produces specification/domain-selection.yaml. Used by product-keeper during /codify.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/kapilvirenahuja/garura --skill infer-domain-selection-from-code命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | infer-domain-selection-from-code |
| description | Infer which product domains a brownfield codebase evidences by cross-walking scan-index.json signals (framework idioms, dependencies, config, readme, CI) against Garura's canonical domain KB. Produces specification/domain-selection.yaml. Used by product-keeper during /codify. |
| version | 0.1.0 |
| user-invocable | false |
| model | sonnet |
| allowed-tools | Read, Write, Grep, Glob |
| deprecated | true |
| deprecated_note | #434 ProductOS realignment — superseded by the command model; retained for Phase E reference, not installed |
Called by product-keeper during /codify. Produces specification/domain-selection.yaml at {stm_base}/{issue}/evidence/codify/proposals/specification/domain-selection.yaml.
/codify bootstraps product LTM for a brownfield repo. This skill answers:
which product domains does this codebase actually evidence? It reads
scan-index.json (produced by scan.py), walks Garura's canonical domain
catalog (core/components/memory/knowledge/domain/*.md), scores each canonical
domain against the scan signals, and surfaces scan-only domains as proposals
for a new canonical entry (per /reap C7a).
Tier 2. learning_category: domain. sub_category: null (domain is a flat
parent).
Received from the /codify play via JSON contract. All paths absolute.
| Field | Required | Description |
|---|---|---|
scan_index_path | yes | scan-index.json produced by scan.py. |
stm_base | yes | STM root, from stm.base-path in .garura/core/config.yaml. |
issue | yes | Issue number (STM namespace). |
output_path | yes | {stm_base}/{issue}/evidence/codify/proposals/specification/domain-selection.yaml. |
decision_manifest_path | yes | Sibling decision-manifest-infer-domain-selection.yaml. |
ltm_context | yes | Resolution Protocol context from product-keeper (product_base, kb_base, resolved LTM paths). |
resolution_trace_path | yes | Sibling resolution-trace-infer-domain-selection.yaml. |
kb_domain_dir | yes | core/components/memory/knowledge/domain/. Walked non-recursively; files starting with are metadata, not domains. |
_scan_index_path must be valid JSON with schema_version, scan_status,
manifests, patterns, config_files, docs, frontend_detection.
Missing → scan_index_missing.kb_domain_dir must exist and hold ≥ 1 non-underscore .md. Missing →
kb_domain_dir_missing.output_path_unwritable.Per core/components/memory/standards/rules/resolution.md, record every probe
in resolution_trace_path.
codify-bootstrap.{product_base}/specification/domain-selection.yaml.
If present AND LOCKED: scoring still runs, but emit a proposal only if scan
evidence conflicts (scan-only domains, or canonical domains scoring high
that the locked artifact rejected). No conflict → alignment_confirmed: true
and no proposal body. Absent (typical /codify): proceed.{kb_domain_dir}/*.md, excluding basenames starting with
_. For each, extract the slug and capture declared search patterns /
feature IDs as the matching vocabulary for Step 3.inferred_domain_new.For each R3 canonical domain compute an evidence bundle. Every hit MUST record a concrete path (file, manifest entry, or scan-index JSON pointer) — no score without evidence.
Signal → domain mapping:
patterns.framework_idioms (case-insensitive substring on the idiom tag):
express|nestjs|fastify|koa|fastapi|flask|django|spring|gin|echo|laravel|rails → backend-apireact|vue|svelte|angular|solid|nextjs|nuxt|remix → frontend-webreact-native|expo|flutter|swiftui|jetpack-compose → mobilegraphql-schema|apollo|graphql-yoga → graphql-apigrpc-service|protobuf → rpcopenapi-spec|swagger → api-contractsconfig_files.docker (Dockerfile, docker-compose.*) → containerizationconfig_files.infra (terraform, pulumi, cloudformation, cdk) → infrastructure-as-codeconfig_files.deploy (helm, k8s manifests, serverless.yml) → deploymentconfig_files.ci (github workflows, gitlab-ci, circle, jenkins) → engineering-experience (CI is that domain's anchor)patterns.test_framework_signals non-empty (jest, vitest, pytest, rspec, go test, junit) → testing (count feeds confidence, not just presence)docs.readme_preview whole-word case-insensitive keyword sweep:
finance|banking|trading → financehealth|medical|patient|clinical → healthcareecommerce|checkout|catalog|cart → commercelogistics|shipping|fleet|warehouse → logisticscms|content|publishing → content-managementanalytics|dashboard|metrics|bi → analyticslearning|course|lms|education → educationmanifests dependency names (union across all manifests):
stripe|braintree|adyen|razorpay → paymentsnext-auth|auth0|clerk|passport|keycloak-* → user-managementtensorflow|torch|transformers|langchain|llamaindex|openai|anthropic → ml-aielasticsearch|meilisearch|typesense|algolia → searchgit.churn_top is NOT a domain signal (hotspot only).Confidence tiers per canonical domain:
high — ≥ 2 distinct signal categories, OR a strong-marker dep (payment SDK, auth lib, ML framework).medium — 1 category with ≥ 2 evidence items (e.g., two idioms in same family), OR a single strong dep.low — 1 category with 1 evidence item (e.g., readme-only hit with no code confirmation).rejected_domains.Any domain slug produced by the mapping that is NOT in the R3 canonical set
becomes a proposed new canonical domain: set inferred_domain_new: true and
learning_category_proposed: true; author taxonomy_justification with
evidence_path (scan-index pointer or file path), excerpt (≤ 200-char
literal excerpt), and reasoning (1–3 sentences on why this is distinct vs.
folding into an existing entry). Confidence capped at medium.
selected_domains = every canonical domain scoring ≥ low, plus every proposed
new domain. rejected_domains = every canonical R3 domain with zero evidence,
each with source_path (KB file) and a one-line reason.
Decision manifest is written BEFORE the primary artifact (Decision Surfacing Discipline — a crash mid-write still leaves the audit trail). Every inferred decision (confidence pick, new-domain proposal, locked-artifact conflict flag) gets one manifest entry.
specification/domain-selection.yamlmeta:
source_type: "inferred_from_code"
evidence: ["<scan-index JSON pointer or file path>", "..."]
confidence: "high" | "medium" | "low" # aggregate over selected_domains
learning_category: "domain"
sub_category: null
tier: 2
selected_domains:
- id: "<canonical domain slug from KB>"
evidence:
- signal: "framework_idiom"
match: "express"
path: "patterns.framework_idioms[3]"
- signal: "manifest_dep"
match: "stripe"
path: "manifests[0].dependencies.stripe"
confidence: "high" | "medium" | "low"
inferred_domain_new: false
- id: "<new-domain-slug>"
inferred_domain_new: true
learning_category_proposed: true
taxonomy_justification:
evidence_path: "<path>"
excerpt: "<≤200-char literal excerpt>"
reasoning: "<1–3 sentences>"
evidence: [...]
confidence: "low" | "medium"
rejected_domains:
- id: "<canonical domain slug>"
source_path: "core/components/memory/knowledge/domain/<slug>.md"
reason: "<one-line rationale>"
alignment_confirmed: false # true only when R2 hit LOCKED and no conflict
Standard Garura shape: schema_version, skill, generated_at, decisions[].
One entry per inferred decision, each with decision_id, decision_type,
tier (high/mid/low per Decision Surfacing Discipline), grounding_source,
recommendation, alternatives, chosen_reason.
R1..R3 probes with source, path, outcome (hit|miss|skipped), and any
extracted payload. R4 omitted.
status: success
artifact_path: "<output_path>"
decision_manifest_path: "<decision_manifest_path>"
resolution_trace_path: "<resolution_trace_path>"
selected_domain_count: <int>
new_domain_proposal_count: <int>
rejected_domain_count: <int>
overall_confidence: "high" | "medium" | "low"
alignment_confirmed: <bool>
status: failure
what_failed: "<code>"
detail: "<specific error>"
evidence:
offending_path: "<file path if applicable>"
offending_field: "<field name if applicable>"
Codes:
scan_index_missing — scan_index_path absent or invalid JSON.kb_domain_dir_missing — directory absent or zero non-underscore .md.ltm_resolution_failed — R2 hit but file is invalid YAML, or product_base resolves outside the repo sandbox.insufficient_signal — scan present but zero manifests, zero framework idioms, AND zero readme preview.output_path_unwritable — parent directory uncreatable.scan_index_path and kb_domain_dir.inferred_domain_new domain into the KB is a human decision handled downstream by /garura:enrich.scan.py; stale scan_index_path is the orchestrator's concern.{stm_base}/{issue}/evidence/codify/proposals/; never modifies .garura/product/ in place.