Claude Code 插件 auto-support：基于公开文档的引用式问答与升级

Draft cited answers to product questions from public documentation, with path and content checks and escalation requests for unsupported questions.

English | 中文版

Design Philosophy

Product support needs a useful answer without expanding access to private implementation material. The selected public documentation is therefore an allowlisted knowledge boundary. Retrieval and the host hook check paths outside the model, and the draft must retain matching source excerpts and citations. A prompt instruction alone cannot enforce that boundary.

This design accepts a narrower answer set: the shipped CLI can refuse a question that a free-form generator would try to answer. It also refuses when document scope cannot be established. The llmcall interpretation used for large prose sections remains fallible, so exact source ranges prove provenance without proving the interpretation correct. Unsupported paraphrasing needs a separately validated integration.

Drafting, escalation and delivery are separate outcomes. A neutral refusal can request escalation, but only a confirmed relay receipt establishes a send. Private durable state prevents an uncertain send from becoming an automatic duplicate. Host hook coverage and live delivery still require deployment checks; the plugin does not install an operating-system sandbox.

Read the full design philosophy.

Scope

This Claude Code plugin supports product Discord questions through public-document retrieval, cited drafts and founder review or relay. It is deployed through the product root's .claude/ configuration. The shipped CLI returns local drafts; it does not automatically post replies.

Questions requiring source code, secrets or other material outside the public allowlist are refused and referred for review. General conversation and code explanation are outside this scope; missing evidence cannot be replaced with answers from memory.

How it works, four fail-closed gates

Discord msg ─▶ entry (injection+intent, spotlighted) ─▶ retrieval (allowlist only, secret-scrubbed)
            ─▶ grounding (retrieval-conf × faithfulness) ─▶ egress (schema + DLP + canary + citation)
            ─▶ draft ─▶ founder review ─▶ approve ─▶ user      (any gate fails ⇒ neutral refusal + escalate)

Knowledge boundary is allowlist-first, default-deny, denylist-wins: retrieval and the hook check both the requested path and its resolved target within the selected public root. Detected credentials and PII are removed before notification or persistence. State reuses the schedule-reminder base; escalation reuses the machine Discord relay with SRE-style dedup.

Retrieval retains local preambles separately and follows explicit same-document section references. Separating large prose chapters requires an exhaustive semantic interpretation through the installed llmcall interface, using its current default judge routing. Missing, uncertain or incorrectly bound interpretations cause refusal. Exact ranges prove where the evidence came from; semantic judgments remain fallible and need independent validation. See source scope.

Install

/plugin install github:DaizeDong/auto-support

Or clone manually:

git clone --recurse-submodules https://github.com/DaizeDong/auto-support.git ~/.claude/plugins/auto-support

Quick start

From the repository root, initialize an empty companion outside the tool checkout:

python scripts/init_config.py --slug example --out ../auto-support-config

Set products/example/product.json's product_root to the absolute directory of the product's public documentation. The generated policy accepts document formats throughout that directory, including a flat usage.md; deny rules still apply. Then run:

python scripts/verify_config.py --config-dir ../auto-support-config
python skills/auto-support/scripts/answer_pipeline.py --policy ../auto-support-config/products/example/policy.json --query "How do I install the SDK?"

This produces a cited draft with no Discord credentials. Keep real configuration and runtime records in a PRIVATE versioned companion. The optional delivery integration must separately configure the host hook, relay and approvals; this initializer does not ship an apply.py.

Config

Store per-product configuration and runtime records in a separate PRIVATE versioned auto-support-config companion. Each product has its own policy.json. CONFIG.md owns discovery order, policy selection, initialization and switching; the schema reference describes the layout.

When switching companions, clear stale higher-priority selectors and update both the doctor/state selection and AUTO_SUPPORT_POLICY for the draft CLI and hook. The policy must belong to the selected companion. DRAFT READY covers configuration and the local documentation root only.

The current credential setup uses Mode B: secrets/* is excluded from Git and credentials need an approved separate backup. Runtime records, including escalation state, remain versioned. Optional @secret:... references require a separately configured delivery adapter.

Before persisting a turn, follow reminder setup to select and initialize a PRIVATE scheduler database. Follow DATA.md for state retention and recovery; uncertain delivery locks must remain until the outcome is reconciled.

How to invoke

Say “Answer this product question from its public docs.” The skill resolves the selected product, asks once for a missing docs root or ambiguous product, validates its policy, then returns a cited draft or a refusal. Use --demo --root <public-docs> only for an explicit policy-free demo. Automatic Discord intake requires a separately installed listener.

Example output

A passing turn returns a grounded draft with citations (public-faq/faq.md:4); a blocked/unsure turn returns one neutral line (这个问题我无法确定，请联系团队进一步确认。). The JSON decision can request escalation; it is not a delivery receipt. escalate.py --dry-run reports a plan; only a successful live relay marks sent=true and starts cooldown. An uncertain response reports sent=null and reconciliation_required=true, retains its private dispatch lock, and requires inspection before any retry, including critical alerts.

Limitations

The shipped CLI accepts only unchanged retrieved excerpts and matching citations, including outputs from a custom generator. Paraphrasing requires a separate verified integration. Scope interpretation depends on the configured model when large prose chapters are separated; deterministic test responses do not establish that model's effectiveness. Discord listeners, secret provisioning, host hooks and live delivery need deployment validation; a passing unit suite does not enable auto-post. The reminder bridge requires schedule-reminder. Runtime writes require Git and authenticated gh to prove PRIVATE visibility for every configured remote's effective fetch and push URLs, including URL rewrites. Default and branch push selections must refer to verified named remotes. PUBLIC or unknown destinations block writes, with no fallback into the tool repository. No OS sandbox is installed.

Languages

English (README.md, authoritative) · 中文 (README_CN.md)

Roadmap · Contributing · License

See ROADMAP.md · CONTRIBUTING.md · LICENSE (MIT).