routing: try a group's rules on a message before saving them #2829

A real case: a user routes with a group, intent rules and a small classifier (Clef on Cloudflare Workers AI, say). The intents are written in English, as the templates suggest: a hard task: designing, debugging, a change across several files, or planning. Then they are stuck: their own message — 帮我把这个函数改成并发安全的 ("make this function concurrency-safe") — will the classifier call it hard or easy? The rule doesn't say: its behavior is whatever the small model answers for that message. Today the only way to learn it is to save the draft, run a real agent conversation, and read classWhy on the Routing page. Every wording change walks the loop again.

Proposal

A try box under the rules in the group editor. Paste a message and see, in under a second in the GUI, which intent the classifier picked, which rule matched, and which member goes first — for an unsaved draft.

Implementation

• One decision function for try and live traffic: the part of ruleFor that keeps nothing is split out as turnDecide. A live turn calls it and keeps what it decided under turnRules, as before; TryRule only calls it. No dry-run flag crosses the hot path.
• A try asks the classifier by itself (directAnswer): no shared cache read or write, no rest for a classifier that failed, no failure note on the group's card, and no request to any member.
• Three entries: the editor's try box, POST /api/groups/try, and magpie group try <id> <message…> — the CLI decides beside any running magpie, asking the classifier itself from the same files, for a saved group.
• The answer is the trace's own RuleHit, rendered by the Routing page's existing ruleWhy/classWhy.

Verified

• 14 cases on real machines, macOS and Windows, all passing. On Windows the GUI's real classifier calls took 548–791 ms and the CLI 0.79–1.08 s (process start included); on macOS the CLI decided in 654 ms. No member model took a request the whole time. A try's classifier call is recorded like any classifier call (agent: magpie, kind: routing) and shows on the Routing page with the classifier badge.
• Branch ggbdpq-try-rule: 22952132 is the feature, b96469dc fixes the single-intent wording. Both commit bodies carry the full record: guard tests shown red when broken and green restored; go vet and go build on darwin, linux and windows; the GUI in Chromium and WebKit — routing-rule-try.test.cjs 16/16, and the 17 group-editor regression files it runs with, 194 tests across both engines.

Savings (an estimate)

Tuning a group of 2–3 intent rules usually takes 5–20 tries. Each try today is a real agent conversation (about 10k–50k tokens) plus 0.5–2 minutes of hands-on time; a try is about 150 tokens and under a second.

Boundaries

• Purely additive: remove it and no existing path changes.
• It serves users who write intent rules and iterate on their wording. A group with no deciding rules says so at once and asks the classifier nothing.
• A try is a preview: a conversation's first turn, no reasoning, images or folder, and it knows nothing of affinity or rests.
• One observation: with a single intent in the group, small models occasionally lean toward 1 (one "hey, what's up?" came back as a hard task). A try shows this; it doesn't fix it.

Open question

On the Routing page, a try's classifier call looks the same as a live one. Whether to tell them apart (a field of its own) is the maintainer's call; the branch doesn't do it.

Attribution: If these commits are applied, please keep their author (ggbdpq) or add Co-authored-by: ggbdpq <ggbdpq@gmail.com>.

AI: written with AI assistance — Generated-by: GLM-5.3-Flash (ZCode).