fix(mods): jev-model-router leaves a bare slash command unclassified

[meesp123]

Before: a prompt that is only a slash command (/simplify) was sent to the decision model like any other. The model sees just the command's name, never the skill or command it runs, so it judged the name. On TypeSafe, three calls each:

Prompt	Jev's answer
/simplify	fast 0.80, effort 0.5
/run	fast 0.88, effort 0.1
/github	fast 0.67, effort 0.1
/code-review	balanced 0.96, effort 1.7
/security-review	deep 0.91, effort 1.5

On Opus 5.5 a fast answer with effort 0.1 moves the main loop to low, so a multi-step skill started with a bare command ran at low effort.

After: a prompt that is a slash command and nothing else is not classified. The log says a command with nothing after it; leaving the turn alone, its turn keeps the session's model and effort, and a null decision is put so an earlier prompt's cannot reach that turn. With text after the name (/code-review high, /simplify the retry loop) the prompt is classified as before. The check is /^\/[^\s/]+$/ on the trimmed text, so a path like /tmp/log.txt is still classified.

Live (claude -p, Claude Code 2.1.280, TypeSafe backend, Opus 5.5):

Prompt	main	this branch
/about-user	jev: tier fast (0.81) · effort 0.1 → low, then main loop → effort low (twice)	a command with nothing after it; leaving the turn alone, then main loop: no decision
/about-user who am I	classified	classified (jev: tier fast (0.95), main loop → effort low)

Checks

• cd cli-tool/components/mods && npx -y -p typescript@5 tsc -p tsconfig.json: clean
• bun test productivity/jev-model-router/tests: 42 pass (1 new: bare commands, and prompts that are not)
• claude plugin validate cli-tool/components/mods/productivity/jev-model-router: passed
• Hook-level tests under claude plugin test (kept locally until jev-model-router: run the tests on the built-in testing kit instead of bun:test #958 lands): a bare command sends nothing and its turn keeps high; with text after it, it is sent and routed; a bare command after a routed prompt keeps the session's effort. Both pass on this branch and fail on main.
• component-reviewer: no blocking issues. Taken from it: the README paragraph moved after the failure list (it is a policy, not a failure), and the doc comment notes that a one-segment path alone (/etc) matches too.

Version not bumped: several open PRs touch this mod. Catalog not regenerated (external PR).

Related

• fix(mods): jev-model-router applies a prompt typed mid-turn and skips text nobody typed #968–feat(mods): jev-model-router reads a follow-up with the reply it answers, sends a long prompt as its two ends, and names the risky acts #973, fix(mods): jev-model-router lets Claude Code's overload fallback stand for the rest of a turn #977 and fix(mods): jev-model-router drops the decision of a prompt a settings hook blocks #978 touch the same file; this one branches from the same main commit. fix(mods): jev-model-router applies a prompt typed mid-turn and skips text nobody typed #968 restructures prompt.submit; whichever lands second, I'll rebase.

[github-actions]

👋 Thanks for contributing, @meesp123!

This PR touches cli-tool/components/** and has been marked review-pending.

What happens next

1. 🤖 Automated security audit runs and posts results on this PR.
2. 👀 Maintainer review — a human reviewer validates the component with the component-reviewer agent (format, naming, security, clarity).
3. ✅ Merge — once approved, your PR is merged to main.
4. 📦 Catalog regeneration — the component catalog is rebuilt automatically.
5. 🚀 Live on aitmpl.com — your component appears on the website after deploy.

⚠️ Do not commit generated catalog files

Please make sure your PR does not include docs/components.json or anything under dashboard/public/
(components.json, counts.json, search-index.json, components/*.json, component-content/**).
Those files are regenerated automatically by scripts/generate_components_json.py after merge, and committing them causes merge conflicts.
If you already ran the script, revert them with: git checkout origin/main -- docs/components.json dashboard/public/

While you wait

• Check the Security Audit comment below for any issues to fix.
• Make sure your component follows the contribution guide.

This is an automated message. No action is required from you right now — a maintainer will review soon.

[greptile-apps]

Confidence Score: 4/5

The PR appears safe to merge, though a hook-level regression test would better protect its central behavior.

Findings

1.  Hook behavior remains untested ▶

Summary

The router now leaves bare slash commands at the session’s model and effort while continuing to classify commands with arguments.

• Adds a predicate and an early hook path that records a null pending decision.
• Documents the policy and adds predicate tests; the hook-level behavior remains untested in the checked-in suite.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[prompt.submit] --> B{Bare slash command?}
  B -- Yes --> C[Put null decision]
  B -- No --> D[Classify and put decision]
  C --> E[Continue prompt]
  D --> E
  E --> F[First turn.step takes pending decision]

Loading

_{Reviews (1) · Last reviewed commit: "fix(mods): jev-model-router leaves a bar..."}

[greptile-apps]

Hook behavior remains untested

This test checks only whether bareCommand() matches a string. It does not exercise the new prompt.submit branch or the later turn.step that reads its pending decision. A regression that still classifies a bare command, or applies a decision to the wrong turn, could therefore pass the checked-in tests. Please add a hook-level test for that flow.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[meesp123]

Written as hook-level tests: a bare /simplify sends nothing and its turn keeps the effort it came with (high), then /simplify the retry loop in payments.ts is sent and routed (low); and a bare command right after a routed prompt keeps the session's effort, so the earlier decision does not reach its turn. 2 of 2 pass on this branch's head and both fail on main. They run locally with claude plugin test (the engine's claude-code/testing kit) and are not in this PR: the kit loads a plugin with its manifest defaults only, so the backend path needs a copy of the mod with a key patched into plugin.json, and this suite still runs under bun test until #958 lands. I can add them once #958 is in, if you want them in the repo.

[greptile-apps]

Thanks—that addresses the behavioral gap: the hook-level checks cover suppression of a bare command, prevention of a prior decision leaking into its turn, and routing commands with trailing text. Given the current test-kit limitation and dependency on #958, I’m fine with keeping them out of this PR. Please add them to the repository once #958 lands; no further change is needed here.