refactor(router): clarify routing prompt within token budget

[Light-code11]

Summary

• shorten the router catalog scaffolding in a dedicated prerequisite so feature PRs do not alter every skill indirectly
• preserve strong Do NOT guidance for exclusions
• distinguish action examples from alternative-tool counterexamples with Actions: and Other tools: headings
• replace the previous vacuous assertion with direct checks for every rendered routing section and target namespace

Why

The Victor routing surface was already approximately 5,998/6,000 tokens before adding another skill. This prerequisite creates room for the Meeting Room skill without bundling a global router change into #225.

The compact prompt still tells the model what each value means:

• Use: contains the skill's positive routing condition
• Do NOT: contains the skill's exclusion condition
• Actions: maps examples to actions within the current skill
• Other tools: maps counterexamples to a different tool

Validation

• router unit, catalog, and hermetic routing gates: 15 passed
• CI-equivalent security suite: 493 passed
• deterministic routing evaluation: 242 cases, 0 misroutes, no baseline regressions
• catalog size: public 5,081.5 approximate tokens; Victor 5,731.5 approximate tokens
• full repository: 842 passed, 20 failed; the same 20 unrelated failures reproduce on feat: add private meeting room booking skill #225
• Python compilation and git diff --check passed

The live OpenAI-backed routing evaluation was not run locally because no real OPENAI_API_KEY is available. The deterministic gate and all local suites above were run; the live command remains python scripts/run_routing_eval.py --mode v2 --check when a key is available.

Companion feature: #225

[Light-code11]

Updated the router prerequisite to address the prompt-semantics concern directly.

• positive examples are labeled Actions: because their values are actions inside the current skill
• counterexamples are labeled Other tools: because their values select a different tool
• exclusion guidance remains the explicit Do NOT: directive
• tests assert the rendered action and alternative-tool mappings, rather than only checking that headings exist

Validation: 15 focused router/catalog tests passed, the 493-test CI-equivalent suite passed, and 242 deterministic cases produced 0 misroutes with no baseline regression. Public and Victor catalogs are approximately 5,081.5 and 5,731.5 tokens. A live OpenAI-backed eval could not be run locally because no real API key is available.

[callumholt]

PR review

I reviewed the router prompt-format changes, generated tool schemas, catalog lint/budget gates, and routing evaluation behavior at commit 526f3f4591e77582f465bc4dc965cdc33e126c75.

Findings

No blocking correctness, security, or routing-contract issues found.

• The compact format preserves each routing dimension: Use: for positive scope, Do NOT: for exclusions, Actions: for examples mapped to actions in the current skill, and Other tools: for counterexamples mapped to alternate tools.
• Action enums, parameter schemas, required fields, channel filtering, event-only filtering, validation, and retry behavior are unchanged.
• The revised tests directly assert all rendered sections and mappings, replacing the previous weak conditional assertion.
• Generated catalog sizes are approximately 5,081.5 tokens public, 5,731.5 tokens Victor, and 1,911.2 tokens admin, all below the 6,000-token gate.

Validation

• Focused router, catalog, and hermetic routing gates: 15 passed.
• Deterministic routing evaluation: 242 cases, 0 misroutes, no baseline regressions.
• GitHub security-checks: passed (493 tests).
• Python compilation and git diff --check: passed.

The deterministic evaluation only exercises exact-match routing and therefore cannot measure how the live model interprets the changed descriptions. The live OpenAI-backed evaluation was not run because no authorized real API key was available. This is a documented residual prompt-quality risk, but I consider it non-blocking because the semantic mappings remain explicit, the schemas are unchanged, and the structural/budget gates pass.

Conclusion: Safe and ready to merge.