feat: reduce skill startup context without workflow changes

[Sebastiangmz]

Objective

Apply the progressive-disclosure and context-load pattern described in Matt Pocock's Writing for Agents to OMP Pantheon. Execution turns receive only relevant skill catalog entries while OMP discovery, skill:// reads, mandatory trigger policy, project-local filtering, session behavior, and every non-skill system-prompt byte retain their previous behavior.

Implementation

• The final before_agent_start handler parses the existing <skills> block from event.systemPrompt; it never scans a second registry.
• The parser accepts both OMP catalog forms: standard list entries with multiline continuation text/blank separators, and custom <skill name="…">…</skill> blocks.
• A dedicated, tool-free call to the active model selects strict, known skill names at temperature 0.
• Certain selections render original catalog blocks in original order and accumulate during the session.
• Switch, branch, and shutdown clear accumulated state; stale in-flight routes cannot repopulate a reset session.
• Uncertain, malformed, unknown-name, model, credential, timeout, provider, thrown-router, or session-change failures return the exact original prompt.
• Existing Skill[] discovery, resources_discover, manual commands, project-local skills.includeSkills, and skill:// resolution remain unchanged.

Compatibility evidence

The original six-scenario evaluation remains recorded in evals/evidence/progressive-skill-routing-2026-08-19.json: five deterministic cases match the full-catalog baseline skill set exactly; the ambiguous case restores the exact original prompt.

Correction-specific runtime evidence:

• Neutral-CWD OMP probe routed and read using-superpowers plus playwright, opened https://example.com, and returned Example Domain; no routing fallback was logged.
• A second turn retained both selected catalog entries and returned PERSISTENCIA_OK; no routing fallback was logged.
• The same captured standard catalog transformed from 294 entries / 121,390 characters / approximately 30,348 tokens to 2 entries / 473 characters / approximately 119 tokens.
• The non-skill SHA-256 digest remained identical before and after routing: 539747d80f861aa8359a7da1d9014e25c57f9e57602bccda54284bccdfba04f7.
• A worktree-CWD probe with the existing project-local allowlist still exposed only herdr; skill://playwright remained unavailable. Its uncertain decision fell back to that one-entry local catalog, not the global catalog.

No raw prompts, credentials, private traces, or local machine paths are committed as evidence.

Verification

• bun test: 539 pass, 1 skip, 0 fail; 1,802 assertions.
• bun run typecheck: pass, including both extensions.
• bun run format:check: pass; 107 files checked.
• bun run lint: pass; 107 files checked.
• Focused routing suite: 46 pass, 0 fail; 91 assertions.
• EvalFly: run-smoke-20260820075730, suite smoke, 12 pass, 0 fail, 0 critical regressions, privacy sanitized, SpecSafe slice SPEC-20260820-SKILL-ROUTING-XML, commit range main..HEAD.
• Installer smoke: pass in an isolated temporary HOME; extension/runtime and existing skill symlinks verified, temporary tree removed.
• SpecSafe: correction slice closed PASS.
• Final five-lens review: goal PASS, local code quality PASS, security PASS, hands-on QA PASS, repository consistency PASS. Remote-head publication is verified after the consent-gated push.

Tradeoff

Each execution turn adds one active-model routing call. The fail-open design prefers the original project-visible catalog whenever selection is uncertain, so latency/cost can increase without reducing previous skill availability or changing user-turn behavior.