(CODUX-202) Centralize persistent skill runtime context

[numnx]

🧩 Task: Centralize persistent skill runtime context

Automated task PR opened by Code UX for T01.

📋 Summary

Sprint: Sprint 202 · Persistent Skills + Agents UX Cleanup · Task: T01

Complete persistent skill runtime context and deliver a modular, accessible Agents settings workflow with discoverable storage contents and management.

🤖 Model & Provider

Provider	Codex CLI
Model	gpt-5.6-sol

⏱️ Timing

Started	2026-07-10 20:51:45 UTC
Finished	2026-07-10 21:03:28 UTC
Duration	11m 43s

📊 Token Usage

Metric	Value
Input tokens	207,564
Cached input tokens	3,300,096
Output tokens	16,131
Total tokens	3,523,791
Invocations	1
Tool calls	44
Active time	11m 10s

Included usage estimate (subscription/local login): $3.17

✅ QA Review

⏳ QA review runs after this PR is opened. Check the task card in the dashboard for the outcome once it lands.

📝 Full Task Prompt

## Objective
Make persistent skill prompt guidance, MCP retrieval access, and isolated storage mounts a single reusable runtime composition for every invocation path that resolves an opted-in agent.

## Scope
- `src/services/persistent-skill-context.ts` to own shared resolution and prompt composition
- `src/services/skill-service.ts` and `PersistentSkillStorageRuntime`
- `src/services/provider-execution-service.ts`
- `src/services/worker-inbox-reply-service.ts`
- `tests/backend/services/provider-execution-service.test.ts`
- `tests/backend/services/worker-inbox-reply-service.test.ts`
- `tests/backend/integration/persistent-skills-runtime.test.ts`

## Implementation Requirements
1. Extract the shared enabled-agent, project-ownership, attachment, prompt-section, and mount-resolution behavior into one typed helper without changing the default-off contract.
2. Use the helper from both the canonical `ProviderExecutionService` path and the direct `WorkerInboxReplyService` text-provider path so task coding, planning, QA, follow-up, dashboard, clarification, CI, and merge invocations receive the same context when an agent preset is known.
3. Preserve agent-scoped `search_skills` access, MCP agent identity, isolated mounts, retry behavior, and invocation prompt persistence while preventing duplicate context on retries or resumed sessions.
4. Add regression coverage for enabled attached storage, disabled retrieval, empty attachments, cross-project agent IDs, direct worker-inbox replies, and unchanged prompts for unscoped runs.

## Constraints
- Do not enable unrelated Code UX tools merely because persistent skill retrieval is enabled.
- Do not mount project workspaces or arbitrary user paths as persistent storage.
- Preserve existing provider prompts, memory injection, provider routing, and optional-service fallback behavior.

## Verification
- Run `pnpm exec vitest run tests/backend/services/provider-execution-service.test.ts tests/backend/services/worker-inbox-reply-service.test.ts tests/backend/integration/persistent-skills-runtime.test.ts`.
- Run `pnpm run lint` and `pnpm run build`.
- Confirm enabled attached agents receive one persistent-skill section, scoped MCP identity, and storage mounts; all disabled, unattached, mismatched, and unscoped cases remain unchanged.

🌿 Branch Info

Base: feature/CODUX-202-persistent-skills-agents-ux-cleanup
Head: task/feature-codux-20-t01-codex-ac0cbbe2-mrfevfu3

---

🤖 Generated by Code UX