feat(dream): add unified self-improvement agent

[xiami762]

Summary

• Add a unified Dream self-improvement Agent that reviews incremental experience and makes evidence-backed updates to durable Memory or managed user Skills.
• Support both explicit /dream execution and scheduled background execution through the same implementation.

Key Changes

• Add the hidden self-improve Agent with the standard Agent prompt-building flow and read, write, edit, search, Skill, and Bash capabilities.
• Add a single Dream pipeline that consumes bounded Session text/tool traces and recent Daily evidence, then applies precise Memory or managed Skill changes.
• Add /dream command handling with visible progress, disabled-state, completion, no-change, and failure output in direct chat and WebUI flows.
• Add the scheduled Dream runner, lifecycle state/watermarks, catch-up behavior, and no-op handling when no durable change is justified.
• Restrict Skill edits to user Skills marked with metadata.managed_by: flocks, with read-before-write and post-change validation.
• Guard Daily Memory and non-target files across file tools, sandbox paths, and Dream execution.
• Add memory.dream configuration with enablement, interval, and recent-Daily window settings.

Impact Scope

• User-visible behavior: /dream runs self-improvement for the active Session and reports what changed; scheduled Dream can run in the background.
• Compatibility / migration: Replaces the previous split Dream/Learn evolution path; no state migration is required for the new evolution state.
• Configuration / environment: Adds memory.dream.enabled, interval_hours, and recent_daily_days; Dream is enabled by default with a 24-hour interval.
• Dependencies: No new third-party dependencies.
• Performance / resources: A scheduled run may invoke one bounded background Agent execution per eligible interval. Input and tool steps are capped.
• Security / permissions: The Agent can use Bash and file tools, but writes are constrained to the exact Memory targets and curator-managed user Skills. Daily and installed/source Skills remain protected.

Business Logic to Review

• Manual and scheduled trigger eligibility, watermark advancement, catch-up limits, and the distinction between failed, no-op, and successful runs.
• Evidence construction and budgeting: Session user/assistant text, bounded tool traces, recent Daily files, and Skill catalog serialization.
• Prompt routing between USER, Global Memory, Project Memory, and Skills, including canonical ownership and NO_CHANGES behavior.
• Skill ownership checks, read-before-write enforcement, path/sandbox guards, and post-run change validation.
• Foreground /dream status/output behavior versus silent scheduled execution.

Why This Approach

• Uses one Agent and one comprehensive prompt for both Memory curation and Skill evolution, avoiding divergent Dream and Learn implementations.
• Runs through the standard Agent flow with explicit write boundaries, keeping the implementation small while preserving normal tool and sandbox enforcement.

Test Plan

• uv run pytest tests/command/test_evolution_commands.py tests/memory/test_evolution.py tests/memory/test_evolution_agent_runner.py tests/sandbox/test_sandbox_file_tools.py tests/server/test_input_dispatcher.py tests/server/test_lifespan.py tests/session/test_lifecycle_hooks.py tests/session/test_prompt_tokens.py tests/session/test_status.py tests/skill/test_skill.py -q (190 passed)
• npm run test:run -- src/components/common/SessionChat.test.ts src/pages/Session/index.test.tsx (213 passed)
• npm run lint
• npm run build

Compatibility, Migration & Rollback

• No migration is required for the new Dream evolution state.
• Set memory.dream.enabled to false to stop manual and scheduled Dream execution without changing Memory files or Skills.
• Reverting the PR removes the runner and command path; existing Markdown Memory and generated Skills remain ordinary files.