[FEATURE] Auto-dream: 3 gaps — identity, accuracy, transparency

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

After enabling auto-dream and inspecting its first run, I found that the output requires manual review and correction before it can be trusted. The memories it writes can be inaccurately named, factually unverified, and impossible to audit — which undermines the purpose of automated memory consolidation.

Specifically:

Dream named a memory file after the session topic rather than the project it wrote into (the project had been renamed — dream didn't notice)
The old project folder's memory files were left orphaned with no detection
Dream wrote a specific count ("18 of 21 items resolved") without reading the file it referenced to verify
Existing memory files were left untouched even when potentially stale
There's no log of what dream created, modified, or removed — the only indication it ran is "last ran Xs ago" in /memory

Proposed Solution

Improve auto-dream's output reliability across three areas:

Identity — Derive the canonical project name from the memory folder path, not the session content. Check for sibling project folders with overlapping path segments and flag or consolidate orphaned memory from renamed projects.

Accuracy — Cross-reference existing memories against recent session activity and update or remove stale entries. When writing claims that reference specific files or counts, read those files to verify before writing. Optionally allow dream to read a configurable set of project files (e.g. a task tracker) for consistency checks.

Transparency — Write a brief changelog (e.g. .dream-log.md in the memory folder) summarizing what was created, modified, or removed each run.

Alternative Solutions

Currently I manually inspect the memory folder after each dream run and correct inaccuracies by hand. This works but defeats the purpose of automation. I'm also considering building a custom skill that cross-references memory against project files, but that shouldn't be necessary if dream handles its own output quality.

Priority

Low - Nice to have

Feature Category

File operations

Use Case Example

I renamed a project from my-old-project to my-new-project several sessions ago
Auto-dream runs after a long session
It creates project_old_feature_status.md in the my-new-project memory folder — wrong name
It writes "18 of 21 items resolved" without reading the actual file to confirm the count
It leaves a stale blocking-issue memory from 4 days ago untouched
The old my-old-project memory folder still has 3 orphaned files — dream doesn't notice
I have no changelog to see what dream actually did, so I have to diff the folder manually
With the proposed fixes, dream would name the file correctly, verify its claims, clean up stale entries, flag the orphaned folder, and leave a log I can review in 10 seconds.

Additional Context

Existing auto-dream issues (#38461, #38426) are about the /dream command not working. This request is about improving the quality of dream's output when it does run successfully.