wiki-dream

wiki-dream

While you sleep, the wiki re-evaluates which old pages have become relevant again. The dreamer reads new ingest entries from log.md since the last dream watermark, identifies entities/tags/citations they bring up, and scores every frozen/archived page against that signal. Pages above the threshold become candidates — surfaced for your review, never auto-promoted by default.

Auto-dreaming is the only thing in kata that runs without you. It is intentionally conservative: dry-run by default, capped at 10 candidates per run, and silently rejecting bad suggestions costs nothing — they just don't appear next time.

Implementation

The algorithm is in plugin/scripts/wiki_dream.py. The script is the source of truth; the prose below explains its behavior. The skill calls the script and renders the result.

# Default — read watermark from log.md, write candidates to dreaming/{date}.md
python {plugin_root}/scripts/wiki_dream.py --wiki {wiki_path}

# Override window — dream over a specific period
python {plugin_root}/scripts/wiki_dream.py --wiki {wiki_path} \
    --since 2026-04-01

# Explain why one specific page did or didn't qualify
python {plugin_root}/scripts/wiki_dream.py --wiki {wiki_path} \
    --explain mosaic

# Apply — write tier_override:active to selected candidates
python {plugin_root}/scripts/wiki_dream.py --wiki {wiki_path} \
    --apply --pages 1,2,3

The candidate file is written to dreaming/{YYYY-MM-DD}.md (one per run, preserved historically). The skill should also surface the JSON to the user.

Inputs the dreamer reads (and ignores)

Reads (only):

• log.md entries since the last ## [date] dream | ... watermark
• Wiki pages with ingested_at or updated ≥ that watermark
• Frozen and archived pages (the candidate pool)
• SCHEMA.md dreaming: config block

Does NOT read:

• Claude Code session history
• External APIs, embeddings, web
• Anything outside {wiki_path}

If a user expects "dream from what we just talked about" — explain that the dreamer is filesystem-only by design (architecture, not bug). To get session content into the dreamer's view, file it back to the wiki via /wiki-query --file first.

Modes

Default: dream

$ /kata:wiki-dream

[Dreaming] since 2026-04-21 (last weekly run)

Candidate pool: 36 (archived + frozen)
Fresh pages this period: 8
Resurgent tags: trend

Found 9 candidates above threshold 0.6:

1. [[mixture-of-experts]]   (frozen, score 0.90)
   - shares entities ['glam', 'moe-foundational-paper'] with new ingests
   - directly linked from new pages ['queries/deepseek-v3-paper-released.md']

2. [[mosaic]]                (archived, score 0.90)
   - shares entities ['mosaic', 'mpt-7b'] with new ingests
   - directly linked from new pages ['queries/databricks-acquires-mosaic.md']

... etc

[Suggested next]
→ /kata:wiki-dream --apply --pages 1,2,4    (promote selected)
→ /kata:wiki-dream --explain mosaic          (full scoring breakdown)

--apply --pages <list>

Writes tier_override: active to the selected candidates' frontmatter, plus a reason and timestamp. Appends a ## [date] dream | applied N candidates log entry advancing the watermark. Never auto-runs — must be invoked explicitly with --apply.

--explain <page>

Re-scores one specific page (regardless of whether it crossed the threshold), printing each component (entity / tag / citation), the reason strings, and the increment summary that drove it. Useful for debugging "why didn't X get promoted?"

Scheduled runs

Recommended weekly cadence. Set up via Claude Code /schedule:

claude /schedule "0 23 * * 0" "/kata:wiki-dream"

The wiki-init skill writes this cron line during interactive setup if the user enables dreaming. The candidates land in dreaming/{date}.md for review the next morning; nothing is applied without explicit user action.

Configuration

All knobs live in SCHEMA.md dreaming::

Key	Default	Effect
enabled	true	Set false to disable entirely
strategy	co-occurrence	Only co-occurrence in v1.6
cadence	weekly	Drives the cron line
confidence_threshold	0.6	Lower for more candidates, higher for less noise
max_repromote_per_run	10	Hard cap
weights.entity	0.5	Weight for entity overlap
weights.tag	0.2	Weight for tag resurgence
weights.citation	0.4	Weight for direct inbound link from a fresh page
resurgence.dormancy_window_days	180	Tag must be silent this long before counting as resurgent
resurgence.min_count	3	Tag must appear in ≥ N fresh pages to count

Tune via /wiki-config --set dreaming.confidence_threshold 0.55.

Failure modes / no-op cases

• No watermark in log.md → dreamer looks back 30 days as a fallback.
• No fresh pages since watermark → emits 0 candidates with explanation; log entry still advances the watermark.
• Wiki has no frozen/archived pages → candidate_pool_size is 0; candidates is empty. Explain to the user: tier system needs to be enabled and pages need to age out for dreaming to have anything to do.

Notes for the agent

• Always show BOTH the candidate list AND a 1-line summary of why dreaming had this signal (the increment summary). Without context the candidates look arbitrary.
• When the user says "yes promote X" — translate to --apply --pages N, not direct frontmatter edit. The script handles log entry + reason capture.
• Reject signals are not stored anywhere v1.6; the user just doesn't apply candidates they don't want. v1.7+ may add a feedback loop.