Shadow Frog Meditate

ShadowFrog Meditate

Shadow hygiene — deduplicate, merge, and resolve conflicts across the entire .shadow/ knowledge base. Prerequisite: .shadow/ exists with discoveries.

Why Meditate?

Over time, shadows accumulate noise:

• Duplicates: the same insight written differently by different sessions
• Near-duplicates: one discovery is a subset of another
• Conflicts: two discoveries contradict each other (code may have changed,

or one was wrong)

• Cross-scope duplicates: a per-file discovery and a _cross/ entry

saying the same thing

This noise confuses downstream agents and dilutes signal. Meditate cleans it up.

Phase 1: Scan

Use parallel subagents to scan the shadow. Each subagent handles a batch of shadow files.

Scope Optimization

Not every file needs scanning. To reduce cost:

• Skip files with 0-1 discoveries — they can't have internal duplicates
• Focus on files modified since last meditate — check _meta/state.json

last_update_at against file modification times

• Always scan files with 5+ discoveries — highest duplicate risk

For the first meditate after a large dream run, most files will need scanning. For incremental meditation after small updates, this can reduce scope by 80%+.

Per-File Scan

For each per-file shadow (e.g., src/auth.py.md):

1. Read all discoveries under each ## symbol heading
2. For each pair of discoveries under the same symbol, classify:

• Duplicate: same behavioral claim, different wording
• Near-duplicate: one discovery is a subset/refinement of the other
• Conflict: the two discoveries make contradicting claims
• Distinct: genuinely different insights — no action needed

1. Record each finding as a structured action (see below)

Scan Output Format

Subagents must output findings as one JSON object per line so the orchestrator can auto-apply resolutions. This is critical for automation — prose recommendations require manual interpretation.

{"action": "merge", "file": "src/auth.py.md", "symbol": "authenticate_user", "keep": "- silently returns None on expired tokens...", "remove": "- returns None when token expires...", "merged": "- authenticate_user() silently returns None on expired tokens instead of raising. 3 of 7 callers don't check.\n  _(verified, source: exploration)_", "reason": "duplicate: same claim, different wording"}
{"action": "merge", "file": "src/db.py.md", "symbol": "connect", "keep": "- connection pool exhaustion...", "remove": "- pool runs out...", "merged": "...", "reason": "near-duplicate: first extends second"}
{"action": "conflict", "file": "src/auth.py.md", "symbol": "validate_token", "entry_a": "- raises ValueError...", "entry_b": "- returns False...", "resolution": "verified_a", "reason": "code inspection: line 42 raises ValueError"}
{"action": "conflict", "file": "src/cache.py.md", "symbol": "invalidate", "entry_a": "...", "entry_b": "...", "resolution": "escalate", "reason": "both claims have evidence, needs user input"}
{"action": "move_to_cross", "file": "src/auth.py.md", "symbol": "validate_token", "entry": "- all validators share...", "cross_slug": "shared-validation-pattern", "reason": "cross-scope: involves 4 files"}

Fields:

• action: merge | conflict | move_to_cross | move_from_cross
• file: shadow file path relative to .shadow/
• symbol: the ##/### heading the discovery lives under
• keep: the discovery text to keep (for merge)
• remove: the discovery text to delete (for merge)
• merged: the final merged text (for merge)
• resolution: verified_a | verified_b | escalate (for conflict)
• reason: human-readable explanation

The orchestrator collects all lines, applies merge and conflict actions automatically, and presents escalate items to the user.

Cross-Scope Scan

After per-file scanning:

1. Collect all per-file discoveries into a flat list
2. For each _cross/*.md discovery, check if any per-file discovery

makes the same or overlapping claim

1. For each _prefs.md preference, check if any per-file discovery

or _cross/ entry duplicates it

1. Record cross-scope findings the same way

Scanning Guidelines

• Compare claims semantically, not just textually. "Returns None on

expired tokens" and "Silently returns None when token expires" are duplicates.

• Two discoveries about the same function but covering different

behaviors are distinct, not duplicates. E.g., "returns None on expired tokens" vs "uses constant-time comparison" — these are unrelated observations about the same function.

• Pay attention to Also involves: — two discoveries with overlapping

Also involves: refs are more likely related.

Phase 2: Resolve

Process each finding by type.

Duplicates → Merge

Combine into a single discovery:

• Keep the richer wording (more detail, more context)
• Keep the stronger trust: source: user > source: interaction > source: exploration
• Keep the stronger status: verified > uncertain > refuted
• Merge Also involves: refs (union of both)
• Delete the weaker entry

Example:

BEFORE (two entries under same symbol):
- authenticate_user() returns None on expired tokens.
  _(verified, source: exploration)_
- When the token is expired, authenticate_user silently returns None
  instead of raising. 3 of 7 callers don't check.
  _(verified, source: exploration)_

AFTER (merged):
- authenticate_user() silently returns None on expired tokens instead
  of raising. 3 of 7 callers don't check the return value.
  _(verified, source: exploration)_

Near-Duplicates → Absorb

The broader discovery absorbs the narrower one:

• Expand the broader entry to include any extra detail from the narrower
• Delete the narrower entry
• Preserve the stronger trust/status between the two

Conflicts → Investigate

When two discoveries contradict each other:

1. Read the actual source code at the file::symbol location
2. Trace the logic to determine which claim is correct
3. If needed, write and run a short test script to verify
4. Mark the correct claim verified, the incorrect one refuted
5. If the incorrect claim was once true but code changed, update it

to reflect the current behavior and mark it verified

If investigation takes more than a few minutes without resolution:

• Keep both discoveries
• Add a note: (conflict unresolved — needs user input)
• Ask the user to clarify at the end of the meditate session

Cross-Scope Duplicates → Place Correctly

When the same discovery exists in both a per-file shadow and _cross/:

• If it involves 3+ files → keep in _cross/, remove from per-file
• If it involves 1-2 files → keep in per-file, remove from _cross/
• Update cross-references in both directions after moving

When a per-file discovery duplicates a _prefs.md entry:

• If it's truly project-wide (not tied to a specific symbol) → keep

in _prefs.md, remove from per-file

• If it's specific to that symbol but happens to match a pref → keep

both (they serve different purposes)

Phase 3: Report

After all resolutions, print a summary:

Meditate Summary
================
  Files scanned:       42
  Duplicates merged:   7
  Near-dupes absorbed: 3
  Conflicts resolved:  2
  Conflicts escalated: 1
  Cross-scope fixed:   2
  Total entries removed: 12

Escalated (needs your input):
  src/auth.py::validate_token
    - "raises ValueError on invalid format" vs "returns False on invalid format"
    Both claims have evidence. Which behavior is correct?

Parallelism

For large shadows (20+ files), use parallel subagents:

1. Partition shadow files into batches of ~10 files each
2. Launch one subagent per batch for Phase 1 (scan)
3. Each subagent outputs JSON-per-line findings (see Scan Output Format)
4. Orchestrator collects all JSON lines from all subagents
5. Auto-apply merge actions: use edit tool with remove as old_str,

replace keep with merged

1. Auto-apply conflict actions where resolution is verified_a or

verified_b — mark the loser refuted

1. Collect escalate items for user review
2. Run Phase 3 (report)

For smaller shadows, run everything in a single pass — the scan output format is still useful for traceability.

Dream Archive Hygiene (_dreams/)

Meditate performs consistency checks and field repair on _dreams/. It does NOT delete or rewrite report prose (those are historical records), but it DOES fix missing/wrong metadata in the index.

Structural Checks

1. Index consistency — verify every folder in _dreams/ has a row in

_dreams/_index.md, and every row in the index has a matching folder. Fix mismatches (add missing rows, remove orphaned rows).

1. Report completeness — each _dreams// should contain at minimum

a report.md. Flag any empty directories.

1. Stale patches — if base_commit in a report's frontmatter is more

than 100 commits behind current HEAD, add a note to the index: ⚠️ patch may not apply cleanly. Check with: ``bash git rev-list ..HEAD --count 2>/dev/null ``

1. Cross-reference integrity — if a per-file discovery has a

Dream report: _dreams// reference, verify that dream folder exists. Remove dangling references.

Index Field Repair (run after Structural Checks)

Dream subagents sometimes write incomplete index rows (e.g., unknown category/verdict, generic titles). meditate-repair.py (below) auto-resolves these by reading each experiment's report.md frontmatter, manifest.json, and verdict-section signals.

Do NOT hand-edit 50+ rows. Use the script. The agent's job is to surface what the script CAN'T auto-fix:

• Corrupted reports — when report.md's dream_id (frontmatter or

body) doesn't match the folder name, the report was copy-pasted from another experiment. The script prints these to stderr and skips them. Do NOT auto-fix corruption — the content is wrong, not just the ID. Log them in the meditate summary for user review.

• Ambiguous parent branches — when no manifest.json exists for an

experiment, the script falls back to slug heuristics (-extend, -fix, -deeper, -improve, -integration, -cleanup, -metrics, -remaining strongly suggest compounding from a sibling). If the heuristic match is not high-confidence, flag for user review rather than guessing.

Applying Index Repairs

SKILL_DIR=""
for DIR in .github/skills/shadow-frog-meditate \
           .claude/skills/shadow-frog-meditate; do
    [ -d "$DIR" ] && SKILL_DIR="$DIR" && break
done

if [ -n "$SKILL_DIR" ] && [ -x "$SKILL_DIR/meditate-repair.py" ]; then
    python3 "$SKILL_DIR/meditate-repair.py"
else
    echo "meditate-repair.py not found; falling back to manual scan." >&2
fi

What it does:

• Backs up .shadow/_dreams/_index.md to .bak first
• Detects corrupted reports (frontmatter dream_id != folder name) and

prints them to stderr — these rows are skipped, you resolve manually

• For every other row with unknown/empty category/verdict/title,

resolves the canonical value from the report's frontmatter, manifest, or verdict section signals

• Title repair replaces generic forms (raw slug, "Dream Report: ",

"Dream t##: ") with the first # H1 or ## Summary line

Verdict detection order is manifest > VERDICT_SECTION signals > whole-body signals. Dead-end signals are checked BEFORE useful signals so not useful doesn't match useful.

Idempotent — safe to rerun until output reports 0 repaired.

Do NOT delete dream reports during meditate — only the user decides what to keep or discard (via Phase 7 review or manual cleanup).

Rules

• Never delete a source: user discovery without asking — user

knowledge is the highest trust. If it conflicts with source: exploration, investigate thoroughly before concluding the user was wrong.

• Preserve Also involves: refs — when merging, take the union.
• Update _index.md after removing entries (discovery counts change).
• Update _meta/state.json — set last_update_type: "meditate".
• Don't touch _prefs.md placement unless a pref is clearly

duplicated verbatim in a per-file shadow.

Format Compliance

Every merged or rewritten discovery must exactly follow the canonical format in /shadow-frog (Discovery Format, Cross-Cutting, Preferences). Re-read both the original entries and the spec before writing — a malformed discovery is worse than a duplicate; it breaks the viewer parser and downstream agents.

Meditate-specific rules:

• When merging, take the union of labels: [...] from both entries.
• When merging, preserve every Also involves: file::symbol from both

entries (union, not intersection).

• Dream report: _dreams// references must survive the merge —

re-attach to the merged entry if either original had one.

• Status (verified/uncertain/refuted) is taken from the stronger

source: user ≻ interaction ≻ verified exploration ≻ uncertain.

• Discovery text stays behavioral, not a code summary — preserve the

more behavioral wording when entries differ in style.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: microsoft
• Source: microsoft/ShadowFrog
• License: MIT

Install and usage instructions live in the source repository linked above.