pk

name	pk
description	Disabled: pk is now invoked explicitly through the /pk prompt template. Kept only as archived documentation for the former skill workflow.
disable-model-invocation	true

pk

Structured project knowledge — intake, search, recall, and audit.

Use the pk CLI for all operations. Output is JSON by default; use --pretty for human-readable. Never read or write knowledge files directly.

MANDATORY READ references/knowledge-model.md when: creating a note type you haven't used before, unsure which folder a type belongs in, validating frontmatter, or unsure which status values are valid.

Commands

# Orient at session start
pk synthesize --session-start

# Search before creating anything
pk search "query"
pk search "query" --type decision --status open --tag infra --limit 5

# Read full content
pk read /abs/path/from/search

# Create — prints path; read skeleton, fill sections, then write back
pk new note "Title" --tags auth,api
pk new note "Title" --tags temp --expires 2026-12-31
pk new decision "Title"
pk new question "Title"
pk new source "Title"

# Write back (always pk read first; never change id, type, or created)
pk write /abs/path <<'EOF'
---frontmatter---
## Section
Content.
EOF

# Validate (run after every write)
pk lint
pk lint path1 path2

# History, tags, index
pk history --limit 20 --filter-type decision
pk vocab
pk index

# Lifecycle maintenance (run daily/weekly via cron or manually)
pk gc              # decay scoring + archive stale + purge expired
pk gc --dry-run   # preview without changes
pk gc --compress  # find similar notes for interactive merge

source vs note: source = raw input (meeting notes, transcripts, external docs). note = stable synthesised fact you've derived. Extract notes from sources; never dump raw input into a note.

Contradiction Protocol

After creating a note that supersedes existing knowledge:

1. pk search "<topic>" — find the old note
2. pk read <old-note-path> — confirm the conflict
3. pk write <old-note-path> — change status: superseded, add to ## Related: Superseded by [[<new-note-id>]] — <reason>.
4. pk lint <old-note-path> — validate

Lifecycle

Knowledge has a lifecycle. Notes decay, expire, and get superseded.

• Decay: notes not accessed fade over time. pk gc updates decay scores and archives notes below threshold.
• Expiry: notes with an expires date (frontmatter) are archived automatically when the date passes. Use --expires YYYY-MM-DD on pk new for time-bounded knowledge (deadlines, temporary blockers).
• Compression: pk gc --compress identifies similar notes for interactive merging. Run weekly.
• Contradictions: see protocol above. The agent marks old notes as superseded.

NEVER

• NEVER skip pk search before pk new. Instead: Search by topic, likely title, and key tags; update an existing note when it already captures the claim. Why: Duplicates fragment knowledge silently; future searches return noise and agents may trust the wrong copy.

• NEVER dump raw input into a note or decision. Instead: Create a source for raw material, then extract stable claims into notes or decisions. Why: Raw input and synthesized knowledge age differently; mixing them makes later recall look more authoritative than it is.

• NEVER silently overwrite a conflicting claim. Instead: Follow the Contradiction Protocol and mark older knowledge superseded with a link to the replacement. Why: Preserving the conflict trail lets future agents understand whether a change was correction, reversal, or context shift.

• NEVER commit when pk lint returns errors. Instead: Fix frontmatter/body structure and rerun pk lint before treating the note as usable. Why: Invalid notes may disappear from search or break lifecycle tooling, which is worse than having no note.