CLI commands

wdpkr is a single binary with a handful of subcommands. All output is JSON on stdout; errors go to stderr.

wdpkr <command> [options]

Command	Purpose
search	Conceptual search — returns tiered JSON
index	Build or refresh the index (full or incremental)
reinforce	Mark documents as freshly used so decay stops sinking them
delete	Remove vectors from the index by path glob
init	Set up wdpkr for a repository
config	Manage configuration
eval	Measure search quality and compression
decision	Record and manage architectural decisions (decision recall)

search

Embed a natural-language query, search the vector store, and return matching files with their top symbols.

wdpkr search "release commission payments"
wdpkr search "how is rate limiting implemented" --pretty
wdpkr search "auth flow" --scope src/auth/ -k 10

Flag	Default	Description
<query>	(required)	Natural-language query
-k, --top-k	5	Max file-level results
--symbols-per-file	3	Max symbols returned per file
--no-symbols	off	File-level results only, omit symbol nesting
--scope <path>	—	Limit search to subtree(s); repeatable; case-insensitive
--filter <glob>	—	Glob filter on result paths; repeatable (OR logic); case-insensitive
--tap <name>	all	Limit search to these tap sources (e.g. files, linear, notion, decision); repeatable. --provider is a deprecated alias
--terse	off	Compact output: paths + one-sentence summaries
--pretty	off	Human-readable output instead of JSON
--no-decisions	off	Disable decision recall: don’t search decisions or attach governed_by

--scope and --filter match paths case-insensitively, as do decision areas globs. Path matching is a comparison, not a display concern, so --filter "*.MD" matches README.md and --scope Src/auth/ finds src/auth/ — useful when a case-insensitive filesystem tab-completes a path in the wrong case.

--filter is applied to results after the vector search ranks them, so a narrow filter can match files that rank below the initial fetch window. Search widens the fetch automatically until it has -k matches, the index runs dry, or an internal cap is reached — a filtered query that comes up short costs an extra round-trip or two but does not silently return nothing.

Results from a non-files tap carry a source field (e.g. "source": "linear") and a scheme-prefixed path (e.g. linear://ENG-123). See Taps for indexing data sources beyond code, and Decay and reinforce for how per-tap age affects ranking.

When decisions exist, code results carry a governed_by array listing the architectural decisions whose areas match that file — so a search for the code also surfaces why it works the way it does. Decisions also appear as their own results ("source": "decision", path decision://<id>). --tap decision searches only decisions; --no-decisions turns the whole feature off for a query.

index

Walk the repo, chunk with tree-sitter, summarize, embed, and upsert to the vector store. Incremental by default — only changed files are reprocessed.

wdpkr index              # incremental
wdpkr index --full       # rebuild everything
wdpkr index --dry-run    # estimate cost, no API calls
wdpkr index --tap linear # index only the configured Linear tap
wdpkr index --tap notion --doc <page-id-or-url>  # index specific Notion pages

Flag	Default	Description
--full	off	Ignore the high-water mark and reindex everything
--dry-run	off	Estimate cost without API calls or writes
--concurrency <n>	4	Bound parallel file processing
--from <sha>	—	Override the starting SHA for the diff (manual recovery)
--max-cost <usd>	—	Abort if estimated remaining cost exceeds this cap
--skip-summaries	off	Re-chunk and rebuild call-graph edges with zero API calls, reusing existing vectors
--tap <name>	all	Run only this configured tap
--doc <id-or-url>	—	Target document(s) for a targeted tap (the Notion tap); repeatable. Additive — only the named documents are (re)indexed

reinforce

Mark one or more documents as freshly used so per-tap decay stops sinking them in the rankings. Bumps each document’s last_used_at to now — a cheap metadata write, no re-embedding.

wdpkr reinforce notion://<page-id>
wdpkr reinforce notion://<page-id> linear://ENG-123   # several at once

Argument	Description
<id>...	One or more document ids (a result’s path). The tap is inferred from the URI scheme; a bare path targets the files namespace

Reinforce the specs an agent actually relied on so the next search ranks them higher. Only meaningful for taps with decay enabled.

delete

Remove indexed vectors whose paths match a glob.

wdpkr delete "src/legacy/**"                  # from the files namespace
wdpkr delete --tap notion "notion://<id>*"    # from a tap's namespace

Flag / Argument	Default	Description
<pattern>	(required)	Glob matching paths to remove. For non-files taps the paths are tap URIs (e.g. notion://<id>*)
--tap <name>	files	Delete from this tap’s namespace instead of the base (files) namespace

init

Set wdpkr up in a repository — writes a CLAUDE.md section, a .wdpkrignore, and a CI workflow that reindexes on merge to main.

wdpkr init

config

Manage the configuration that the other commands resolve from. See Configuration for the full key reference.

wdpkr config init                       # write a default config file
wdpkr config list                       # show all values and their sources
wdpkr config get embedder.model         # read one value
wdpkr config set indexer.concurrency 16 # set one value
wdpkr config path                       # print the resolved config file path
wdpkr config edit                       # open the config file in $EDITOR

Subcommand	Purpose
init	Write a default config file to ~/.config/wdpkr/config.yaml
get <key>	Get a config value by dotted key
set <key> <value>	Set a config value by dotted key
list	Show all config values and their sources
path	Print the resolved config file path
edit	Open the config file in $EDITOR

eval

Run the evaluation suite to measure search quality and how much context wdpkr saves versus reading source directly.

wdpkr eval

decision

Record and manage store-native architectural decisions — authored memory that captures why the code is the way it is, surfaced back in search. Decisions live in the store’s <namespace>--decision namespace (not files on disk); their metadata rides in a registry inside that namespace. See Decision recall for the full guide.

wdpkr decision add "Half-up rounding for commission" \
  --context "Payouts were off by a cent" \
  --decision "Round half up at 2 decimals" \
  --area 'src/finance/**' \
  --tap notion --doc <page-id-or-url> \
  --supersedes 3
wdpkr decision list
wdpkr decision edit 7 --status deprecated
wdpkr decision delete 4

decision add <title>

Flag	Description
<title>	(required) Short title of the decision
--context <text>	Why the decision was needed (Context section)
--decision <text>	What was decided (Decision section)
--consequences <text>	Resulting trade-offs (Consequences section)
--area <glob>	Code path glob this decision governs; repeatable. Drives governed_by recall
--tap <name>	Pull provenance snapshots from this configured tap (e.g. notion)
--doc <ref>	Document ref(s) to pull from --tap; repeatable. Snapshot + source URI are recorded
--supersedes <id>	Decision id(s) this one replaces (marks them superseded); repeatable
--overrides <id>	Decision id(s) this one wins over in overlapping areas; repeatable
--relates-to <id>	Related decision id(s); repeatable
--author <name>	Author (defaults to git config user.name)
--status <status>	proposed | accepted | superseded | deprecated (default accepted)

decision edit <id>

Takes the same field flags as add; only the fields you pass change. --area, --supersedes, --overrides, and --relates-to replace their lists when given; --tap/--doc append new provenance. Content changes re-embed the decision.

decision delete <id>...

Delete one or more decisions by id (aliased rm) — removes their vectors and registry entries and scrubs any dangling links (supersedes/overrides/relates-to) in other decisions. Mirrors the top-level delete verb.

decision list

Print the decision registry as JSON (or --pretty for a human-readable table).