菜单
Curate the web-capture index. Use when the capture backlog grows, captures sit unprocessed at seedling/pending, or to surface stored research during work.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/athola/claude-night-market --skill palace-index-curator命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Turn hunches into accepted results: worthiness score, evidence bar, research-to-rules. Use when vetting ideas. Not for QA; use night-market-validation-and-qa.
Generates and self-executes a diff-derived test plan for a PR. Use when validating PR changes before merge. Do not use for code review; use sanctum:pr-review.
Shapes turns action-first: action leads, steps numbered, state restated. Use for ADHD-friendly output. Do not use to trim tokens; use response-compression.
基于 SOC 职业分类
正在显示 SKILL.md
| name | palace-index-curator |
| description | Curate the web-capture index. Use when the capture backlog grows, captures sit unprocessed at seedling/pending, or to surface stored research during work. |
| alwaysApply | false |
| category | governance |
| tags | ["knowledge-management","capture-index","curation","promotion","analytics"] |
| dependencies | ["memory_palace.corpus.index_analytics","memory_palace.corpus.index_promoter"] |
The web-research hooks auto-capture every WebFetch and WebSearch into
hooks/memory-palace-index.yaml, storing each as a markdown file and an
index entry. Captures land at the defaults routing_type: pending,
maturity: seedling, importance_score: 50, and nothing advances them.
Left alone, the index becomes a write-only graveyard: the majority of
entries are never incorporated, analyzed, or surfaced.
This skill drains that backlog and keeps it drained. It wires the capture index to the corpus tooling the plugin already ships (, , ) through three commands: a read-only report, a dry-run-first promotion engine, and a SessionStart surfacing hook.
decay_modelkeyword_indexmarginal_valuepending
entries the drain held back.pending.knowledge-intake.knowledge-locator.digital-garden-cultivator.uv run python scripts/memory_palace_cli.py index report
Reports total entries, the inert ratio, orphaned captures (entries whose backing file is gone), the largest topic clusters by domain, and the top promotion candidates. Writes nothing.
# Dry run: prints promote/archive proposals, writes nothing.
uv run python scripts/memory_palace_cli.py index promote
# Apply: backs up the index under data/backups/, then persists.
uv run python scripts/memory_palace_cli.py index promote --apply
Each pending entry is classified into one action:
seedling -> growing.archived rather than promoted, following the
principle that unused captures should drain, not accumulate.pending with no change.Applying is idempotent: promoted and archived entries are no longer
pending, so a second run proposes nothing new. The dry-run diff is
always shown before --apply writes.
Running these by hand is the exception. --apply runs on every commit
from scripts/precommit_palace_maintenance.sh, so the backlog drains
continuously rather than in occasional sweeps. Reach for the commands
above when a commit is blocked, or when you want the dry-run diff before
the hook decides for you.
The commit that carries the index must carry it with zero pending
entries. scripts/check_capture_index_drained.py runs at the end of the
maintenance hook and fails the commit otherwise, and
tests/test_capture_index_artifact.py re-checks the same invariant in
CI so a bypassed hook does not land a backlog.
Two things make that gate reachable rather than a standing block:
hooks/shared/deduplication._stage_index). Without it, pre-commit
reverts the unstaged write before any hook runs, so the drain reads a
tree the fresh capture is missing from and converges on a fixed point
that excludes exactly the entries it exists to process. That is how 47
captures accumulated behind a drain that reported nothing to do.hold survives
it, so a blocked commit means a specific capture needs a person to
score or archive it. The gate names the keys.A SessionStart hook (hooks/index_surfacer.py) names the highest-value
promoted captures at the start of a session. It is disabled by default.
Enable it in memory-palace-config.yaml:
feature_flags:
context_injection: true
The hook only speaks when promoted entries clear the importance floor, and it exits silently on any error so it can never block a session.
The three steps above all operate on the capture index at
hooks/memory-palace-index.yaml. Retrieval reads a different file:
data/indexes/keyword-index.yaml, built from the staging captures and
consumed by cache_lookup. Curating one does nothing to the other.
That keyword index is derived data and is not tracked in git, so a fresh checkout has none at all. Rebuild it with:
# Report what would be indexed, writing nothing.
uv run python scripts/build_indexes.py --dry-run
# Write data/indexes/keyword-index.yaml.
uv run python scripts/build_indexes.py
The builder refuses to write an empty index over a populated one. An
empty corpus is reported with "wrote": false and any existing index
is left untouched. Writing entries: {} over real data is how the
corpus went dark in 1.5.0, and it stayed dark because the regeneration
script named in that stub file had never been written.
cache_lookup / keyword_index), and
embeddings are not required at the current corpus scale. BM25 is the
workhorse up to ~5000 documents. Embeddings add value only for
vocabulary-mismatch discovery.content_hash) then MinHash with k-shingling for near-duplicates
(Broder, 1997). SimHash is preferable only at tens of thousands of
documents.relevance = w1 * centrality + w2 * decay(t) + w3 * usage. The plugin ships all three terms (graph_analyzer
PageRank, decay_model, usage_tracker).build_indexes.py --dry-run reports a non-zero entry count and
leaves data/indexes/keyword-index.yaml byte-identical.build_indexes.py against an empty corpus reports "wrote": false and leaves an existing populated index untouched.index report runs and prints the inert ratio and orphan count
for the live index.index promote (no flag) prints proposals and writes nothing
(the index file is byte-identical afterward).index promote --apply creates a timestamped backup under
data/backups/ before persisting, and a re-run proposes nothing.context_injection: true, a SessionStart event surfaces the
top promoted captures, and with the flag off it stays silent.check_capture_index_drained.py exits 0 against the committed
index and exits 1 naming the keys when one is left pending.update_index appears in git diff --cached
without anyone staging it by hand.