Adopt: directory keeps its NN- ordering prefix; skill `name:` is the clean form
without it (dir 16-seo-schema-validator → name: seo-schema-validator). Nicer to
invoke, matches the original desktop/SKILL.md names, still globally unique.
- 71 root SKILL.md: name: NN-foo → name: foo (flat skills + reference-curator suite).
Plugins (mac-optimizer/multi-agent-guide/dintel-bootstrap) already clean; 95 already clean.
- scripts/migrate_skill_root.py: derive name = dirname minus NN- prefix (skill_name()).
- CLAUDE.md + SKILL-MIGRATION-GUIDE.md: document the dir-prefix / clean-name convention.
verify_skills.py: 0 name collisions across all renamed skills. (The ~/.claude/skills
symlinks were re-pointed to the clean names separately — filesystem only.)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Run the additive migration pass from SKILL-MIGRATION-GUIDE: generate a root SKILL.md
for every skill that lacked one, copied from its desktop/SKILL.md (or code/SKILL.md),
with name set to the directory name and description + body preserved verbatim.
- scripts/migrate_skill_root.py: the reusable, non-destructive migrator (dry-run default).
- 61 new root SKILL.md (desktop source for most; code/SKILL.md for 61/62/92).
- Untouched: 16/17/95 (already had root); desktop/ and code/ packaging left intact.
- All 64 root SKILL.md validate: frontmatter <=1024, kebab name, description present.
Still MANUAL (no SKILL.md source — commands/README only), need hand-authored root SKILL.md:
81-mac-optimizer, 90-reference-curator, 91-multi-agent-guide, 94-dintel-bootstrap.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
aiohttp was listed in requirements.txt but never imported by any of the
six Python scripts in custom-skills/31-notion-organizer/code/scripts.
The async HTTP work is done via notion-client 2.2.1, which uses httpx
internally (confirmed via `pip show notion-client`).
Removing the unused pin will stop future dependabot churn for a library
this skill doesn't depend on.
Verified before this change: aiohttp 3.13.4 (just merged via #6) was
inert — all 30 tests in test_notion_search.py pass, and all 6 scripts
import cleanly without aiohttp present.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Final review caught that both --filter example locations used simplified
JSON ({"Status": "Done"}) that Notion's data_sources.query API rejects
with a 400. The script passes --filter verbatim, so users copy-pasting
the example would hit a confusing error.
Replace with Notion's actual filter shape:
{"property": "Status", "status": {"equals": "Done"}}
Also added a compound (and/or) example in CLAUDE.md so users have a
reference for combining filters.
30/30 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Slash command at custom-skills/31-notion-organizer/commands/notion-search.md
documents the CLI surface and JSON output schema. CLAUDE.md gains a
Semantic Search section explaining the 4-stage pipeline and env var
requirements. requirements.txt notes the optional anthropic SDK
dependency (the skill falls back to the claude CLI if missing).
Final task of Phase 3b-i. 30 tests passing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Inline code review polish:
- --filter is only meaningful in per-database mode (workspace search
doesn't accept Notion filter objects). Previously a user passing
--filter without --databases would have it silently parsed and
ignored. Now emit a stderr warning and clear the filter.
- cache_kwargs dict was built twice in run_search (once for cache_get,
once for cache_put). Build once before the rerank call.
30/30 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Builds a rerank prompt with title + flattened properties + excerpt for
each candidate, calls Claude, parses JSON, sorts by score descending,
takes top N. On any failure (LLM error, missing JSON, parse error,
non-list shape), falls back to candidates in input order with null
relevance/snippet.
4 tests: ordering, limit, parse-error fallback, exception fallback.
26 passing total.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Code review polish:
- Drop the unused `name` parameter from `_flatten_property` (it was
reserved for future per-property-name special-casing but never used).
- Add a regression test pinning that checkbox=False and number=0 are
preserved in the enriched output. The existing empty-value filter is
`if value not in (None, [], "")` which keeps falsy-but-meaningful
values, but the contract wasn't tested.
22/22 tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
For each candidate page, extract the title, flatten common property
types (status/select/multi_select/date/checkbox/number/url/etc.) to
display values, and fetch the first text-bearing block as a 200-char
excerpt. Empty excerpt is acceptable when the page has no leading text.
4 tests: title+properties, paragraph excerpt, empty fallback,
truncation. 21 passing total.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Code review caught that without exception handling around
notion.search() and notion.data_sources.query(), any transient API
hiccup (rate limit, 5xx, network blip) would crash the whole search
and lose candidates already accumulated from earlier variants/DBs.
Wrap both calls in try/except, mirror the resolver's pattern: stderr
warning + continue. Same query against another variant or another DB
still has a chance to succeed.
Also tightened response.get("results", []) → response.get("results")
or [] so a hypothetical {"results": null} response doesn't crash the
inner loop.
17/17 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
For each expanded query variant: hit Notion's workspace search OR
per-database data_sources.query (when --databases is specified).
Union and dedupe by page ID, cap at 30 candidates total. Filters out
non-page objects (databases) from workspace search results.
Property filters (--filter JSON) pass through to data_sources.query
when in per-database mode.
5 tests: workspace dedup, 30-cap, DB-mode dispatch, page-only filter,
property-filter passthrough. 17 passing total.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Code review polish:
- Add stderr warning for the fourth failure path (LLM returned valid
JSON but it's not a list-of-strings). Previously this fell back
silently, making debugging harder when Haiku produces unexpected
shapes.
- Drop unused f-string prefixes from two warnings (no interpolation).
- Type hint: Callable[..., str] = None → Optional[Callable[..., str]] = None
for strict type-checker compatibility.
12/12 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Generates up to 5 query variants (synonyms + cross-language KR↔EN) so
later Notion API search can union over them. Permissive failure modes:
LLM error or non-JSON response falls back to [original] with stderr
warning. Dedupes and caps variants.
4 tests: variants list, dedup+cap, JSON-parse fallback, exception
fallback. 12 passing total.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Code review polish:
- Type hints: List → List[Dict[str, Any]] for cache_get return and
cache_put results parameter, since rerank produces dict-shaped entries.
- TTL=0 short-circuits to None deterministically. Previously the test
for ttl_seconds=0 passed only because clock motion between cache_put
and cache_get made `time.time() - cached_at > 0` true. Now the
semantics match the docstring intent.
8/8 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cache layer for rerank results. Key is SHA256 of query + sorted
candidate IDs, so changing the candidate set automatically invalidates.
1-day TTL by default; corrupted cache files are silently dropped.
5 tests cover hit, miss, different-candidates, TTL expiry, corruption.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Code review caught that the bare alias 'claude-haiku-4-5' works in the
Claude Code CLI but may not resolve in the Anthropic SDK. Pin the
default to 'claude-haiku-4-5-20251001' (the full dated ID) via a
DEFAULT_MODEL constant so the SDK path doesn't silently break the
first time it hits the real API.
Also document the max_tokens-ignored behavior on the CLI path in the
public call_claude docstring (not just the private _call_via_cli),
and drop the unused LLMCaller type alias (dead code).
3/3 tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First task of Phase 3b-i (notion semantic search). Adds _search_llm.py
that dispatches to the anthropic SDK when ANTHROPIC_API_KEY is set,
falls back to `claude -p` CLI otherwise, and raises a clear setup
error if neither works. Symlinks _notion_compat.py from 32-notion-writer
so both skills share one source of truth.
3 tests cover the three dispatch paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pipeline Orchestrator:
- Add 07-pipeline-orchestrator skill with code/CLAUDE.md and desktop/SKILL.md
- Add /reference-curator-pipeline slash command for full workflow automation
- Add pipeline_runs and pipeline_iteration_tracker tables to schema.sql
- Add v_pipeline_status and v_pipeline_iterations views
- Add pipeline_config.yaml configuration template
- Update AGENTS.md with Reference Curator Skills section
- Update claude-project files with pipeline documentation
Skill Format Refactoring:
- Extract YAML frontmatter from SKILL.md files to separate skill.yaml
- Add tools/ directories with MCP tool documentation
- Update SKILL-FORMAT-REQUIREMENTS.md with new structure
- Add migrate-skill-structure.py script for format conversion
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
