knowledge-explorer

Manages entries in research/ — a knowledge base of verified research on tools, libraries, and frameworks. The script is research/knowledge-explorer.py (PEP 723, run via uv run).

KB root: research/ relative to repo root. Each entry is a .md file inside a category subdirectory (e.g., research/agent-frameworks/agno.md).

Script invocation

uv run research/knowledge-explorer.py [--verbose] <command> [args]

--verbose / -v prints tracebacks on error.

Commands

list

Browse all KB entries grouped by category with freshness metadata.

uv run research/knowledge-explorer.py list [--layer 0|1|2]

Options:

• --layer / -l: Filter by SDLC layer (0=process, 1=language, 2=stack). See [plugins/development-harness/docs/sdlc-layers/](../../../plugins/development-harness/docs/sdlc-layers/).

Observed output (2026-02-22, truncated):

research/
+-- agent-frameworks/  (8 entries, 0 overdue)
|   +-- agno.md  v2.4.7  · verified 2026-01-31  · review 2026-05-01
|   +-- bmad-method.md  6.0.0-Beta.4  · verified 2026-02-01  · review 2026-05-01
+-- developer-tools/  (19 entries, 0 overdue)
|   +-- github-cli.md  v2.64.0  · verified 2026-02-20  · review 2026-05-20

Each entry shows: filename, version, verified date, next-review date, and [OVERDUE] when next_review < today. Categories show total entry count and overdue count. Running the script with no subcommand also invokes list.

show-template

Print the skill-spec frontmatter template for new entries.

uv run research/knowledge-explorer.py show-template

Observed output (2026-02-22):

---
name: kebab-case-identifier
description: >-
  Verified reference for <topic>. Use when configuring or working with
  <topic> in deployments. Max 1024 chars.
license: MIT
metadata:
  topic: kebab-case-identifier
  category: category-dir-name  # one of: agent-frameworks, agent-infrastructure, ...
  source_url: https://...
  github: owner/repo
  version: "1.0.0"
  verified: "2026-02-22"
  next_review: "2026-05-23"
  tags: "tag1,tag2"
---

# Display Name

[Body content here]

Use this template as the starting point for any new entry before passing it to add.

Valid categories (verified from source lines 59-86):

agent-frameworks, agent-infrastructure, ai-design-tools, ai-observability,
ai-research-tools, ai-writing-tools, api-frameworks, async-libraries,
code-auditing, coding-agents, context-management, data-infrastructure,
developer-tooling, developer-tools, documentation-tools, evaluation-testing,
installer-tools, llm-infrastructure, low-code-platforms, mcp-ecosystem,
ml-infrastructure, python-runtimes, research-agent-patterns, rust-python-bindings,
skill-generation-tools, task-management

fetch-github

Fetch README and docs/ from a GitHub repository via the gh CLI and produce a draft KB entry.

uv run research/knowledge-explorer.py fetch-github <owner/repo> [--output <path>] [--category <cat>]

Requires: gh CLI installed and authenticated (gh auth login).

Options:

• --output / -o PATH — write draft to file instead of printing to stdout
• --category / -c TEXT — override inferred category (must be a valid category name)

Observed example (2026-02-22):

uv run research/knowledge-explorer.py fetch-github anthropics/claude-code \
  --output /tmp/test-fetch.md
# Output: Draft written to /tmp/test-fetch.md

The produced draft (observed):

---
name: claude-code
description: Claude Code is an agentic coding tool that lives in your terminal...
metadata:
  topic: claude-code
  category: UNCATEGORIZED
  source_url: https://github.com/anthropics/claude-code
  github: anthropics/claude-code
  version: "v2.1.50"
  verified: "2026-02-22"
  next_review: "2026-05-23"
---

# claude-code

> Claude Code is an agentic coding tool ...

<!-- DRAFT: Review category and tags, then run: ./knowledge-explorer.py add <this-file> -->

[README content follows]

What it fetches (verified from source lines 361-408):

• Repository metadata via gh api repos/{slug}
• Latest release tag via gh api repos/{slug}/releases/latest (404 = no releases, silently skipped)
• Root directory listing to detect docs/ or doc/ subdirectory
• README content via gh api repos/{slug}/readme (base64-decoded)
• If docs dir found: lists files in comment block inside the draft

Category inference: Matches GitHub repo topics against VALID_CATEGORIES; falls back to UNCATEGORIZED when no match. Always review and correct category before running add.

Workflow after fetch-github:

1. Review draft: check category, tags, description
2. Edit body if needed
3. Run: uv run research/knowledge-explorer.py add <draft-file>

add

Route a frontmatter entry file to the correct category directory and update README.md.

uv run research/knowledge-explorer.py add <file>
# or pipe from stdin:
cat entry.md | uv run research/knowledge-explorer.py add

What it does (verified from source lines 1774-1847):

1. Reads file (or stdin if omitted)
2. Validates format is frontmatter (not inline-header)
3. Parses entry and validates all required fields
4. Auto-generates description from body first paragraph if description is empty (warns)
5. Validates topic slug (1-64 chars, lowercase alphanumeric + hyphens, no leading/trailing/consecutive hyphens)
6. Validates category is in VALID_CATEGORIES
7. Checks for topic conflicts at target path
8. Writes entry to research/<category>/<topic>.md
9. Updates research/README.md table (warns if update fails, does not abort)

Required frontmatter fields: name, description, metadata.topic, metadata.category, metadata.sourceurl, metadata.verified, metadata.nextreview

Exit codes: 0 = success, 1 = parse/write error, 2 = validation error (invalid topic slug or category)

<example>

# Typical add workflow
uv run research/knowledge-explorer.py fetch-github some-org/some-tool \
  --output /tmp/some-tool-draft.md
# Edit /tmp/some-tool-draft.md: set correct category, review description
uv run research/knowledge-explorer.py add /tmp/some-tool-draft.md

</example>

update-append

Append a dated update section to an existing KB entry. Opens $EDITOR for the update content.

uv run research/knowledge-explorer.py update-append <topic-slug>

What it does (verified from source lines 1719-1766):

1. Searches all KB files for the entry matching the topic slug
2. Migrates entry to frontmatter format in-place if it was inline-header
3. Opens $EDITOR with placeholder <!-- Replace this with your update content -->
4. If editor closed with unchanged placeholder or empty content: aborts with exit 0
5. Appends a dated section: ## Update: YYYY-MM-DD\n\n<content>
6. Updates verified to today and next_review to today + 90 days
7. Writes updated entry atomically

Editor interaction: This command requires interactive terminal access. It uses typer.edit() which invokes $EDITOR (or system default). When running in a non-interactive context (e.g., scripted agent loop), set EDITOR to a script that writes content to the temp file programmatically, or use VISUAL.

Topic not found: If slug is not found, suggests up to 3 alternatives using Levenshtein distance <= 2.

<example>

uv run research/knowledge-explorer.py update-append agno
# Opens $EDITOR → write update content → save → exits
# Result: ## Update: 2026-02-22\n\n<content> appended to agent-frameworks/agno.md

</example>

generate-descriptions

Generate or repair frontmatter description fields for KB entries that are missing, empty, truncated, or auto-generated from prose.

When to run: User asks to "fix descriptions", "generate descriptions", or "repair KB descriptions", or the add command warns about an empty description.

Orchestrator workflow:

# Step 1 — get only the entries that need fixing (JSON array)
uv run research/knowledge-explorer.py list-candidates

Output is a JSON array of entries that fail the bad-description heuristics. Empty array means nothing to do. Each object contains topic, filepath, name, category, tags, currentdescription, body.

# Step 2 — spawn one Haiku subagent per entry in parallel
# Pass each entry object + the description rules below to the agent.
# The agent generates the description and writes it directly:
uv run research/knowledge-explorer.py set-description <topic> "<description>"
# Exit 0 = success (silent). Exit 1 = topic not found. Exit 2 = invalid description.

The orchestrator only hears back from an agent on non-zero exit. Generated descriptions never route through the orchestrator.

--all flag: include every entry regardless of current description quality.

uv run research/knowledge-explorer.py list-candidates --all

Description rules (pass verbatim to each subagent):

• Single line, no newlines
• Max 1024 characters
• Front-load what the KB covers, then when an agent should load it
• Include specific trigger keywords (tool name, alternatives, use cases)
• Format: {what it covers}. Use when {trigger scenarios}.
• Output ONLY the description string — no quotes, no YAML, no explanation
• After writing, validate with uvx skilllint@latest check --fix <file>

<example>

Subagent receives:

{
  "topic": "dasel",
  "name": "dasel",
  "category": "developer-tools",
  "tags": ["cli", "json", "yaml", "toml", "csv", "xml"],
  "body": "Dasel (Data Selection) is a command-line tool and Go library..."
}

Subagent generates and writes:

uv run research/knowledge-explorer.py set-description dasel "Dasel CLI and Go library for reading and writing JSON, YAML, TOML, CSV, and XML via a unified selector syntax. Use when querying or transforming structured config files, scripting data pipelines, or replacing jq/yq with a single multi-format tool."

</example>

migrate

Migrate old-format entries to skill-spec frontmatter in-place across the entire KB.

uv run research/knowledge-explorer.py migrate [--dry-run]

Options:

• --dry-run — show what would change without writing (safe to run at any time)
• --all — migrate all entries (default; no effect to omit)

Source formats handled (verified from source lines 1956-2008):

• inline-header: # Heading + bold/table field pairs before ## Body — converted to skill-spec frontmatter
• flat frontmatter: top-level KB fields (topic, name, etc.) — moved into metadata sub-mapping
• skill-spec frontmatter: already has metadata.topic — skipped

Summary output (from source lines 1998-2007):

Migrated: N  Already done: M  Failed: P
  FAILED relative/path.md: reason

Run with --dry-run first to preview changes before committing.

Frontmatter schema

Skill-spec format (canonical, written by all write operations):

---
name: kebab-case-tool-name          # top-level: Agent Skills spec name
description: "..."                  # top-level: max 1024 chars
license: MIT                        # top-level: optional SPDX identifier
metadata:                           # KB tracking fields
  topic: kebab-case-tool-name       # kebab-case slug; must match filename stem
  category: developer-tools         # must be in VALID_CATEGORIES
  source_url: https://...           # primary reference URL
  github: owner/repo                # optional: owner/repo slug only
  version: "1.2.3"                  # optional: quoted string
  verified: "2026-02-22"            # ISO date string, quoted
  next_review: "2026-05-23"         # ISO date string, quoted; 90 days after verified
  tags: "tag1,tag2"                 # optional: comma-separated, quoted
---

Key constraints (verified from source lines 1321-1340, 1467-1503):

• topic slug: 1-64 chars, [a-z0-9][a-z0-9-]*[a-z0-9] or single char, no --
• description max: 1024 chars
• category must be exact match in VALID_CATEGORIES
• Date fields must be quoted strings (not YAML date scalars) — the script handles this via DoubleQuotedScalarString

Error handling

• ExternalCommandError: gh not on PATH or returned non-zero. Hint shown in error panel. Fix: gh auth login or install gh CLI.
• TopicNotFoundError: topic slug not found; up to 3 Levenshtein suggestions shown.
• TopicConflictError: target path exists with a different topic slug.
• ParseError: file format unrecognisable or required field missing.
• FrontmatterValidationError: frontmatter present but fails schema (missing required fields).

All errors render as a Rich panel on stderr. Use --verbose to include full traceback.

Source reference

Script: research/knowledge-explorer.py (1911 lines, verified 2026-02-24)

Key line ranges:

• Constants and valid categories: lines 46-88
• fetch-github command: lines 1637-1711
• update-append command: lines 1719-1766
• add command: lines 1774-1847
• migrate command: lines 1956-2008
• Frontmatter schema serializer: lines 1261-1300
• Name validation rules: lines 1321-1340