LLM Wiki

██╗     ██╗     ███╗   ███╗    ██╗    ██╗██╗██╗  ██╗██╗
██║     ██║     ████╗ ████║    ██║    ██║██║██║ ██╔╝██║
██║     ██║     ██╔████╔██║    ██║ █╗ ██║██║█████╔╝ ██║
██║     ██║     ██║╚██╔╝██║    ██║███╗██║██║██╔═██╗ ██║
███████╗███████╗██║ ╚═╝ ██║    ╚███╔███╔╝██║██║  ██╗██║
╚══════╝╚══════╝╚═╝     ╚═╝     ╚══╝╚══╝ ╚═╝╚═╝  ╚═╝╚═╝

github.com/nvk/llm-wiki

LLM-compiled knowledge bases for any AI agent. Parallel multi-agent research, collector catalogs, automated session capture, feedback curation, thesis-driven investigation, source ingestion, wiki compilation, truth-seeking audits, querying, and artifact generation. Ships as a Claude Code plugin, an OpenAI Codex plugin, an OpenCode instruction file, or a portable AGENTS.md for any other LLM agent. Obsidian-compatible.

Install · Quick Start · Sessions · Commands · How It Works · Research Modes · Thesis Research · Query Depths · Linking · Obsidian · Architecture · Nono Sandbox · Upgrade · Changelog · Credits

Changelog

v0.16.0 — Query Lite and token benchmarks. Adds a compact read-only query protocol across Claude Code, Codex, Pi, DS4, OpenCode, and portable agents; reduces Claude /wiki:query instructions by 72.54%; adds explicit $wiki-query, read-only Pi launchers, static context budgets, corpus-identity gates, and reproducible Codex/Claude/DS4 benchmark lanes. OpenCode live-model behavior remains best effort.

v0.15.0 — Topic guides. Reframes schema.md as a human-owned topic guide, adds the friendlier llm-wiki schema adopt helper, keeps schema migrate --apply as a compatibility alias, and updates librarian docs to talk about conventions proposals instead of database-style schema migrations.

v0.14.0 — Default topic guides. Adds advisory schema.md by default, safe adoption helpers for older wikis, proposal-only librarian convention updates, deterministic docs/lint checks, and clarifies the optional index/server layer as a rebuildable non-authoritative cache.

v0.12.0 — Feedback curator. Captures high-signal corrections, preferences, approvals, and plan acceptance as redacted feedback candidates under HUB/.sessions/feedback/; ignores generic acknowledgements; adds review/manual-capture/promotion workflows; and keeps promotion into topic raw/notes/ explicit.

v0.11.1 — Session helper compatibility. Small hotfix so automated session capture works on Python 3.9/macOS system Python as well as newer Python runtimes.

v0.11.0 — Automated session capture. Default-on redacted checkpoints under HUB/.sessions/, opt-out via session disable, compact context rehydration, explicit digest promotion, and a tested Codex hook bundle.

Install

Claude Code (native plugin):

claude plugin install wiki@llm-wiki

OpenAI Codex (marketplace plugin):

Install from GitHub:

codex plugin marketplace add nvk/llm-wiki
codex plugin add wiki@llm-wiki
# Start a new Codex thread, then use @wiki or type $ to select wiki-query

Install from a local checkout with the managed bootstrap helper:

./scripts/bootstrap-codex-plugin.sh --scope user --verify

Or register the local checkout manually:

codex plugin marketplace add /absolute/path/to/llm-wiki
codex plugin add wiki@llm-wiki

Canonical explicit invocation:

$wiki-query "What does the wiki say about hardware wallet threat models?"
@wiki research "hardware wallet threat models"
@wiki collect "bitcoin memes" --wiki memes-bitcoin
@wiki ingest https://example.com/article
@wiki audit --project coldcard-threat-model
@wiki session status
@wiki feedback list --unpromoted
@wiki session disable   # optional opt-out
@wiki ll "codex plugin install gotchas"

Upgrade:

codex plugin marketplace upgrade llm-wiki
codex plugin add wiki@llm-wiki

Remove:

codex plugin remove wiki@llm-wiki
codex plugin marketplace remove llm-wiki

Troubleshooting:

• codex plugin marketplace add registers the catalog; codex plugin add wiki@llm-wiki installs and enables the cached plugin non-interactively.
• Open /hooks to review and trust the bundled hooks if you want automated session capture. The @wiki skill works without hook trust.
• $wiki-query is the small, explicit, read-only skill for lookups. In Codex CLI/IDE, type $ or open /skills and select it. It never activates implicitly or changes wiki files.
• @wiki is the full research and maintenance entry point. Natural-language wiki requests can still auto-activate it.
• Restart Codex after changing config if an existing session does not pick up the new plugin state.
• If you run Codex under a sandbox wrapper like nono, see Nono Sandbox Permissions — Codex needs r+w to $HOME/.codex for plugin install.

OpenCode (instruction file):

Add to your opencode.json (project-level or ~/.config/opencode/.opencode.json for global):

{
  "instructions": ["https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md"],
  "permission": {
    "external_directory": {
      "~/.config/llm-wiki/**": "allow",
      "~/Library/Mobile Documents/com~apple~CloudDocs/wiki/**": "allow"
    }
  }
}

OpenCode fetches the URL fresh on every session start — no manual updates needed. If you prefer a local copy instead:

curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md > ~/.config/opencode/AGENTS.md

For a smaller read-only setup, use the best-effort query preset instead:

{
  "instructions": ["https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-query/SKILL.md"]
}

The OpenCode profile is sync- and budget-tested, but not tied to one model, so it does not have a provider-specific live quality gate. Treat it as a portable best-effort preset and keep OpenCode's write and shell permissions disabled for query-only sessions.

The external_directory permission is required because the wiki hub lives outside the project directory. Set the paths to match your hub location. Alternatively, use --local mode (.wiki/ in the project) to skip permissions entirely.

Web search requires export OPENCODE_ENABLE_EXA=1.

Pi (skill file, best for local models):

Pi's minimal system prompt leaves room for on-demand wiki workflows on local models. Load the full skill for research and write-capable maintenance:

pi --skill path/to/llm-wiki/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md

Invoke it as /skill:wiki-manager, or let Pi load it when the request clearly matches its description.

For fast read-only queries with Pi's currently configured provider, use the generic launcher. It disables discovery and write tools and loads the compact shared query protocol:

./scripts/pi-wiki-query

For DS4, the provider-specific launcher additionally creates an isolated Pi state directory and pins the local model settings:

./scripts/pi-ds4-wiki-query

Set PI_CLI, DS4_BASE_URL, or LLM_WIKI_PI_DS4_STATE_DIR only when your local setup differs from the defaults. Both launchers accept --dry-run to show the exact command. The equivalent generic Pi settings are:

pi \
  --append-system-prompt path/to/llm-wiki/profiles/query-lite/SKILL.md \
  --tools read,grep,find,ls \
  --no-extensions --no-skills --no-prompt-templates --no-themes

The DS4 query profile is intentionally unable to write. Switch to the full skill for ingest, research, compile, lint, or other mutating workflows. See profiles/ds4/README.md and the reproducible benchmarks/README.md DS4 lane.

Any LLM Agent (portable instruction file):

# Read-only queries: small default
cp profiles/query-lite/SKILL.md ~/your-project/AGENTS.md

# Research and maintenance: complete protocol
cp AGENTS.md ~/your-project/AGENTS.md

The query-lite profile works with agents that can read and search files. The root AGENTS.md contains the complete write-capable protocol for agents that can also edit files and search the web.

Claude-First, Multi-Runtime

Claude Code is the principal user. Keep one shared behavior layer and thin packaging layers per runtime:

• claude-plugin/ is the primary distribution target and UX surface.
• claude-plugin/skills/wiki-manager/ is the behavioral source of truth.
• plugins/llm-wiki/skills/wiki/ is the generated Codex packaging target behind @wiki.
• claude-plugin/skills/wiki-manager/references/query-lite.md is the canonical read-only query protocol.
• profiles/query-lite/SKILL.md and generated wiki-query skills expose that protocol without the full research context.
• plugins/llm-wiki-opencode/ is the OpenCode and Pi packaging target.
• .agents/plugins/marketplace.json makes the Codex plugin installable from this repo.
• AGENTS.md is the portable single-file protocol for any other LLM agent.

Query and research presets:

Sizes are checked-in UTF-8 bytes, not provider token estimates. Use query mode for lookups, evidence checks, lists, and inventory status. Use the full profile for research, ingestion, compilation, lint fixes, or any workflow that writes.

Both runtime mirrors are generated, not hand-maintained. Rebuild from the Claude source of truth:

./scripts/sync-codex-plugin.sh      # regenerates plugins/llm-wiki/
./scripts/sync-opencode-plugin.sh   # regenerates plugins/llm-wiki-opencode/
./scripts/sync-query-lite-profile.sh # regenerates profiles/query-lite/

Each sync script:

• copies claude-plugin/skills/wiki-manager/SKILL.md into the target tree and reapplies a small list of runtime-specific wording patches
• copies references/ from the Claude source — references are runtime-neutral and shared verbatim (previously a symlink, now a real copy so Codex marketplace caching works)
• generates compact wiki-query packages from the canonical query-lite reference
• (Codex only) recreates agents/openai.yaml for Codex UI metadata and syncs the plugin version

Drift is caught by ./tests/test-codex-sync.sh and ./tests/test-opencode-sync.sh, which run the sync scripts and fail (with self-healing fix instructions) if the generated directories differ from HEAD.

Practical rule: design workflows first for Claude commands and behavior, but keep the underlying knowledge model and references runtime-neutral. Runtime wrappers adapt invocation and metadata, not wiki logic.

Nono Sandbox Permissions

If you run any AI coding agent inside a nono sandbox, the wiki needs filesystem access beyond the default profile.

Claude Code / OpenCode

{
  "extends": "claude-code",
  "policy": {
    "add_allow_read": [
      "$HOME/.config/llm-wiki"
    ],
    "add_allow_readwrite": [
      "$HOME/Library/Mobile Documents/com~apple~CloudDocs/wiki"
    ]
  }
}

Replace "extends": "claude-code" with "opencode" for OpenCode.

Codex

Codex needs r+w to its own $HOME/.codex directory for plugin install, marketplace cache, state, and skill registration:

{
  "extends": "codex",
  "policy": {
    "add_allow_read": [
      "$HOME/.config/llm-wiki"
    ],
    "add_allow_readwrite": [
      "$HOME/.codex",
      "$HOME/Library/Mobile Documents/com~apple~CloudDocs/wiki"
    ]
  }
}

Path reference

Hub resolution and ~/wiki

Hub resolution checks ~/.config/llm-wiki/config.json first and only falls back to ~/wiki when no config exists. If your wiki lives on iCloud or any non-default path, set the config and you don't need $HOME/wiki in the sandbox at all:

# Set once — agents will resolve from config, never touch ~/wiki
/wiki config hub-path "~/Library/Mobile Documents/com~apple~CloudDocs/wiki"

Use a portable hub_path with ~ for iCloud-shared hubs. Older configs may contain resolved_path; llm-wiki treats it as a fallback cache because an absolute /Users/<name>/... path from one Mac will not be valid on another. The shared wikis.json should likewise store hub topic paths as topics/<slug>, not absolute user-home paths.

If you prefer ~/wiki as a symlink to iCloud, nono's Seatbelt follows symlinks — the target path must be allowed, not the symlink itself.

Diagnosing access issues

Without the right permissions, Seatbelt or macOS privacy controls can block file access. A useful diagnostic pattern is: stat succeeds for the iCloud wiki path, but reading wikis.json or listing topics/ fails with Operation not permitted. That means the configured hub_path is correct; grant Full Disk Access or iCloud Drive access to the exact app launching the agent, restart it, and do not switch to a machine-local resolved_path or ~/wiki fallback. Use nono why to diagnose sandbox rules:

nono why --path ~/.config/llm-wiki --op read
nono why --path ~/Library/Mobile\ Documents/com~apple~CloudDocs/wiki --op readwrite

OpenCode also needs the external_directory permission in opencode.json (see Install) — nono and OpenCode have independent sandboxes that both need the same paths allowed.

Upgrade

Agents and sandboxed sessions should use GitHub CLI web login with HTTPS git transport, not SSH. This avoids SSH host-key prompts and known_hosts writes inside nono:

gh auth login --web --git-protocol https
gh auth setup-git

Claude Code — if claude plugin update pulls the latest correctly:

git -C ~/.claude/plugins/marketplaces/llm-wiki remote set-url origin https://github.com/nvk/llm-wiki.git
claude plugin update wiki@llm-wiki
# Restart Claude Code to apply

If the update command doesn't pick up the new version (stale marketplace cache), sync manually from the repo:

# Clone or pull the latest
git clone https://github.com/nvk/llm-wiki.git  # or: git -C ~/llm-wiki pull

# Sync plugin files to Claude Code's plugin cache
REPO=~/llm-wiki/claude-plugin
DEST=~/.claude/plugins/cache/llm-wiki/wiki
VERSION=$(grep '"version"' "$REPO/.claude-plugin/plugin.json" | grep -o '[0-9.]*')
rm -rf "$DEST"/*
mkdir -p "$DEST/$VERSION"
cp -R "$REPO/.claude-plugin" "$REPO/commands" "$REPO/skills" "$DEST/$VERSION/"

# Restart Claude Code to apply

Codex — upgrade from the marketplace:

codex plugin marketplace upgrade llm-wiki
codex plugin add wiki@llm-wiki

OpenCode — if using the GitHub URL in instructions, updates are automatic (fetched every session). If using a local copy:

curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/plugins/llm-wiki-opencode/skills/wiki-manager/SKILL.md > ~/.config/opencode/AGENTS.md

AGENTS.md — just pull the latest and replace:

curl -sL https://raw.githubusercontent.com/nvk/llm-wiki/master/AGENTS.md > ~/your-project/AGENTS.md

Check your installed version:

• Claude Code: look for the version in /wiki status output or check ~/.claude/plugins/installed_plugins.json
• Codex: run ./scripts/verify-codex-plugin.sh --scope user and confirm the installed cache path and marketplace source match this repo

New to a topic? One command, from anywhere:

/wiki:research "gut microbiome" --new-topic --min-time 1h

Creates a topic wiki, launches parallel agents, and keeps researching for an hour — drilling into subtopics each round finds. Come back to a fully compiled wiki.

Quick Start

/wiki:research "nutrition" --new-topic            # Create wiki + research in one shot
/wiki:research "gut-brain axis" --wiki nutrition   # Add more research to existing wiki
/wiki:research "fasting" --deep --min-time 2h     # 8 agents, keep going for 2 hours
/wiki:research "keto" --retardmax                 # 10 agents, max speed, ingest everything
/wiki:research "What makes long form articles go viral?" --new-topic  # Question → decompose → playbook
/wiki:thesis "fiber reduces neuroinflammation via SCFAs"  # Thesis-driven: evidence for + against → verdict
/wiki:thesis "cold exposure upregulates BDNF" --min-time 1h  # Deep thesis investigation
/wiki:collect "bitcoin memes" --wiki memes-bitcoin  # Find, dedupe, download media, catalog, and optionally inventory artifacts
/wiki:collect "bitcoin memes" --scale medium --media reference --inventory corpus  # Catalog media without binary downloads
/wiki:query "How does fiber affect mood?"         # Ask the wiki
/wiki:query "compare keto and mediterranean" --deep  # Deep cross-referenced answer
/wiki:query --resume                              # Where did I leave off?
/wiki add https://example.com/article             # Fuzzy router detects URL → ingest
/wiki what do we know about CRISPR?               # Fuzzy router detects question → query
/wiki:ingest https://example.com/article          # Manually ingest a source
/wiki:ingest --inbox                              # Process files dropped in inbox/
/wiki:ingest-collection https://github.com/bitcoin/bips --wiki bitcoin  # Bulk import spec repos
/wiki:ingest-collection https://dump.bitcoin.it/dump_20260429_en.xml.bz2 --wiki bitcoin  # Import MediaWiki dumps
/wiki:ingest-collection messages.csv --adapter csv-messages --wiki bitcoin  # Split message archives
/wiki:ingest-collection "https://example.com/*" --adapter wayback-cdx --from 20100101 --to 20200101  # Import archived snapshots
/wiki:inventory add ingest-candidate "Bitcointalk archive" --wiki bitcoin  # Track source queues and next actions
/wiki:inventory add item "TRX-4M ring and pinion" --wiki trx4m-1-18  # Track actual parts, tools, hosts, or assets
/wiki:inventory list --view actions --limit 10   # Compact chat table of current inventory next actions
/wiki:inventory scan-outputs --dry-run          # Preview queues/backlogs before any inventory pivot
/wiki:dataset add "Bitcointalk Temporal Graph" --location https://figshare.com/articles/dataset/BitcoinTemporalGraph/26305093 --wiki bitcoin  # Index data that stays external
/wiki:dataset list --view schema --limit 10      # Compact chat table of dataset schema/readiness state
/wiki:dataset scan-outputs --dry-run            # Find legacy data reports that could become dataset manifests
/wiki:archive topic old-interest --reason "No longer active"  # Preserve a topic but hide it from normal context
/wiki:archive list --archived                   # Show active and archived topic wikis
/wiki:archive restore old-interest              # Bring an archived topic back
/wiki:compile                                     # Compile any unprocessed sources
/wiki:audit --project gut-brain-playbook          # Truth-seeking audit across outputs + wiki + fresh research
/wiki:output report --topic gut-brain             # Generate a report
/wiki:output slides --retardmax                   # Ship a rough slide deck NOW
/wiki:assess /path/to/my-app --wiki nutrition     # Gap analysis: repo vs wiki vs market
/wiki:lint --fix                                  # Clean up inconsistencies

When working from this repository checkout, you can also run deterministic checks without an agent:

./scripts/llm-wiki lint /path/to/wiki
./scripts/llm-wiki lint --fix /path/to/wiki
./scripts/llm-wiki schema status /path/to/wiki
./scripts/llm-wiki schema adopt /path/to/wiki          # add starter topic guide
./scripts/llm-wiki schema adopt --dry-run /path/to/wiki # preview without writing
./scripts/llm-wiki archive --hub /path/to/hub topic old-interest --reason "No longer active"
./scripts/llm-wiki archive --hub /path/to/hub list --archived
./scripts/llm-wiki archive --hub /path/to/hub restore old-interest
./scripts/llm-wiki-session --hub /path/to/hub status
./scripts/llm-wiki-session --hub /path/to/hub disable   # optional opt-out
./scripts/llm-wiki-session --hub /path/to/hub enable --mode balanced --tool-events 50
./scripts/llm-wiki-session --hub /path/to/hub rehydrate --cwd "$PWD"
./scripts/llm-wiki-session --hub /path/to/hub feedback list --unpromoted

This local helper covers structural checks and safe migrations that do not require an LLM. The agentic /wiki:lint workflow remains the full protocol for editorial and deep verification passes. The local schema helper creates only a starter advisory topic guide (schema.md); run librarian conventions advice for established wikis before adopting topic-specific vocabulary. The local archive helper performs the deterministic folder move plus wikis.json, hub index, and log updates.

Session Memory

Long agent runs can lose useful context when a chat compacts, a terminal closes, or work moves between runtimes. LLM Wiki keeps a separate operational-memory layer under HUB/.sessions/ (or .wiki/.sessions/ for a local wiki) so a later turn can recover the thread without copying full transcripts into the knowledge base.

/wiki:session status
/wiki:session list --limit 10
/wiki:session show codex:abc123
/wiki:session rehydrate --cwd "$PWD"
/wiki:session promote codex:abc123 --topic meta-llm-wiki
/wiki:feedback list --unpromoted
/wiki:feedback promote fb-abc123 --topic meta-llm-wiki

This harness-session layer is distinct from .research-session.json and .thesis-session.json crash recovery, and from .session-events.jsonl and .session-checkpoint.json workflow provenance. See the session reference for the storage layout, privacy defaults, capture modes, and adapter contract.

Commands

All commands accept --wiki <name> to target a specific topic wiki and --local to target the project wiki. Archived topic wikis are skipped by default; commands that support --include-archived require that explicit flag before reading or writing archived material. Commands that generate content (query, output, plan) also accept --with <wiki> to load supplementary wikis as cross-wiki context — e.g., --with article-writing applies writing craft knowledge when generating output from a domain wiki.

/wiki:librarian is the focused wiki-maintenance tool. /wiki:audit is broader and may perform fresh research to decide whether the current knowledge or generated outputs are still trustworthy.

How It Works

Architecture

~/wiki/                                 # Hub — lightweight, no content
├── wikis.json                          # Registry of all topic wikis
├── _index.md                           # Lists topic wikis with stats
├── log.md                              # Global activity log
├── .sessions/                          # Optional session capture + feedback candidates
└── topics/                             # Each topic is an isolated wiki
    ├── nutrition/                      # Example topic wiki
    │   ├── .obsidian/                  # Optional Obsidian vault config
    │   ├── inbox/                      # Drop zone for this topic
    │   ├── inventory/                  # Lazy: durable tracking records + derived views
    │   ├── datasets/                   # Lazy: manifests for large/external data
    │   ├── raw/                        # Immutable sources
    │   ├── wiki/                       # Compiled articles
    │   │   ├── concepts/
    │   │   ├── topics/
    │   │   └── references/
    │   ├── output/                     # Generated artifacts
    │   ├── _index.md
    │   ├── config.md
    │   ├── schema.md                  # Human-owned topic guide
    │   └── log.md
    ├── woodworking/                    # Another topic wiki
    ├── .archive/                       # Archived topic wikis, hidden by default
    └── ...

The hub is just a registry — no content directories, no .obsidian/. All content lives in topic sub-wikis with isolated indexes, articles, and a human-owned advisory topic guide (schema.md). Init creates the core wiki skeleton first; optional inventory and dataset layers are created when you use them. Queries stay focused. The multi-wiki peek finds overlap across topics when relevant.

The Flow

1. Research a topic — parallel agents search the web, ingest sources, and compile articles in one command
2. Ingest additional sources — URLs, files, text, tweets (via Grok MCP), or bulk via inbox
3. Collect catalogs of discoverable things — examples, media, memes, tools, projects, entities, or source candidates — then optionally inventory them
4. Inventory items, candidates, entities, corpora, watch lists, and next actions that should persist; the agent tells you when inventory is the wrong layer
5. Index datasets that are too large for markdown — manifests, profiles, samples, and query recipes
6. Archive whole topic wikis that should stay preserved but quiet
7. Compile raw sources into synthesized wiki articles with cross-references and confidence scores
8. Query the wiki — quick (indexes), standard (articles), or deep (everything active, archived indexes separated)
9. Session capture — automatically preserve redacted Codex/Claude/OpenCode/Gemini checkpoints under .sessions/ and rehydrate future turns
10. Feedback curator — capture reviewable correction/preference/approval candidates under .sessions/feedback/ and promote only what matters
11. Lessons learned — extract knowledge from the current session (errors, fixes, gotchas) into the wiki
12. Assess a repo against the wiki — gap analysis: what aligns, what's missing, what the market offers
13. Lint for consistency — broken links, missing indexes, orphan articles, archive registry drift
14. Output artifacts — summaries, reports, slides — filed back into the wiki

Key Design

• One topic, one wiki — each research area gets its own sub-wiki with isolated indexes. No cross-topic noise.
• Canonical collection slugs — collection families use kind-first topic names such as memes-bitcoin, memes-ethereum, or tools-bitcoin so related catalogs group naturally as the hub grows.
• Parallel research agents — 5 standard, 8 deep, 10 retardmax. Each agent searches from a different angle.
• Collector workflow — search-driven catalogs for objects, media, and examples; saves a provenance map first, then inventories only the durable subset.
• _index.md navigation — every existing wiki-managed directory has an index. Claude reads indexes first, never scans blindly.
• Articles are synthesized, not copied — they explain, contextualize, cross-reference.
• Raw is immutable — once ingested, sources are never modified.
• Multi-wiki aware — queries peek at sibling wiki indexes for overlap.
• Archive-aware — archived topic wikis stay preserved under topics/.archive/ but are hidden from default query/compile/research/collect/output and maintenance workflows.
• Dual-linking — both [[wikilinks]] (Obsidian) and standard markdown links on every cross-reference. Works everywhere.
• Confidence scoring — articles rated high/medium/low based on source quality and corroboration.
• Structural guardian — auto-checks wiki integrity after operations, fixes trivial issues silently.
• Activity log — log.md tracks every operation, append-only, grep-friendly.
• Opinionated inventory — durable tracking gets records; one-off sources stay ingest/query; large row-like data becomes datasets or collection ingests. Big pivots start with a sample table before records are written.
• Media-safe catalogs — bounded public binary media is cached under output/assets/collect-<slug>/ by default; raw sources remain textual evidence, and large media sets become dataset manifests. Bulk downloads use timeouts, file-size caps, content-type checks, and IPv4 retry when needed.
• Feedback is candidate memory — corrections, preferences, approvals, and plan acceptance are saved as redacted candidates under .sessions/feedback/; generic acknowledgements are ignored, and promotion into topic notes is explicit.
• Zero dependencies — runs entirely on built-in tools (Claude Code, OpenCode, or Codex).

Research Modes

Smart input detection — /wiki:research auto-detects whether you're passing a topic or a question:

Question mode produces a playbook (actionable output artifact) and suggests testable theses derived from the findings.

Modifiers (combine with any mode):