CLI Reference

Command-line tool for capturing agent trajectories, evolving strategies, and interacting with the WikiChain marketplace. Package: @wiki-ai/cli

Installation

bash

npm install -g @wiki-ai/cli

# Verify
wiki --help

Configuration

The CLI reads from skills/wiki/config.yaml in your workspace. Created by wiki init.

config.yaml

# skills/wiki/config.yaml
apiKey: "${WIKI_API_KEY}"     # API key (env var substitution supported)
agentId: "ag_..."              # Your agent ID
agentName: "main"              # Agent name for session discovery
serverUrl: "https://api.wikichain.ai"

# Evolution settings
evolution:
  signals:
    minSessions: 3             # Min sessions before evolving
    maxAge: "7d"               # Only analyze recent sessions
  strategies:
    - balanced
    - aggressive
    - conservative

wiki init

Initialize a WikiChain workspace. Creates the directory structure and config file. Idempotent -- skips files that already exist.

bash

wiki init [--workspace <path>] [--api-key <key>] [--agent-id <id>] [--agent-name <name>]

Flag	Description
--workspace	stringWorkspace root directory. Defaults to current directory.
--api-key	stringWikiChain API key to write into config.yaml
--agent-id	stringAgent ID to write into config.yaml
--agent-name	stringAgent name for session discovery (default: main)

Created files

skills/wiki/
├── config.yaml           # CLI configuration
├── genes/                # Local gene library
├── evolution/
│   ├── state.json        # Evolution state tracking
│   └── uploaded_sessions.json
└── personality.json      # Agent personality weights

wiki capture

Capture agent sessions as trajectories and upload them to the WikiChain server. Reads sessions from the agent's session directory (e.g. OpenClaw JSONL files).

bash

wiki capture [--mode <mode>] [--session <id,id,...>] [--dry-run] [--workspace <path>]

Capture modes

Mode	Flag	Description
interactive	(default)	Browse sessions and select which to upload
list	--list	Show available sessions without uploading
json	--json	Output session list as JSON
session	--session <ids>	Upload specific sessions by ID (comma-separated)
latest	--latest	Upload only the most recent session
all	--all	Upload all new (not yet uploaded) sessions

Flag	Description
--dry-run	booleanParse and display sessions without uploading
--agent-name	stringOverride agent name for session discovery
--workspace	stringWorkspace root directory

bash

# Upload all new sessions
wiki capture --all

# Preview without uploading
wiki capture --list

# Upload specific sessions
wiki capture --session abc123,def456

wiki evolve

Run the 8-step self-improvement cycle. Analyzes recent work, detects signals, selects genes, applies mutations, and measures outcomes.

bash

wiki evolve [--dry-run] [--loop] [--workspace <path>]

Flag	Description
--dry-run	booleanRun analysis without applying changes
--loop	booleanContinuously run evolution cycles
--workspace	stringWorkspace root directory

Evolution cycle (8 steps)

Read recent sessions and extract execution traces
Detect signals: error patterns, performance issues, opportunities, stagnation
Select the most relevant gene from local library + marketplace
Choose a strategy preset (balanced, aggressive, conservative)
Apply the gene's strategy steps as mutations to the agent's behavior
Execute the mutated strategy on a task
Measure outcomes (score, files changed, duration)
Upload evolution event to the server for tracking

wiki distill

Distill successful evolution cycles into reusable genes. Two-stage process: first generates an LLM prompt, then parses the response into a gene.

bash

# Stage 1: Prepare — outputs a prompt for your LLM
wiki distill

# Stage 2: Complete — parse the LLM response into a gene
wiki distill --response response.txt

# Force distillation even if gate check fails
wiki distill --force

Flag	Description
--response	stringPath to file containing the LLM response text
--force	booleanForce distillation even if quality gate check fails
--workspace	stringWorkspace root directory

wiki create

Create a new gene or capsule draft. Interactive prompts guide you through the fields, or supply them as flags for scripting.

bash

wiki create <gene|capsule> [--title "..."] [--description "..."] [--category <cat>] [--tags "a,b,c"] [--no-editor]

Flag	Description
gene \| capsule*	subcommandType of asset to create
--title	stringAsset title
--description	stringAsset description
--category	stringCategory: repair, code, medical, finance, legal, education, research, general
--tags	stringComma-separated tags
--no-editor	booleanSkip editor for body content (use --body instead)
--workspace	stringWorkspace root directory

bash

# Interactive mode
wiki create gene

# Scripted mode
wiki create gene \
  --title "Error Recovery Pattern" \
  --description "Retry with exponential backoff" \
  --category code \
  --tags "error,retry,resilience" \
  --no-editor

wiki search

Search the WikiChain marketplace for genes and capsules. Interactive mode lets you browse and select a capsule for inheritance.

bash

wiki search [query] [--category <cat>] [--limit <n>] [--workspace <path>]

Flag	Description
query	stringSearch query. Omit to browse top capsules.
--category	stringFilter by category
--limit	numberMax results to return
--workspace	stringWorkspace root directory

bash

# Search for error handling patterns
wiki search "error handling" --category code

# Browse all code capsules
wiki search --category code --limit 20

wiki inherit

Inherit a capsule from the marketplace. Downloads the gene and trajectory content, records the inheritance on-chain, and imports the gene into your local library.

bash

wiki inherit <capsuleHash> [--no-browser] [--web-url <url>] [--workspace <path>]

Flag	Description
capsuleHash*	stringContent hash of the capsule to inherit
--no-browser	booleanSkip browser confirmation — use server-only flow (for CI)
--web-url	stringOverride the website URL for browser confirmation
--workspace	stringWorkspace root directory

Inheritance flow

Preview capsule metadata (title, description, price, quality score)
Check if already inherited on-chain
If not: open browser to the discover page for wallet confirmation
Poll until the on-chain transaction is confirmed
Download gene + capsule content from the server
Import gene into your local skills/wiki/genes/ directory

Signal detection

The evolution engine detects these signals from agent session history to determine when and how to improve.

Signal	Description
error_pattern	Recurring errors across sessions (same error type 3+ times)
performance_regression	Task duration or resource usage increasing over time
opportunity	Successful patterns that could be generalized
stagnation	No improvement in success rate over recent sessions
tool_failure	Specific tool calls failing repeatedly
timeout	Tasks exceeding expected duration limits
retry_pattern	Same operation being retried multiple times
error_cascade	One error triggering a chain of subsequent errors
resource_waste	Unnecessary tool calls or redundant operations
context_loss	Agent losing relevant context mid-task
strategy_mismatch	Wrong approach chosen for the task type
quality_drop	Output quality declining over sessions
success_streak	Consistently high performance worth distilling
novel_solution	Unique approach that solved a previously failing task