CLI reference

Cortex, memory, and provider commands

Search verified repository memory and route a bounded context pack through Claude, Codex/OpenAI, or Ollama.

Ask an AI provider

Codex client

Uses your installed Codex authentication. Cortex launches an ephemeral, read-only client run.

agentdiff cortex agent ask \
  --provider codex-cli \
  --task "Plan the smallest safe parser fix"

OpenAI / Codex API

export OPENAI_API_KEY="..."
agentdiff cortex agent ask \
  --provider openai-api \
  --model gpt-5.6-terra \
  --task "Review the authentication recovery plan"

Continue a Responses API chain with:

agentdiff cortex agent ask \
  --provider openai-api \
  --previous-response-id resp_123 \
  --task "Now minimize the proposed diff"

Claude Code client

Uses your installed Claude Code authentication. Cortex uses non-persistent print mode and plan permissions.

agentdiff cortex agent ask \
  --provider claude-cli \
  --task "Find the safest recovery boundary"

Anthropic API

export ANTHROPIC_API_KEY="..."
agentdiff cortex agent ask \
  --provider anthropic-api \
  --model claude-sonnet-5 \
  --task "Review the evidence capsule design"

Ollama API or client

The model is always explicit because installed local models differ by machine.

agentdiff cortex agent ask \
  --provider ollama-api \
  --model qwen3.6 \
  --task "Plan the parser repair"

agentdiff cortex agent ask \
  --provider ollama-cli \
  --model qwen3.6 \
  --task "Review the rollback logic"

Use --no-memory for a provider-only request, --max-memories to change the default limit of four evidence cards, and --format json for the normalized provider response and usage fields. --endpoint, --executable, and --api-key-env support self-hosted or non-default configurations without putting a secret value on the command line.

Search trajectory memory

agentdiff cortex memory stats
agentdiff cortex memory search "authentication session regression"
agentdiff cortex memory search "src/auth/session.py" --limit 3 --format json

The search command is offline by default. It ranks compressed evidence cards by shared task/path terms, exact paths, recency, and policy risk.

Add local semantic vectors

ollama pull embeddinggemma
agentdiff cortex memory index --model embeddinggemma
agentdiff cortex memory search \
  "authentication session regression" \
  --embedding-model embeddinggemma

memory index sends compressed memory cards to the configured Ollama embedding endpoint and stores the returned vectors locally. Re-run it after adding episodes or changing the embedding model.

To use semantic memory automatically during an AI request:

agentdiff cortex agent ask \
  --provider ollama-api \
  --model qwen3.6 \
  --embedding-model embeddinggemma \
  --task "Plan a safe session middleware refactor"

Pack context without calling a provider

agentdiff cortex context pack --task "Fix payment gateway timeout"

The output includes matched skills, fragile paths, relevant verified runs, and the rule that rejected runs are warnings rather than successful examples.

Generate a reusable skill card

agentdiff cortex skill list
agentdiff cortex skill generate <run-id> --title "Postgres Connection Pooling"

The generated .agentdiff/skills/<skill-slug>.md remains traceable to its source capsule.

Generate remediation advice

agentdiff cortex advise <run-id>
agentdiff cortex advise <run-id> --format json

The payload identifies collateral paths and the conflict-safe rollback command. It does not execute recovery, modify files, or retry an agent.