Choosing the classifier CLI

Pending mail is classified by a CLI you select in classifier.json: the GitHub Copilot CLI (default) or Claude Code. The same tool also hosts the interactive review skills; neither is the hourly scheduler. The reasoning and the measurements behind the Claude Code defaults are in ADR 0002.

Select at install time

bash scripts/setup-macos.sh --cli claude    # installs Claude Code, sets "cli", runs claude auth login
bash scripts/setup-macos.sh --cli copilot   # the default

make doctor checks whichever CLI classifier.json selects.

Select later

classifier.json selects the CLI with "cli" and keeps each CLI's binary and models under "providers":

{
  "cli": "claude",
  "providers": {
    "copilot": {"binary": "copilot", "cheap_model": "gpt-6-luna", "escalation_model": "gpt-6.1-sol",
                "disabled_mcp_servers": []},
    "claude": {"binary": "claude", "cheap_model": "claude-haiku-5-5",
               "escalation_model": "claude-sonnet-5-5", "auth": "subscription",
               "max_budget_usd_per_call": 1.0}
  }
}

Switch with an editor or with:

python3 runner.py --select-cli claude   # validates the file, rewrites only "cli"
python3 runner.py --cli claude          # one paid dry run with Claude Code; file unchanged

Every providers block is validated on each run, so a mistake in the inactive one is reported before you switch to it. Switching invalidates same-day cached classifications (a different classifier produced them).

GitHub Copilot CLI

Copilot calls have no available tools, run in a temporary working directory, suppress project custom instructions and remote session export, and receive mail through stdin, not command-line arguments. Built-in MCP servers and the configured disabled_mcp_servers names are disabled. Refused environments: COPILOT_PROVIDER_BASE_URL and COPILOT_PROVIDER_TYPE.

The installed Copilot CLI must support --available-tools=, --no-custom-instructions, --no-remote-export, --output-format json, and --usage-output-file. Check copilot --help and copilot help permissions.

Claude Code

Each classification is one claude -p --output-format json call with:

• no tools (--tools ""), no MCP servers (--strict-mcp-config with an empty config), no skills (--disable-slash-commands) and no saved session (--no-session-persistence);
• --setting-sources project in an empty temporary directory, so your user settings, hooks, plugins and permissions are not loaded;
• the classifier instructions and tag definitions as the system prompt (replacing Claude Code's coding-agent prompt) and the mail on stdin only;
• a per-call cap, --max-budget-usd = max_budget_usd_per_call.

The answer must be a single successful, tool-free turn; anything else is recorded as an error and no decision is applied.

Claude Code was verified with version 2.1.284; it must support -p, --output-format json, --tools, --strict-mcp-config, --setting-sources, --no-session-persistence and --max-budget-usd (check claude --help).

Setting	Default	Meaning
providers.claude.cheap_model	claude-haiku-5-5	First-pass classification
providers.claude.escalation_model	claude-sonnet-5-5	Ambiguous cases only
providers.claude.auth	subscription	subscription or api_key (see below)
providers.claude.max_budget_usd_per_call	1.0	Runaway guard, in the CLI's own cost estimate

Keep pinned model IDs. On Claude Code 2.1.284 the haiku alias resolved to the older Claude Haiku 4.5, so an alias can silently change both classifications and cost.

Auth. subscription (default) uses your claude auth login and removes ANTHROPIC_API_KEY/ANTHROPIC_AUTH_TOKEN from the call's environment so billing stays on that login. With these flags, a synthetic probe showed that your global ~/.claude/CLAUDE.md and project instructions are not sent; Claude Code still adds a short account section containing your sign-in email. api_key adds --bare, which also skips that and every other automatic context, but accepts only ANTHROPIC_API_KEY; the scheduled job cannot use it.

Refused environments. Like Copilot's custom-provider variables, the runner refuses to call Claude Code when ANTHROPIC_BASE_URL, ANTHROPIC_BEDROCK_BASE_URL, ANTHROPIC_VERTEX_BASE_URL, CLAUDE_CODE_USE_BEDROCK, CLAUDE_CODE_USE_VERTEX or CLAUDE_CODE_USE_FOUNDRY is set. A terminal inside a Claude Code session sets ANTHROPIC_BASE_URL for that session's own proxy, so run runner.py from a normal terminal or the schedule, not from inside Claude Code.

max_budget_usd_per_call is a per-call guard against runaway output, measured in the CLI's own estimate, not a budget. See Costs for why that estimate is unreliable for these models.

Data retention

The selected CLI's provider retention policy applies to data sent to it: Copilot (GitHub and the model provider) or Claude Code (Anthropic; terms differ between a Claude subscription and API-key use).