The claude runner cannot run on a Claude subscription, so a person without an API key has no route to any role abcd runs from a plain terminal: the planning interview, a build's implementer and validators, the drain's judge.

The claude runner cannot run on a Claude subscription, so a person without an API key has no route to any role abcd runs from a plain terminal: the planning interview, a build's implementer and validators, the drain's judge.

internal/core/runner/claude.go:116-157 always starts claude --print --bare …. The vendor's headless documentation (https://code.claude.com/docs/en/headless and https://code.claude.com/docs/en/authentication, read 2026-10-24) says --bare never reads OAuth credentials or the keychain and ignores CLAUDE_CODE_OAUTH_TOKEN, so on a subscription-only login every routed role fails at launch, with no early check naming why. Surfaced 2026-10-24 when abcd intent interview refused a person whose account is a subscription and the only remedy it offered was a runner they cannot run.

The restriction is ruling RN2 of the technical facilitator (dcn-2610020000000016, 2026-10-02: "runner keeps --bare with an API key … subscription-only owners run the role in their own session"), restated in the brief (.abcd/development/brief/04-surfaces/34-build.md:465-470, "The claude runner stays bare, on an API key alone") and built on by dcn-2610161624025493, whose headless shell approver is designed around --bare ("dropping --bare would let the repository's hooks, plugins and CLAUDE.md run"). RN2 rests on two grounds, and both have to fall for the remedy to stand:

1. Isolation. RN2's research (the cli-reference, read 2026-10-02) found that "no flag set short of the bare flag seals project instructions, plugins, skills and agents". The documented flags read on 2026-10-24 (https://code.claude.com/docs/en/cli-reference) cover settings sources, MCP servers (--strict-mcp-config --mcp-config), inline --settings, tools (--tools/--allowedTools) and plugins (--plugin-dir), but no page read states that any of them stops CLAUDE.md, skills, agents or auto-memory loading. This ground is not shown to fall: it has to be proven by a live probe that starts claude -p without --bare under the candidate flag set in a repository carrying a CLAUDE.md, a skill, an agent, a plugin and a hook, and reads the run's system/init event for what loaded, before anyone is asked to rule.
2. Terms. The vendor's legal page as read on 2026-10-24 (https://code.claude.com/docs/en/legal-and-compliance) bars a developer from routing requests through Free, Pro or Max credentials on behalf of their users and from collecting or storing Claude.ai tokens, and states that it does not prevent an end user from signing in to the unmodified Claude Code binary with their own subscription; Pro and Max limits assume ordinary individual use. On that reading, abcd starting the person's own claude on their own machine under their own login is the permitted case. A second reviewer could not fetch the page, so the reading is one session's and is to be re-read when the ruling is asked.

Reversal flag (advisory, for the technical facilitator to rule): the remedy reverses RN2 for a route with no API key, and the brief paragraph and the --bare premise of dcn-2610161624025493 with it; dcn-2610161624025493's approver, which reaches the run through --mcp-config, is unaffected by the flag itself. Nothing is applied until the probe has run and the ruling is given.

The runner has no notion of a key today (no ANTHROPIC_API_KEY reference in internal/core/runner/ outside tests), so "a route with no API key" needs a stated test: the presence of the variable in the runner's environment, never its value, or an explicit machine setting naming the credential the route uses.

Probe 2026-10-24: the isolation ground, tested live

Claude Code 2.1.310 on macOS, a subscription login with ANTHROPIC_API_KEY and CLAUDE_CODE_OAUTH_TOKEN unset, model haiku, in a throwaway repository under ~/.abcd.noindex/probes/<root-sha>/safe-mode-2610241000/ carrying a decoy CLAUDE.md (an instruction to say a fruit word), a project skill, a project agent, a project SessionStart hook and a project .mcp.json. Each run read the system/init event and the result.

What it shows: a subscription login runs claude -p without --bare; --safe-mode seals the project but also drops abcd's own approver and hook, so it cannot replace --bare for a shell role; the second set, with auto-memory turned off in the inline settings, sealed every decoy the probe planted while keeping the approver and abcd's hook. What it does not show: the user layer's own CLAUDE.md and skills (the probe planted none there), output styles, and behaviour on another Claude Code version; the runner should still refuse a run whose system/init lists anything outside the expected set.

Deferral 2026-10-27

Deferred past v0.15.1: Waits on the technical facilitator's ruling on reversing RN2 (dcn-2610020000000016) after a live isolation probe; subscription owners keep running roles in their own session, as v0.15.1 documents (ruled 27 October 2026).