TokenPak v1.3.20

Fixed — Restore the Apr 15-18 subprocess companion bridge as the default

The v1.3.13-19 sequence iteratively refined an HTTP header-injection approach to OpenClaw → Claude Code routing. Investigation against the maintainer's working 4/15/2026 fleet build + commit history showed that approach is architecturally wrong: the Apr 15-18 build that shipped the working companion bridge used subprocess dispatch through the claude CLI, not HTTP header rewriting.

The HTTP path gives OpenClaw traffic valid OAuth but loses everything else tokenpak's companion workflow provides — tool use, MCP, CLAUDE.md, skills, proper session continuity, Claude Max billing pool membership. The Apr 15-18 build routed OpenClaw's request body to claude --resume <session_uuid> --print --output-format json with a clean subprocess environment + agent workspace as cwd; Claude CLI handled everything else because it's Claude Code by definition.

Restored

Subprocess dispatch is now the DEFAULT for tokenpak-claude-code. Flipped from opt-in (TOKENPAK_COMPANION_SUBPROCESS=1) to opt-out (TOKENPAK_COMPANION_SUBPROCESS=0 disables). When the platform bridge resolves to tokenpak-claude-code, the proxy dispatches via AnthropicOAuthBackend.dispatch → claude subprocess instead of the HTTP header-injection path.

AnthropicOAuthBackend.dispatch enhanced to match the Apr 15-18 behavior point-by-point:

• Model pass-through: --model <model> forwarded from the request body so Claude CLI doesn't pick its own default (OpenClaw's /model selection is now respected end-to-end).
• Clean subprocess env (_resolve_env): strips ANTHROPIC_BASE_URL / OPENAI_BASE_URL so Claude CLI doesn't loop through this proxy; sets DISABLE_PROMPT_CACHING=1 so the CLI's cache_control doesn't collide with tokenpak's compression pipeline; sets TOKENPAK_COMPANION_BARE=1 so the companion hook skips injecting the CLI's own CLAUDE.md when the caller is already carrying context.
• Workspace cwd (_resolve_workspace): reads X-OpenClaw-Workspace header, falls back to OPENCLAW_WORKSPACE env, then <config-dir>/workspace. cwd matters because Claude CLI reads CLAUDE.md + settings from the cwd tree and tool_use operations are relative.
• Session continuity via session_mapper (already added in v1.3.14, now actually exercised by default): first turn spawns fresh + captures session_id from the CLI's JSON output; subsequent turns pass --resume <uuid>.
• Timeout bumped to 300s (was 120s) to match Claude CLI's typical long-response upper bound.

HTTP credential injection retained for tokenpak-openai-codex + tokenpak-anthropic — those backends don't have a subprocess equivalent (Codex hits chatgpt.com/backend-api; Anthropic direct byte-preserve is correct for api-key and SDK OAuth callers).

Live verification

OpenClaw-shape curl (Anthropic JS SDK UA + x-api-key placeholder + X-TokenPak-Backend: claude-code):

HTTP 200
{
 "id": "msg_claude_e68da47c-6c40-4a1e-94b9-75199e39b2ef",
 "model": "claude-haiku-4-5",
 "content": [{"type": "text", "text": "subprocess-companion-restored"}],
 "usage": {"input_tokens": 137333, ...}
}

• msg_claude_<uuid> prefix = subprocess dispatch (not HTTP forward).
• 137k input tokens = full CLAUDE.md + companion context loaded, confirming the companion workflow is active.
• Model honored from request body.

Tests

339 regression green (services + proxy + conformance + cli). Import contracts clean. Ruff clean.