TokenPak v1.4.0

This release consolidates the work merged on main since v1.3.22 (Apr 25). The previously-staged but never-tagged 1.3.23 content (Codex format adapter) is included here under "Codex format adapter restored." Three large initiatives shipped on top of it: Intent Layer, Native Client Concurrency Parity (NCP), and OpenClaw Path C session-binding via governor failover.

Added — Intent Layer (PRs #46–#55, #57–#60; closeout #56, #61)

A new observation→advisory→opt-in-suggestion→opt-in-injection pipeline for understanding and (optionally) shaping prompt-side intent before traffic leaves the local boundary. All phases are dry-run or opt-in; nothing is forced.

• tokenpak intent report (Phase 1) — observation-only intent reporting CLI surface.
• Intent dashboard / read-model (Phase 1.1).
• Intent policy engine — design (Phase 2), dry-run engine (Phase 2.1), explain / report / dashboard preview (Phase 2.2).
• Opt-in suggest mode — spec (Phase 2.4), PolicySuggestion builder + telemetry (Phase 2.4.1), display surfaces (Phase 2.4.2), config (Phase 2.4.3).
• Confirmation mode — design / spec (Phase 2.5).
• PromptPatch (Intent Prompt Intervention) — design (PI-0), PromptPatch builder + intent_patches table (PI-1), Claude Code Companion intent preview surface (PI-2), opt-in companion-side PromptPatch injection (PI-3).
• Closeout milestone docs for Intent Advisory MVP and Claude Code Intent Guidance Injection MVP.

Added — Native Client Concurrency Parity (NCP) (PRs #62–#72, #77–#78, #80–#81)

Diagnostic, instrumentation, and advisory mitigation for native vs. tokenpak-proxied concurrency parity gaps observed under Claude Code's multi-lane traffic.

• Diagnostic + standard proposal (NCP-0).
• Claude Code parity A/B test harness + operator protocol (NCP-1).
• OAuth / subscription parity redesign — auth-plane scoping correction (NCP-1r), launcher-path fix (NCP-1r-fix).
• 1v1 baseline + hypothesis re-rank + A/B/C/D plan (NCP-1a iter-1); 2-TP retry + 1-native healthy concurrent observation (iter-3); retry localizes to post-tool-result continuation (iter-4).
• Session-lane diagnostic plan + iter-2 update + read-only inspection harness (NCP-3).
• In-proxy parity-trace instrumentation v1 (NCP-3I), activation verification + H10 stream-integrity (NCP-3I-v2), pre-dispatch lifecycle hooks + stream_abort (NCP-3I-v3).
• Terminal early-return events for adapter_detected → before_dispatch death cohort (NCP-3A enrichment).
• tp_parity_trace canonicality for wire-side completion (issue #73).
• stream_abort phase classification (issue #74 phase 1).
• Advisory circuit-breaker mode for anthropic (NCP-4 phase A) — opt-in, observe-only by default; flips to enforce only after operator-configured soak.

Added — OpenClaw Path C session-binding (PR #82)

Governor failover provisioning batch (5 commits) shipping the Path C filesystem-rendezvous pattern for OpenClaw v2026.3.23-2 (which has no outbound-request mutation surface).

• tokenpak/integrations/openclaw/hooks/openclaw-adapter/ — HOOK.md (manifest), handler.js, tests/test-active-json.js.
• tokenpak/services/routing_service/platform_bridge.py — _openclaw_extract reads <config-dir>/sessions/active.json for cross-platform session_id attribution.
• integrations/openclaw/tokenpak-inject.sh — installer that adds the openclaw-adapter hook bundle non-destructively.

Added — OpenAICodexResponsesAdapter (Codex format adapter restored)

Restores the OpenAICodexResponsesAdapter deleted in the Apr-16 wave-4 cleanup, ported as a functional reiteration against the current modular tree (not a verbatim copy of the Apr-12 file).

New module tokenpak/proxy/adapters/openai_codex_responses_adapter.py:

• Extends OpenAIResponsesAdapter. Wire format is identical to /v1/responses; only the upstream, path, and a few payload constraints differ.
• Detection (priority 270): matches /v1/responses requests whose Authorization header carries a ChatGPT OAuth JWT (Bearer eyJ…). The standard OpenAIResponsesAdapter (priority 260) still matches the API-key (sk-…) case, so existing direct-OpenAI traffic is untouched.
• Upstream rewrite: get_default_upstream → https://chatgpt.com/backend-api, get_upstream_path → /codex/responses. SSE format reuses openai-responses-sse.
• Payload constraints enforced in denormalize (the ChatGPT Codex backend rejects requests that don't conform):
• stream: true (always)
• store: false (no server-side conversation persistence)
• max_output_tokens is dropped (the backend computes its own cap)
• input is promoted to a list of message objects when the parent emits a string

Relationship to credential_injector (v1.3.15): the existing CodexCredentialProvider handles the platform-bridged path — when OpenClaw stamps X-TokenPak-Backend / X-TokenPak-Provider: tokenpak-openai-codex, the proxy's pre-forward hook reads ~/.codex/auth.json, injects headers, rewrites the URL, and normalises the body. That runs before adapter selection. This adapter handles the direct path — when a Codex CLI client (or any caller carrying a JWT directly) hits the proxy at /v1/responses. No double injection because the bridge path returns its own response before the adapter registry runs.

Registered in build_default_registry at priority 270, between AnthropicAdapter (300) and OpenAIResponsesAdapter (260).

Tests (tests/test_openai_codex_responses_adapter.py, 22 cases):

• JWT detection: eyJ + . shape, Bearer prefix stripping (case-insensitive), sk- rejection, empty-header rejection
• Adapter detect: path filter, auth-header casing, missing-auth rejection
• Upstream + path + SSE format overrides
• denormalize: forces stream=true + store=false, drops max_output_tokens, promotes string input to structured list, preserves existing list/dict input via deep-copy
• Registry integration: JWT routes to Codex adapter, sk- falls through to standard OpenAIResponsesAdapter