commit aefa0f7b7f4b2b1a1e2882c03ef8fb5f2f4c32aa
from: Isaac Meerleo <isaac@itm.works>
date: Tue Jul 21 12:47:21 2026 UTC

Reconcile Claude subscription acceptance state

commit - bbc238db63c343eb395c9f65d934517481871a51
commit + aefa0f7b7f4b2b1a1e2882c03ef8fb5f2f4c32aa
blob - bda673d6d23f38f558b7765541d516402735e4ff
blob + e1cf48f8a91c091d9e4ae6dba7170dd0dcc786c8
--- .scratch/experimental-claude-subscription/spec.md
+++ .scratch/experimental-claude-subscription/spec.md
@@ -1,6 +1,6 @@
 # Experimental pure-C Claude subscription compatibility
 
-Status: ready-for-agent
+Status: needs-info
 
 ## Problem Statement
 
@@ -422,24 +422,21 @@ Anthropic applications and directs product developers 
 cloud authentication. The experimental label is therefore substantive, not a
 quality-stage euphemism.
 
-Suggested implementation sequence:
+Implementation status:
 
-1. Record the architectural decision and update domain/security/dependency
-   language to the pure-C Custodian design.
-2. Remove the transitional Node/Agent SDK surface and restore the package and
-   Custodian sandbox to the direct-provider baseline.
-3. Implement the credential-free compatibility profile, request shaping,
-   attribution, Tool mapping, and hermetic contract fixtures.
-4. Route `claude` through the existing direct TLS/SSE request lifecycle, add
-   response-name validation, and run the complete regression and hardened
-   sandbox gates.
-5. Run the explicit Owner credential live check, record the profile and date,
-   and ship the route with its experimental warning and supported direct-API
-   alternative.
+1. Commit `d938d90` recorded the architectural decision, removed the
+   transitional Node/Agent SDK surface, implemented the pure-C compatibility
+   profile and direct TLS/SSE route, and added hermetic, sandbox, package, and
+   live-check coverage.
+2. The combined OpenBSD tree at `12670a1` passed the normal and hardened gates
+   (44 suites and 17,515 checks in each run), plus isolated package, linkage,
+   fake-installation, and unchanged-PLIST checks.
+3. The remaining acceptance gate requires the Owner to run the explicit
+   non-root credentialed check with a real setup token, including a nondefault
+   Effort value, and record scrubbed text, harmless Tool-continuation, usage,
+   and credential-non-reflection evidence.
 
-The feature is complete only when a stock OpenBSD package has no Node/Python/
-Claude runtime dependency, `fugu-api` demonstrably cannot open files or
-execute in subscription mode, a real setup token completes text and Tool
-Generations, the full hermetic and hardened suites pass, and every Owner-facing
-description consistently says one Owner and experimental subscription
-compatibility.
+The implementation is therefore no longer agent-ready work. It remains
+`needs-info` and intentionally has no `Resolution: completed` until the
+Owner-credentialed live evidence is recorded. The feature is complete only
+when that evidence joins the already-passing hermetic and hardened proof.