CaPU × ATMAN: loopback HTTP recovery with a separate read-only observer - #104
CaPU × ATMAN: loopback HTTP recovery with a separate read-only observer#104safal207 wants to merge 2 commits into
Conversation
…shes Add an isolated recovery laboratory with separate controller/device SQLite stores, five process-exit cut points, 55 tests, and an equal-guarantee conventional FSM comparison. No changes to native runtime or hardware. Preserve UNKNOWN and A7 receipt-sequence semantics. Keep production crypto, power-loss durability, universal exactly-once, and CPU superiority out of scope.
Preserve PR #103 sources unchanged. Add one loopback device process, controller subprocesses and a stdlib-only read-only ledger observer. Exercise lost responses, external-observation-triggered process kill, UNKNOWN preservation, restart, receipt binding and non-idempotent bypass. Keep real deployment, hardware, independent attestation and production crypto out of scope. Candidate must remain draft and unmerged.
|
Important Draft PR not reviewedDraft PRs are not automatically reviewed by default.
To automatically review draft PRs, update your CodeRabbit configuration: reviews:
auto_review:
drafts: trueThanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
@codex review Please review exact candidate head Focus on the v2 delta from unchanged #103 ( Explicit non-claims: one trusted host and loopback HTTP; the database append IS the effect; no real payment/physical device/WAN; no OS permission boundary or bypass resistance; public fixture keys and synthetic A7 tags; no general exactly-once/liveness, power-loss/NVRAM or independent attestation. The direct-call control intentionally demonstrates two effects for two unguarded requests. Verified CI at this exact head: https://github.com/safal207/CaPU/actions/runs/33940511054 — both Python 3.11 and 3.13 jobs passed v1 and v2 and uploaded artifacts. The separate v1 laboratory, FCRP Credential Boundary and Validate Examples workflows also passed. Local v1 reproduced its prior 55-test result digest exactly. CI and local PASS are not review approval. |
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
1 participant
Scope — keep draft and unmerged
Adds
experiments/capu_atman_http_recovery_v2/and one read-only path-scoped workflow. Onlysafal207/CaPUis changed. No deployment, main publication, ready-for-review transition, RTL, Bardo/COSMIC or ATMAN-repository changes.Exact candidate head:
8a2f2a37023a50aeac52cb8c8aed84b2eeceec88.Parent: unchanged #103 head
977864167c65f161e6db87b3d14257a11a67516f. This standalone branch includes unmerged v1; the delta from #103 is seven new files. #103 and existing v1 code are not modified. Candidate code has not changed after the review request.Executed validation
CI run: https://github.com/safal207/CaPU/actions/runs/33940511054 . The separate v1 laboratory, FCRP Credential Boundary and Validate Examples workflows also passed at this head.
Local v1 digest:
7e0d4e80d71bbfb27dfda5939809992d39a993e70097779db8a911568f6ce38f.Local v2 digest:
1eb8e50a35df8cb9312616c0c21f6182c10592db10e0ba77e96943b9d22128d5.The digest includes source hashes and environment. It is not an independent timestamp, trust anchor or completeness proof. CI artifact bytes were not downloaded; CI execution/upload status was checked separately from local evidence.
28 tests means 14 named scenarios run against each arm, not 28 independent guarantees. The ordinary FSM shares HTTP/SQLite I/O, ATMAN verification and synthetic receipt authentication, but not native CaPU lifecycle transitions. Both controller state and observer-read device rows are compared. No superiority over the equal-guarantee FSM is established.
What actually ran
A controller subprocess receives only its state directory and a loopback URL. A separate HTTP device appends a deliberately non-idempotent row in its own SQLite ledger; that durable insertion IS the effect. A third stdlib-only observer opens the device ledger read-only without importing controller code or reading controller state.
The device commits its effect and holds the response. Only after the observer sees that effect does the harness kill the still-waiting controller. A new controller retains UNKNOWN and refuses a repeat. After a policy change it accepts the exact historical positive receipt. The observed device ledger remains at one request, one effect before and after reconciliation.
Connection loss before/after the effect, timeout, HTTP 503 and malformed ACK do not release the reservation. Missing receipt records remain UNKNOWN, never invented NOT_COMMITTED. Additional cases cover device-process restart, tampered/replayed receipts, full-token lookup, stale authority and competing dispatch workers.
The bypass control intentionally sends two identical unguarded HTTP requests and observes two effects. The device has no deduplication constraint. Its ambiguous duplicate ledger returns CONFLICT rather than a fabricated clean receipt. This also explicitly demonstrates that bypass resistance is NOT provided.
Non-claims and review boundary
Real loopback HTTP/process separation on one trusted host, not a third-party service, real payment, physical device, WAN, OS permission sandbox or independent attestation. No distributed atomicity, production authentication, physical power-loss/NVRAM, Byzantine-device or storage-rollback proof. Public fixture keys and synthetic A7 rotate/XOR tags remain unchanged. One trusted controller/device/lineage and one unresolved attempt. UNKNOWN may remain blocked indefinitely; no universal exactly-once/liveness, speed/energy, full ATMAN or Bardo/COSMIC claim.
Authorization linearizes at durable admission, not physical actuation. Historical reconciliation does not renew old execution authority.
Native Codex is the only review gate. One request was made for this exact head; the bot reported exhausted review quota. No substitute reviewer or manual approval. Do not mutate the candidate after review and do not merge automatically. Local/CI PASS is not code-review approval.
Reproduce