the-custodian/docs/assessments/2026-09-08-helixforge-factory-critical-path.md
codex ce984af48d
All checks were successful
CI Smoke / host-smoke (push) Successful in 0s
CI Smoke / container-smoke (push) Successful in 1s
Record native factory sender proof and deployment defect correction
2026-09-11 14:26:23 +02:00

16 KiB
Raw Blame History

Factory critical-path execution — 2026-09-08

Continued on 2026-09-09: issuer pin complete and custody reviews ready.

Implemented and published prerequisite corrections in ops-warden, key-cape, approval-engine and audit-core, followed by local runtime installation and an attended OpenBao capability preflight, then verified the actual upstream issuer. The canonical integration workplan and next admission sequence remain in prj-helixforge-factory.

Delivered Evidence Effect
Warden preserves explicit policy refusal Code 31d9b66; 429 tests passed, four integration tests excluded by the normal suite configuration; Ruff passed. Existing SSH probe still HTTP 200/ALLOW. Installed credential request now HTTP 403 / exit 4 before transport. Replaces misleading “unavailable / fail-open” behavior with the actual caller/owner binding refusal. WARDEN-WP-0039-T03 owns that live admission.
Installer refreshes the actual Warden package 19bb75c; initial --force reused a cached wheel, while --reinstall-package ops-warden installed the corrected source. Installed module and behavior checked. A published fix now reaches the executing CLI.
KeyCape requires genuine issuer denial dcebd46; complete Go suite and vet passed. Negative checks reject timeouts, server errors, invalid tokens and unrelated refusals. Image pulled by immutable digest. Prevents false successful scope/predecessor proofs. KEY-WP-0013-T06 is done.
Approval audit scope is exact Approval b46b0f2, Audit bbe8c15; 81 receiver and six sender checks passed. Exact tenant:platform, source approval-engine, write-only/load-bearing, explicit redact. Resolves both tenant/redaction input waits and removes the proposed wildcard tenant scope. Credential and live admission remain open.
KeyCape rollout is concrete Pinned candidate plus two credential refs, /readyz, and single-instance replacement passed Railiance server dry-run. Rollback restores the compatible config/image pair. The next attended rollout starts from a reviewed patch, with known configuration and credential prerequisites. No production deployment performed.

The human approver UI/callback is now its own KEY-WP-0013-T05 task. Its absence does not block the independent service registrations or approval service startup. It still blocks evidence requiring a real human approver. Together with the two audit inputs, three source-input waits are removed from the service-startup path; no live security gate is marked complete by that separation.

The project now records thirteen source-pinned supplying records, including RPF-WP-0035-T05 / proposed CCR-2026-0017/0018 and WARDEN-WP-0039-T03. The two CCRs explicitly cover verifier-side delivery only. Client-side read lanes and linked approval audit receiver/sender custody remain distinct returns.

The immediate sequence is: ensure the now-verified upstream issuer is pinned; obtain the named custody reviews and run the contained attended first provision; deploy and verify KeyCape; admit audit/consumer credentials; prove deployed approval claim/consume and native model credential delivery. Exact policy binding and required human-approval evidence remain explicit. Runtime placement preparation can proceed alongside this chain. The factory's exact operating grant and enforceable budget still precede paid execution.

All four modified owner repos returned zero consistency failures. The checks retain historical/advisory warnings where applicable; canonical reconciliation then projected published source. Five new task assignees were empty despite source metadata and were corrected through the supported API with readback. HFACT-WP-0001-T02 retains that projection parity issue.

Published evidence records revisions, validations and the dry-run. These changes reduce ambiguous handoffs and unreliable checks. Time saved and autonomous throughput are not measured. No secrets were provisioned, no new policy binding granted and no paid model run admitted; G0G5 remain open and the fourteen-day window has not started.

Runtime installation and custody authority follow-through

SAND-WP-0015-T06 is done. The existing pinned runtime is installed on bnt-lap001 in its owner's mode-0700 content-addressed store. The installer checks the expected digest before writing, verifies the copied artifact, rejects unsafe ownership/permissions and changed destinations, and publishes atomically. Reusing an intact installation is idempotent. It is protected by the local owner boundary; it is not root-owned or a Railiance installation.

Sand-boxer source df5e941 passed the complete check: 188 tests and Ruff. The installed-path namespace smoke started Claude 2.1.263, rein and the adapter; proved read-only runtime, private persistent HOME, isolated networking, clean worktree and teardown. It used no model credentials or provider request. Sand-boxer published main 3e49a98 and Glas 4499160 carry the receipt. T04 and GLAS-WP-0012-T02 still own execution configuration, real-model acceptance and Railiance placement. No repeat of the unchanged candidate build is needed.

RPF-WP-0035-T05 now has a successful first-provision capability proof. Through the native founder-attended openbao-platform-admin-login envelope, the platform diagnostic checked create/update capability on all six exact policy, Kubernetes-role and KV data paths for CCR-2026-0017/0018. The successful session self-revoked and removed its helper. Source a94d06d preserves the metadata-only diagnostic and receipt. No secret value was read, custody seeded, policy/role changed or production workload deployed.

The receipt preserves all attempts: the first direct launcher failed before the child command and left revocation unconfirmed; the second native-wrapper session revoked successfully but its query used unsupported multi-path syntax; the final six single-path queries passed and that session revoked successfully. The later success does not establish revocation for the first failed attempt.

The project consumes both returns in its thirteen-record dependency view. Factory T03 and T04 remain waiting for their outstanding acceptance evidence; installation and authority discovery no longer appear as unresolved inputs. Source, Forgejo and the primary State Hub are reconciled at published revisions.

This removes two concrete setup uncertainties and makes installation and the custody preflight reproducible. Throughput, hours saved and autonomous delivery remain unmeasured. The next sequence is actual upstream signed-token issuer proof, named CCR reviews, attended custody activation and compatible KeyCape rollout. Separate client/audit/native credential admission, the exact factory grant and enforceable spend limits still precede a real model run. G0G5 and the fourteen-day observation window remain open.

Runtime and custody continuation receipt retains source revisions, validations, projection readbacks and progress IDs.

Executable upstream issuer proof prepared

KEY-WP-0013-T07 closes the missing diagnostic preparation. KeyCape source 6f33abd adds a one-shot upstream verifier; source 1b9793f pins the tested image and execution packet. The probe verifies a real authorization-code response inside the adapter and emits only a verified issuer and check metadata. It issues no downstream token.

The complete Go race suite and vet passed. Five packet/receipt tests and seven real Traefik 3.6.7 route cases passed; ordinary callbacks remain on their normal backend. Railiance server dry-run accepted the exact five-resource packet. The immutable container also produced the expected one-second idle-timeout receipt as UID 65534 with a read-only filesystem and no network.

The temporary Job has a ten-minute deadline and no retry; its route, service and two network policies must carry the real Job UID for bounded cleanup. It uses only the existing config YAML, without the signing-key mount or a Kubernetes API token. That YAML contains credential data, so live execution requires the deployment owner's admission. KEY-WP-0013-T02 retains that decision, the actual signed issuer receipt, contained configuration pin and subsequent custody/rollout reviews. No live issuer result is inferred from preparation.

This replaces an unspecified manual observation with a tested command and bounded execution packet. The current receipt is key-cape/docs/evidence/upstream-issuer-probe.json.

Live issuer observation completed

The user admitted the prepared ten-minute probe. At 2026-09-08 21:44:44 UTC it verified an actual signed upstream token with issuer exactly https://auth.coulomb.social. Signature, audience, validity window and nonce checks passed. The pinned container exited 0. This closes the unknown-issuer observation in KEY-WP-0013-T02 and RPF-WP-0035-T05; the earlier preparation-only state above is superseded by this live return.

Cleanup removed the dedicated route, Job/Pod, Service and both temporary network policies. Deletion used the recorded object UIDs; every temporary resource is absent. Normal KeyCape Deployment/config Secret metadata and image remained unchanged. The probe retained no token and issued no downstream credential. It did not prove downstream MFA/application login or activate custody.

KeyCape 41f6916 and Platform 8f40d73 published the owner returns. Project 89041ec consumes the live proof and records HFACT-DEC-2026-002 (Hub decision 1c0d9fd1-4790-4d9a-be6c-0ed4c4890549). The concrete probe approval is resolved; its pending human-needed flag is cleared. HFACT-WP-0001-T03 remains wait for the configuration owner to ensure authelia.issuer equals the verified HTTPS value, named CCR-2026-0017/0018 reviews, and the admitted custody/compatible KeyCape rollout. Separate client-side/audit/native delivery remains open.

The efficiency gain is one fewer unknown on the activation path and a retained repeatable check for relevant provider/configuration changes. No renewed token observation is needed for this unchanged proof context. Autonomous throughput, spend and the fourteen-day factory observation are still unproved.

Final readback found stale KeyCape T02 and Platform T05 task descriptions despite applied repository reconciliation receipts. Both were repaired from the published source and verified. HFACT-WP-0001-T02 retains the underlying source/projection parity work; these repairs do not establish that automatic synchronization is fixed.

Live issuer and synchronization receipt records the proof, cleanup, source/Hub parity and progress IDs.

2026-09-11 retry: sender credentials provisioned

The user requested another attended login attempt. It succeeded: CCR21/22 created independent version-1 custody and advanced the shared sender registry with CAS 7→8, preserving existing senders and fields. The seed wrapper exited 0 and confirmed self-revocation for that session. This clears first provisioning for Approval Engine and Informed Decision.

Platform's delivery helper now pins the reviewed manifests and verifies native reader/namespace restrictions, exact ESO projection and receiver-registry equality before reload. Seven delivery tests and 18 custody tests pass, including 11 actual disposable OpenBao cases. A fresh attended delivery login then failed before command handoff (exit 5; remote revocation unconfirmed), so no delivery or reload is claimed. RPF-WP-0035-T08 waits for delivery login and AUDIT-WP-0009-T09/T11 retain real producer acceptance. Do not repeat the approved CCRs or reseed their credentials.

Factory attempts and paid calls remain zero. The immediate sequence remains delivery/receiver proof, native approval and Informed Decision policy/service/ human admission, exact credential delivery and runtime/spend admission, then the single bounded reuse-surface worker attempt. CCR19's existing operator-group question remains separate; the customer UI product remains vergabe-teilnahme. The successful seed and failed delivery attempt are separate metadata receipts in the platform, Audit Core and factory project evidence directories.

2026-09-11 08:27 UTC: native audit delivery completed

The user was available for the renewed attended login. It succeeded; the reviewed delivery command completed in 28 seconds and Warden exited 0 with self-revocation confirmed for this session. Both independent version-1 sender credentials now reach only their intended ESO-owned producer Secrets. Native reader/path/namespace denials and the coding-agent boundary all passed. Temporary reader sessions were revoked and the denied namespace was removed.

Audit Core reloaded only after its registry JSON matched authoritative version 8. Receiver c82e0442 remains 1/1 Ready and reports operational/durable storage. RPF-WP-0035-T08 clears its human login wait and returns to progress for native producer acceptance. AUDIT-WP-0009-T09/T11 retain accepted/duplicate events, source/tenant/read refusals, independent readback and audit bearer-revocation proof. Reader-session revocation is a different control. Both CCRs stay applied until their full acceptance criteria are proved.

This supersedes the earlier delivery wait; no provisioning or login retry is needed for the unchanged successful delivery. Attestation/offsite operation remains AUDIT-WP-0009-T12, and tamper_evidence=false remains the accurate claim. Factory attempts and paid calls are still zero. The next useful step is actual producer/receiver acceptance, followed by the existing service/policy/human admission chain and the bounded reuse-surface factory attempt.

Native producer return — 2026-09-11 10:42 UTC

Both actual source adapters used their own ESO-delivered credential inside a bounded native Job. Each synthetic event received 202, retained its durable outbox after an injected lost receipt, and received duplicate 200 after a fresh process reopened the outbox. Each source's reconciliation reported exactly one record for its unique probe class. Wrong source/tenant, seven archive read routes, sibling reconciliation and invalid-bearer requests were refused.

The first Informed Decision probe exposed a private-store startup defect: fsGroup volumes add inherited setgid to a newly created 0700 directory. Informed Decision 1a12223 corrects creation without repairing unsafe existing data. The regression failed before the fix; the available suite now passes 335 tests with 39 optional checks skipped. Native retry verifies the exact corrected helper. Six probe tests and eleven rebuilt-container checks pass, including actual setgid-volume startup, restart/restore and cleanup.

All temporary probe Jobs, Pods, ConfigMaps and NetworkPolicies are verified absent. Receiver c82e0442 remains operational/durable; tamper_evidence=false. Producer verification requires no operator registry read or new OpenBao login. Independent readback is a separate attended command consuming the successful receipt; it must not rerun the producer jobs. Both attended verification-login attempts this session ended before handoff (exit 5; remote revocation unconfirmed), so independent archive readback remains pending.

This proves synthetic outbox delivery/recovery with native sender credentials, not domain-transaction atomicity, deployed producer services or human binding. An invalid bearer is not a revoked formerly valid bearer. AUDIT T09/T11 and RPF T08 remain progress for independent readback and audit bearer-revocation acceptance; CCR21/22 remain applied. Service/policy/human admission and native runtime/spend gates remain separate. Factory attempts and paid calls remain 0.

The practical efficiency gains are a native startup defect found before UI cutover, a container fixture corrected to exercise real volume permissions, and producer checks separated from privileged readback so a failed attended login no longer discards completed verification. These are setup/verification gains; factory throughput and the fourteen-day founder-load target remain unmeasured.

Retained native proof and attempt history and exact readback resume command preserve the next step without reconstructing the session.