Skip to content

Build the durable alpha controller and scoped review publisher #176

Closed as not planned

Description

@HemSoft

Description

Planning backlog under the alpha specification. The owner's October 19 merge-mission invocation authorizes scoped offline implementation and guarded merges. The retired project remains frozen; credentials, App installation and live activation retain separate authority and evidence gates.

Implement the private polling controller and publication path in the alpha specification, using fake GitHub/provider transports first. Integrate the protocol and owner command parser. Live operation depends on eligibility/isolation evidence and explicit activation authority.

Implementation prerequisites: Build both subscription adapters and the isolated source reader, Implement adversarial rounds and typed per-PR configuration.

Why

Duplicate events, downtime, changed revisions and ambiguous GitHub writes can waste subscription quota or publish stale reviews. A job for request A must never consume a later request B.

Goal

A recoverable single-job service admits only the owner and this repository, snapshots each request, and publishes/reconciles advisory reviews without blind retries or hidden uncertainty.

Trigger decision — October 21, 2026

Franz chose @sfl review plus automatic PR-open/head-push review on October 21, 2026 at 01:19 EDT, via Discord relayed by the Chief. Implement both offline with fake transports and normal review/merge gates. Discovery stays polling-based; no webhook, machine user, live inference, subscription use, deployment or paid usage is authorized by this decision. Native-picker requests are no longer required. PR #222 records the failed native experiment; #221 is closed unmerged.

Acceptance Criteria

  • Enroll only repository ID 1169772257 and the verified selected-repository installation. Reject wrong repo/installation/requester IDs before provider work or writes. No org wildcard or inherited retired IDs. Poll PRs, original @sfl review comments and PR-open/head changes with complete provenance. Only the spec-authorized commenter can issue a manual request; automatic work uses the enrolled repository policy and PR event/observation attribution. Ambiguous attribution never launches inference.

  • Persist SQLite request identities, immutable author/event IDs, generations, config/snapshot hashes, leases, phase artifacts and publication intent/outbox. Recover duplicate discovery, downtime, fresh-comment re-request, automatic head transitions, cancellation, restarts and expired leases. Admit one job and sequential provider sessions.

  • Honor command admission hashes, next-request config semantics and explicit cancel. Configuration commands alone never run review. An original authorized @sfl review comment requests review. PR opened/head synchronize creates automatic work, including after completion; deduplicate replay of the same request/observation. Head/base movement fences stale work; base-only movement does not create a new automatic request.

  • Recheck access, open state, requester, request generation, head and base-tip before/after publication. Either revision change invalidates unpublished results. Never mark request B fulfilled by A's review. COMMENT publication never consumes a later comment/event request. Persist and reconcile both identities and any in-flight uncertainty. Never dispatch A after cancellation or replacement is known to the publisher. Failure to establish an acceptable recovery path blocks live publication.

  • Only the publisher holds scoped GitHub write credentials; it has no provider credentials. Resolve repo/PR/diff anchors from trusted job state. Validate paths/lines, sanitize output, bound comments and redact secrets. Use advisory COMMENT reviews, never approvals/merges.

  • Reconcile uncertain POSTs using persisted intent and remote author/snapshot/request markers before retrying. Delayed visibility, access loss and crashes produce publication_unknown until resolved, never blind duplicate POSTs. Document GitHub's non-atomic review boundary; every initial result names its reviewed snapshot. Record supersession locally and annotate remotely when access permits.

  • Render complete, disputed, unverified, quota/incomplete, cancelled and superseded status honestly. Alpha never publishes substantive partial findings; status may report INCOMPLETE and its cause. Logs retain bounded sanitized metadata/findings, never public raw transcripts.

  • A configuration acknowledgment identifies its durable version. Requests created after that acknowledgment use that version; earlier requests retain the preceding version regardless of discovery order. Ambiguous ordering defers admission. Recovery reconstructs the same mapping from persisted records.

  • Record omitted paths, missing history, truncation, binaries, submodules and reader-limit exhaustion. Unreadable or omitted material never counts as clean. Preserve source-valid findings that cannot anchor in the diff as summary findings.

  • Render model text as untrusted content: neutralize mentions, prohibit remote images/raw HTML and allow only explicitly approved links. Generate source links from trusted snapshot/path data.

  • Ignore non-owner commands without replying or making GitHub writes. Rate-limit owner acknowledgments and record each job's head repository ID, including fork heads, while preserving inert-source isolation.

  • Persist configuration acknowledgment intent/version/unique remote marker before dispatch. An uncertain acknowledgment write defers dependent admission until paginated remote reconciliation recovers its identity/time. Never blindly duplicate an acknowledgment; test a crash after acceptance but before local response persistence.

  • Polling is reconciliation, not webhook delivery: pin PR identity/head repository/head/base and retain stable observation generations. Coalesce obsolete heads discovered after downtime; never invent unavailable push actors, timestamps or event IDs. Repeated identical snapshots do not rerun. A visible A→B→A transition is a new generation; document the inherent polling gap for unseen intermediate pushes. Fresh manual comments remain a durable explicit re-request. Preserve owner-only cancellation and config-ack ordering, including uncertain acknowledgments and equal timestamps.

  • When trigger implementation is complete, publish an exact checklist of remaining live-activation blockers: provider public-target/account permission, subscription/auth and paid-overflow evidence, completed containment/recovery/control/benchmark qualification, pinned trusted-release promotion, required GitHub rendering proof and explicit live owner authorization. No trigger test or merge grants activation.

Definition of Done

  • Document durable states, schema migrations, reconciliation/operator recovery, backoff/retention, permissions and enrollment. Preserve frozen workflows/old controller.
  • Provide fake-transport fault tests and runtime packaging with no automatic deployment or promotion. Link eligibility limitations and remaining live gates.

Validation

  • Run go -C reviewer vet ./... and go -C reviewer test -race -count=1 ./... once the module exists.
  • Inject crashes before/after POST and outbox persistence, delayed API visibility, duplicate discovery, cancellation after dispatch, access loss, changed head/base and request A/B races. Expected: no blind duplicates, no new request marked fulfilled by an old job, with explicit recovery for ambiguous request or publication state, visible uncertainty and correct generation fencing.
  • Assert wrong identity makes zero provider/write calls and models cannot control destinations or obtain publisher credentials.
  • After live gates and authority pass, embed current-head GitHub screenshots for inline findings, dispute/late summaries, incomplete and superseded states as uploaded attachments rendered in PR Validation; add a recording where practical or explain omission. Before then use offline renderer fixtures and mark live rendering blocked.
  • Test reversed command/request discovery, downtime spanning both events and an acknowledgment while an older request awaits admission. Reconstruct the same config-to-request mapping after restart.
  • Test hostile Markdown, mentions, HTML, remote images and model-provided links, plus non-diff anchors and omitted content. Expected: safe output, trusted source links and visible coverage limits. Inject A/B replacement during in-flight publication and prove B remains unresolved with correct reconciliation, independently of the Reviewers sidebar.
  • Record validation output/evidence at the exact PR head SHA. For code changes, the reviewer reruns applicable commands locally; GitHub Actions CI is currently disabled.

Offline evidence reconciliation, October 21, 2026

Checked validation items cover fake providers/GitHub transports and local synthetic fixtures. Durable controller tests cover uncertain ACK/POST persistence, A/B replacement, head/base/cancel fences, leases, no blind replay and wrong-owner/no-write behavior. HTTP publication tests add actual local TLS framing/pagination/unknown-write reconciliation. Selected-trigger tests cover original owner comments, PR-open/head polling, downtime, observed A-B-A, unchanged snapshots, configuration ordering and cancellation. Rendering tests cover hostile content and honest incomplete/coverage output.

PR #232 records exact-head 7dd5d612016b9ea0bd396230f0f5fc6342943d88 vet/full opt-in race PASS, main 243.811s. Its reviewed tree equals merged main bd5cd59ec873ec18c849ad96bd179d5242bef9cf. Service/startup/release components are delivered in #228, #229 and #231; a source merge does not promote or activate them.

Actual selected-App trigger/permission and GitHub-rendering qualification, aggregate protected host recovery, independent trusted promotion, provider/account evidence and explicit live authority remain unproved. PR #233 merged the exact gate checklist as ab260ad. No live rendering, model inference or activation is claimed. This issue remains open.

Merged activation checklist, October 21, 2026

PR #233 merged the dated six-gate checklist as ab260ad2d6883ddc4a22446a58ebfb93161296c6, after clean required review of e4c4a82a3b2178a4e9346c2955d0cdcbd3f00c38 and five addressed threads. It separates implemented mechanisms from unproved physical and live evidence and prepares the selected-trigger provider questions without sending them.

The six live gate groups remain open: applicable provider permission; account/auth/paid-overflow evidence; complete aggregate containment/recovery; actual comparative evaluation plus independently qualified trusted release/plan and pilot; final selected App/trigger/rendering qualification; separate explicit owner activation. PR #234 merged as 917d854 after a clean exact-head review. Its four controlled synthetic supervisor/driver-loss cases prove original-unit stop/recovery at those points, not arbitrary host crashes or aggregate qualification. The original host noswap/sealed-exec failures remain failures. PR #235 merged as 00d6a708 after a clean exact-head review and all 15 threads resolved. It delivers an inert measured VM prerequisite diagnostic and a design for a distinct VM source adapter; it grants no activation. This issue remains open.

Source Definition of Done reconciliation, October 21, 2026

The two checked Definition of Done items cover delivered source and documentation. Controller states and recovery, polling ownership/backoff and additive schedule schema, explicit startup packaging and separate promotion/rollback decisions document the implemented mechanisms, migration refusals, enrollment and remaining gates. The retired central-reviewer and .github/workflows trees are unchanged from the frozen planning baseline 6ce330bc1f83690d225d05f7e0452819ec7f4ff9. Actions remains disabled, verified as HemSoft against repository ID1169772257.

Durable fault tests exercise lost ACK/POST responses, restart, ambiguous ordering, retention and generation fences. Publication HTTP tests add local TLS/pagination and unknown-write reconciliation, and polling tests cover persisted backoff, cancellation, qualification refusal and shutdown ownership. Their inspected source/test blobs match merged PR #234's final reviewed df3ec524348b8d18caaf2a0c59c7287e16c25106, where vet and the full race suite passed, 195.255s wall. No redundant test run was needed for this issue-text reconciliation.

These checks do not establish actual selected-App trigger/permission/rendering behavior, aggregate protected-host qualification, independent release approval or live authority. Those acceptance criteria remain open. Source construction does not deploy, promote or activate the service, and this issue stays open.

PR #236 merged as 0f333755 after clean exact-head review, all four threads resolved and verified delivery/cleanup. PR #237 merged as 60d64113 with the guest-role original-descriptor fix, clean exact-head review, full vet/race and six passing synthetic-client VM runs covering normal completion and cancellation. PR #238 is under review for equivalent original-instance cleanup in two inert diagnostics. Its 48 Python tests, actual escaped-child/replacement checks and KVM fixture pass; the inert host sealed-exec comparison still reports incompatibility. Production source-VM acquisition and startup integration, aggregate qualification and separate live gates remain open. Source work is not exhausted.

Activity

  1. HemSoft commented on Oct 19, 2026

    @HemSoft
    CollaboratorAuthor

    Offline controller/publisher work is prepared in PR #183, currently awaiting required review. This is not a deployed alpha and does not close the live acceptance criteria.

    Two independent routes were exercised. The SQLite/fake-transport route tests atomic command/request discovery, durable ACK/session/lease state, lost POST responses, complete post-publication history, cancelled/superseded jobs, moved revisions and both observed and previously unseen A/B request races. The native-platform route checks GitHub's bot request schema and authenticated App installation inventory. The fresh inventory contains the existing Codex connector and Vercel, with no supplied/authorized selected-repository alpha App identity. A schema and fake tests cannot prove native picker eligibility or actual polling/completion semantics.

    Exact remaining need: applicable support/account evidence from #173/#174 and explicit new-App/external-credential/live-activation authority. Then implement and qualify the real HTTP transport and worker supervisor against the exact App/Bot/installation IDs, perform the model-free native request/completion/removal/re-request experiment, and exercise the real race/recovery cases with rendered GitHub evidence. The current library intentionally supplies trusted offline seams, not a production transport or daemon. No App, credential or activation was created. No API billing, private-mirror or machine-user fallback is inferred.

  2. HemSoft commented on Oct 19, 2026

    @HemSoft
    CollaboratorAuthor

    PR #183 is merged as 3cc932d. The one owner-authorized retry completed with3P2/noP1 at36fdb420; all three findings are fixed in579fb35 and all17threads resolved. Vet and the full race suite passed. The merge used the invoked bounded-review rule; corrected579fb35 was tested but not freshly clean-reviewed. No GitHub CI pass is claimed. Actions remain disabled and no live reviewer was activated.

    Remaining acceptance: Real scoped HTTP transport, stopped-worker supervisor and App/native-history/publication/rendered recovery qualification, behind #173/#174.

    Attempt1: Shipped and race-tested SQLite/fake transport failure injection, unknown-write reconciliation, A/B history, retention and durable refund/restart recovery. CONTROLLER.md explicitly states no real transport/daemon/launcher exists.

    Attempt2: Revisited bot/schema and authenticated installation inventory independently; no exact authorized alpha App/install/bot/credential tuple or live native-picker experiment is supplied. Existing native Codex review does not qualify a custom App.

    Exact next owner/external action: Pass #173/#174 and authorize/provision the exact App credentials, selected-repository installation and live experiment; then implement and qualify the real transport and supervisor.

    Implementation/qualification work remains for the agent after those gates. The owner retains external identity and activation authority; current fake-only instructions do not permit silently supplying a live service. This issue remains open; offline source completion is not live acceptance.

  3. HemSoft commented on Oct 19, 2026

    @HemSoft
    CollaboratorAuthor

    PR #184 merged as 67077bd. Exact head8c3bdb3 passed vet/race and a clean connected-Codex review with no findings. Issue #175 is closed after its explicit fake-first protocol/configuration contract was fully audited. The new source broker and actual fake-worker executable path address the prior read-capture implementation gap. This does not qualify official-client confinement or activate the alpha.

    Remaining: Exact App/install/bot credentials, real HTTP/native-history/picker/publication and production stopped-worker qualification.

    Attempt1: Merged controller/fake publisher failure injection in183 and source-audit integration184. Tests exercise durable ACK/review uncertainty, A/B races, leases, pruning, failed refunds and audit recovery without provider replay.

    Attempt2: Fresh authenticated installation inventory and source inspection still provide no authorized alpha App tuple, real transport/daemon or production launcher/supervisor. The test-only process bridge is explicitly not that implementation.

    Exact next owner/external action: Pass173/174 and authorize/provision the exact App credentials, selected-repository install and live experiment, then implement/qualify the real transport and supervisor.

    Owner controls external identity/activation. Further agent implementation and live acceptance remain; fake transport evidence is not a deployed service. This issue stays open.

  4. HemSoft commented on Oct 21, 2026

    @HemSoft
    CollaboratorAuthor

    The offline trigger implementation merged in PR #223, commit 758ada030323dcd66ec73c2b6199bc3825494e98. It implements original authorized @sfl review comments plus automatic PR-open/observed-head-change polling, with durable requester/event attribution, config ordering, cancellation and head/base/generation fences. Polling cannot recover an entire unseen ABA push sequence and never invents a pusher or push time. The native picker is superseded; the machine-user prohibition remains.

    The corrected head passed vet and fresh race tests, 84.730 seconds. All three P2 findings were fixed with failing-before regressions and resolved. Merge used the merge-mission policy after two finding-fix rounds; the last reviewed head had no P1. The corrected head is not claimed clean-reviewed. Its tree exactly matches merged main. Actions remain disabled, and the merge triggered no delivery workflows.

    The exact remaining live-activation gates are:

    1. Evidence that both pinned official Codex and Claude clients/accounts permit this public-target, unattended subscription use, including model entitlement and refresh/expiry behavior.
    2. Account-specific subscription/quota and billing evidence that API-key or paid-overflow usage cannot occur. A private polling host does not establish that permission.
    3. Finished credential-owner/acquisition/runtime/controller composition and qualification: whole-acquisition generation ownership, exact manager/unit and cgroup binding, complete mount/device/descriptor and filter checks, bounded hostile/resource tests, authenticated source publication after broker drain and physical stop, and restart/uncertain-attempt recovery. The bounded Git acquisition/process-evidence slice PR #224 merged at 403d47fa430134a5765f2e45574f51c2917a7747. The installation journal PR #220 merged at 7ddf30132b820170d72111d0ab0c3e5d0682b669. The credential-generation owner PR #225 merged at 4258d0e1fcf068bb4fc9ee333fd34b61f441ae44, including the final receipt and publisher-state fences. Its final corrected head passed vet/race and merged under the documented P2-only bounded-review rule; it is not claimed clean-reviewed. Git generation/early-terminal recovery PR #226 merged at 57f4cb6d1f1151f23b0edc90726e020d3341d204 after clean required review of 00237dc, zero unresolved threads, vet and full race passing in 127.305 seconds. It includes active-exited birth/physical-stop evidence, zero-invocation refusal and existing cached-worker verification for older receipts. The composed acquisition factory, mount/descriptor checks, broker drain, authenticated source publication, restart recovery and crash-safe terminal retention remain in PR #227, now targeting main at b157c18e43c332c32d1d58f93a84ef28595a8337. Its whole tree is identical to the source validated at 2b2acc1 by vet and a full descriptor/namespace opt-in race suite, 174.072 seconds. Twelve owned commits were replayed onto the actual parent squash without source changes. Corrections cover endpoint bounds, older release/worker restoration and retirement, idle collection, nonblocking active-capture polling and work/journal alias refusal. Required review of the new head and guarded merge remain pending. Active/exited manager metadata can remain after uncaptured physical-stop proof and requires separately coordinated host/resource qualification. Trusted service/release startup wiring and full physical qualification remain unfinished. The full fixed tmpfs/systemd launch and sealed inner-worker execution probes on mini remain failed, without fallback.
    4. Completed native control workers and durable six-arm runner, comparative evaluation/quality evidence, pinned trusted-release promotion and rollback, and bounded pilot preparation. The frozen corpus remains 14 clean and 6 defect cases, with zero unlabeled; those labels are not comparative model results.
    5. Final selected App/installation/permission verification, model-free qualification of the new triggers, and current-head rendered GitHub evidence for the required review and status cases under the applicable authority. The native-picker experiment in PR #222 does not qualify the replacement triggers.
    6. Explicit owner authorization for provider credential use, public-repository model inference and live activation. The trigger decision and source merges supply none of these grants. The owner-only two-week pilot and stop/rollback evidence precede any organization expansion.

    These are unsatisfied gates. Offline implementation continues under the existing merge mission. The merged trigger contract and checklist retain the detailed boundaries. No live inference, credential recreation, paid usage or retired automation was enabled.

  5. HemSoft commented on Oct 23, 2026

    @HemSoft
    CollaboratorAuthor

    Reconciliation as of October 23, 2026 (US Eastern). Offline progress, all against a fake GitHub and fake providers:

    • #250 (merged 3:05 PM) runs the whole path on mini: polling discovery, admission, source capture, four native sessions and an outbox COMMENT. close-publish makes the fake GitHub commit a review but withhold the response. After reopen, the reviewer reconciles it without a duplicate POST. Each GitHub credential role is held to its own endpoint allowlist. Admission is one job at a time, so isolation is shown between sequential jobs only (E2E-QUALIFICATION.md).
    • #251 (6:32 PM) adds explicit operator recovery of a stopped provider attempt (GIT-STARTUP.md). Before it, closing the service during a session left the controller unable to admit any further request.
    • #253 (2:39 AM) serializes polling observation with the worker's admission. The race could discard and redo a finished acquisition.

    Still open, with the gate for each:

    • Selected-App trigger, permission and installation qualification. App installation and credentials are a separate gate. #222 supplied App 5406096 only for the model-free native experiment. The @sfl review and automatic polling triggers have no live or model-free App qualification.
    • GitHub rendering. The Validation item needs current-head screenshots of inline, dispute and late, incomplete and superseded reviews. This needs live activation. Offline renderer fixtures exist, and live rendering stays blocked.
    • Provider gates. Provider permission (#173) is blocked. Live Codex is blocked by #248.
    • Trusted release promotion and owner live authority. Both are owner decisions.
    • Not done offline. There is no bounded shutdown grace. A process kill during a session, before its stopped audit receipt, stays fenced, and operator recovery cannot clear it.

    Mapping I could not do confidently: several acceptance items read as purely offline. They cover coverage limits, untrusted rendering, ignoring non-owner commands and config acknowledgment ordering. Their Validation tests are already checked. The October 21 reconciliation left the acceptance items unchecked until live App qualification. A separate per-item audit should decide which ones can be checked from offline evidence alone. This draft does not check any.

    Recommendation: keep open. The live App, rendering and activation criteria are not met.

    🤖 Generated with Claude Code

  6. HemSoft commented on Oct 24, 2026

    @HemSoft
    CollaboratorAuthor

    Per-item audit of the unchecked acceptance criteria, October 23, 2026, 9:08 PM EDT. This follows up the October 23 reconciliation. It was read-only and ran against main at a3f665b. The rule: an item counts as offline-complete only if merged code and tests satisfy its exact wording. Any item whose wording needs the selected App, real installation scope, live permissions or GitHub's own rendering is not.

    Tests run: GOMAXPROCS=2 go -C reviewer test -count=1 -p 2 -v -run '<render, trigger, durable, publication, discovery, polling, installation, reader, snapshot, provenance, engine, config tests>' . on go1.26.7, without -race. Result: 221 top-level tests passed in 18.1 s. One was skipped: TestGitHubInstallationJournalChild, a subprocess helper that only runs when its parent test invokes it. Every test named below passed in that run.

    # Criterion Verdict Evidence / gate
    1 Enroll only repo 1169772257 and the verified selected installation; provenance; owner-only manual requests GATED Offline: exact enrollment (durable.go#L27-L32). Wrong scope is refused before any provider call or write: TestWrongScopeAndNonOwnerHaveNoProviderOrWriteEffect, TestPollingControllerRejectsWrongEnrollmentAndQualificationBeforeWork, TestGitHubPublicationWrongScopeAndIntentMakeZeroHTTPCalls, TestGitHubPublicationWrongRemoteRepositoryIDPreventsWrite, TestGitHubInstallationRejectsUnattestedIdentityAndScope. Ambiguous provenance defers: TestGitHubDiscoveryRejectsAmbiguousOrPartialProvenance. Still needed: the "verified selected-repository installation" is the selected App/installation qualification (gate 5). The installation, App and bot IDs only need to be positive; no code refuses retired IDs.
    2 SQLite identities/leases/outbox; recover duplicates, downtime, re-request, head transitions, cancel, restarts, expired leases; one job PARTIAL Offline: TestSQLiteConcurrentAdmissionAcrossHandles, TestLeaseDoesNotStealExpiredWorkerAndCrashConsumesBudget, TestDurablePhaseResumesWithoutRepeatingValidatedWork, TestCrashAfterCompletionRecoversWithoutProviderWork, TestPollingControllerOwnerLockAndUncertainLeaseSurviveRestart, TestPollingTriggersRestartDoesNotRerunAndEditsDoNotRewrite. Not done offline: if the host process is killed during a session before the stopped audit receipt is written, the lease stays fenced and RecoverStoppedProvider refuses (GIT-STARTUP.md). There is also no bounded shutdown grace. Gate 3 covers aggregate physical stop/recovery qualification.
    3 Command hashes, next-request config, cancel; config alone never reviews; @sfl review; open/synchronize incl. after completion; dedup; head/base fences, base-only no request OFFLINE-COMPLETE Commands are parsed by config.go#L110-L150 and applied by durable.go#L353-L430, which never creates a request. Triggers: triggers.go#L89-L91 and #L199-L297. Tests: TestPollingTriggersOpenPushAndManual, TestPollingTriggersRunCompletedPushAndExplicitRerequest, TestPollingTriggersRetainManualConfigAndCancel, TestPollingTriggersBaselineBaseAndEpoch, TestPollingTriggersCommentAuthorizationAndCoalescing, TestMovedBaseAndCancelFenceDispatch, TestPollingControllerSnapshotMoveDoesNotCreateReplacementRun, TestOwnerConfigurationIsTypedAndHasNoExecution.
    4 Recheck access/open/requester/generation/head/base before and after publication; A never fulfills B; no dispatch after known cancel/replacement OFFLINE-COMPLETE publication.go#L244-L355 runs remote and local fences before the POST. #L412-L475 runs the postcheck over complete history. Tests: TestPublicationARetainsRequestBWhenGitHubClearsB, TestPostHistoryFindsBThatWasNeverLocallyObserved, TestLocalCancelRetiresPreparedOutboxWithoutPost, TestClosedPRRetiresPreparedReviewButAccessLossDefers, TestGitHubDiscoveryFencesNewCancellationBeforePublisherPOST, TestGitHubDiscoveryAWriteCannotFulfillDisappearingB, TestPollingTriggersStalePublicationDoesNotConsumeOrStrandB, TestPollingTriggersUnknownPublicationBlocksBWithoutRepostingA, TestPollingTriggersAccessLossDoesNotReleaseOldB. Live publication stays blocked. The live race experiment remains gate-5 work in GITHUB-PUBLICATION.md, but this item's wording does not require it.
    5 Only the publisher holds scoped write credentials, with no provider credentials; trusted anchors, sanitize, bound, redact; COMMENT only GATED Offline: TestGitHubInstallationActualTLSAndRoleBoundary, TestGitHubPublicationRejectsRemoteDocumentAndIdentityMutations, TestKnownSecretsAreRedactedAfterRenderingTransforms, TestPublicationBoundsRetainOversizeLedger, TestGitHubPublicationMaximumEscapedDocumentFitsWireLimit. "Scoped" credentials need the real App token's permissions (gate 5). Separating the publisher from workers needs measured aggregate isolation (gate 3).
    6 Reconcile uncertain POSTs via intent and markers; publication_unknown, never blind duplicates; document the non-atomic boundary; results name their snapshot; supersession OFFLINE-COMPLETE Unique marker: durable.go#L330. Reconciliation requires exactly one match from a complete scan: publication.go#L199-L243. A failed POST becomes unknown: #L330-L345. The summary names head, base and merge base: render.go#L68. The boundary is documented in CONTROLLER.md#L85-L97. Tests: TestOutboxUnknownPublicationAndDelayedVisibility, TestGitHubPublicationUnknownWriteReconcilesWithoutAnotherPost, TestGitHubPublicationReceiptGET422RemainsUnknown, TestGitHubPublicationPartialInlinePaginationCannotConfirm, TestGitHubPublicationIntentScanRejectsDuplicateMarkerOnLaterPage, TestGitHubPublicationACKAndIdempotentSupersession, TestPollingControllerPersistsBackoffAndNeverRepostsUnknown. Caveat: a direct POST 422 returns the intent to prepared, which relies on GitHub's documented validation semantics.
    7 Render complete/disputed/unverified/quota-incomplete/cancelled/superseded status honestly; no partial findings; sanitized logs PARTIAL Offline: the renderer's sections and INCOMPLETE: <state> with no findings (render.go#L78-L82), the supersession annotation (publication.go#L467), TestQuotaWaitResumesOnlySameCapturedJob, TestExactDuplicateObjectionBlocksInline. Not done offline: the publisher accepts only complete outcomes (publication.go#L126-L128), and the only write kinds are ack and review. Quota, cancelled, auth or expired outcomes are therefore never published, and cancelled renders as a generic INCOMPLETE: cancelled. Gate 5 covers live rendering.
    8 ACK names its durable version; later requests use it, earlier keep the prior version regardless of discovery order; ambiguity defers; recovery rebuilds the mapping OFFLINE-COMPLETE The ACK body names the version and hash (durable.go#L419). The effective time is the remote receipt time, applied FIFO (publication.go#L175-L187). Admission needs every ACK confirmed, and equal timestamps defer (durable.go#L546-L561). Tests: TestUnknownAcknowledgmentSurvivesRestartAndOrdersRequest (command-first and request-first, with restart), TestEarlierAndEqualRequestConfiguration, TestFIFOConfigAcknowledgmentsAndResetDefaults, TestGitHubDiscoveryOwnerCommandsUseOriginalACKOrdering, TestPollingTriggersConfigAcknowledgmentStillGatesManualRequest.
    9 Record omitted paths, missing history, truncation, binaries, submodules, reader-limit exhaustion; omitted is never clean; non-diff findings stay in the summary PARTIAL Offline: per-path omissions (snapshot_evidence.go#L147-L188), reader limits (reader.go#L90-L136), and the "not clean" wording plus summary-only non-diff findings (render.go#L90-L92, #L146). Tests: TestSnapshotEvidenceRetainsSizeOmissionsAndEnforcesInventoryCap, TestSnapshotEvidenceOmissionsSurviveDurableFakeReviewAndRecovery, TestReaderRecordsByteBudgetApartFromMissingObjects, TestSourceSearchRetainsScannedObjectsAndOmissions, TestDefaultAdversarialPhasesAndDeterministicPublication. Not done offline: missing or ambiguous history and inventories over 128 files fail capture outright (snapshot_evidence.go#L376). The poll loop then returns before admission (polling.go#L445-L448) and records only a generic deferred backoff (#L150-L162).
    10 Model text untrusted: neutralize mentions, no remote images or raw HTML, only approved links; source links from trusted data PARTIAL Offline: render.go#L22-L47 redacts credentials, drops URLs, converts @ to a fullwidth sign, and escapes HTML and Markdown; source URLs are built from the snapshot head and the fixed repo. Tests: TestRendererNeutralizesUntrustedContent, TestRendererRedactsRecognizableCredentialStrings, TestAdapterChecksSourceInsteadOfTrustingModelAnchors. Not done offline: GitHub's reference autolinks (#123, GH-123, owner/repo#123, owner/repo@sha, bare commit SHAs) are not neutralized beyond a backslash before #, and no test covers them. Cross-repo references can add timeline events in other repositories. Gate 5 covers live confirmation.
    11 Ignore non-owner commands with no reply or write; rate-limit owner ACKs; record head repo ID incl. forks; inert-source isolation PARTIAL (test gap) Offline: non-owner text is skipped (durable.go#L362, github_triggers.go#L108). One ACK per flush (polling.go#L321-L328). The head repo ID is captured (github_triggers.go#L193). Tests: TestPollingTriggersCommentAuthorizationAndCoalescing (stranger case: zero writes), TestWrongScopeAndNonOwnerHaveNoProviderOrWriteEffect, TestPollingControllerAcknowledgmentsAreFIFOAndRateLimited, TestGitHubDiscoveryCapturesActualBaseAndDurableRequest (fork ID 5555, legacy native reader), TestGitReaderUsesImmutableObjectsWithoutCheckoutOrHooks. Missing: a fork-head test on the selected trigger path. Real fork-head fetch and aggregate isolation are gates 3 and 5 and are not claimed.
    12 Persist ACK intent, version and unique marker before dispatch; uncertain ACK defers admission until paginated reconciliation; no blind duplicate; crash-after-acceptance test OFFLINE-COMPLETE The intent is written in the capture transaction (durable.go#L419-L425) with a unique marker (#L330). ACKs dispatch FIFO (publication.go#L293-L303), and admission defers until all are confirmed (durable.go#L548-L551). TestUnknownAcknowledgmentSurvivesRestartAndOrdersRequest: the remote accepts, the response is lost, the store reopens, a partial scan defers, a complete scan confirms, and there is one POST. Also TestGitHubPublicationACKAndIdempotentSupersession and TestPollingTriggersConfigAcknowledgmentStillGatesManualRequest.
    13 Polling is reconciliation: pin PR/head repo/head/base; coalesce after downtime; never invent actors/times/IDs; identical snapshots don't rerun; visible ABA is a new generation; document the gap; owner-only cancel and ACK ordering PARTIAL (test gap) Offline: triggers.go#L199-L297 and github_triggers.go#L83-L131. The polling gap is documented in GITHUB-TRIGGERS.md#L68-L76. Tests: TestPollingTriggersOpenPushAndManual (ABA), TestPollingTriggersSameSecondHeadChangeAndLateComment, TestPollingTriggersClosedHistoryPreservesHeadForReopen, TestPollingTriggersCanceledCommentCannotSuppressNewHead, TestPollingTriggersScopeIdentityAndClockChangesRefuse, TestPollingTriggersAcceptActualGitHubPullMetadata. Untested: head-repository pinning on the selected path (github_triggers.go#L129, triggers.go#L231).
    V Validation: current-head GitHub screenshots of inline, dispute/late, incomplete and superseded states GATED Gates 5 and 6. The only offline fixture is adversarial-review.golden.md, which covers peer-supported findings. An incomplete state cannot be published today (see item 7).

    Checked off in this update, as offline-complete:

    • Item 3, commands and triggers. Every clause is local controller semantics with direct fake-transport tests.
    • Item 8, ACK version mapping. The version is named, the mapping is ordered by receipt time, equal timestamps defer, and restart is tested in both discovery orders.
    • Item 12, ACK intent persistence. It is persisted before dispatch, and crash-after-acceptance reconciliation without a second POST is tested.

    Offline-complete by their wording but left unchecked for the owner: item 4 (publication rechecks) and item 6 (uncertain-POST reconciliation). Both are fully tested offline, but GITHUB-PUBLICATION.md plans a live publication race experiment as gate-5 work, so checking them is Franz's call.

    Offline work that remains queued:

    1. Status publication for non-complete outcomes (item 7). Add a status-only write kind through the existing outbox, marker and reconciliation path for quota_wait, cancelled, auth_failed, invalid_output and incomplete_expired. Give cancelled and superseded their own wording. Add golden fixtures for disputed/late, incomplete and superseded states.
    2. Durable capture-refusal record (item 9). Persist a per-request reason when capture fails for missing or ambiguous history, or an inventory over 128 files. Surface that reason as INCOMPLETE status instead of generic poll backoff. Test that a permanently refused PR does not hold back other PRs.
    3. GitHub reference neutralization (item 10). Neutralize #N, GH-N, owner/repo#N, owner/repo@sha and bare SHAs in model text, and add hostile-input tests.
    4. Fork and head-repository tests on the selected trigger path (items 11, 13). Cover a PR whose head.repo.id differs from 1169772257, a head-repo-only change, and a head-repo change during a read.
    5. Recovery after a process kill before the stopped audit receipt (item 2). This includes a bounded shutdown grace.

    No GitHub write, App action, credential use, inference or VM run was performed for this audit.

    🤖 Generated with Claude Code

  7. HemSoft commented on Oct 28, 2026

    @HemSoft
    CollaboratorAuthor

    Park polling, SQLite jobs, auth leases, custom App writes and promotion. Native Codex publishes its own review.

    The parked work stops. Revival requires the original unresolved gates, as stated in the approved plan.

  8. added
    parkedWork parked pending its documented revival condition
    on Oct 28, 2026
  9. HemSoft commented on Oct 29, 2026

    @HemSoft
    CollaboratorAuthor

    Closed per Franz (2026-10-29 19:34 ET, set-it-free-loop post: 'Agreed'): SFL stays inside pmd (project-manager). The standalone alpha controller, adapters and publisher are not being built. Next SFL work is project-manager #75 (review status) and #76 (changes between rounds). Reopen if that direction changes.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestparkedWork parked pending its documented revival condition

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions