Skip to content

Follow-up ledger: residuals from the 2026-11-18 maintainer merge batch (#369, #385, #388, #400) #402

Open

Description

@gzuuus

Residual items carried out of the four maintainer-merged PRs (review warnings/devin findings deliberately not fixed on the branches — bounded-rounds rule). Each is an edge-class item, not a blocker for the landed fixes.

From #369 (resume in-flight template deploys — merged 4a59570)

  • Fail-open in-flight status allowlist (template_deploy_retry.rs::deployment_is_in_flight matches pending|applying|watching|creating|deploying): a future PaaS live status would let an old-CLI retry treat a live rollout as settled and sign a second deployment — the exact cap#292 race. Pin the server's status vocabulary in a generated fixture/contract test so a PaaS addition trips CI here (Hermes warning).
  • Retirement rules mirror PaaS internals across repos (template_instance_rejection_retires_attempt code allowlist; unstamped_cap_rejection_is_permanent mirroring is_permanent_cap_rejection): no shared source of truth; drift either deletes a replayable resume file or wedges the name until the 23h window passes. Add a fixture/comment naming the owning code on both sides (Hermes warning).
  • Changed --log-key/--generate-log-key silently ignored on replay (Devin 🟡): the Replay branch is byte-identical by design; a diverging retry should refuse naming the flags instead of ignoring them.
  • Recovery-seed error blocks a usable saved bootstrap key (Devin 🟡): peek_password_bootstrap_key propagates the seed error before trying the saved per-app key; prefer the saved key when it matches.
  • ours_inflight permits replay beyond the 23h cutoff (Devin 🔍): verify whether PaaS extends idempotency retention for live rows; if not, bound the replay window explicitly.
  • PaaS-side pin for the destroy slug-fold (maintainer verification on fix(cli): resume in-flight template deploys and retire consumed keys (#292) #369): destroy Shell→shell holds because cli_delete_app runs validate_slug (server/src/http/apps/lifecycle.rs:475, lowercase at validation.rs:12) before lookup, but no test pins case folding on the delete path. Add one in enclava-paas.

From #385 (workload-publication recovery — merged 229d81f)

  • Recovery probe holds an open DB tx up to 10s (routes/deployments.rs:1878, routes/deployments/rollback.rs:536, routes/unlock.rs:1030): move the probe before tx.begin() where ordering allows, or bound it (Hermes warning; mitigations already in place — BeforeSupersession short-circuit + the 10s cap).
  • WORKLOAD_DEPLOYMENT_ID_LABEL (deploy.rs:7260) is the 4th hardcoded copy of enclava.dev/deployment-id — export the constant from the engine crate and consume it in all readers (Hermes suggestion).
  • publication_stamp_candidate duplicates plan_unpublished_workload_spec probe plumbing and the recovery rows load twice; the reconciler could call plan(..., BestEffort) (Hermes suggestion).
  • recover_deleting_workload_publications probes deleting apps sequentially (10s cap each, N apps stall a policy reconcile pass) — bound the per-pass budget or probe concurrently (Hermes suggestion).

From #388 (owner-wait probe lane — merged ba404d6)

  • Probe-predicate evaluations per shared claim: the candidate CTE evaluates the EXISTS per head-of-queue row (N past-budget rows cost N probes before reaching a fresh deploy) — add the metric/comment before signup widens the backlog (Hermes suggestion; the MAX_OWNER_WAIT_PROBE_WORKERS doc already names the queue-depth metric).
  • dispatcher_claim_wrappers_route_owner_wait_probes can lease foreign leftover rows in a shared test DB (harmless under --test-threads=1) — document drain-and-ignore or filter to fixture orgs (Hermes suggestion).
  • EXTENDED_OWNER_WAIT_PROBE_SLICE < poll_interval asserted against ApplyConfig::default() only — a runtime poll_interval below 3s silently degrades the one-poll property; debug_assert or doc note (Hermes suggestion).

From #400 (init stage/error pairing — merged 4265da2)

  • Stage-before-error write order is deliberate (the inverse names a stage the failure never reached) — add the one-line comment at record_failure_file (Hermes suggestion).
  • report_failure sets diagnostic.stage post-diagnose() by convention — make "stage only from last_stage()" structural via a builder/param (Hermes suggestion).

Activity

  1. enclavalabs commented on Nov 18, 2026

    @enclavalabs
    Contributor

    PR #403 clears the bounded-risk, in-repo residuals from this ledger (12 of 16 items), reviewed (grok-4.7, approve-with-comments) with sabotage-proven regression tests for the two behavior fixes (#369-3 diverging log-key flags on replay, #369-4 saved bootstrap key vs recovery-seed error).

    Remaining items, for whoever picks them up next:

    Also: the AGENTS.md required_checks list still needs bash scripts/gen-template-deployment-status-vocabulary.sh added (protected file; the CI step landed in #403).

  2. gzuuus commented on Nov 18, 2026

    @gzuuus
    ContributorAuthor

    Residuals from paas#429 (merged 2026-11-18, settle-replay + truthful pre-CAP reads): (1) add a direct claim-release-mid-flight regression test for the retry takeover arm, including the app-delete authority-transaction variant (current test covers the observe-settled arm; takeover is protected by the stable CAP idempotency key but only indirectly exercised); (2) observe the bounded 20 s claim-wait under a wedged lane — it deliberately changes previously immediate 409s into up-to-20 s waits, with a 500 ms claim UPDATE + reload poll per waiter; tune/back off only if production evidence shows latency or DB pressure. cap#369 (CLI cross-run resume) is already merged, so the timeout-alignment dependency is satisfied in source; both still await the next staging tuple.

  3. gzuuus commented on Nov 18, 2026

    @gzuuus
    ContributorAuthor

    Residual from paas#375 (merged 2026-11-18, locked-teardown destroy deferral + out-of-band rejection handling): add a teardown-waiver forwarding assertion to cli_delete_abandon_teardown_supersedes_deferred_locked_teardown — the reference CAP delete mock ignores request bodies, so the test pins lane supersession but not that abandon_teardown: true actually reaches CAP in the forwarded request (devin analysis finding on the merged head; the forwarding path is live-exercised by the e2e abandon-teardown row, so this is coverage hygiene, not a correctness gap).

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

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions