Skip to content

bug: installed plugin was 5 releases stale with no signal — agents enforced superseded rules while reporting compliance #52

Closed

Description

@cdeust

Reproduction (observed 2026-07-24, this repo, a live session)

$ grep -m1 '"version"' ~/.claude/plugins/marketplaces/zetetic-marketplace/.claude-plugin/plugin.json
  "version": "2.29.0",

$ git -C ~/.claude/plugins/marketplaces/zetetic-marketplace rev-list --count HEAD..origin/main
31

The installed plugin was 31 commits and 5 releases behind origin/main (2.29.0 vs 2.34.0). Nothing in the session surfaced this. The session ran for its full duration believing it was operating under current rules.

What was silently missing from the agents' loaded standards:

Release Content absent from the running plugin
2.34.0 §15 No-Deviation Rule — the task definition is the contract
2.33.0 redaction checker full mechanical inventory (CONTRAST / SETUP / PUFFERY groups)
2.32.1 redaction first-party identity rename
2.32.0 no-slop as a default output-quality gate (#43)
2.31.0 Advisor/Orchestrator multi-model loops

Why this is a defect and not an operator error

This plugin's entire product is rule currency. Its agents enforce rules/coding-standards.md, and §10 lists §1/§2/§7/§8/§13/§14/§15 as always-on at every stakes level. A silently stale copy does not degrade gracefully — it enforces a superseded standard while reporting compliance, which is worse than enforcing nothing, because the compliance report is now false. An agent on 2.29.0 will confidently pass a change that §15 would have blocked, and say so.

This is the §13 F2 failure class exactly: a degraded mode that is not explicit, not named in the output, and silently defaulted to. The repo requires that of the code it reviews; it does not currently require it of itself.

It is also the problem CBM solves in src/daemon/version_cohort.h ("crash-safe exact-build admission across CBM processes", CBM_VERSION_COHORT_CONFLICT, stable HELLO conflict envelope) — the parity pass against DeusData/codebase-memory-mcp (HEAD 97ce23f, v0.9.0) is what surfaced this, tracked by cdeust/enterprise-backlog#26. CBM treats which build am I as a first-class, checked property; we treat it as an assumption.

Acceptance criteria

  1. Every agent and hook run can determine the plugin version it is operating under, and that version is available to the agents' own output.
  2. A staleness check runs at session start, compares the installed version against the published release, and emits a visible notice when behind — naming the installed version, the latest version, and the count of releases missed.
  3. The check is fail-open and non-blocking: no network, no release metadata, or a rate-limited/offline host degrades to silence, never to a broken session or a hang. The offline path is tested.
  4. When the gap crosses a release that changed rules/coding-standards.md, the notice says so specifically — a rules change is the case that matters, a docs-only bump is not worth interrupting for.
  5. Compliance reports emitted by agents carry the rules version they were evaluated against, so a report read later is interpretable (§8 — a claim without provenance is not one).
  6. Every arm asserted (§13 A3), each with a named test: up to date (silent — a negative assertion per G4), behind by N, behind across a rules change, version undeterminable, network unavailable, malformed release metadata.
  7. Documented update path in the README: the exact command a user runs when the notice fires, verified end-to-end and quoted in the PR.

Non-goals

Auto-updating. The plugin does not mutate its own installation — it reports, the user decides. Silent self-update would replace one invisible state with another.

Definition of Done

Completion Ledger (§13.2), one row per arm in criterion 6 with its asserting test, plus the transcript from criterion 7. The notice-emission path itself must be asserted, not a downstream side effect (§13.4).

Activity

  1. cdeust commented on Jul 24, 2026

    @cdeust
    OwnerAuthor

    Root cause found — and it is worse than "the user did not update"

    The original report said the installed plugin sat at v2.29.0 while the repo was at v2.34.0. Follow-up investigation shows no user action could have fixed that, and the version-drift is structural.

    What is actually loaded

    ~/.claude/settings.json enables zetetic-team-subagents@cortex-plugins — not @zetetic-marketplace. The cortex-plugins marketplace is a clone of cdeust/Cortex, and its .claude-plugin/marketplace.json pins this plugin by version:

    { "name": "zetetic-team-subagents",
      "source": { "source": "github", "repo": "cdeust/zetetic-team-subagents" },
      "version": "2.29.0" }

    That pin is v2.29.0 on cdeust/Cortex@origin/main today, six releases after 2.29.0 shipped. Verified by fast-forwarding the marketplace clone to current origin: the pin does not move.

    $ git -C ~/.claude/plugins/marketplaces/cortex-plugins pull --ff-only origin main
    $ ... marketplace.json -> zetetic-team-subagents=v2.29.0
    

    The same manifest pins cortex-viz at v2.6.3 while that repo has released v2.7.1.

    Why this matters more than the original report suggested

    Releasing this repo does not ship it. Cutting v2.30.0 … v2.34.0 delivered nothing to any user, because delivery is gated by a version string in a different repository's manifest that nothing updates. Five consecutive releases — including §15, the redaction gates, and the Fable loops — went out to zero installs, and no signal anywhere reported that.

    That reframes the defect. It is not "installs can go stale"; it is a cross-repo release coupling with no automation and no alarm. Every release since 2.29.0 has been, in effect, unpublished.

    It also means the fix for this issue cannot live entirely here: the staleness notice (criteria 1–7 below) still belongs in this repo, but the pin must be bumped by the release process of cdeust/Cortex — filed as cdeust/Cortex#179.

    Consequence for the criteria

    Criterion 2 changes meaning. Comparing the installed version against "the published release" is not sufficient, because the marketplace pin is what determines what a user can even install. The check must compare three values, not two:

    • what is installed
    • what the marketplace pins
    • what this repo has released

    and it must distinguish the two failure modes, because they have different owners and different fixes:

    Case Meaning Fix owner
    installed < pin the user has not updated user
    pin < latest release the release was never delivered cdeust/Cortex#179

    The second case is the one that was silently true for six releases and is invisible to a two-value check. A notice that says "you are up to date" because the install matches a stale pin is a false compliance statement — the same defect class this issue is about, one level up.

    Added criterion

    1. The staleness check reports pin lag as a distinct, named condition from install lag, with the owning repo identified in the message. A test asserts the pin-lag arm specifically — the arm that six releases of real evidence say is the one that actually fires.
  2. added a commit that references this issue on Jul 25, 2026
  3. added a commit that references this issue on Jul 27, 2026
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

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions