Skip to content
TheChatArchiveThe Chat Archive
gpt-5.6-sol· openaiHugging Face ibragim-bad/swe_rebench_07_2026_trajectories

Stale-detection timestamp-field allowlist misses spec-declared date fields Retro finding…

  1. User

    Stale-detection timestamp-field allowlist misses spec-declared date fields Retro finding · priority: P2 · component: generator · source: Trello CLI generation (printing-press 4.9.0)

    What happened

    The generated internal/cli/pm_stale.go matches inactivity against a hardcoded staleTimestampFields allowlist (updatedAt, updated_at, updatedDate, modifiedAt, modified_at, lastModified, last_modified, lastEditedTime, last_edited_time, updated, last_updated). Trello cards use dateLastActivity, which is absent. The flagship stale command returned "No stale items found" while a hand-built health command (reading dateLastActivity) counted 93 stale cards on the same hydrated data. Silent-empty flagship command, no error.

    Root cause

    The stale template hardcodes a timestamp-name allowlist instead of also deriving candidate timestamp fields from the spec (properties with format: date-time).

    Durable fix (proposed)

    Build the stale timestamp-field set as hardcoded allowlist ∪ spec-declared date-time property names for the synced resources. Guard: union is additive only — never removes existing allowlist entries, so no regression for currently-working APIs.

    Test

    • Positive: a spec with a dateLastActivity (format: date-time) card property produces a stale that finds rows older than N days.
    • Negative: an API using updated_at is unchanged.

    Repro / evidence

    trello-pp-cli stale --days 14 → empty; trello-pp-cli health → 93 stale on the same mirror. Fixed by hand-adding dateLastActivity to the allowlist. The allowlist approach will keep silently breaking the flagship command for the next API with an unlisted activity field.

    Interface

    No new interfaces are introduced.

  2. Assistant

    I’ll trace how synced resource schemas feed the stale template, then add the date-time field union at generation time with focused regression coverage. I’ll also run formatting, the relevant package tests, the full Go suite, and golden verification because generated output may change.

  3. Assistant

    rg isn’t installed in this workspace, so I’m using targeted find/grep searches instead. I’m first locating the template data model and existing generator tests to keep this a systemic Printing Press fix.

  4. Assistant

    The stale command reads the generic synced-resource JSON, so the fix belongs in workflow rendering: identify the profiler-selected sync endpoints, collect their response fields marked format: date-time, and append only non-default names to the existing ordered allowlist. This preserves every current default and avoids duplicate updated_at entries.

  5. Assistant

    The implementation and regression test are in place. I’m validating in offline mode; the first Go invocation attempted module resolution, so I stopped it and will keep all remaining test commands constrained with GOPROXY=off to honor the no-internet requirement.

  6. Assistant

    The generator-level assertions pass far enough to confirm rendering, but the nested generated-CLI test cannot compile offline because its separate module dependencies (Cobra, TOML, SQLite) are not cached. I’m keeping the end-to-end regression test for CI and adding a dependency-free unit test around field selection so the core behavior can still be verified locally without weakening coverage.

  7. Assistant

    Focused offline tests now pass, including deterministic selection, additive default preservation, generated source, and the end-to-end test fixture construction. I’m moving to repository-wide formatting/tests and golden verification; any unavailable uncached dependency will be reported distinctly from a code failure.

  8. Assistant

    The first full-suite attempt hit the container’s process limit (fork/exec … resource temporarily unavailable), not test assertions. I’m rerunning with Go parallelism capped so the suite can complete within the environment. Golden verification is separately blocked because this image lacks jq.

  9. Assistant

    The implementation is clean and the Printing Press binary builds offline. I’ve kept the generated-CLI end-to-end test unconditional for normal CI rather than weakening it for this container; locally, its compile step remains blocked only by uncached dependencies, while the dependency-free regression tests pass.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.