Repository navigation
Mixed-version teams: tests_update (2026.10.12+) writes per-test updated_by/updated_at that pre-10.12 hosts drop from hashes — valid history fails verify and CI goes red #1458
Description
Problem
tests_update, added in 2026.10.12 ("Edit linked-test metadata in place with preserved provenance"), writes new per-test fields: metadata.tests[i].updated_by and metadata.tests[i].updated_at. Hosts older than 2026.10.12 don't know these fields. When they canonicalize the item for hashing they drop them, so they compute a different item hash and report history corruption on a perfectly valid history.
In a fleet where agents use a newer global pm while the repository pins an older @unbrained/pm-cli, which is the normal state between Dependabot bumps, a single linked-test edit turns the repo's CI red.
Real occurrence (public repo unbraind/pm-review, PR #1, pins 2026.10.10)
An agent ran pm test <id> --update … --timeout 600 with a global 2026.10.13 host. History entry 23 of pm-review-round:
{"op":"tests_update","patch":[
{"op":"add","path":"/metadata/tests/0/timeout_seconds","value":600},
{"op":"add","path":"/metadata/tests/0/updated_by","value":"codex-agent"},
{"op":"add","path":"/metadata/tests/0/updated_at","value":"2026-10-13T12:56:35.894Z"},
{"op":"replace","path":"/metadata/updated_at","value":"2026-10-13T12:56:35.894Z"}]}Same files, two hosts:
pm 2026.10.13 history pm-review-round --verify → ok:true, current_matches_latest:true
pm 2026.10.10 history pm-review-round --verify → ok:false
errors: verify_failed:after_hash_mismatch:entry_23, verify_failed:current_item_hash_mismatch
CI (pinned 2026.10.10) then fails pm health with history_drift_chain_mismatch:pm-review-round and history_drift_hash_mismatch:pm-review-round.
A second agent in another repo (pm-context) hit the same thing. The pinned host also rewrote the tests block and dropped updated_by/updated_at when it next touched the item, so the "fix" an agent reaches for is data loss: re-anchoring history with the older host.
Why this matters
The CLI's core promise is that many agents on many branches can work on one tracker and merge losslessly. Mixed host versions are unavoidable in that world: global installs, pinned devDependencies and CI images all drift. This is the same class as #976, #1040 and #1162, now triggered by a new feature's fields rather than by hashing changes.
Proposal
- Forward-compatible hashing: older hosts hash unknown nested fields verbatim instead of dropping them, or the canonical form is declared field-agnostic for
metadata.tests[]entries. - Or gate new fields behind a declared format/feature version: a newer host writing to a tracker whose
settings.json/ format_version predates the feature either refuses or writes the legacy shape, unless the tracker opts in (for examplepm workspace upgrade). - Diagnostics: when verification fails on an op the verifying host doesn't recognize, or on fields it doesn't know, report
history_written_by_newer_host(naming the op and fields) rather than generic chain/hash drift, and recommend upgrading the verifying host. Never suggest a re-anchor that drops the newer data. - A cross-version test matrix: history written by N verifies under N-1 and N-2 for every new op and field.
Activity
unbraind commentedon Oct 13, 2026
Second independent occurrence, in a different repo (unbraind/pm-github PR #115). The repository pins
@unbrained/pm-cli2026.10.9; an agent wrote with a global 2026.10.13. Both reportitem_hash_version: 3for the same TOON document, yet 2026.10.9 cannot reproduce thebefore_hash/after_hashvalues 2026.10.13 stored. CI then reportshistory_drift_chain_mismatchandhistory_drift_hash_mismatchat every write boundary, although every patch replays cleanly with no content loss.Two consequences worth designing for:
fix(pm): restore the rewritten fmrf history and bump the pinned pm-cli
unbraind commentedon Oct 13, 2026
Tracked by pm-gh1458. The linked-test editor and published-artifact matrix are linked predecessors. This owner captures both reported occurrences, nested-field/hash preservation and safe verifier-upgrade diagnostics; independent reproduction remains pending.
feat: revalidate live GitHub blockers without unblocking items (#115)
unbraind commentedon Oct 14, 2026
Independent synthetic published-artifact reproduction now confirms this existing report for 2026.10.10 → 2026.10.13 → 2026.10.10.
Provenance/scope: both were existing installed official npm artifacts, with file bytes checked against their retained tarballs (10.10: 6,375 files; 10.13: 6,411 files). No fetch or upgrade was used. The test used disposable unpinned Git worktrees, an isolated HOME/global PM root, disabled telemetry and a deny-network preload. This does not claim that the existing project-pin guard is absent, or that a 10.14 package was tested.
Minimal sequence (PM path and JSON output applied consistently):
pm test <id> --add 'command=node --version'.history <id> --verifyreturnok:true(two records).pm test <id> --update 'node --version' --set-json '{"timeout_seconds":600}' --author newer-editor. This adds the nestedupdated_by/updated_atstamps.verify_failed:after_hash_mismatch:entry_3andverify_failed:current_item_hash_mismatch. The old read omits the two stamps, but these read-only commands do not change the files.comments <id> --add 'Synthetic older-host ordinary mutation'succeeds and removes both nested stamps from the item. It appends a fourth event with an incompatible before-hash; both versions then reject verification. Original history records remain a byte-identical prefix, so a rewrite/re-anchor is not needed to trigger the incompatible chain.Negative controls:
The 10.14
item-format.tsGit blob is identical to 10.13, but that is only source evidence; this reproduction does not establish 10.14 installed behavior. Suggested next acceptance remains the existing pm-gh1458 scope: per-operation/nested-field matrix coverage and a preserve-or-refuse boundary before unsupported writes, with diagnostics pointing to a compatible verifier rather than destructive history repair.