fix: route reports by booted image - #1010
kubestellar-hive[bot] wants to merge 4 commits into
Conversation
Signed-off-by: Copilot <223556219+Copilot@users.noreply.github.com>
hanthor
left a comment
hanthor
left a comment
There was a problem hiding this comment.
CI is genuinely red on this head commit: the PR's own new regression test test_bonedigger_report.bats #4 ("image info prefers booted image reference over stale build metadata") fails in CI with an output mismatch — a real failure, not an infra flake (build/E2E jobs otherwise pass). Root cause looks like a shell-quoting bug in the test itself (a single-quoted JSON literal embedded inside an outer single-quoted bash -c '...' string), which needs fixing before merge regardless of whether the underlying read_boot_status/read_image_info reordering fix (which does look correct on inspection) is sound.
Generated by Claude Code
castrojo
changed the title
![sec-check[bot]](https://github.githubassets.com/images/gravatars/gravatar-user-420.png){kind=avatar}
…ng (projectbluefin#1045) ## Refactor **Cluster:** image-name/tag → upstream GitHub repository routing. **Files:** `system_files/shared/usr/libexec/ublue-image-repo` (new), `system_files/bluefin/usr/libexec/bonedigger-report` (`route_issue_repo`), `system_files/bluefin/usr/share/ublue-os/just/changelog.just` (`changelogs`), `tests/test_image_repo.bats` (new), `tests/test_bonedigger_report.bats`, `tests/test_changelog.bats`. `common` decided "which upstream repo owns this image" in two places with two different grammars: | site | dakota match | unknown image | |---|---|---| | `bonedigger-report:route_issue_repo` | `dakota*` glob | `projectbluefin/common` | | `changelog.just:changelogs` | `== "dakota"` exact | `projectbluefin/bluefin` | They had already drifted. `dakota-nvidia` is a published image (`docs/skills/image-registry.md`), matched by the glob but not by the exact comparison — so `ujust changelogs` on `dakota-nvidia` fetched release notes from `projectbluefin/bluefin`. This extracts the grammar into `/usr/libexec/ublue-image-repo` in the **`shared/`** overlay layer (the decision is variant-agnostic; both copies previously sat in the GNOME-desktop-only `bluefin/` layer). The one legitimate per-caller difference — the fallback repo for unrecognised images — is passed in as `--default`, so each call site keeps its existing fallback and no unrelated behaviour changes. Both consumers become one-line delegations. `tests/test_image_repo.bats` covers the grammar directly, including a `dakota-nvidia` regression test and a structural invariant asserting no consumer restates `projectbluefin/bluefin-lts` / `projectbluefin/dakota` inline. **Verification:** `bats tests/test_image_repo.bats tests/test_bonedigger_report.bats tests/test_changelog.bats` — 46/46 pass. **Overlap check:** projectbluefin#1010 also touches `bonedigger-report`, but only `read_boot_status` (how `IMAGE_NAME`/`IMAGE_REF` are *sourced*); this PR touches only `route_issue_repo` (how they are *routed*) — disjoint hunks, complementary. projectbluefin#1042 moves `00-entry.just` into `shared/` and does not touch `changelog.just`. Refs projectbluefin#1044 --- *Filed by architect agent (ACMM L5 — hold-gated mode). Hold-gated: human review required.* — hive: agent=architect backend=copilot model=claude-opus-5 Signed-off-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com> Co-authored-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
…luefin#1034) ## Test Improvement Adds `tests/test_shared_just.bats` — 14 BATS cases covering `system_files/shared/usr/share/ublue-os/just/shared.just`, which had **zero** test coverage. ### Why this file `powerwash` is the most destructive recipe in the repo: it ends in `sudo bootc install reset --experimental`, an irreversible factory reset. It is guarded by two `gum choose` confirmations plus a `bctl powerwash` fast path. None of that gating was verified by any test. ### How The `powerwash` body is a `#!/usr/bin/bash` shebang recipe with no `just` interpolation, so the test extracts it verbatim and **executes** it against stubbed `bctl`/`gum`/`sudo` binaries. This exercises real control flow rather than grepping recipe text, and requires no `just` binary in CI. Asserted behavior: - `bctl` present → delegates to `bctl powerwash`, and `gum`/`sudo` are never invoked - declining the first confirmation → cancels, only one prompt, no `sudo` - declining the second confirmation → cancels, no `sudo` - non-affirmative (`maybe`) and empty answers → treated as declines - two prompts are always required; a single "yes" never wipes - both confirmations → `sudo bootc install reset --experimental`, exactly once - both prompts carry their warning text - `toggle-tpm2` uses the absolute `/usr/bin/luks-tpm2-autounlock` path All 14 pass locally (bats 1.14.0). Mutation-checked: deleting the second confirmation from `shared.just` fails 3 of the tests. ### Scope / disjointness Claimed ground: **`tests/test_shared_just.bats` only** — one new test file, no production code touched. Deliberately does **not** touch `.github/workflows/unit-tests.yml` (claimed by open PRs projectbluefin#942 and projectbluefin#961) and does **not** touch `system.just` (projectbluefin#942), `apps.just` (projectbluefin#1007), `default.just` (projectbluefin#1003), `bonedigger-report` (projectbluefin#1010), `hooks.py` (projectbluefin#1011), or `ublue-fastfetch` (projectbluefin#1005). CI registration of this and the other test files not yet wired into `unit-tests.yml` is tracked separately by projectbluefin#968. --- *Filed by quality agent (hold-gated mode). Human review required. Do not merge on my behalf.* — hive: agent=quality backend=copilot model=claude-opus-5 Signed-off-by: kubestellar-hive[bot] <kubestellar-hive[bot]@users.noreply.github.com> Co-authored-by: kubestellar-hive[bot] <kubestellar-hive[bot]@users.noreply.github.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…th invalid JSON
The two image-info tests embedded their BOOTC_JSON fixture inside the
single-quoted 'bash -c' body:
run bash -c '
...
BOOTC_JSON='{"status":{"booted":...}}'
...
' _ "$BONEDIGGER_SCRIPT" "$WORKDIR/stale-image-info.json"
The inner quotes close the outer single-quoted string, so the JSON was
left unquoted and the outer shell stripped every double quote from it:
{status:{booted:{image:{image:{image:ghcr.io/projectbluefin/dakota:stable}}}}}
jq could not parse that, returned empty, and read_image_info silently
fell back to the stale image-info.json. That made
'image info prefers booted image reference over stale build metadata'
fail, and made 'image info falls back when no booted image is reported'
pass for the wrong reason — it exercised the fallback because the
fixture was corrupt, not because booted was null.
Pass the fixture as a positional argument instead so it reaches the
script intact. Both tests now exercise the branch they name; mutating
the jq query in read_image_info turns the first one red, which it did
not before.
Signed-off-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
|
Status after pushing the test fix. The original red check on this PR was Root cause was in the test, not in BOOTC_JSON='{"status":{"booted":...}}'The inner quotes closed the outer single-quoted string, so the outer shell stripped every double quote and Fixed by passing the fixture as a positional argument. Remaining red check is unrelated to this diff —
|
|
Closing: the fix this PR implements already landed on main through #1020 (read the booted tag from bootc, not baked image-info) and #1030 (fall back to the boot-time snapshot when bootc status fails). The tracking issue #1009 is closed as completed. Thanks for the investigation — it matched what was merged. |
Fix
Use the booted image reference from
bootc status --jsonwhen generatingujust reportmetadata and repository routing. Keep build-timeimage-info.jsonas fallback when no booted reference is available, and add regression coverage for stale metadata.Refs #1009
Filed by scanner agent (ACMM L5 — hold-gated mode). Hold-gated: human review required.
— hive: agent=scanner backend=copilot model=gpt-5.6-luna