version: "1.0.1" name: payload-snapshot description: Snapshot OpenShift payload data (release controller, PR diffs, comments, CI jobs, JUnit results, regression tracking) to a local directory for offline analysis

Payload Snapshot

This skill downloads all data needed to analyze an OpenShift payload into a local directory tree. The resulting snapshot can be navigated entirely via file reads — no live API calls required during analysis.

When to Use This Skill

Use this skill when you need to:

• Analyze a rejected payload and want all data available locally before starting
• Create a reproducible snapshot of payload state at a point in time
• Track test failure regressions across multiple payloads
• Work offline or reduce API calls during payload analysis

Prerequisites

1. Python 3 (3.10 or later)

• Uses only standard library (no external dependencies)

1. GitHub CLI (`gh`) — for PR diff, comment, and job data

• Install: brew install gh (macOS) or see https://cli.github.com
• Authenticate: gh auth login
• Without gh, release controller data is still fetched; PR data is skipped

1. Google Cloud SDK (`gcloud`) — for JUnit test result download

• Install: brew install google-cloud-sdk (macOS) or see https://cloud.google.com/sdk
• Authentication is not required — the CI artifact buckets are public and

are read anonymously when no account is configured

• Without gcloud entirely, JUnit data is skipped and the snapshot is

reported as incomplete (see Data completeness)

1. Container runtime (`podman`) — for RHCOS RPMDB extraction

• Install: available in most Linux distributions; brew install podman on macOS
• Requires access to the release image registry (e.g., registry.ci.openshift.org)
• Without podman, RPMDB data is skipped; all other data is still collected

1. `rpm` CLI — for the RPM changelog diffs between payloads

• Queries the already-extracted RPMDBs locally; no network, no root
• Without rpm, changelog diffs are skipped; the RPMDBs themselves are still extracted

1. Network access to:

• *.ocp.releases.ci.openshift.org (release controller)
• sippy.dptools.openshift.org (historical payload fallback)
• api.github.com (via gh CLI)
• storage.googleapis.com (via gcloud CLI)

Implementation Steps

Step 1: Run the Snapshot Script

script_path="plugins/ci/skills/payload-snapshot/scripts/payload_snapshot.py"
 
# Snapshot a specific payload
python3 "$script_path" 4.22.0-0.nightly-2026-02-25-152806
 
# Custom output directory
python3 "$script_path" 4.22.0-0.nightly-2026-02-25-152806 --output-dir .work/snapshot
 
# Limit chain depth
python3 "$script_path" 4.22.0-0.nightly-2026-02-25-152806 --max-chain 5
 
# Skip JUnit download (faster, still generates job structure and summary)
python3 "$script_path" 4.22.0-0.nightly-2026-02-25-152806 --no-junit
 
# Skip RPMDB extraction
python3 "$script_path" 4.22.0-0.nightly-2026-02-25-152806 --no-rpmdb
 
# Skip the RPM changelog diffs (RPMDBs are still extracted)
python3 "$script_path" 4.22.0-0.nightly-2026-02-25-152806 --no-rpm-changelogs
 
# Force Sippy for all payload metadata (normally fallback is automatic)
python3 "$script_path" 4.22.0-0.nightly-2026-02-25-152806 --sippy

The script will:

1. Parse the payload tag to determine version, stream, and architecture
2. Probe all available streams for the version (nightly, ci, across architectures)
3. Chain backwards through Sippy's time-ordered release tag list until finding one where all blocking jobs passed, restoring tags garbage collected from the release controller
4. For each payload in the chain, prefer release controller data and changelogs; fall back per historical tag or cross-tag diff to Sippy payload, PR, and job data
5. Split jobs into blocking/informing directories with metadata and GCS browser links
6. For each failed blocking job, download and parse JUnit XML test results
7. For each failed blocking job, download build-log.txt from GCS and extract error/warning lines + log tail
8. Track test failure regressions — when did each failure first appear?
9. Track per-job failure streaks — consecutive failures, originating payload, failure pattern
10. For each unique PR across all changelogs, fetch the git diff, comments, and CI jobs via gh
11. For each payload in the chain, extract the full RPM database from RHCOS images via podman
12. Diff the target payload's RPM changelogs against every older payload in the chain, using the extracted RPMDBs
13. Generate summary.json with comprehensive triage data, plus AGENTS.md/CLAUDE.md for agent orientation

Step 2: Navigate the Snapshot

The output directory is structured for easy navigation:

payload/
  <version>/
    <stream>/
      summary.json                         # START HERE — full triage data
      CLAUDE.md                            # Imports AGENTS.md for Claude Code
      AGENTS.md                            # Dynamic snapshot orientation doc
      streams.json                         # All streams for this version
      <tag>/                               # Each payload in the chain
        payload.json                       # Release controller API response
        changelog.json                     # PRs that changed vs. previous payload
        regressions.json                   # Test failure regression tracking
        jobs/
          blocking/
            <job-name>/
              job.json                     # Job metadata (state, URLs, GCS link, retries)
              build_log.json               # Error/warning lines + log tail (failed only)
              junit/                       # Only for failed jobs
                junit_operator.xml         # CI phase results
                junit-aggregated.xml       # Aggregated jobs only
                results.json               # Parsed test failures (full output)
          informing/
            <job-name>/
              job.json                     # Job metadata only (no JUnit/build log)
        <component>/                       # e.g., machine-config-operator
          prs/
            <pr_number>/
              code.diff                    # Git diff of the PR
              comments.json                # PR comments and reviews
              jobs.json                    # CI check runs
        rpmdb/                             # RPMDB from RHCOS images
          rhel-coreos/                     # queryable with rpm --dbpath
            rpmdb.sqlite
          rhel-coreos-10/
            rpmdb.sqlite
        rpm-changelogs/                    # Target payload only
          rhel-coreos-10/                  # One diff per older payload
            <older-payload-tag>.md

Step 3: Use the Data

Find failed blocking jobs (with streaks):

jq '.blocking_jobs.failed_jobs[] | {name, state, streak: .streak.streak_length, pattern: .streak.failure_pattern}' payload/<version>/<stream>/summary.json

Check test failures and when they started:

jq '.[] | {test: .test_name, first_failed: .first_failed_in, payloads: .payloads_failing, jobs: .jobs}' payload/<version>/<stream>/<tag>/regressions.json

List PRs in a payload:

jq '.changeLogJson.updatedImages[].commits[] | {component: .name, pr: .pullURL, subject: .subject}' payload/<version>/<stream>/<tag>/changelog.json

Read a specific PR's diff:

cat payload/<version>/<stream>/<tag>/<component>/prs/<number>/code.diff

Check JUnit failures for a specific job:

jq '.[].name' payload/<version>/<stream>/<tag>/jobs/blocking/<job-name>/junit/results.json

Query RHCOS RPM packages (using the extracted rpmdb.sqlite):

rpm -qa --dbpath $(pwd)/payload/<version>/<stream>/<tag>/rpmdb/rhel-coreos

See what changed in the RHCOS RPMs since the baseline (inline in summary.json — no file read needed):

jq '.rpm_changelogs[] | select(.diff) | {variant, compared_tag, diff}' payload/<version>/<stream>/summary.json

Find which payload in the chain introduced a package bump:

cat payload/<version>/<stream>/<target-tag>/rpm-changelogs/<variant>/<older-tag>.md

CLI Reference

python3 payload_snapshot.py <payload_tag> [OPTIONS]
 
Positional:
  payload_tag          Payload tag (e.g., 4.22.0-0.nightly-2026-02-25-152806)
 
Options:
  --output-dir DIR     Base output directory (default: payload)
  --max-chain N        Maximum backward chain depth (default: 20)
  --workers N          Parallel workers for API calls (default: 8)
  --no-junit           Skip JUnit download and regression tracking
  --no-rpmdb           Skip RHCOS RPMDB extraction
  --no-rpm-changelogs  Skip the RPM changelog diffs between the target payload
                       and the older payloads in the chain
  --sippy              Force Sippy for all payload metadata instead of using
                       the automatic release-controller-first fallback
  --fail-on-incomplete Exit 1 if any requested data could not be collected

Output Files

streams.json

Lists all available streams for the payload's version.

summary.json

Comprehensive stream-level triage data — start here. Contains:

• Payload metadata: payload_tag, phase, release_url, source, architecture, stream, version
• Chain data: chain_length, baseline_tag, hours_since_baseline
• blocking_jobs.failed_jobs[] — detailed objects with name, state, prow_url, gcs_url, and relative path job_json. May include: rhcos_version, streak (with streak_length, originating_payload, is_new_failure, failure_pattern), build_log_errors, test_failure_count, and relative paths junit_results, build_log
• informing_jobs.failed_jobs[] — job name strings
• test_failures.blocking[] — gating failures only: test_name, jobs, first_failed_in, payloads_failing, failure_message, failure_text (full, not truncated). These are the failures that can fail a job and therefore reject the payload.
• test_failures.informing[] / test_failures.flakes[] — test_name, jobs. Neither can fail a job. No onset is tracked for them, because an onset implies there is a culprit to find.
• payloads[] — per-payload entries with tag, phase, source, changelog_source, relative file paths, prs[] with component/diff/comments paths, rhcos_changes[] with RPM diffs per RHCOS variant (package versions only — no changelog text), and, on every payload but the target, rpm_changelogs[] pointing at the report holding that text
• rhcos_rpms[] — RPMDB metadata for the target payload's RHCOS variants: tag, name, pullspec, rpmdb (relative path to rpmdb.sqlite)
• rpm_changelogs[] — one RPM diff per (RHCOS variant, older payload in the chain): variant, compared_tag, is_baseline, changed/added/removed counts, and changelogs (relative path to the full report). The oldest surviving comparison per variant — normally the chain baseline — also carries diff inline, so target-vs-baseline needs no file read: diff.changed[] with package, old, new and changelog (the entries that version added), plus diff.added[] / diff.removed[] with package and version. The intermediate hops are a subset of that diff and stay behind their changelogs path; read them to find which hop introduced a given package bump. If the true chain baseline's RPMDB couldn't be read for a variant, the next-oldest readable comparison takes over is_baseline/diff instead, flagged with baseline_rpmdb_missing: true so consumers know the diff doesn't reach all the way back to the chain's actual start.
• data_complete — true when all requested data was ultimately collected, including via a fallback after an initial read failed. false means some requested data could not be read at all.
• collection_errors[] — every read failure encountered. Each entry has reason, command, and optionally detail, stage, job, payload_tag, recovered. Reasons: auth, timeout, gcloud_missing, command_failed, junit_unavailable (nothing readable), junit_missing (nothing discovered), junit_unparseable (corrupt XML), junit_partial (some files unread), build_log_unavailable.
• recovered: true means a fallback subsequently obtained the data. These entries are diagnostic only (useful for spotting a timeout that needs tuning) and do not make data_complete false.
• data_complete is false only when at least one error was not recovered.

<a id="data-completeness"></a>

Only gating results count as failures

A test result falls into exactly one of three categories, and only the last can fail a job or reject a payload:

Informing tests are run to stabilize them and are not expected to gate. A missing lifecycle attribute means the test does gate — the attribute exists only to opt a test out.

results.json records all three so nothing is hidden, each entry carrying status (failed, error, flake) and test_lifecycle (blocking, informing).

test_failure_count on a failed job counts only gating results — it is the failure count, and flakes and informing tests are not added to it. Those two are listed by name under test_failures.flakes[] and test_failures.informing[]. Regression onset (first_failed_in) is derived from gating failures alone.

"Informing job" vs "informing test" — two unrelated concepts

The word "informing" appears in two places with completely different meanings. Confusing them produces wrong analysis:

An informing test can run inside a blocking job. An informing job can contain blocking tests. They are orthogonal. Never combine them in the same section or count.

gcloud credentials are not required

The CI artifact buckets are public. When gcloud has no active account it is run in anonymous mode automatically, so an unauthenticated environment still produces a complete snapshot. Authenticate only if you also need private buckets.

Missing data is absent, never empty

When a collection step fails, the affected file is not written and the corresponding summary field is omitted — it is never emitted as an empty list or a zero count. Specifically:

• If JUnit could not be read for a job, results.json is not created, the

job entry has no test_failure_count and junit_results, and instead carries junit_collection_failed: true.

• An absent test_failure_count therefore means unknown, whereas 0 means

verified clean.

• If some JUnit files were read and others were not, results.json holds

the real results that were obtained and the job entry adds junit_collection_partial: true. test_failure_count is then a lower bound, not a total.

• A failed job with no readable JUnit at all — none discovered

(junit_missing), or nothing that parsed (junit_unparseable) — is left without a count entirely. A job whose tests may never have run is unknown, not clean. A parse failure alongside other readable files is the partial case above, not this one.

Consumers must distinguish these. Treating unreadable data as "no test failures" makes a broken job look like it failed for some other reason, which misdirects root-cause analysis. Check data_complete before drawing any conclusion from an absence of test failures. Use --fail-on-incomplete in automation to exit non-zero rather than emit a partial snapshot.

State is persisted to collection_errors.json beside summary.json, so a later process can tell that an existing results.json came from an incomplete read.

AGENTS.md / CLAUDE.md

Dynamic orientation document generated at snapshot time. Contains the specific payload tag, chain, failed jobs, file layout, key concepts, and summary.json schema. CLAUDE.md imports AGENTS.md via @AGENTS.md.

payload.json

Full release controller response including blockingJobs, informingJobs, and asyncJobs with their states, Prow URLs, and retry attempt URLs.

changelog.json

Release controller diff response with changeLogJson.updatedImages listing every PR that changed between this payload and its predecessor. Also contains nodeImageStreams at the top level — RPM diffs for each RHCOS variant showing which packages changed between payloads, as package names and versions only.

For the target payload, _rpm_changelogs[] is injected alongside _source: variant, compared_tag (the predecessor, the same comparison this file describes) and changelogs, a path relative to the payload directory pointing at the report that carries the changelog text nodeImageStreams lacks. payload.json gets the same key.

rhcos_changes[] (in payloads[] within summary.json)

Per-payload RHCOS RPM diff data extracted from the changelog's nodeImageStreams. Each entry contains:

• name: Human-readable RHCOS version name (e.g., "Red Hat Enterprise Linux CoreOS 10.2")
• tag: RHCOS image stream tag (rhel-coreos for RHCOS 9, rhel-coreos-10 for RHCOS 10)
• changed: Dict of {package_name: {"old": old_version, "new": new_version}}
• added: Dict of newly added packages {package_name: new_version} (when present)
• removed: Dict of removed packages {package_name: old_version} (when present)

Only variants with non-empty rpmDiff are included. When the snapshot is created from Sippy-based changelogs (which lack nodeImageStreams), this field is absent.

rpmdb/<variant>/rpmdb.sqlite

The rpmdb.sqlite file extracted from RHCOS images via podman. One directory per RHCOS variant (e.g., rpmdb/rhel-coreos/, rpmdb/rhel-coreos-10/), each containing rpmdb.sqlite. Queryable with rpm -qa --dbpath <absolute-path-to-variant-dir> or directly with sqlite3.

Extracted for every payload in the chain. Skipped when podman is unavailable or --no-rpmdb is passed.

rpm-changelogs/<variant>/<older-payload-tag>.md

Written for the target payload only, one file per (RHCOS variant, older payload in the chain) — including the baseline. Each file answers "what do this variant's RPMs have in the target that they did not have in that older payload?", by querying both already-extracted RPMDBs with rpm locally. No API calls and no image pulls are involved, and rhcos_changes[] (which the release controller computes) is not consulted.

Each file has three sections:

• Changed packages — <old version> -> <new version> plus only the changelog entries the new version added. RPM changelogs are additive and grow from the top, so the cut is at the first entry the old version already had. (Not a common-suffix subtraction: rpm trims stale entries off the bottom at build time, so two builds of the same package weeks apart routinely disagree about their oldest entries.) A version bump with no new entry is reported as a rebuild.
• New packages — packages the older payload did not have at all, with their full changelog.
• Removed packages — name and version only. Nothing to read a changelog from.

Deliberate simplifications, because the consumer is an LLM that can read a full dump when it has to:

• Each older payload is compared against the target independently. No attempt is made to follow a package's evolution hop by hop. A package removed and later re-added shows up as "no difference" against the baseline and as some difference against an intermediate payload — that is the honest answer to each of those two questions, and no further reconciliation is done.
• A changelog whose history was rewritten degrades to a full dump. When the two versions share no entries, the whole changelog is emitted rather than guessing a cut point from version strings.
• A variant missing from the older payload is skipped, rather than reported as several hundred new packages.

Indexed by summary.json's rpm_changelogs[], whose baseline entry repeats the same content inline under diff. Read these files for the intermediate hops, or when the prose layout is easier than the JSON.

Skipped when rpm is unavailable, --no-rpm-changelogs is passed, the RPMDBs were not extracted, or the chain has no older payload (chain_length < 2).

regressions.json

Per-payload regression tracking data. For each failing test in the target payload:

• test_name: the failing test
• jobs: which jobs it fails in
• first_failed_in: the earliest payload in the chain where it was failing
• payloads_failing: how many consecutive payloads it has been failing
• failure_message: the error message
• failure_text: full failure output

job.json

Per-job metadata including name, state, lifecycle (blocking/informing), Prow URL, GCS browser URL (gcs_url), retry count, whether it's an aggregated job, GCS bucket path, and rhcos_version.

The rhcos_version field is determined from the job name and OCP version:

• rhcos9_10 — heterogeneous cluster (mixed RHCOS 9 and 10 node pools)
• rhcos10 — RHCOS 10 only
• rhcos9 — RHCOS 9 (explicit)
• rhcos9-default — no explicit fragment; defaults to RHCOS 9 for OCP 4.x installs (including major upgrades to 5.x)
• rhcos10-default — no explicit fragment; defaults to RHCOS 10 for OCP 5.x fresh installs

build_log.json (failed blocking jobs only)

Extracted from build-log.txt in GCS (handles gzip decompression). Contains:

• total_lines: total line count of the build log
• error_warning_count: number of lines matching error/warning patterns
• error_warning_lines[]: each with line_number and text
• tail_start_line, tail_lines[]: last 20% of the log for context

results.json (in junit/ subdirectory)

Parsed JUnit test failures for a specific job. Only includes failed/error tests. For aggregated jobs, includes per-run pass/fail/skip data with Prow URLs for each run.

code.diff, comments.json, jobs.json

PR artifacts from GitHub (unchanged from previous version).

Chain Logic

The script chains backwards from the target payload until it finds a payload where all blocking jobs succeeded. This is stricter than the Accepted phase — a payload can be force-accepted with failed blocking jobs, which does not count as a stop point.

Sippy's release tag list, sorted by release_time, is used to identify every preceding assembled payload. If a tag is still retained, its payload details and changelog come from the release controller. If it has been garbage collected, the script constructs compatible payload.json and changelog.json files from Sippy's release tags, pull requests, and job runs APIs. A changelog that crosses a garbage-collected tag also comes from Sippy because the release controller can no longer compute that diff.

The generated source and changelog_source fields expose this provenance. Sippy-backed data is intentionally partial: RHCOS nodeImageStreams, async jobs, and previousAttemptURLs are unavailable.

For terminal payloads (Accepted/Rejected), jobs showing Pending on the release controller are cross-checked against the actual Prow prowjob.json artifact to get their real state.

Aggregated Jobs

Aggregated jobs run the same underlying test multiple times with statistical analysis. The script:

• Detects aggregated jobs by the aggregated- name prefix
• Downloads junit-aggregated.xml which contains per-run pass/fail/skip data
• Parses the YAML in <system-out> to extract individual run URLs

Error Handling

• Tag not found: Exits with code 2 and a descriptive error
• Release controller unreachable: Exits with code 1
• Historical tag missing from release controller: Automatically uses Sippy
• `gh` not authenticated: Prints a warning and continues without PR data
• `gcloud` not available: Warns, skips JUnit download, and records

gcloud_missing so the snapshot is reported incomplete

• `gcloud` not authenticated: Reads the public buckets anonymously
• Individual job/PR fetch failure: Logs a warning, records a collection

error, and continues

• Idempotent: Re-running skips files that already exist — except JUnit

output a previous run recorded as incomplete, which is discarded and re-collected so a partial snapshot cannot be inherited as complete

Notes

• The script uses only Python standard library — no pip dependencies
• PR data is deduplicated across payloads — each PR is fetched once
• JUnit and build-log download are scoped to failed blocking jobs only (informing jobs get job.json but no JUnit or build log)
• The --workers flag controls parallelism for all subprocess calls (default 8)
• Summary is always regenerated on re-run (not skipped like other files)
• Progress is printed to stderr; the script produces no stdout output

See Also

• Related Skill: fetch-payloads (fetches recent payloads from the release controller)
• Related Skill: fetch-new-prs-in-payload (fetches PRs new in a specific payload)
• Related Skill: payload-analysis (analyzes a payload snapshot for revert candidates)