Repository navigation
ci(release): publish Edge release metadata for the dashboard - #28
Conversation
The dashboard will soon offer each new chart release to customers as soon as it is published, with no review of its own. Helm ignores values keys a chart does not read, so if the chart renamed or dropped one the dashboard sets, the install would still render and quietly lose the token, cluster name, netobs switch or a source. Each case sets the dashboard's keys and asserts the value reaches the rendered config, not only that the render succeeds. Refs: NOFireAI#27 Refs: NOFireAI/engineering#1983 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
Write edge/<version>.json and edge/latest.json after each release, so the dashboard can read the current chart version instead of pinning it. - The versioned file is written once, with If-None-Match. - latest.json only moves to a higher version and never to a pre-release. The current one is read from S3 directly, so a cached copy cannot hide a newer version. - The steps run on every run and each write is idempotent: a re-run changes nothing, and a manual run backfills a release made before this change. - latest.json waits until GitHub Pages serves an index.yaml that lists the version. Otherwise `helm repo update` could miss it. - Nothing runs until the RELEASE_METADATA_ROLE_ARN repository variable holds the role from NOFireAI/terraform#295. Refs: NOFireAI#27 Refs: NOFireAI/engineering#1983 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
…chart The dashboard's step 5 and disconnect commands run `helm upgrade --reuse-values`, and once they follow the published chart version they render each new chart's templates against an older release's values. The new suite only covers fresh values. This case is 0.9.0's values.yaml plus what the dashboard set on that release (token, cluster name, netobs, edge-proxy and one source), and checks those still reach config.json. Refs: NOFireAI#27 Refs: NOFireAI/engineering#1983 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
|
@claude-fleet review |
|
Review task queued. Task: 0de6f291-5046-4ef2-9aff-75c65a48548e |
There was a problem hiding this comment.
Review complete. Working tree restored (nothing left modified).
Request changes: two should_fix findings in the publish step. The test additions are solid — I verified both the unit suite and the render case actually fail on a silent key rename, so the "gate" claim holds.
Findings:
- should_fix, .github/workflows/release.yml:201 — a 403 on
latest.jsonis treated as "no latest.json yet", so an access failure silently downgrades it - should_fix, .github/workflows/release.yml:209 — read-modify-write of
latest.jsonwith noconcurrency:group and no conditional put - nit, .github/workflows/release.yml:129 —
yqis used but never installed in this workflow, while ci.yaml installs it explicitly
Verified:
- The versioned object is written once:
--if-none-match '*'withPreconditionFailedtolerated, anything else fatal — release.yml:178-188 latest.jsonnever moves backwards and never to a pre-release — release.yml:193-207; I ran the comparison for0.9.9vs0.10.0, equal versions, empty current: all behave as described- The wait step blocks on
index.yamllistingname+version, 40 × 15 s = 10 min — release.yml:127-136;yq -e "…select(.version == …)"exits 1 on a missing version, 0 on a present one (ran against a stub index) - All three steps are gated on
vars.RELEASE_METADATA_ROLE_ARN != '', and none is gated onsteps.check.outputs.exists, so a re-run and a manual run both execute as the PR says — release.yml:120,139,146 id-token: writeadded for the OIDC role — release.yml:17- Metadata generation:
image.tagis set in values.yaml, soedgeisnofireai/edge:v3.10.0andedgeProxyfalls back to the same tag (edgeProxy.image.tagis commented out, values.yaml:402) — ran theyq/jqpipeline against this checkout, output is well-formed - The reuse case values are 0.9.0's values.yaml verbatim apart from four dashboard-set keys, each marked
# dashboard:—git show 19313bb:values.yamldiffs to 22 lines, all of them the header comment plus those four keys - Notes claim about
image: v3.9.0under reuse: 0.9.0's values.yaml:18 pinstag: "v3.9.0", carried into the case values at line 18
Checks run:
helm unittest .→ 5 suites, 36 tests, all pass (helm v3.18.0, helm-unittest v1.2.1, same as CI)tests/render/run.sh→ 67 passed, 0 failed- Renamed
.Values.config.publisher.apiKeyin templates/configmap.yaml →dashboard install commandfails 1 of 6 (install puts the token…); restored - Renamed
.Values.onPremConnectionsin templates/edge-proxy-configmap.yaml → the same suite fails 2 of 6; restored - Appended
{{ fail "boom" }}to templates/edge-proxy-configmap.yaml →dashboard-reuse-from-0.9.0fails, so the case does cover render errors in templates it does not output; restored
Smaller things:
- The reuse case sets
edgeProxy.enabled: trueand anonPremConnections.prometheusblock, buttests/render/run.shrenders with-s templates/configmap.yamlandexpect.jqonly seesconfig.json. With theonPremConnectionsrename above in place, the case still passes. Those values only prove the proxy templates render; the unit suite is what asserts their content. Worth a line in the case's header comment so the next person does not readexpect.jqas covering step 5. metadata.json,latest.jsonanderr.txtare created in the chart root. Harmless today becausehelm package .runs earlier in the job, and.helmignoredoes not list them — if the steps ever move above packaging, they ship inside the chart.- If
latest.jsonever lacks.chart.version,jq -ryields the stringnull,sort -Vranks it above any number, andlatest.jsonthen freezes forever. One[ "$current" = null ] && current=""closes it.
Not checked:
app/dashboard/activation/installCommand.tsin NOFireAI/dashboard — I cannot reach that repo, so "sets each value the dashboard sets" is unverified; the suite asserts what it sets, not that the set matches the dashboard's.- The IAM policy in NOFireAI/terraform#295 (whether the role has
s3:GetObject, which finding 1's fix needs) and the bucket/Cloudflare public-read configuration. - Whether
yqand anawsCLI new enough for--if-none-matchare present on the currentubuntu-latestimage.
Not checked:
- app/dashboard/activation/installCommand.ts in NOFireAI/dashboard — cannot reach the repo, so the claim that the suite covers every value the dashboard's install/sources/disconnect commands set is unverified
- The IAM policy in NOFireAI/terraform#295, including whether the role has s3:GetObject for the fix suggested in finding 1
- The releases.nofire.ai bucket policy, Block Public Access settings and Cloudflare configuration — whether an anonymous GET of edge/latest.json returns 200 today
- Whether the current ubuntu-latest runner image provides yq and an aws CLI new enough to accept --if-none-match
Task 0de6f291-5046-4ef2-9aff-75c65a48548e on rp1, model claude-opus-5, effort medium.
Review on NOFireAI#28 found two ways latest.json could go wrong without a failed step: - The anonymous read took any 403 as "no latest.json yet". S3 also answers 403 when the file exists but cannot be read, so a backfill could replace a newer version with an older one. The role now reads it from S3 and can list the bucket, so only NoSuchKey means absent and any other error fails the step. - Two releases close together both read the same latest.json and the last writer won, even when it was the older version. The write now carries the ETag it read (or If-None-Match when there was none); the loser of a race reads again and compares again, up to five times. Also install yq pinned as ci.yaml does instead of relying on the runner image, keep scratch files in RUNNER_TEMP so they can never be packaged, treat a latest.json without chart.version as empty, and name the render case alongside the unit suite as the gate. Needs s3:GetObject and s3:ListBucket on the role (NOFireAI/terraform#301). Refs: NOFireAI#27 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
expect.jq checks config.json only; the edge-proxy values in the case prove those templates render, and the unit suite asserts their content. Refs: NOFireAI#27 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
|
@claude-fleet review |
|
Review task queued. Task: 174b18be-b775-4a31-bac6-219a2319e108 |
There was a problem hiding this comment.
Previous round at 9ca4bd2: 3 findings.
| Previous finding | Classification | Cited line |
|---|---|---|
.github/workflows/release.yml:201 (should_fix): HTTP 403 from the anonymous read is treated as "latest.json does not exist", but 403 is also what S3 returns when the object exists and the caller may not read it (bucket policy change, Block Public Access turned on, object written without public read). The comment at release.yml:192 only justifies 403 for a missing key. Failure: edge/latest.json holds 0.12.0 and public read is off. The next release, a 0.10.1 backfill, gets 403, sets current="", skips the "only moves forward" check at release.yml:204 and overwrites latest.json with 0.10.1. The dashboard then offers customers an older chart than the one already shipped, and no step failed. Evidence: release.yml:197-203 — the 403 and 404 arms are collapsed into current=""; release.yml:204 is skipped whenever current is empty. Fix: the role's credentials are already configured at this point (release.yml:139-144), so read with aws s3api get-object --bucket "$BUCKET" --key edge/latest.json latest.json and treat only NoSuchKey as absent, anything else as fatal. That also reads from S3 directly, which is what the comment asks for. |
fixed | .github/workflows/release.yml:217 |
.github/workflows/release.yml:209 (should_fix): The latest.json update is a read-modify-write (read at release.yml:197, write at release.yml:209) with no concurrency: group on the workflow and no precondition on the put, unlike the versioned object which uses --if-none-match. Failure: two Chart.yaml bumps land on main close together, e.g. 0.11.0 then 0.11.1. Both jobs read latest.json as 0.10.0 (the wait loop at release.yml:127-136 holds each for up to 10 minutes, so the overlap window is large), both decide to write, and whichever finishes last wins. If that is the 0.11.0 job, latest.json ends on 0.11.0 while 0.11.1 is released, and nothing retries — the next release is the only thing that repairs it. Evidence: no concurrency key anywhere in .github/workflows/release.yml (grepped); release.yml:209-211 puts unconditionally. Fix: add concurrency: {group: release-metadata, cancel-in-progress: false} to the job, or pass the ETag from the read as --if-match and retry the read-compare-write on PreconditionFailed. |
fixed | .github/workflows/release.yml:231 |
.github/workflows/release.yml:129 (nit): yq is used in the wait step and three times in the publish step, but this workflow never installs it — it only installs Helm (release.yml:24-25). ci.yaml:17-20 installs yq explicitly, pinned to v4.44.3, so the repo does not otherwise assume it is on the runner. Failure: if the ubuntu-latest image stops shipping yq, or ships a version whose -e/-r flags differ, the wait step fails after the chart tarball, the GitHub release and the gh-pages index have already been published. The release is out and latest.json is stale, which is exactly the split this PR exists to avoid. Evidence: release.yml:129,154,155,157,165,166 use yq; no install step in the file; ci.yaml:17-20 for the contrasting treatment. Fix: copy ci.yaml's install step into this job, or read the three fields with the grep/awk style already used by the Extract chart metadata step (release.yml:28-32). I did not check whether the current runner image includes yq. |
fixed | .github/workflows/release.yml:119 |
Previous round's three findings are all fixed in 2f9163c. I re-ran the two new test suites locally (helm 3.18.0, helm-unittest v1.2.1, yq v4.44.3) and proved both catch a values-key rename.
Approve: no blocking or should_fix finding.
Findings:
- nit, .github/workflows/release.yml:193 — versioned put accepts only
PreconditionFailed, notConditionalRequestConflict
Verified:
- Previous #1 fixed: the read is
aws s3api get-objectwith the role, and onlyNoSuchKeysetscurrent=""; anything else prints the error and exits 1 — release.yml:213-223. - Previous #2 fixed: the put carries
--if-match <etag read>or--if-none-match '*', andPreconditionFailed|ConditionalRequestConflictre-reads and re-compares up to five times — release.yml:212-244. Walked 0.11.0 against 0.11.1 in both orders through the loop: the loser re-reads, loses the version comparison at release.yml:224-228, and exits 0 without writing. - Previous #3 fixed: "Install yq" at release.yml:119-123, same pin (v4.44.3) and same commands as ci.yaml:17-20, gated on the same variable as the steps that use yq.
- Scratch files are under
$RUNNER_TEMP, so the chart package cannot pick them up — release.yml:162-163;tests/is ignored by .helmignore anyway. - "the 0.9.0 chart's values.yaml, verbatim except for what the dashboard set":
diffofgit show 19313bb:values.yamlagainst the case file shows only the header comment and the five# dashboard:lines (clusterName, netobs.enabled, publisher.apiKey, edgeProxy.enabled, onPremConnections.prometheus). - "the edgeProxy and onPremConnections values prove the edge-proxy templates still render (a render error fails this case)" — tests/render/cases/dashboard-reuse-from-0.9.0/values.yaml:8-10. run.sh renders with
-s templates/configmap.yaml, which still executes the whole chart: appending{{- fail "boom" }}to templates/edge-proxy-configmap.yaml failed the case. - "ci.yaml runs both" — ci.yaml:24-26 runs
helm unittest .(picks up tests/dashboard_install_test.yaml) andtests/render/run.sh. - The metadata's image strings match what the chart renders: script defaulting at release.yml:165-168 (
image.tagelse appVersion;edgeProxy.image.tagelseimage.tag) mirrors _helpers.tpl:68 and _helpers.tpl:122. - The suite's regex patterns match the templates' literal key order, including
"type"before"enabled"in a connection — templates/edge-proxy-configmap.yaml:37-39; the chart defaultconfig.netobs.enabledis false (values.yaml:291), so that assertion really depends on theset.
Checks run:
helm unittest -f tests/dashboard_install_test.yaml .→ 6 passed.tests/render/run.sh dashboard→ 1 passed.- Rename check (the analogue of the revert check, since this PR adds tests rather than a code fix): pointed templates/configmap.yaml:92 at
config.netobs.onand line 77 atconfig.publisher.apiKeyRenamed→ the install test fails anddashboard-reuse-from-0.9.0fails onexpect.jq. Restored;git status --porcelainshows only the untrackednode_modules.
Smaller things:
- release.yml:176 —
yq -r '.edgeProxy.image.repository'printsnullif that key is ever removed from values.yaml, so the metadata would carrynull:v3.10.0rather than failing; _helpers.tpl:121 defaults it toedge-proxyinstead. Only reachable if someone deletes the key. - release.yml:175-176 — these
yqcalls are command substitutions in an argument, not assignments, soset -edoes not stop a failure there; the earlier reads at 164-167 do stop. Not reachable while 164-167 succeed. - tests/dashboard_install_test.yaml:81-89 asserts only that the edge-proxy Deployment exists, unlike the header's "each case asserts the value comes out the other side". It is the switch itself, so there is nothing further to assert; the header sentence is slightly wider than the suite.
Not checked:
- That the suite covers every value the dashboard sets: NOFireAI/dashboard's
installCommand.tsis not reachable from here. - The role's
s3:GetObject/s3:ListBucket(NOFireAI/terraform#301), so I could not confirm a missing key surfaces asNoSuchKeyrather thanAccessDeniedin the real bucket. Without that grant the step fails loudly and leaves latest.json alone, which is the safe direction. - Whether the aws-cli on ubuntu-latest supports
--if-matchonput-object; if it did not, the error text would not match the retry grep at release.yml:236 and the step would fail loudly. - helm lint, the other unittest suites and the rest of the render cases: unchanged by this PR and run by CI.
Not checked:
- NOFireAI/dashboard installCommand.ts — cannot confirm the suite covers every value the dashboard sets
- NOFireAI/terraform#301 role grants — cannot confirm a missing key surfaces as NoSuchKey rather than AccessDenied
- aws-cli version on ubuntu-latest supports --if-match on put-object
- helm lint and the unchanged test suites/render cases (CI runs them)
Task 174b18be-b775-4a31-bac6-219a2319e108 on ravel-clickbench-ubuntu-2, model claude-opus-5, effort medium.
S3 answers a concurrent conditional write with 409 ConditionalRequestConflict rather than 412. The latest.json write already accepted both; the versioned write failed the run on 409 and never reached latest.json. Refs: NOFireAI#27 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
|
@claude-fleet review |
|
Review task queued. Task: cceeeb30-2eb1-43e0-9f9f-2757f78da438 |
There was a problem hiding this comment.
Previous round at a3d886c: 4 findings.
| Previous finding | Classification | Cited line |
|---|---|---|
.github/workflows/release.yml:217 (should_fix): HTTP 403 from the anonymous read is treated as "latest.json does not exist", but 403 is also what S3 returns when the object exists and the caller may not read it (bucket policy change, Block Public Access turned on, object written without public read). The comment at release.yml:192 only justifies 403 for a missing key. Failure: edge/latest.json holds 0.12.0 and public read is off. The next release, a 0.10.1 backfill, gets 403, sets current="", skips the "only moves forward" check at release.yml:204 and overwrites latest.json with 0.10.1. The dashboard then offers customers an older chart than the one already shipped, and no step failed. Evidence: release.yml:197-203 — the 403 and 404 arms are collapsed into current=""; release.yml:204 is skipped whenever current is empty. Fix: the role's credentials are already configured at this point (release.yml:139-144), so read with aws s3api get-object --bucket "$BUCKET" --key edge/latest.json latest.json and treat only NoSuchKey as absent, anything else as fatal. That also reads from S3 directly, which is what the comment asks for. |
fixed | .github/workflows/release.yml:217 |
.github/workflows/release.yml:231 (should_fix): The latest.json update is a read-modify-write (read at release.yml:197, write at release.yml:209) with no concurrency: group on the workflow and no precondition on the put, unlike the versioned object which uses --if-none-match. Failure: two Chart.yaml bumps land on main close together, e.g. 0.11.0 then 0.11.1. Both jobs read latest.json as 0.10.0 (the wait loop at release.yml:127-136 holds each for up to 10 minutes, so the overlap window is large), both decide to write, and whichever finishes last wins. If that is the 0.11.0 job, latest.json ends on 0.11.0 while 0.11.1 is released, and nothing retries — the next release is the only thing that repairs it. Evidence: no concurrency key anywhere in .github/workflows/release.yml (grepped); release.yml:209-211 puts unconditionally. Fix: add concurrency: {group: release-metadata, cancel-in-progress: false} to the job, or pass the ETag from the read as --if-match and retry the read-compare-write on PreconditionFailed. |
fixed | .github/workflows/release.yml:216 |
.github/workflows/release.yml:119 (nit): yq is used in the wait step and three times in the publish step, but this workflow never installs it — it only installs Helm (release.yml:24-25). ci.yaml:17-20 installs yq explicitly, pinned to v4.44.3, so the repo does not otherwise assume it is on the runner. Failure: if the ubuntu-latest image stops shipping yq, or ships a version whose -e/-r flags differ, the wait step fails after the chart tarball, the GitHub release and the gh-pages index have already been published. The release is out and latest.json is stale, which is exactly the split this PR exists to avoid. Evidence: release.yml:129,154,155,157,165,166 use yq; no install step in the file; ci.yaml:17-20 for the contrasting treatment. Fix: copy ci.yaml's install step into this job, or read the three fields with the grep/awk style already used by the Extract chart metadata step (release.yml:28-32). I did not check whether the current runner image includes yq. |
fixed | .github/workflows/release.yml:122 |
.github/workflows/release.yml:193 (nit): The versioned write treats only PreconditionFailed as "already exists", while the latest.json write 40 lines down also accepts ConditionalRequestConflict for the same S3 conditional-write mechanism. Failure: two runs publish edge/0.11.0.json at the same time (a push plus a manual dispatch, or a re-run overlapping the original). S3 answers the loser with 409 ConditionalRequestConflict rather than 412, the grep -q PreconditionFailed at release.yml:193 misses, the else arm exits 1, and that run never reaches the latest.json block — so the release ships with latest.json still on the previous version until someone re-runs the workflow. Evidence: release.yml:193 greps PreconditionFailed alone; release.yml:236 greps PreconditionFailed|ConditionalRequestConflict for the same condition on latest.json. Fix: use the same grep -Eq 'PreconditionFailed|ConditionalRequestConflict' at line 193. |
fixed | .github/workflows/release.yml:193 |
No longer carried to the next round, settled in rounds 2 and 3:
- .github/workflows/release.yml:217 (should_fix): HTTP 403 from the anonymous read is treated as "latest.json does not exist", but 403 is also what S3 returns when the object exists and the caller may not read it (bucket policy change, Block Public Access turned on, object written without public read). The comment at release.yml:192 only justifies 403 for a missing key. Failure:
edge/latest.jsonholds 0.12.0 and public read is off. The next release, a 0.10.1 backfill, gets 403, setscurrent="", skips the "only moves forward" check at release.yml:204 and overwrites latest.json with 0.10.1. The dashboard then offers customers an older chart than the one already shipped, and no step failed. Evidence: release.yml:197-203 — the 403 and 404 arms are collapsed intocurrent=""; release.yml:204 is skipped whenevercurrentis empty. Fix: the role's credentials are already configured at this point (release.yml:139-144), so read withaws s3api get-object --bucket "$BUCKET" --key edge/latest.json latest.jsonand treat onlyNoSuchKeyas absent, anything else as fatal. That also reads from S3 directly, which is what the comment asks for. - .github/workflows/release.yml:231 (should_fix): The latest.json update is a read-modify-write (read at release.yml:197, write at release.yml:209) with no
concurrency:group on the workflow and no precondition on the put, unlike the versioned object which uses--if-none-match. Failure: two Chart.yaml bumps land on main close together, e.g. 0.11.0 then 0.11.1. Both jobs read latest.json as 0.10.0 (the wait loop at release.yml:127-136 holds each for up to 10 minutes, so the overlap window is large), both decide to write, and whichever finishes last wins. If that is the 0.11.0 job, latest.json ends on 0.11.0 while 0.11.1 is released, and nothing retries — the next release is the only thing that repairs it. Evidence: noconcurrencykey anywhere in .github/workflows/release.yml (grepped); release.yml:209-211 puts unconditionally. Fix: addconcurrency: {group: release-metadata, cancel-in-progress: false}to the job, or pass the ETag from the read as--if-matchand retry the read-compare-write on PreconditionFailed. - .github/workflows/release.yml:119 (nit):
yqis used in the wait step and three times in the publish step, but this workflow never installs it — it only installs Helm (release.yml:24-25). ci.yaml:17-20 installs yq explicitly, pinned to v4.44.3, so the repo does not otherwise assume it is on the runner. Failure: if the ubuntu-latest image stops shipping yq, or ships a version whose-e/-rflags differ, the wait step fails after the chart tarball, the GitHub release and the gh-pages index have already been published. The release is out and latest.json is stale, which is exactly the split this PR exists to avoid. Evidence: release.yml:129,154,155,157,165,166 use yq; no install step in the file; ci.yaml:17-20 for the contrasting treatment. Fix: copy ci.yaml's install step into this job, or read the three fields with thegrep/awkstyle already used by the Extract chart metadata step (release.yml:28-32). I did not check whether the current runner image includes yq.
Approve: all four previous findings are fixed; one nit on the new code.
The ETag read-compare-write loop (release.yml:212-242) is sound in each ordering I traced: two runs racing from NoSuchKey both send --if-none-match '*', the loser gets 412/409, re-reads, and the "only moves forward" check at release.yml:224-228 stops it; a same-version re-run compares equal and exits without writing. The image strings the metadata publishes match what the chart actually renders (_helpers.tpl:68 and :122 resolve tags in the same order as release.yml:165-168).
Findings:
- nit, .github/workflows/release.yml:176 — repository reads have no
// ""guard, so a missing key publishesnull:<tag>into an immutable object
Verified:
- The read now uses the role and treats only
NoSuchKeyas absent, any other error fatal — release.yml:213-223 - The versioned and latest writes now grep the same two conditional-write errors — release.yml:193, release.yml:236
yqis installed pinned to v4.44.3, identical to ci.yaml:17-20 — release.yml:119-123- Scratch files are under
$RUNNER_TEMP, outside the checkout, sohelm package .cannot pick them up — release.yml:162-163 - The edge/edgeProxy image strings mirror the chart's own resolution order (
edgeProxy.image.tag→image.tag→ appVersion) — release.yml:165-168 vs templates/_helpers.tpl:122 dashboard-reuse-from-0.9.0/values.yamlis 0.9.0's values verbatim plus five annotated dashboard overrides —diffagainst19313bb:values.yamlshows only the header comment and the five# dashboard:lines- The case header's claim that edge-proxy templates still render:
tests/render/run.sh:52uses-s templates/configmap.yaml, which filters output but still executes every template, so a proxy template error fails the case helm unittest .picks up the new suite via the defaulttests/*_test.yamlpattern — ci.yaml:23- The new suite's assertions match the templates they target:
apiKeySecret.enableddefaults false at values.yaml:316, so the"apiKey": "dash-test-token"assertion exercises the plaintext branch at templates/configmap.yaml:77
Smaller things:
workflow_dispatchhas no branch restriction, so a dispatch from a feature branch with a higherversion:would now move customer-facinglatest.json, not just cut a GitHub release. Pre-existing trigger, new blast radius.- The sources test (tests/dashboard_install_test.yaml:54) sets
config.publisher.apiKeybut asserts nothing about it on the edge-proxy config; templates/edge-proxy-configmap.yaml:12 renders it. The install test covers the key name, so this is only an unasserted input. - The yq install step is byte-identical to ci.yaml:17-20. Two copies to keep pinned together; fine for workflows, worth a note if a third appears.
Not checked:
- helm, helm-unittest and yq are not in this workspace, so neither
helm unittest .nortests/render/run.shwas run; CI runs both. - The aws CLI version on
ubuntu-latest—put-object --if-matchneeds a build with S3 conditional overwrite support. It fails loudly (unknown option goes to thecat/exit 1arm) rather than silently if it is too old. - NOFireAI/terraform#301, so I could not confirm the role has
s3:ListBucketand therefore getsNoSuchKeyrather thanAccessDeniedon a missing key. app/dashboard/activation/installCommand.tsin NOFireAI/dashboard is not reachable here, so I could not confirm the suite covers every value the dashboard sets (e.g. whether it also setsconfig.publisher.heartbeat.urlalongsidegraph.url).
Not checked:
- helm unittest . and tests/render/run.sh were not run: helm, helm-unittest and yq are absent from this workspace
- The aws CLI version on ubuntu-latest, i.e. whether put-object supports --if-match/--if-none-match there
- NOFireAI/terraform#301: that the role has s3:ListBucket, which is what makes a missing key return NoSuchKey instead of AccessDenied
- NOFireAI/dashboard app/dashboard/activation/installCommand.ts: whether tests/dashboard_install_test.yaml covers every value the dashboard sets
Task cceeeb30-2eb1-43e0-9f9f-2757f78da438 on ravel-clickbench-ubuntu-2, model claude-opus-5, effort medium.
pmoust
left a comment
There was a problem hiding this comment.
We download arbitrary binaries from github in a repo we don't control in this change, and we have a write token in the same release pipeline.
That's risky. At minimum audit what we do, and pin a sha256 with it.
However, I also think we are overcomplicating this change, we have a bunch for shellscripts nastily escaped in the pipeline, and a test I don't really get how it helps in the future.
Ask claude to have a pass with my comment.
Simplify, don't load unnecessary external dependencies (esp random binaries0, write code if the shellscript is over 7-8 lines, and execute the code itself as a script.
Etc.
edge/<version>.json is written once and never replaced. Without a guard, a values.yaml missing an image repository would publish "null:<tag>" into it for good. Read appVersion and both repositories with `// ""` and fail the step before any write if one is empty. Refs: NOFireAI#27 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
|
@claude-fleet review |
|
Review task queued. Task: ed99313e-1a08-41b5-a8b2-8d19b1a963f1 |
There was a problem hiding this comment.
Previous round at e81e47b: 2 findings.
| Previous finding | Classification | Cited line |
|---|---|---|
.github/workflows/release.yml:193 (nit): The versioned write treats only PreconditionFailed as "already exists", while the latest.json write 40 lines down also accepts ConditionalRequestConflict for the same S3 conditional-write mechanism. Failure: two runs publish edge/0.11.0.json at the same time (a push plus a manual dispatch, or a re-run overlapping the original). S3 answers the loser with 409 ConditionalRequestConflict rather than 412, the grep -q PreconditionFailed at release.yml:193 misses, the else arm exits 1, and that run never reaches the latest.json block — so the release ships with latest.json still on the previous version until someone re-runs the workflow. Evidence: release.yml:193 greps PreconditionFailed alone; release.yml:236 greps PreconditionFailed|ConditionalRequestConflict for the same condition on latest.json. Fix: use the same grep -Eq 'PreconditionFailed|ConditionalRequestConflict' at line 193. |
fixed | .github/workflows/release.yml:203 |
.github/workflows/release.yml:176 (nit): The two repository reads have no // "" fallback, unlike the tag reads two lines up, so a missing key publishes the literal string null as part of an image reference. Failure: a later values.yaml drops edgeProxy.image.repository (templates/_helpers.tpl:121 defaults it to edge-proxy, so the chart still renders correctly). yq -r '.edgeProxy.image.repository' values.yaml prints null, and edge/<version>.json is published with "edgeProxy": "null:v3.10.0". No step fails, and the versioned object is written with --if-none-match '*', so it can never be corrected. Evidence: release.yml:175-176 read .image.repository and .edgeProxy.image.repository bare; release.yml:165,167 use // "" for the tags; templates/_helpers.tpl:121-122 is the chart's own resolution. Fix: read them into variables with // "" and fail the step when either is empty, matching the chart's defaults. |
fixed | .github/workflows/release.yml:173 |
No longer carried to the next round, settled in rounds 3 and 4:
- .github/workflows/release.yml:193 (nit): The versioned write treats only
PreconditionFailedas "already exists", while the latest.json write 40 lines down also acceptsConditionalRequestConflictfor the same S3 conditional-write mechanism. Failure: two runs publishedge/0.11.0.jsonat the same time (a push plus a manual dispatch, or a re-run overlapping the original). S3 answers the loser with 409 ConditionalRequestConflict rather than 412, thegrep -q PreconditionFailedat release.yml:193 misses, the else arm exits 1, and that run never reaches the latest.json block — so the release ships with latest.json still on the previous version until someone re-runs the workflow. Evidence: release.yml:193 grepsPreconditionFailedalone; release.yml:236 grepsPreconditionFailed\|ConditionalRequestConflictfor the same condition on latest.json. Fix: use the samegrep -Eq 'PreconditionFailed\|ConditionalRequestConflict'at line 193.
No findings.
Approve: both previous findings are fixed, and I found nothing new.
Findings:
- none
Verified:
- Finding 1 fixed: the versioned write now greps
PreconditionFailed|ConditionalRequestConflict, matching the latest.json write — release.yml:203 vs release.yml:246 - Finding 2 fixed: both repositories and
appVersionread with// "", and the step exits 1 before any PUT when one is empty — release.yml:164,169-178.proxy_tagis absent from that loop but cannot be empty: it falls back totag, which the loop already requires (release.yml:167-168) - The script's image resolution matches the chart's:
edgeProxy.image.tagdefaults toimage.tagthenappVersion, same order as templates/_helpers.tpl:121-123; values.yaml:401 leavesedgeProxy.image.tagunset, so metadata publishesnofireai/edge-proxy:v3.10.0 - "latest.json only moves forward, never to a pre-release": pre-releases exit at release.yml:211-214; I ran the
sort -Vcomparison at release.yml:234-235 for 0.9.9/0.10.0 both ways, equal versions, and 0.10.0/0.11.0 — forward-only in every case - Only
NoSuchKeycounts as absent; any other read error exits 1 before touching latest.json — release.yml:227-233 - The reuse case's values.yaml is 0.9.0's verbatim plus six dashboard-set lines, each marked
# dashboard:—diffagainstgit show 19313bb:values.yamlshows only the header comment and those six - The case header's claim that the edge-proxy values "prove the edge-proxy templates still render":
helm template -srenders the whole chart and only filters the output, so a template error there does fail the case helm unittest .picks the new suite up: ci.yaml:22 runs it overtests/*_test.yaml
Checks run:
tests/render/run.sh dashboard-reuse→1 passed, 0 failedhelm templatefor each of the six cases in tests/dashboard_install_test.yaml → every asserted pattern is present in the rendered output (apiKey/clusterName/netobs, graph url, name label, both connections withindex: "logs-*", edge-proxy Deployment count 1, disconnected connection still listed with"enabled": false)actionlint .github/workflows/release.yml→ clean;bash -non the extracted publish script → ok
Smaller things:
- The yq install (release.yml:122) pulls a release asset by tag with no checksum, and the binary then runs in a job that holds AWS write credentials for releases.nofire.ai. It copies ci.yaml:17-20, so this is a repo-wide choice, not a regression in this PR; pinning a sha256 in both places would be better.
- The index wait (release.yml:133-140) fetches index.yaml through the Pages CDN with no cache-buster. If an edge cache holds a pre-release copy for near its full TTL, the 10-minute loop could expire and fail the job after the release is already out. I did not measure the Pages cache TTL, so this is a suggestion (
?t=$(date +%s)), not a finding.
Not checked:
- The stub-run results in the Testing section (no
aws/stub here), and whether the step's script is byte-identical to the one tested. - The IAM grants in NOFireAI/terraform#295 and #301, and the dashboard side (NOFireAI/dashboard#1445) — not reachable from this repo. If #301 has not been applied when
RELEASE_METADATA_ROLE_ARNis set, the latest.json read fails the step; that fails closed. helm unittestitself was not run locally (plugin not installed); I verified the suite's assertions against realhelm templateoutput instead.
Not checked:
- The aws/gh stub run described in the PR Testing section, and whether the step script is byte-identical to the one tested
- IAM policy contents of NOFireAI/terraform#295 and #301
- The dashboard side (NOFireAI/dashboard#1445)
- helm-unittest was not run locally; assertions were verified against helm template output instead
- GitHub Pages CDN cache TTL, for the index-wait timeout suggestion
Task ed99313e-1a08-41b5-a8b2-8d19b1a963f1 on pmoust-mbp, model claude-opus-5, effort medium.
…oads Review on NOFireAI#28: the job downloaded an unpinned yq binary next to a write token, and the publishing logic was some 70 lines of escaped shell in the workflow. - The release job is back to what it is on main. A new publish-metadata job, after it, holds the only id-token: write and has contents: read, so the AWS role's token never shares a job with the release's write token. configure-aws-credentials is pinned to a commit. - The logic moves into .github/scripts/publish_release_metadata.py, standard library only, calling tools the runner image already ships (aws, gh, helm). It reads the chart name, version and appVersion from Chart.yaml, and waits for the version with `helm show chart --version`, the lookup a customer's `helm repo update` makes. - The metadata drops the image names: the dashboard reads only chart.version, and from 0.6.0 the images follow appVersion. Behaviour is otherwise unchanged: edge/<version>.json written once, latest.json read with the role and written conditionally on its ETag, only forward, never to a pre-release. Refs: NOFireAI#27 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
It copied 0.9.0's values.yaml to show that a --reuse-values upgrade still renders, but its expect.jq only re-checked config.json, which tests/dashboard_install_test.yaml already asserts. It would also need a new copy for every chart version worth testing from. Refs: NOFireAI#27 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
|
@pmoust thanks, reworked in a4cfaa9 and 2f68def. External code and tokens, per job
Simplified
Tests
One part you may want cut
|
|
@claude-fleet review |
|
Review task queued. Task: 21e74506-5d0e-4031-9845-190a9356cd03 |
There was a problem hiding this comment.
Previous round at e7d903a: 1 finding.
| Previous finding | Classification | Cited line |
|---|---|---|
.github/workflows/release.yml:173 (nit): The two repository reads have no // "" fallback, unlike the tag reads two lines up, so a missing key publishes the literal string null as part of an image reference. Failure: a later values.yaml drops edgeProxy.image.repository (templates/_helpers.tpl:121 defaults it to edge-proxy, so the chart still renders correctly). yq -r '.edgeProxy.image.repository' values.yaml prints null, and edge/<version>.json is published with "edgeProxy": "null:v3.10.0". No step fails, and the versioned object is written with --if-none-match '*', so it can never be corrected. Evidence: release.yml:175-176 read .image.repository and .edgeProxy.image.repository bare; release.yml:165,167 use // "" for the tags; templates/_helpers.tpl:121-122 is the chart's own resolution. Fix: read them into variables with // "" and fail the step when either is empty, matching the chart's defaults. |
new-context-gone | none |
Approve: no blocking or should_fix finding. The shell block became a Python script since the last review, so I re-read the whole job and script rather than only the delta.
Findings:
- nit, .github/scripts/publish_release_metadata.py:127 — an unparseable
chart.versionin latest.json is treated as absent, so an older release overwrites it
Verified:
- the
releasejob is unchanged from main — the whole net diff to release.yml is the 31 added lines ofpublish-metadata(release.yml:107-136), andid-token: writeis gone from the release job (release.yml:14-15) - the AWS token never shares a job with the release's write token —
contents: readplusid-token: writeat release.yml:118-120 - only
NoSuchKeycounts as "no latest.json"; anything else fails the step — publish_release_metadata.py:122-125 - the write is conditional on the ETag that was read, with a re-read on a lost race — publish_release_metadata.py:118-137
- latest.json never moves to a pre-release:
parse_versionreturns None for anything but a plainx.y.z, and the script returns before touching it — publish_release_metadata.py:47-52, 107-110 - the versioned object is written once, and a conflicting write is not an error — publish_release_metadata.py:101-105
- the new suite is a real gate, not just a render check: renaming the values keys it sets makes it fail (see Checks run)
tests/dashboard_install_test.yamlis picked up by the defaulthelm unittest .glob, so CI runs it without a config change- every value the suite sets differs from the chart default (
netobs.enabled: falsevalues.yaml:291,edgeProxy.enabled: falsevalues.yaml:398,onPremConnections: {}values.yaml:520), so an assertion cannot pass on the default alone
Checks run:
helm unittest .→ 5 suites, 36 tests, all pass- renamed
.Values.config.publisher.apiKeyand.Values.config.netobs.enabledchart-wide (templates + values.yaml) → dashboard suite: 1 failed, 2Expected to matchassertions. Separately renamed.Values.onPremConnectionsin templates/edge-proxy-configmap.yaml → 2 failed. Both restored;git status --porcelainshows only the pre-existingnode_modules. python3 -m py_compile .github/scripts/publish_release_metadata.py→ ok (local interpreter is 3.11, not the 3.12 the description names)- loaded the script and called its helpers against the real Chart.yaml →
name=nofire-edge version=0.10.0 appVersion=v3.10.0;parse_versionreturns None for0.10.0-rc1,0.10,v0.10.0,0.10.0.1
Smaller things:
edge.images.{edge,edgeProxy}is no longer in the payload (publish_release_metadata.py:92-98) although earlier commits in this PR published it, whileschemastays1. The description's sample matches the code, so nothing here is stale — but NOFireAI/dashboard#1530 has to agree that the published shape has no image references.- Chart.yaml is parsed two different ways:
grep '^version:' | awk '{print $2}'in the release job (release.yml:31-32) andline.partition(":")with quote-stripping in the script (publish_release_metadata.py:38-44). A quoted version, or one with a trailing comment, makes them disagree, and the script then looks up agh release viewtag the release job never created — failing after the chart has already shipped. Reading the two values from the release job's outputs, or having the release job use the script's parse, removes the second copy. - This job installs no Helm, unlike ci.yaml:13 and release.yml:25-26 (both
azure/setup-helm), and relies on the runner image forhelm,aws,ghandpython3(release.yml:112-113). That is the same class as the earlieryqnote; I did not verify the current image ships Helm. wait_for_helm_repodiscards the result ofhelm repo add(publish_release_metadata.py:57). A failure there only shows up as the 10-minute timeout with a message about the chart, not the repo.- The "sources deploy edge-proxy with each connection" case sets
config.publisher.apiKey(tests/dashboard_install_test.yaml:54) but asserts nothing about it, even though templates/edge-proxy-configmap.yaml:12 renders it. The first case covers the same values key for templates/configmap.yaml, so a rename is still caught; the set here is dead weight unless you assert it.
Not checked:
- The 11 stub cases in the Testing section — no AWS stand-in or stubbed gh/helm is available here, so the racing-writers and AccessDenied behaviour is read from the code only.
- actionlint — not installed in this environment.
- NOFireAI/dashboard's app/dashboard/activation/installCommand.ts, so whether tests/dashboard_install_test.yaml covers every value the install, sources and disconnect commands actually set.
- Whether the ubuntu-latest runner image ships helm, aws, gh and python3 as release.yml:112-113 asserts.
- The role policy in NOFireAI/terraform#301 (s3:GetObject/PutObject on edge/*, s3:ListBucket).
- Whether the GitHub Pages index and the S3 bucket behave as assumed at release time (no network access to either from this review).
Task 21e74506-5d0e-4031-9845-190a9356cd03 on pimox5, model claude-opus-5, effort medium.
The chart repo's tooling is shell and Go (tests/render/edgecheck), and the backend is Go; nothing here uses Python. Same logic, standard library only, still calling the runner's aws, gh and helm, so nothing is downloaded at build time. setup-go is pinned to a commit and uses the Go the runner image caches. Refs: NOFireAI#27 Signed-off-by: Stelios Charmpalis <stelios.charmpalis@nofire.ai>
|
Follow-up: ported the script to Go in 0db32f1 ( |
|
@claude-fleet review |
|
Review task queued. Task: 8d39d594-d5ba-454e-be4a-1466a18af03b |
There was a problem hiding this comment.
Previous round at 2f68def: 2 findings.
| Previous finding | Classification | Cited line |
|---|---|---|
.github/workflows/release.yml:173 (nit): The two repository reads have no // "" fallback, unlike the tag reads two lines up, so a missing key publishes the literal string null as part of an image reference. Failure: a later values.yaml drops edgeProxy.image.repository (templates/_helpers.tpl:121 defaults it to edge-proxy, so the chart still renders correctly). yq -r '.edgeProxy.image.repository' values.yaml prints null, and edge/<version>.json is published with "edgeProxy": "null:v3.10.0". No step fails, and the versioned object is written with --if-none-match '*', so it can never be corrected. Evidence: release.yml:175-176 read .image.repository and .edgeProxy.image.repository bare; release.yml:165,167 use // "" for the tags; templates/_helpers.tpl:121-122 is the chart's own resolution. Fix: read them into variables with // "" and fail the step when either is empty, matching the chart's defaults. |
new-context-gone | none |
.github/scripts/publish_release_metadata.py:127 (nit): A chart.version in latest.json that parse_version cannot read is treated the same as no latest.json at all, so the forward-only check is skipped and the write proceeds. Failure: edge/latest.json holds 0.12.0-rc1 (or v0.12.0, or a four-part version) because it was backfilled or written by hand. A later 0.10.1 release reads it, parse_version returns None at publish_release_metadata.py:127, the held is not None guard at line 128 is false, and latest.json is overwritten with 0.10.1 — an older chart offered to customers, with no step failing. The shell version this replaced compared with sort -V, which kept the newer string. Evidence: publish_release_metadata.py:47-52 (parse_version returns None unless the string is exactly three numeric parts; confirmed by running it on 0.10.0-rc1, 0.10, v0.10.0, 0.10.0.1 — all None); line 128 skips the comparison whenever held is None. Fix: treat a non-None current that does not parse as fatal, the same way a non-NoSuchKey read error is at line 125 — the script never writes such a value, so finding one means something it does not understand owns the file. |
open | .github/scripts/publish-release-metadata/main.go:194 |
No longer carried to the next round, settled in rounds 5 and 6:
- .github/workflows/release.yml:173 (nit): The two
repositoryreads have no// ""fallback, unlike thetagreads two lines up, so a missing key publishes the literal stringnullas part of an image reference. Failure: a later values.yaml dropsedgeProxy.image.repository(templates/_helpers.tpl:121 defaults it toedge-proxy, so the chart still renders correctly).yq -r '.edgeProxy.image.repository' values.yamlprintsnull, andedge/<version>.jsonis published with"edgeProxy": "null:v3.10.0". No step fails, and the versioned object is written with--if-none-match '*', so it can never be corrected. Evidence: release.yml:175-176 read.image.repositoryand.edgeProxy.image.repositorybare; release.yml:165,167 use// ""for the tags; templates/_helpers.tpl:121-122 is the chart's own resolution. Fix: read them into variables with// ""and fail the step when either is empty, matching the chart's defaults.
Approve: three nits, and one earlier nit carried over into the Go port.
The port is faithful on the parts that matter: the wait loop, the --if-none-match '*' versioned write, the ETag read-compare-write retry and the forward-only check all behave as the Python did. go run <file>.go does work with no go.mod in the tree (verified), and gofmt -l and go vet are clean on the file here.
Findings:
- nit, .github/scripts/publish-release-metadata/main.go:143 — JSON output is not byte-identical to the Python, contrary to the description
- nit, .github/scripts/publish-release-metadata/main.go:71 —
parseVersionnow rejects zero-padded parts the Python accepted - nit, .github/workflows/release.yml:142 — the job runs
helmwithout installing it, unlike thereleasejob above it - Previous finding 2 (unparseable
chart.versionskips the forward-only check) came across unchanged; it now sits at main.go:194.
Verified:
chartFieldmatches the Python'spartition-based reader for this Chart.yaml:name,version,appVersionare top-level and unindented, and the indentedmaintainers[].namecannot match — Chart.yaml:1-17, main.go:60-65.gh release view --json publishedAt,urlunmarshals intostruct{ PublishedAt, URL string }(Go field matching is case-insensitive) — main.go:134.- Only
NoSuchKeycounts as absent; every other read error is fatal — main.go:188-191. - The conditional write retries on both
PreconditionFailedandConditionalRequestConflict, for both objects — main.go:121-124, 156, 206. - The test regexes match the template's actual field order:
"type"then"enabled"in the connections loop, andurl/indexin the non-postgres arm — templates/edge-proxy-configmap.yaml:38-59;netobs.enabledandapiKey— templates/configmap.yaml:77, 91-92. - No reference to the deleted
publish_release_metadata.pyremains (grepped the workflows and docs).
Checks run:
gofmt -landgo veton main.go → cleango run <file>.gofrom a directory tree with nogo.mod→ runs, fails only on the missing Chart.yaml, so the invocation form at release.yml:142 is sound- Ran the ported
parseVersionand the metadata marshalling side by side with the Python originals → results in the findings below
Smaller things:
- The comment at release.yml:113 says "setup-go uses the Go it already caches", but
go-version: '1.26'is a pin: it is a cache hit only while the runner image's cached Go is a 1.26.x. When the image moves on, the step downloads a toolchain — which is fine, but then the comment is the stale part. Dropping the pin togo-version-fileis not possible without ago.mod, so either loosen the comment or accept the download. actions/checkout@v4at release.yml:123 is a tag while the two actions below it are pinned to commits.
Not checked:
helm unittest .— no helm on this machine, so the claim that tests/dashboard_install_test.yaml fails when a values key is renamed is unverified here.- The dashboard's command builders (
app/dashboard/activation/installCommand.ts, NOFireAI/dashboard#1530) are not reachable, so "sets each value the dashboard sets" is taken on the description's word. - Whether the ubuntu-latest image currently ships helm, aws and a 1.26.x Go.
- The stub-harness runs listed under Testing; I did not rebuild them.
Not checked:
- helm unittest . — no helm binary available here, so the claim that tests/dashboard_install_test.yaml fails on a renamed values key is unverified
- The dashboard's install/sources/disconnect command builders (NOFireAI/dashboard, app/dashboard/activation/installCommand.ts) — not reachable, so the set of values the suite must cover is taken from the description
- Whether the ubuntu-latest runner image currently ships helm, the aws CLI, and a cached Go 1.26.x
- The 11 stub-harness cases listed under Testing; not rebuilt or rerun
Task 8d39d594-d5ba-454e-be4a-1466a18af03b on rp1, model claude-opus-5, effort medium.
Summary
Publish Edge release metadata to
releases.nofire.aion every chart release, and gate it on a test of the values the NOFire AI dashboard sets (NOFireAI/engineering#1983).release.yml: thereleasejob is unchanged frommain. A newpublish-metadatajob runs after it withcontents: readandid-token: write, so the AWS role's token never shares a job with the release's write token. It is skipped until theRELEASE_METADATA_ROLE_ARNrepository variable is set..github/scripts/publish-release-metadata/main.go: Go standard library only, run withgo runafteractions/setup-go(pinned to a commit, using the Go the runner caches). It calls theaws,ghandhelmthe runner image ships, so nothing is downloaded at build time. It:appVersionfromChart.yaml;helm show chart nofire/nofire-edge --version <v>resolves, the same lookup a customer'shelm repo updatemakes;edge/<version>.jsononce (If-None-Match);edge/latest.jsonforward only, never to a pre-release. It reads the file with the role, and onlyNoSuchKeycounts as absent. The write is conditional on the ETag it read, so two racing releases cannot leave the older version.tests/dashboard_install_test.yaml: sets each value the dashboard's install, sources and disconnect commands set, and asserts it reaches the rendered config. Helm ignores keys a chart does not read, so without it a rename would render cleanly and lose the token, cluster name, netobs switch or a source. Once the dashboard followslatest.json(NOFireAI/dashboard#1530), this is the only check before a chart reaches customers.latest.json:{ "chart": { "name": "nofire-edge", "repo": "https://nofireai.github.io/edge-helm-chart", "version": "0.10.0" }, "edge": { "version": "v3.10.0" }, "releaseUrl": "https://github.com/NOFireAI/edge-helm-chart/releases/tag/nofire-edge-0.10.0", "releasedAt": "...", "schema": 1 }Testing
aws s3api(objects with ETags, conditional puts) and stubbedghandhelm, in 11 cases:latest.json;latest.jsonwithoutchart.version;AccessDeniedon the read, which fails and leaveslatest.jsonalone;go vetandgofmtare clean, and every case behaves as the earlier Python version did. The JSON holds the same fields and values, but Go writes the keys in alphabetical order, so the files are equivalent rather than byte-identical.actionlintis clean.main. The firstworkflow_dispatchafter merge is the end-to-end check.helm unittest .. A claude-fleet review proved the suite fails when a values key the dashboard sets is renamed.Notes
workflow_dispatch) to publish 0.10.0. A re-run changes nothing.s3:GetObjectands3:PutObjectonedge/*ands3:ListBucketon the bucket.Closes #27
Refs: NOFireAI/engineering#1983, NOFireAI/terraform#301, NOFireAI/dashboard#1530
🤖 Generated with Claude Code