Repository navigation
ci: run jobs that read secrets in environments - #2019
Conversation
Repo-level secrets are readable by any workflow pushed to any branch. Jobs that need them now name an environment, so that once the secrets live in environments with deployment-branch rules, only the branches that need them get them. - run_tests.yml takes an `environment` input for its test job; daily.yml and release.yml pass `integration`, ci.yml passes none. - integration.yml enters `integration` for the mem0 jobs on push only; a pull request gets no MEM0_API_KEY and its mem0 suites skip. - publish-sdk-kotlin.yml's publish and upload jobs use `release`. An environment not yet configured has no rules and sees repo secrets, so nothing changes until the secrets move. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kanat
left a comment
There was a problem hiding this comment.
Reviewed #2019 at 602275e merged onto the current accelerate (03994a6, clean). CI: 53 pass, 4 skipped (Kotlin publish jobs on a PR, CodeRabbit). actionlint on the five files: clean.
Checked and sound:
- At head, every
secrets.*read in.github/is in a job that names an environment:run_tests.yml:95-120,164(test,inputs.environment),integration.yml:162-163(mem0 entries),publish-sdk-kotlin.yml:131-133,172-175(publish,upload,release).ci.yml:52-55callsrun_tests.ymlwith nosecrets:, so itstestjob gets none. integration.yml:41:&&binds tighter than||, so the value isintegrationonly for a push of an entry withmem0: true, and''for any pull request and any entry withoutmem0. The PR run ofintegration (internal/memory/mem0)at this head showsMEM0_API_KEYandRUN_MEM0_SUITEempty, and the deployments API lists no deployment for this sha.- The environment on the job inside the reusable workflow applies to the called job, and its environment secrets take precedence over the inherited ones.
- Trigger refs against the planned rules: the schedule runs with ref
main, av*release withrefs/tags/v*, a Kotlin snapshot withrelease-kotlin-X.Y.Z-SNAPSHOT, a Kotlin release withaccelerate(merge push). All match. Aworkflow_dispatchofdaily.ymlfrom another branch is refused, as the body says. - The secret lists in step 3 and 4 cover every repo-level secret name.
[Should fix] R1.1 The main edits are a precondition for step 3, but they sit outside the numbered steps and no PR into main exists. The schedule runs main's daily.yml (daily.yml:11-18@8e2d3a199), which calls main's run_tests.yml, and v* releases are tagged on main. These jobs read every provider key and name no environment, so after step 3 the nightly and the release integration run lose their keys. Fix: open the main PR now and make it a numbered step before step 3, with its link.
[Should fix] R1.2 The deployment rules only hold as far as the refs they name are protected. The admin list should also have, before step 3: (a) a ruleset on accelerate that requires a pull request and blocks force pushes (the repo's rulesets target the default branch only); (b) a ruleset that limits who may create release-kotlin-* branches, or required reviewers on release. Otherwise anyone who can push can create release-kotlin-x-SNAPSHOT with an edited publish-sdk-kotlin.yml (its push trigger includes that file) and enter release. (c) The v* tag ruleset, which the body mentions only as an aside in step 1, should be a numbered step.
[Should fix] R1.3 Step 4 misses an organization secret: one more is shared with this repo, and no workflow at head reads it. Add it to step 4: remove this repo from its access list.
[Should fix] R1.4 The PR body:
- In the diagram,
push to accelerate -> run_tests (integration) -> provider keysis wrong. A push toacceleratecallsrun_tests.ymlonly fromci.yml, with no environment and no secrets. The jobs that enterintegrationon a push are the mem0 entries ofintegration.yml. - Steps 1-2 say "Create". The first push of a mem0 job after merge creates
integrationwith no rules, so write "create or configure". - In step 3, remove "(with new values as they are rotated)".
[Should fix] R1.5 daily.yml:20-21 (inline).
[Nit] R1.6 Step 2 lets main into release, but no workflow on main reads the Kotlin secrets. Remove main from the rule.
[Nit] R1.7 Nothing checks the new environment expressions: dropping github.event_name == 'push' && at integration.yml:41, or turning && into ||, still passes CI. Only the runs listed in the body show that the expressions behave. Ticket.
Not verified: the positive path, a push that runs a mem0 job inside integration. The first push to accelerate after merge will show it as a deployment to integration.
Verdict: request changes, for R1.1-R1.5 (PR body and one comment; the workflow logic is sound).
| # The provider keys and the Slack webhook live in this environment. | ||
| environment: integration | ||
| # A schedule only fires on the default branch, so it tests accelerate, where the | ||
| # work happens; Run workflow on any other branch tests that branch. |
There was a problem hiding this comment.
[Should fix] R1.5 "Run workflow on any other branch tests that branch" stops being true once step 1 sets rules on integration: a dispatch from a branch outside main/accelerate is refused before it reaches a runner. The PR body says so itself. Change the comment to say that only main and accelerate can run it.
kanat
left a comment
There was a problem hiding this comment.
Delta review of 318ce01 (was 602275e). The code delta is one comment in daily.yml; the rest is the PR body. CI at this head: 53 pass, 4 skipped.
Prior findings
- R1.1 fixed. #2020 is open on
main, it is step 1 with its link, and it holds the edits the body lists (theenvironmentinput andenvironment: ${{ inputs.environment }}inrun_tests.yml,environment: integrationindaily.ymlandrelease.yml). Its CI is green; it is still a draft, so mark it ready before step 1. - R1.2 fixed. Steps 2-4 are the
accelerateruleset, required reviewers onreleaseand thev*tag ruleset. All three come before the environments (5-6) and before any secret moves (8). - R1.3 fixed. Step 7 covers the unused organization secret and keeps the two
STREAM_REPO_*ones on the list until step 8.gh api .../actions/organization-secretsshows three names for this repo: the two read bypublish-sdk-kotlin.yml:170-171and one that no workflow reads. - R1.4 fixed. The diagram no longer has the wrong
accelerate->run_testsline, step 5 and 6 say "create or configure", and the rotation wording is gone. - R1.5 fixed, with one inaccuracy: see N1.
- R1.6 fixed.
releaseisaccelerateandrelease-kotlin-*-SNAPSHOT. - R1.7 left as a follow-up, as agreed ("A ticket will cover a check").
Order of the admin steps is safe: #2020 (1), branch and tag protection (2-4), environments with their rules (5-6), org access (7), secrets (8). Step 7 keeps STREAM_REPO_* until step 8, so the Kotlin publish keeps working.
Secret lists. The repo has 32 secrets. Step 8's integration (27) plus release (5) lists equal them exactly (diffed); the extra TELNYX_API_KEY is the parenthetical, correctly marked as not set at repo level. Every secrets.* name read in .github/workflows is in one of the lists or in the org pair.
Public repo. The body has secret names only (no values, ids, or account details) and nothing about incidents. Safe.
[Nit] N1 daily.yml:21-22 (and the same sentence in #2020): "Only main and accelerate can run this workflow" is not quite what step 5 sets. integration also accepts tag v*, and Run workflow can pick a tag. Say "only main, accelerate and v* tags", or drop the "only".
[Nit] N2 Step 3/6: required reviewers on release also gate the Kotlin publish that runs on every merge push to accelerate that touches it, not only the snapshot branches. If that is the intent, one clause in the body saying so helps the admin.
Not verified: the positive path (a mem0 push job entering integration) and the environments themselves, which are admin steps after merge.
Verdict: GO. No blockers; N1 and N2 are optional.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kanat
left a comment
There was a problem hiding this comment.
Security review (attacker with push access, not admin) of e54f585 and its pair #2020 (96236d5), with every admin step in the body applied as written. CI on the head: 52 pass, 4 skipped, 1 pending (integration (internal/agent), which does not touch these files). actionlint: clean. The earlier reviews' findings are not repeated here.
Closed by the plan. An environment rule is matched against the run's GITHUB_REF (docs):
- A workflow pushed to a new branch,
workflow_dispatchfrom that branch, and a call torun_tests.ymlwithenvironment: integrationfrom it (local or@accelerate) all run onrefs/heads/<branch>and are refused. pull_requestruns onrefs/pull/N/mergeand is refused.pull_request_targetruns the base branch's file: onmain/acceleratethat is onlylabeler.yml, which checks out nothing and reads no secret. Into the attacker's own branch, it runs on that branch and is refused.workflow_run,repository_dispatch,issue_commentandschedulerun only the default branch's files.- A tag push outside
v*is refused. Av*tag, or a release that would create one, is stopped by step 4's create/update/delete rule. - Caches: a PR run writes only its merge-ref scope, so it cannot plant a cache that a push job restores (docs).
uv.lockhas hashes and only the pypi.org index, and installs use--frozen.- The diffs add no environment name an attacker can set:
inputs.environmentcomes only from caller files in the same ref, andintegration.yml:41reads onlyevent_nameand the static matrix. No step widens access compared with today.
What is still open is the merge gate. The nightly runs main's daily.yml, which runs run_tests.yml in integration on accelerate's code (daily.yml:23). So whoever can land code on accelerate gets every integration key the next night.
[Should fix] R1.1 Step 2 (accelerate ruleset) says "pull request required" but sets no approval count, and a ruleset allows 0. With 0, the attacker opens a PR and merges it. Fix: in step 2, require 1 approval. Also turn on "Dismiss stale pull request approvals when new commits are pushed" and "Require approval of the most recent reviewable push": without them, a pusher can push a commit onto someone else's approved PR and merge it. Also restrict deletions, and keep the bypass list to admins. Check that main's ruleset has the same options, because daily.yml and release.yml run from main.
[Should fix] R1.2 Add a step before step 8: under Settings → Actions → General → Workflow permissions, make sure "Allow GitHub Actions to create and approve pull requests" is off, at repo and org level. When it is on, a workflow pushed to any branch can ask for pull-requests: write (a permissions: key overrides the default, docs). It can then approve the attacker's own PR as github-actions[bot], which meets a 1-approval rule (changelog). In the same step, set the default token permissions to read.
[Should fix] R1.3 Steps 3 and 6: also turn on "Prevent self-review" on release, and name a small reviewer group. Without that setting, a reviewer who starts a run on release-kotlin-x-SNAPSHOT can approve that run.
[Should fix] R1.4 Step 9 is worded as a remark, but PyPI is an artifact path just like release. A workflow file named release.yml on any branch, with environment: pypi and id-token: write, matches the PyPI trusted publisher, and pypi has no deployment rule. Make step 9 an action: on pypi, allow only tag v*. Also check that the PyPI trusted publisher names environment pypi. If it names none, any job in a release.yml can publish.
[Nit] R1.5 Steps 5, 6 and 9: give the ref type for each rule. main, accelerate and release-kotlin-*-SNAPSHOT are Branch rules, and v* is a Tag rule. If v* is entered as a Branch rule, any pusher can create a branch v-x and enter integration.
[Nit] R1.6 Step 8, STREAM_REPO_*: organization secrets cannot be read back. Say that the values come from the owner of that credential, or from a new key.
Separate work (tickets, not in this PR):
- [Should fix] T1
release-sdk-js.yml:17-24(npm) andrelease-sdk-dotnet.yml:31-36(NuGet) publish through OIDC from jobs with no environment. Theirif: github.ref == 'refs/heads/accelerate'lives in the file that a branch can edit. npm and NuGet policies match repo + workflow filename + an optional environment, and no ref (npm, NuGet). Fix: give each job an environment limited toaccelerate, with reviewers, then set that environment on the registry policy. - [Nit] T2
main-side hardening thatacceleratealready has (see #2020).
Org pair (STREAM_REPO_*): steps 7-8 already move it into release and remove repo access. GITHUB_TOKEN in the env jobs here: contents: read, except Kotlin upload (contents: write, inside release). No id-token outside build-core, and no cloud federation.
Verdict: NO-GO, for R1.1-R1.4. All four are PR-body edits to the admin steps. The workflow changes are sound.
Their trusted publishers can then require that environment, so a job outside it cannot publish. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Replies to the security review (#2019 (review)). Its R1.1-R1.6, T1 and T2 are tracked as S1.1-S1.8.
Step order, checked again: steps 1-9 only narrow access, and every secret stays where it is today until step 11 moves it. |
kanat
left a comment
There was a problem hiding this comment.
Delta security review of b4f7279 (was e54f585). Verdict: NO-GO until the one Should fix below is in. Everything else is fixed.
Prior findings
- S1.1 (ruleset: 1 approval, dismiss stale, approval of the latest push, restrict deletions, admin-only bypass, same on
main): fixed, step 3. - S1.2 (Actions may not approve PRs; default token read; repo and org): fixed, step 2.
- S1.3 (Prevent self-review on
release, small named team): fixed, step 6. - S1.4 (
pypirule Tagv*, publisher check): fixed, now step 8. - S1.5 (ref type per rule): fixed, steps 6, 7, 8 name Branch or Tag.
- S1.6 (org secret values are write-only): fixed, last bullet of step 11.
- S1.7 (
environment: releaseon the npm and NuGet OIDC jobs): code fixed, theif: github.ref == 'refs/heads/accelerate'guards are kept, the registry-publisher step is step 9. Mutation: reverting b4f7279 puts both OIDC jobs back outside any environment (grep -n "environment:" release-sdk-js.yml release-sdk-dotnet.ymlfinds nothing). The choice of environment leaves a hole, see N1.
New finding
- N1, Should fix (
release-sdk-js.yml:23,release-sdk-dotnet.yml:36, body step 6/9).releaseadmits Branchrelease-kotlin-*-SNAPSHOT(it has to, for the Kotlin snapshot publish). The npm and NuGet publishers now trust "environmentrelease" and check no ref, so a pusher can createrelease-kotlin-x-SNAPSHOT, editrelease-sdk-js.ymlthere (drop theif:) and dispatch it. The only gate left is a reviewer who must notice the branch name. The security review (T1) asked for an environment limited toaccelerate. Fix: give these two jobs their own environment (for examplerelease-sdk), deployment branchaccelerateonly, same reviewers with Prevent self-review, and name that environment in steps 6 and 9. The Kotlin jobs keeprelease.
Attacker model re-run (one member with push, no admin, 11 steps applied)
- Own branch or PR naming
integration,pypi,release: refused by the ref rules (push and dispatch ref is the branch, PR ref isrefs/pull/N/merge).v*tags: closed by step 5. - Nightly runs
acceleratecode inintegration: gate is step 3 (1 approval from someone else, stale approvals dismissed, no bot approval after step 2). - Registry publish from another branch or file: refused by the registry once step 9 is set, except through N1.
- (a) Step order: no step widens access. 1 and 2 to 9 only add rules or narrow publishers; secrets move last, in 11. Step 4 holds no action of its own ("step 6"), so an admin can skip it by reading in order; fold it into step 6 (Nit).
- (b) Yes:
releasereviewers now hold every Kotlinpublishanduploadon a merge toaccelerate, and every npm and NuGet run. Step 4 and step 9 say so. Add to step 6 that the team needs at least two people, since the person who merged cannot approve the run their merge started (Nit). - (c)
main'srelease-sdk-js.ymlhas no environment. Its job is skipped offaccelerateby itsif. The npm trusted-publisher page lists "Environment name (optional)" and advises environments for approvals, but does not state the mismatch rule in words (NuGet's page does: the claim is validated if the policy sets it). An env-less token cannot match a publisher that names one, so the fixer's argument holds. Not checked on npm's settings page (not readable here): after step 9, a dry run from a throwaway branch should be refused. - (e) "Follow-ups" can be dropped, with this reason: once step 7 is applied, a pull-request job that names
integrationhasGITHUB_REF=refs/pull/N/merge, which no rule admits, so the job is refused and CI goes red. Losinggithub.event_name == 'push' &&atintegration.yml:41is then caught by the environment itself, not by a new check. Replace the section with that sentence.
Checks: git diff e54f585e4 b4f727923 touches only the two workflow files (3 lines each). CI: 53 pass, 4 skipped, none failing. #2020 must merge first (step 1).
| runs-on: ubuntu-latest | ||
| # The npm trusted publisher accepts only a token minted in this environment. A run enters | ||
| # it only from the branches its rules name, and only after a reviewer approves it. | ||
| environment: release |
There was a problem hiding this comment.
N1 (Should fix): environment release admits release-kotlin-*-SNAPSHOT branches, and npm checks no ref. Use a separate environment limited to accelerate for the npm and NuGet jobs (same for release-sdk-dotnet.yml:36).
There was a problem hiding this comment.
Fixed in 9a93133: the npm job now uses environment registry-publish, limited to accelerate, with the same required reviewers. release keeps the Kotlin jobs. The PR body steps 5, 6 and 9 say so.
| runs-on: ubuntu-latest | ||
| # The NuGet trusted publisher accepts only a token minted in this environment. A run enters | ||
| # it only from the branches its rules name, and only after a reviewer approves it. | ||
| environment: release |
There was a problem hiding this comment.
N1: same as release-sdk-js.yml:23.
There was a problem hiding this comment.
Fixed in 9a93133: same, the NuGet job uses registry-publish.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
kanat
left a comment
There was a problem hiding this comment.
Delta security review of 9a93133 (was b4f7279). Verdict: GO. One body-wording nit remains.
Prior findings
- S2.1 fixed.
release-sdk-js.ymlandrelease-sdk-dotnet.ymlnow nameregistry-publish;publish-sdk-kotlin.yml(publish,upload) is the only user ofrelease. The table, steps 5, 6 and 9 agree with the code. Planned rule: Branchaccelerateonly, same reviewers, Prevent self-review. The jobs also keepif: github.ref == 'refs/heads/accelerate', so arelease-kotlin-*-SNAPSHOTrun can no longer reach the npm or NuGet token. - S2.4 fixed: the reviewer step is folded into step 5, team of at least two.
- S2.5 fixed: the follow-up line is gone; the last paragraph says CI goes red if a pull-request job names the environment.
Attacker model, final state (org member with push access, not admin; all admin steps applied)
- Repo and org secrets: repo copies deleted (step 11), unread org secrets removed from the access list (step 10). A branch workflow gets only
GITHUB_TOKEN. integration: refused for every branch exceptmainandaccelerate(both PR plus approval, stale-approval dismissal, no force push) and for tags other thanv*(tag ruleset).pypi:v*tags only; tag creation, update and deletion are limited by step 4, so the attacker cannot create a tag or a GitHub release that startsrelease.yml.- npm and NuGet: both registries ignore the ref, but the OIDC token of a job without the environment lacks the environment claim, so the publishers from step 9 refuse it. A job naming
registry-publishfrom another branch is refused before a runner starts. A pull request run (refs/pull/N/merge) is refused by every rule. - Kotlin: the one reachable path is a pushed
release-kotlin-*-SNAPSHOTbranch with an edited workflow; it stops at the required reviewers, and the author cannot approve their own run. That leaves a reviewer approving without reading the branch. The body names this; nothing more to change. labeler.ymlusespull_request_targetbut only runsactions/labeler; it checks out no pull-request code.- Order: each step narrows or leaves access as today.
registry-publish(step 6) exists before the publishers require it (step 9); until step 9 the publishers accept what they accept today. Secrets move last.
Nit
- Body, "After merge": "Steps 1-9 come before any secret moves" should say 1-10, since step 10 also comes before step 11.
CI: all checks pass; the Kotlin jobs are skipped on a pull request, as designed.
Why
Repo-level secrets are readable by any workflow pushed to any branch. Environments with deployment-branch rules limit them to the branches that need them. This PR makes every job that reads secrets name an environment, so the secrets can move into environments in repo settings. The move is an admin step after merge (below).
The PR alone changes nothing for a push: an environment that is not configured yet has no rules and still sees the repo-level secrets. The one change is on pull requests:
integration.ymlno longer givesMEM0_API_KEYto a pull request's code, and the mem0 suites skip there, as they already did for forks.How it works
run_tests.ymltestinputs.environment(new, default'')daily.yml->run_tests.ymlintegrationrelease.ymlintegration->run_tests.ymlintegrationv*ci.ymlunit->run_tests.ymlintegration.ymlmem0matrix entriesintegrationon push,''on PRpublish-sdk-kotlin.ymlpublish,uploadreleaseif)release-sdk-js.ymlrelease(npm, OIDC)registry-publish(new)accelerate(existingif)release-sdk-dotnet.ymlrelease(NuGet, OIDC)registry-publish(new)accelerate(existingif)release.ymlbuild-corepypi(unchanged)v*The environment goes on the job inside the reusable workflow: the caller's job cannot set one, and
secrets: inheritstill has to pass the secret (GitHub docs, reuse workflows).Validation
environment/environment.namewith thegithub,inputs,matrix(and other) contexts, and say nothing about an empty value (workflow syntax). A throwaway branch (deleted since) ran seven jobs: literalenvironment: '',environment: {name: ''}, an expression that yields'', a matrix expression that yields''for one entry, and a reusable workflow'senvironment: ${{ inputs.environment }}with the default and with''. All seven succeeded, the deployments API showed no deployment for that commit, and no environment was created. A job inpypidoes create one (its last deployment has refv0.6.6), so the check can tell. actionlint refuses the literal''forms (string should not be empty), so this PR uses only expressions.run_tests.ymlfromci.ymlwith the empty default, andintegration.ymlonpull_requestwith the expression yielding''.release-sdk-js.ymlthat the base already has.client-sdks.ymlrunsflutter pub publish --dry-runonly, with noid-tokenand no secret.release.ymlbuild-core(PyPI) is the only other job withid-token: write.After merge (repo admin)
Do the steps in order. Steps 1-10 come before any secret moves, and no step widens access compared with today.
Merge the
mainPR, ci: run the jobs that read provider keys in an environment (main) #2020, once its CI is green. The nightly schedule andv*releases run frommain:daily.ymlresolvesrun_tests.ymlinmain's commit, and Python releases are tagged onmain.mainhas diverged for these files, so ci: run the jobs that read provider keys in an environment (main) #2020 holds the same three edits by hand: theenvironmentinput andenvironment: ${{ inputs.environment }}on thetestjob ofrun_tests.yml, andenvironment: integrationindaily.ymland inrelease.yml'sintegrationjob. Without it the nightly and the release integration run lose their keys after step 11.integration.ymlandpublish-sdk-kotlin.ymlare not onmain.Workflow token settings. Under Settings -> Actions -> General -> Workflow permissions, at repo level and at org level:
pull-requests: writeand approve a pull request asgithub-actions[bot], which meets the approval rule of step 3 (docs).permissions:at the top or on every job, so none of them changes.Add a branch ruleset on
accelerate, with the same options onmain's ruleset. The deployment-branch rules below only hold as far as the branches they name are protected, and the nightly runsaccelerate's code inintegration.maintoo.Add a tag ruleset on
v*: limit who may create, update and delete those tags.release.ymlruns on a release whoseGITHUB_REFisrefs/tags/vX.Y.Z, sointegrationandpypiacceptv*tags (steps 7 and 8), and this ruleset is what keeps the path closed.Create or configure environment
release, for the Kotlin jobs only:accelerate, Branchrelease-kotlin-*-SNAPSHOT;Anyone who can push could create
release-kotlin-x-SNAPSHOTwith an editedpublish-sdk-kotlin.yml(itspushtrigger includes that file) and enterrelease. The required reviewers close that path with one setting on the environment that guards the secrets, and hold every Kotlin publish on a merge toaccelerateuntil a reviewer approves the run.Create environment
registry-publish, for the npm and NuGet jobs: Branchaccelerateonly, the same required reviewers asreleasewith "Prevent self-review" on. A separate environment keeps these jobs out of therelease-kotlin-*-SNAPSHOTbranches thatreleaseadmits, because neither registry checks the ref.Create or configure environment
integration(the first push of a mem0 job after merge creates it with no rules): Branchmain, Branchaccelerate, Tagv*. Enterv*as a Tag rule: as a Branch rule it would admit any pushed branch namedv-x.Configure environment
pypi: Tagv*only. Then check on PyPI that the trusted publisher of each published package names environmentpypi. If it names none, a job in anyrelease.ymlon any branch can publish.Make the npm and NuGet trusted publishers require environment
registry-publish: on npm, the trusted publisher of@stream-io/stream-agentsforrelease-sdk-js.yml; on nuget.org, the trusted publishing policy forrelease-sdk-dotnet.yml. Neither registry checks the ref (npm, NuGet), so the environment is what keeps a job on another branch from publishing. A dry run now also waits for a reviewer.An org admin removes Vision-Agents from the access list of every organization secret that no workflow here reads. Check with
gh api repos/GetStream/Vision-Agents/actions/organization-secrets. The twoSTREAM_REPO_*secrets below are read; an environment does not limit an organization secret, so keep them on the access list only until step 11 is done, then remove this repo from their lists too.Move each secret from repo level into its environment, then delete the repo copy:
integration:ANTHROPIC_API_KEY,ASSEMBLYAI_API_KEY,AWS_ACCESS_KEY_ID,AWS_BEARER_TOKEN_BEDROCK,AWS_BEDROCK_API_KEY,AWS_SECRET_ACCESS_KEY,CARTESIA_API_KEY,DEEPGRAM_API_KEY,ELEVENLABS_API_KEY,FAL_KEY,FISH_API_KEY,GOOGLE_API_KEY,HF_TOKEN,INWORLD_API_KEY,MISTRAL_API_KEY,NVIDIA_API_KEY,OPENAI_API_KEY,OPENAI_MODEL,OPENROUTER_API_KEY,SARVAM_API_KEY,STREAM_API_KEY,STREAM_API_SECRET,TURBOPUFFER_API_KEY,TWELVELABS_API_KEY,XAI_API_KEY,SLACK_WEBHOOK_URL,MEM0_API_KEY. (TELNYX_API_KEYis read byrun_tests.ymlbut is not set at repo level.)release:SIGNING_KEY,SIGNING_KEY_ID,SIGNING_PASSWORD,STREAM_REPO_ENDPOINT,STREAM_REPO_BUCKET.STREAM_REPO_ACCESS_KEY_IDandSTREAM_REPO_SECRET_ACCESS_KEYare organization secrets: put them inrelease, then remove this repo from the organization secrets' access lists. Organization secrets cannot be read back, so get the values from the owner of that credential, or create a new key.Once the
integrationrules exist, a pull-request job that names the environment is refused (refrefs/pull/N/merge) and CI goes red, so a droppedgithub.event_name == 'push' &&guard is caught.After the move, Run workflow on
daily.ymlfrom a branch other thanmainoraccelerateis refused by theintegrationrules, instead of testing that branch with the keys.🤖 Generated with Claude Code