Repository navigation
feat(ci): apply the dependency remediation without a model - #3845
Merged
Merged
Conversation
The remediation ran through Claude, and Claude was permitted to reproduce the output of a pure function. The planner already computed the exact result with `applyRemediation`, and the validation step then rejected any candidate whose bytes did not match that result. The model could therefore add nothing that the function did not already produce, and it could not fix anything the planner refused. That step cost the workflow an API key, 175 seconds of retries when the key failed, and a dependency on a credential that company policy now issues for 14 days at a time. A weekly job cannot hold a 14-day key. The planner now writes the manifest itself. - The plan step keeps its read-only behavior, so a dry run is unchanged. - A new step applies the plan in `apply` mode, then proves that the planner created no unexpected file and changed nothing except `package.json`. - The report records `deterministic-planner` as the resolver. The publish gate requires that value for an npm action, where it required `claude-code`. - The Go path is unchanged. The Go toolchain already owned its own change. Removed with the step: - `validate-claude-dependency-remediation.mjs` and its test. The script existed to constrain an untrusted author. There is no untrusted author now, so the byte-match compares a function against itself. - `ANTHROPIC_API_KEY` from the module and from both callers. The module no longer requests it. The safety model is stronger, not weaker. Only deterministic code writes the manifest, and every later gate is unchanged: the file allowlist, the lockfile resolution, the post-install proof that each remediated alert escapes its advisory range, the hash attestation, and a publishing job that runs no package code. `claude.yml` still uses the key for pull request review. This change does not affect that workflow. 77 tests pass.
The apply step refused any untracked file other than the two evidence reports. That guard existed because a model could create a file. Deterministic code writes one path, so the guard protects nothing and it fails for a person who runs the step in a working tree that holds other new files. The check that no tracked file outside package.json changed stays, because it still proves where the planner wrote.
The fixture named claude-code as the resolver. The verification script does not read that field, so the test passed either way, but the value no longer matches what the planner writes.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This was referenced Aug 20, 2026
laurakwhit
added a commit
that referenced
this pull request
Sep 9, 2026
Auto-generated version bump from 2.53.3 to 2.54.0 Bump type: minor Changes included: - [`7a5684a5`](7a5684a) Check if api has been upgraded for minimum minor bump (#3823) - [`af39e483`](af39e48) refactor(events): derive event groups on read in the grouped-event-buffer (#3790) - [`3ce42ebb`](3ce42eb) Add the developer Workflow Catalog (#3785) - [`a1234e36`](a1234e3) chore: split Catalog into a focused Temporal UI plugin (#3835) - [`060eacf7`](060eacf) Get total slots based on available plus used (#3833) - [`283d3cfc`](283d3cf) fix(timezone): use regular font weight for the timezone switcher (#3834) - [`88d61845`](88d6184) Manual on-call dependency review with constrained Claude remediation (#3816) - [`f87d239e`](f87d239) ci: validate the weekly review candidate the way the PR checks do (#3841) - [`dd80eca0`](dd80eca) feat(ci): apply the dependency remediation without a model (#3845) - [`403e5b97`](403e5b9) refactor(icons)!: replace holocene's icon registry with Io icon components (#3822) - [`d6823cc4`](d6823cc) fix(ci): correct the provenance claim in the remediation PR body (#3847) - [`f1b05e99`](f1b05e9) feat(ci): post the weekly review only when it needs a person (#3848) - [`b993c04b`](b993c04) ci: run the on-call review test only for an on-call change (#3850) - [`4307d20f`](4307d20) fix(copyable): pin the copy icon to 16px (#3855) - [`62e3014f`](62e3014) Update SAA heartbeat info (#3856) - [`3a797592`](3a79759) Headers spacing alignment and Small Standalone Activity UI fixes (#3854) - [`41687b1f`](41687b1) Add total Worker count (#3641) - [`68a3bc2f`](68a3bc2) fix(catalog): keep isolated test fixtures out of the repo root (#3859) - [`9a56d3d9`](9a56d3d) fix: don't render an empty selected-count badge in combobox (#3853) - [`4b97f35b`](4b97f35) Make only keyboard focus pin Tooltip open (#3861) - [`2fd63c0e`](2fd63c0) Add Rust SDK to worker heartbeats SDK warning (#3863) - [`4b43f7f9`](4b43f7f) Move Saved Views above Filtering (#3858) - [`cdfd92c9`](cdfd92c) Add option to maximize paginated tables (#3864) - [`ea4d3f14`](ea4d3f1) Fix ConfigurableTableHeadersDrawer for SAA nexus operations (#3867) - [`10f50c8b`](10f50c8) Fix duplicate schedule action keys (#3868) - [`9feeb756`](9feeb75) feat(DT-3944): SignalWithStart system Nexus obfuscation — decode infrastructure + UI (#3356) - [`5af59f03`](5af59f0) feat(workers): let users set the Cloud Run scale-down stabilization window (#3860) - [`bb7d6fbd`](bb7d6fb) Support secondary text on Combobox options (#3866) - [`e0925297`](e092529) Fix paginated table height and mobile scrolling (#3871) - [`6dbd9b96`](6dbd9b9) fix: categorize local activity markers as local-activity (#3872) - [`353fbbc8`](353fbbc) Wrap long filter chips instead of overflowing the filter bar (#3877) - [`f4c71240`](f4c7124) Fix workflow list timestamp formatting (#3880) - [`fa9c51e9`](fa9c51e) Open saved views in a new tab on modifier-click (#3881) - [`376af414`](376af41) feat(workers): warn about Cloud Run deployment latency when creating a version (#3879) - [`1bde3187`](1bde318) Fix extra horizontal scroll from filter bar dropdowns (#3885) - [`dd1dcd9b`](dd1dcd9) Add Next Retry Delay under Retry State (#3886) - [`66c57a40`](66c57a4) Make system saved views additive on custom saved views (#3876) - [`ffa1f2fb`](ffa1f2f) Audit fields for defaults and order (#3887) - [`9ab9c40b`](9ab9c40) fix(server): update labstack/echo/v4 to 4.15.4 (FE-493) (#3874) - [`e4449991`](e444999) Make Build ID filterable and show worker status counts (#3889) - [`0a5e01f5`](0a5e01f) SAA fields audit part 2 (#3890) - [`0d604b5a`](0d604b5) Fix payload input focus loss (FE-643) (#3891) - [`06712381`](0671238) Keep portal menus anchored when the layout shifts (#3893) Co-authored-by: laurakwhit <15069288+laurakwhit@users.noreply.github.com>
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description & motivation 💭
The remediation asked Claude to author
package.json, and Claude was permitted to reproduce the output of a pure function. The planner already computed the result withapplyRemediation, and the next step rejected any candidate whose bytes did not match that result.So the model could add nothing the function did not already produce, and it could not fix anything the planner refused. The planner now writes the manifest itself.
flowchart TD subgraph N["Now"] A2["Plan the remediation"] --> B2["The planner writes package.json"] B2 --> C2["Prove that nothing<br/>outside package.json changed"] end subgraph B["Before"] A1["Plan the remediation"] --> B1["Claude edits package.json.<br/>Needs an API key."] B1 --> C1["Byte-match gate.<br/>Reject any bytes that differ<br/>from the planned result"] endWhat the step cost
Run 32312572416 failed at the Claude step. The cause was not the prompt and not the tools:
Ten retries, 169 seconds, no work done. The same 401 appears in
temporalio/cloud-ui, which last summarized a pull request at 2026-08-18T22:07 and has failed since. A company-wide change deactivated the standing API keys on 2026-08-18.New keys are short-lived by default, at 14 days. A weekly workflow cannot hold a 14-day key, so this step would need a director-approved exception for a credential that adds no capability.
The change
applymode, then proves the planner changed nothing outsidepackage.json.deterministic-planneras the resolver. The publish gate requires that value for an npm action, where it requiredclaude-code.go getandgo mod tidyalready owned that change.Removed:
validate-claude-dependency-remediation.mjsand its test. The script constrained an untrusted author. With no untrusted author it compares a function against itself.ANTHROPIC_API_KEYfrom the module and from both callers.The safety model gets stronger
Only deterministic code writes the manifest. Every later gate is unchanged: the file allowlist, the lockfile resolution, the proof after the install that each remediated alert escapes its advisory range, the hash attestation, and a publishing job that runs no package code.
claude.ymlstill uses the key to review pull requests. This change does not affect that workflow.Testing 🧪
How was this tested 👻
77 tests pass. Prettier reports no problem. All three workflow files parse.
Two tests cover the new resolver: one for an applied plan and one for a plan that holds no action.
The whole apply path was run by hand against the live alert data, in the order the workflow runs it:
^7.6.5.package.jsonand recordedresolver: deterministic-planner.package.jsonchanged.pnpm install --lockfile-only --ignore-scriptsresolved protobufjs to 7.6.5.Steps for others to test: 🚶🏽♂️🚶🏽♀️
Open Actions, select Weekly On-Call Review PR Test, select Run workflow on this branch, and set the mode to
apply. The rehearsal setspublishto false, so no pull request is touched. The run needs no Anthropic credential.Checklists
Merge Checklist
claude.yml, which is a separate workflow and still needs one.Issue(s) closed
None.
Docs
Any docs updates needed?
.github/WORKFLOWS.mdis updated in this change. The safety model section no longer describes a model, andANTHROPIC_API_KEYis removed from the required configuration.