chore(deps): bump claude-agent-sdk to 0.2.121, codex-sdk to 0.125.0#1460
Merged
chore(deps): bump claude-agent-sdk to 0.2.121, codex-sdk to 0.125.0#1460
Conversation
Both SDKs were ~30 patch releases behind. Validation suite passes
(type-check, lint, format, tests across all 10 packages) without code
changes. The only sustained Claude SDK behavior change in the range —
v0.2.111's options.env overlay/replace flap, since reverted to overlay —
is a no-op for Archon, which already passes { ...process.env } as the
SDK env.
|
Caution Review failedPull request was closed or merged during review No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: ⛔ Files ignored due to path filters (1)
📒 Files selected for processing (2)
📝 WalkthroughWalkthroughDependency versions are updated across two package.json files. The root package.json upgrades Changes
Estimated code review effort🎯 1 (Trivial) | ⏱️ ~2 minutes Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
This was referenced Apr 28, 2026
Wirasm
added a commit
that referenced
this pull request
Apr 28, 2026
* refactor(workflows): trust the SDK for model validation Drops cross-provider model inference and hard-coded model allow-lists. The string a workflow author writes in `model:` is forwarded to the SDK unchanged; the SDK and its API decide whether the model exists. Provider identity is the only thing Archon validates at load time — typos like `provider: claud` are caught early; everything else fails at runtime through the SDK's normal error path. Why this matters: a recent run on Sasha showed `provider: claude` + `model: opus[1m]` getting silently routed to Codex (because Codex's isModelCompatible was defined as the complement of Claude's, so anything not literally `sonnet|opus|haiku` matched). Codex then rejected the model as a `⚠️ ` system warning and the node "completed" in 2.1 seconds with empty output, after which the workflow opened a hallucinated PR. Three stacked bugs and two amplifiers; this commit removes all five. Changes: - Delete model-validation.ts entirely (inferProviderFromModel and isModelCompatible are gone). Drop the matching field from ProviderRegistration and from the claude/codex/pi entries. - Replace the resolver in executor.ts and dag-executor.ts (both the per-node and per-loop paths) with a flat `node.provider ?? workflow.provider ?? config.assistant`. Model never influences provider selection; load-time validation is just isRegisteredProvider on the resolved provider id. - Remove the dag-node Zod superRefine that recomputed model-compat — load-time provider validation moved to loader.ts. - Codex provider: stream loop now matches Claude's contract. error events that aren't followed by turn.completed yield `result.isError: true` (subtype `codex_stream_incomplete`) so the dag-executor's existing isError path catches them. turn.failed becomes `codex_turn_failed` with the same shape. Iterator close without a terminal event is itself a fail-stop. MCP-client errors remain filtered (Codex retries those internally). - dag-executor: AI nodes that exit the streaming loop with empty assistant text and no structured output now fail with `dag.node_empty_output` instead of completing silently — the Sasha bug's final amplifier. Bash/script/approval nodes are unaffected. Tests: model-validation.test.ts and isPiModelCompatible block deleted; codex provider tests rewritten to assert the new fail-stop contract; dag-executor empty-output test flipped to assert failure; new tests cover (a) loader rejecting unknown provider, (b) loader accepting any model string with a known provider, (c) executor passing provider+model through without re-routing, (d) executor throwing on unknown provider, (e) Codex synthesizing fail-stop on iterator close. Two cost-tracking tests adjusted to yield non-empty assistant text since their intent was cost accumulation, not empty-output handling. bun run validate: green (check:bundled, type-check, lint --max-warnings 0, format:check, all packages' test suites — 0 fail). End-to-end smoke (.archon/workflows/test-workflows/): - e2e-deterministic: PASS (engine healthy) - e2e-codex-smoke: PASS (Codex sendQuery + structured output work) - e2e-claude-smoke: FAIL with `error: unknown option '--no-env-file'` — this is a regression from the SDK 0.2.121 bump (#1460), not from this redesign. The Claude provider source is unchanged on this branch. To be fixed separately. * fix(workflows): address review on #1463 Critical: - C1: empty-output guard now skips idle-timeout completions. The on-screen message says "completed via idle timeout"; flipping that to a failure contradicted the user-facing log. Added !nodeIdleTimedOut to the guard. - C2: per-node provider identity is now validated at YAML load time. Loader iterates dagNodes after parsing and rejects any unknown provider id with "Node 'X': unknown provider 'Y'. Registered: ...". The dag-executor's runtime check stays as defense-in-depth. Important: - I1: CHANGELOG entry under [Unreleased] > Changed describing the resolver redesign + an explicit migration line for workflows that relied on cross-provider model inference. - I2: restored the dropped mockLogger.error('turn_failed') assertion in the turn.failed-without-error-message test. - I3: empty-output test now also asserts store.failWorkflowRun was called, matching the parallel error_max_budget_usd test pattern. - I4: new test that proves a node yielding zero assistant text but a valid structuredOutput is treated as a successful completion (not caught by the empty-output guard). - I5: rewrote the post-loop comment in codex/provider.ts to be precise about which dag-executor branch catches the synthesized result chunk (the throwing msg.isError branch, distinct from the empty-output guard's { state: 'failed' } return). - I6: removed PR-era "redesign" / "Sasha workflow" references from three test-file comments. - I7: docs sweep for the deleted isModelCompatible field — six files updated (CLAUDE.md, two docs guides, quick-reference, contributing guide, architecture reference). Polish: - S3: dropped the dead sawTerminal flag in streamCodexEvents — both terminal branches `return`, so reaching the post-loop block always means no terminal fired. Pure simplification. - S4: dropped parsePiModelRef and PiModelRef from community/pi/index.ts exports. The parser is consumed only by Pi's provider.ts; making it package-internal narrows the public surface. - S6: new Codex test for the bare-stream-close case (zero events, iterator just ends) — locks in the default fallback message used when no captured non-MCP error is available. - S7: new dag-executor test for per-node unknown-provider at runtime. Bypasses the loader to exercise resolveNodeProviderAndModel's throw, asserts the node_failed event carries the "unknown provider 'claud'" detail (the workflow-level fail message is a generic summary). bun run validate green across all 10 packages. * fix(workflows): address CodeRabbit review on #1463 Two real issues from CodeRabbit's automated pass on db95e8a: 1. Empty-output fail-stop now applies to loop iterations too. The single-shot AI-node guard at executeNodeInternal only covered prompt/command nodes; executeLoopNode has its own streaming path, so a provider that closed cleanly with zero content could pause an interactive loop with a blank gate or burn the full max_iterations budget. Mirrors the contract of the single-shot guard: `fullOutput.trim() === '' && !iterationIdleTimedOut` fails the iteration with a `loop_iteration_failed` event carrying a clear error. Idle-timeout exits remain exempt for the same reason as single-shot nodes — the on-screen "completed via idle timeout" message would otherwise contradict the failure. 2. Unknown loop providers now throw instead of return-failed. The early-return path bypassed the layer dispatch's outer catch at line 2870, so loop nodes with an invalid per-node `provider:` field skipped the standard `node_failed` event, the user-facing message, and the pre-execution log entry. Throwing reuses the common failure path — same shape as resolveNodeProviderAndModel uses for non-loop nodes. Both align with CLAUDE.md's "fail fast, explicit errors, never silently swallow" principle. The third CodeRabbit finding (boundary violation for `@archon/providers` import in loader.ts) is consistent with existing precedent — `dag-executor.ts`, `executor.ts`, and `validator.ts` already import from the same path; the runtime contract (every entrypoint bootstraps the registry before parseWorkflow runs) is already enforced in tests and documented at `loader.test.ts:31`. bun run validate green across all 10 packages.
4 tasks
Merged
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
1 participant
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.
Summary
@anthropic-ai/claude-agent-sdk^0.2.89vs0.2.121;@openai/codex-sdk^0.116.0vs0.125.0). Behind-ness was surfaced while diagnosing a workflow-routing bug where Codex rejected a Claude model alias with a non-fatal warning.claude-opus-4-7[1m], current GPT-5.x variants) are validated server-side by the SDKs. Staying on old pins doesn't break anything per se, but it widens the gap before the next bump and slows iteration on follow-up provider work.package.json+packages/providers/package.json, plus the regeneratedbun.lock.options.envoverlay/replace flap (introduced in 0.2.113, reverted in 0.2.111-line — confirmed against the SDK CHANGELOG) is a no-op for Archon becausebuildSubprocessEnv()already passes{ ...process.env }.UX Journey
No user-facing behavior change. Workflow execution, platform adapters, and CLI surfaces are identical before and after.
Architecture Diagram
No architectural change — pure dependency version bump. Same modules, same edges, same exports.
Connection inventory:
@archon/providers(claude)@anthropic-ai/claude-agent-sdk@archon/providers(codex)@openai/codex-sdk@anthropic-ai/claude-agent-sdkLabel Snapshot
risk: lowsize: XSdependenciesproviders:claude,providers:codexChange Metadata
choremulti(root +@archon/providers)Linked Issue
Validation Evidence (required)
bun run validate # EXIT=0All five gates pass on the bumped SDKs:
check:bundled✓type-check(10 packages) ✓lint --max-warnings 0✓format:check✓0 fail) ✓Security Impact (required)
Patch-level upstream releases only — no new transitive dependencies of note. Vendor SDKs continue to talk to the same Anthropic / OpenAI endpoints with the same auth flow.
Compatibility / Migration
Existing workflow YAMLs,
.archon/config.yamlfiles, and stored sessions continue to work unchanged.Human Verification (required)
bun installresolves cleanly to 0.2.121 / 0.125.0 in the lockfile; fullbun run validatereturnsEXIT=0; the SDK API surface used bypackages/providers/src/claude/provider.tsandpackages/providers/src/codex/provider.ts(query,Options, message events) type-checks against the new.d.ts.options.envoverlay/replace flap, which is a no-op for Archon's{ ...process.env }usage. Codex SDK release notes 0.116 → 0.125 show no documented JS API changes.Side Effects / Blast Radius (required)
dag.node_sdk_error_resultlog event catches result-level errors; provider-level Pino logs (stream_error,turn_failed) catch streaming issues.Rollback Plan (required)
git revert <merge-sha>and re-runbun install. No DB or filesystem state is touched.query()calls or as a TypeScript build break onbun run type-check. Both are caught immediately in CI.Risks and Mitigations
packages/providers/src/claude/provider.test.ts,packages/providers/src/codex/provider.test.ts). Rollback is a one-line revert if a real-world issue surfaces.Summary by CodeRabbit