[dhaern]

Issue for this PR

Closes #60822

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

A step that ends with finish: "content-filter" sets providerFailed, so step.ts never builds llmError and the session.retry hook is skipped. The same refusal sent back as an HTTP ContentPolicy error does reach the hook, which is why a fallback plugin can recover one but not the other, and why a refused subagent job ends as error. In v1 both cases ended up on the same path (prompt.ts turned a content-filter finish into ContentFilterError and published Session.Event.Error, like any API error), so plugins could treat them the same way.

This isn't tied to one provider. content-filter is the normalized finish the protocols already produce for Anthropic refusal, OpenAI Chat and Responses content_filter, Gemini safety blocks, Bedrock guardrails and Mistral, so all of them go through the same branch. The change wraps that finish in a ContentPolicyError and sends it through the retry decision like other provider failures. The hook receives the same provider.content-filter error that step.failed carries today. ContentPolicy isn't retryable, so the proposal stays retry: false and nothing changes without a hook. If a hook switches the model and returns retry: true, the runner reuses the existing Retry outcome when no output was produced, or Continue when part of the answer was already streamed, the same way it handles a mid-stream rate limit. It's 10 lines in step.ts.

How did you verify your code works?

Added two tests in session-runner.test.ts, modeled on "retries an unknown finish before output" and "continues after a mid-stream rate limit": a hook turns a content-filter finish into a retry before output and into a continuation after partial output. Both fail without the change. The existing content-filter tests pass unchanged, bun test test/session-runner.test.ts is green and bun run typecheck in packages/core is clean.

I also ran it on v2.2.4 with the patch cherry-picked and a local OpenAI-compatible stub returning finish_reason: "content_filter", using oh-my-opencode-slim's retry hook. A root session and a background subagent both switched model and finished on the fallback, and the parent got state="completed" instead of state="error". With no fallback configured the step still fails as before.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

---

Added by the opencode bot (AI-generated).

Related: #60586

Public API: session.retry plugin hook now also fires for content-filter finishes (default retry: false; backwards compatible)

[opencodejr]

The fix looks correct to me and I found nothing blocking. I ran the two new tests against the base: both fail there with StepFailedError, and the session.retry hook is never called. With this PR, all 245 tests in session-runner.test.ts pass. Type-checking packages/core with tsc -b tsconfig.tests.json is clean. I couldn't get an end-to-end TUI check working, because the simulated model's first response went to the session title.

I also checked the paths around the change. Without a hook, behaviour is unchanged: the proposal stays retry: false and the content-filter failure is still recorded and reported once, as before. If a hook retries before any output, the step returns before step.failed is published. If a hook continues after partial output, the existing Continue path is used.

Public API: no signatures change. One behaviour change for plugins: the session.retry hook now also fires for content-filter finishes, with error type provider.content-filter and a default decision of retry: false. This is backwards compatible, but a hook that sets retry: true for every error will now also retry refusals.

Related: #60586

[opencodejr]

The new commit only re-wraps lines in step.ts; the behaviour is unchanged. I re-ran session-runner.test.ts (245 pass) and the packages/core type check (clean). The public API impact is the same as before: the session.retry hook now also fires for content-filter finishes, with a default decision of retry: false. This is backwards compatible. I have no new concerns.

[opencodejr]

The new commit only renames the commit message and doesn't change any code, so my previous review still holds and I have no new concerns. As before, the session.retry hook now also fires for content-filter finishes, with a default decision of retry: false. This is backwards compatible.

[GiuseppeBellamacina]

Question on step.ts: the content-filter failure is rebuilt as a ContentPolicyError carrying only the message, and the retry path is gated on !recorded.outputStarted. Does outputStarted cover tool calls that already ran before the content-filter finish? If not, a hook-driven retry could run those tools again.

[dhaern]

@GiuseppeBellamacina yep, that's covered. outputStarted flips to true on tool-input-start, tool-call and tool-input-error, plus whenever text or reasoning starts (publish-llm-event.ts:441-494), and a tool can only run after its tool-call event, so by the time anything has executed in that step the flag is already set and the retry branch at step.ts:215 just gets skipped. The hook still sees the content-filter error, but the step fails the same way it does today and nothing runs twice. Basically a hook-driven retry only ever replays a step that hadn't put anything out yet, same deal as the other provider errors.

[dhaern]

@rekram1-node you added the session.retry hook in #45999, could you take a look?

[mahammedmahieousman]

60836 Sent from Outlook for Android<https://aka.ms/AAb9ysg>

…

________________________________ From: Raxxoor ***@***.***> Sent: Monday, 02 November 2026 00:45:23 To: anomalyco/opencode ***@***.***> Cc: Subscribed ***@***.***> Subject: Re: [anomalyco/opencode] fix(core): let retry hooks handle content-filter finishes (PR #60836) [https://avatars.githubusercontent.com/u/7317522?s=20&v=4]dhaern left a comment (anomalyco/opencode#60836)<#60836 (comment)> @rekram1-node<https://github.com/rekram1-node> you added the session.retry hook in #45999<#45999>, could you take a look? — Reply to this email directly, view it on GitHub<#60836?email_source=notifications&email_token=BVRSBAGPXBHWKHB3MN3TF535W6VHHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTMNJWGQZDSNJYHE3KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#issuecomment-6564295896>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BVRSBAA5THAX4B5PRVH4SGD5W6VHHAVCNFSNUABFKJSXA33TNF2G64TZHM4TONJXGM2DGMJZHNEXG43VMU5TMMJTGI2DENZQGY2KC5QC>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/BVRSBAB5LBNYVELFHQNATH35W6VHHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTMNJWGQZDSNJYHE3KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android<https://github.com/notifications/mobile/android/BVRSBAFGJ3L3STWYUPTXJGD5W6VHHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTMNJWGQZDSNJYHE3KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[opencodejr]

The new commit only merges dev into the branch; the fix is unchanged and still applies cleanly after the reworked overflow handling. With the merge, session-runner.test.ts passes (246 tests), including the compaction content-filter test added on dev, and the packages/core type check is clean. The public API impact is the same as before: the session.retry hook now also fires for content-filter finishes, with a default decision of retry: false. This is backwards compatible. I have no new concerns.

[mahammedmahieousman]

Thanks for your explanation.

…

On Mon, Nov 2, 2026, 2:05 AM opencodejr[bot] ***@***.***> wrote: ***@***.***[bot]* commented on this pull request. The new commit only merges dev into the branch; the fix is unchanged and still applies cleanly after the reworked overflow handling. With the merge, session-runner.test.ts passes (246 tests), including the compaction content-filter test added on dev, and the packages/core type check is clean. The public API impact is the same as before: the session.retry hook now also fires for content-filter finishes, with a default decision of retry: false. This is backwards compatible. I have no new concerns. — Reply to this email directly, view it on GitHub <#60836?email_source=notifications&email_token=BVRSBAG6AWF7A5FXXSLDP5L5W66TPA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKNZTGA3TAMJSGYZ2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#pullrequestreview-5730701263>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BVRSBAC5FVMPIVLWSD3FDUD5W66TPAVCNFSNUABFKJSXA33TNF2G64TZHM4TONJXGM2DGMJZHNEXG43VMU5TMMJTGI2DENZQGY2KC5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/BVRSBAFY7E7UH7VJQWEM7ST5W66TPA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKNZTGA3TAMJSGYZ2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/BVRSBAFQOBXGKI4ABO5GHK35W66TPA5CNFSNUABKM5UWIORPF5TWS5BNNB2WEL2QOVWGYUTFOF2WK43UKJSXM2LFO4XTKNZTGA3TAMJSGYZ2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>

[rekram1-node]

Yeah good idea