fix(devin): recover oversized prompt history after rejection

[riverpilot]

Summary

• Treat zero-output Devin invalid_argument: an internal error occurred trailers as context overflow when the serialized, compactable chatMessagePrompts history is already at least 512 KiB.
• Exclude the system prompt, tool schemas, and other fixed request fields from the recovery threshold so context maintenance is selected only when it can shrink the qualifying bytes.
• Preserve transient handling for small histories, fixed-envelope-heavy requests, partial-output failures, and specific invalid-argument responses.

Related #4218

Screenshots

No visual change.

Testing

• bun test packages/ai/test/devin-large-request-recovery.test.ts — 5 passed
• bun test packages/ai/test/devin-frame-cap.test.ts packages/ai/test/devin-streaming-args.test.ts packages/ai/test/devin-large-request-recovery.test.ts — 8 passed
• bun test packages/coding-agent/test/agent-session-auto-compaction-progress-guard.test.ts — 28 passed
• bun --cwd=packages/ai test — 3,009 passed, 337 skipped, 0 failed
• bun run check:ts

Risk

• The 512 KiB value is a recovery heuristic, not an asserted Devin backend limit. It is consulted only after the server rejects a zero-output request with the observed opaque trailer, so accepted large histories remain unaffected.
• The qualifying size is re-encoded from only the repeated history field after rejection; large system prompts and tool schemas retain transient retry behavior.
• Small intermittent Devin errors and errors after partial output retain their existing transient retry behavior.

Notes

• The history measurement uses the uncompressed protobuf field bytes because the Connect server decodes the gzip frame before request handling.
• The original error text and trace ID remain intact for backend diagnosis.

AI Review Report

Focused review verified that AIError.finalize preserves the structural ContextOverflow flag alongside text-derived Transient, AgentSession prioritizes overflow maintenance over retry, the revised threshold measures only compactable history, and partial-output/fixed-envelope cases remain non-overflow. No substantive findings remain.

Security Audit

• No prompt or tool-result content is added to logs; only uncompressed and compressed byte counts are recorded.
• No new network requests, credential handling, or silent history deletion paths are introduced.

[roboomp]

review:p2 — focused implementation; the six targeted Devin provider tests pass, but the recovery heuristic needs explicit maintainer sign-off.
Blocking: total protobuf size is not equivalent to compactable history, so the opaque intermittent trailer can misroute large system-prompt/tool-schema requests into ineffective context maintenance.
Coverage: no partial-output case protects the new zero-output guard. Maintainer call: accept this heuristic tradeoff or require size-causal evidence/a shrinkable-history measurement.
Thanks for the tightly scoped patch and regression coverage, @riverpilot.

[riverpilot]

Addressed both review findings in 73b9760fe:

• Recovery now measures only serialized chatMessagePrompts, excluding system prompt/tool-schema bytes that context maintenance cannot shrink.
• Added negative coverage for a >512 KiB fixed request envelope with small history.
• Added partial-output coverage using a real Devin text-delta frame before the trailer; it remains non-overflow/transient.

Re-verified: 8 focused Devin tests, 3,009 AI tests (337 skipped), and bun run check:ts pass.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d6ce706662

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 450856e95f

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[riverpilot]

Rebased onto latest main to clear the merge conflict (packages/ai/CHANGELOG.md and related history drift). No behavioral changes in this push.

• New head: cc3b2011b
• Verification: bun test packages/ai/test/devin-large-request-recovery.test.ts → 9 pass
• Merge state after push: clean / mergeable

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cc3b2011b5

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[riverpilot]

Follow-up after rebase: CI Lint, type check & web build failed because the large tool-execution recovery fixture still used a string content + top-level toolCalls shape. Updated it to the current AssistantMessage block-array form.

• New head: dcb04238d
• Verification: bun run --cwd packages/ai check:types and bun test packages/ai/test/devin-large-request-recovery.test.ts → 9 pass

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9cc82b1f03

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Exclude all trailing user active-turn messages

The fresh remaining active-tail case is when the same prompt ends with another user-role companion: this backward scan is skipped unless the final role is developer, but #promptWithMessage can append same-turn messages after the prompt (e.g. file mentions after messages.push(message) in packages/coding-agent/src/session/agent-session.ts:8876-8892), and image file mentions convert to role: "user" in packages/coding-agent/src/session/messages.ts:845-872. For a huge current user prompt followed by a small user companion, the huge live prompt is counted in shrinkablePrompts, so the opaque Devin rejection is mislabeled as ContextOverflow and sent through compaction even though that active tail cannot be shrunk.

Useful? React with 👍 / 👎.

[riverpilot]

Investigated against the existing regression suite. The history contains indistinguishable consecutive user-role records: flushed prior execution history must remain compactable, while same-turn user companions must be excluded. Without an active-turn boundary in the Message/Context contract, excluding every trailing user record breaks the prior-history regression and misclassifies large flushed execution output. The current implementation preserves that safety invariant; this finding needs an explicit active-turn boundary before it can be implemented without regressing the earlier P1/P2 fixes.