Repository navigation
Open
Description
Summary
In an agentic Codex session (Codex Desktop → magpie gateway → agy-cli → gemini-4-carbon), a turn died with:
agy-cli: agy bridge: a request combining a live tool_result with an assistant prefill tail is not supported
Three problems chain together:
- The turn was apparently killed by
TURN_TIMEOUT_MS(5 min of agy silence). The model had already streamed a short commentary sentence and then went silent (most likely while generating a large tool call). About 300 s later the request failed. This step is inferred: no separate timeout line was logged. - magpie then tried to continue the cut-off reply by sending the partial text back as an assistant prefill. In an agent loop, the trailing messages are
… user(tool_result) → assistant(partial text).validateBodyrejects exactly that combination (index.mjs~L5549). magpie only falls back from prefill when the error matches its own "prefill refused" pattern, roughlydoes not support (assistant message )?prefill|requests ending with a model turn are not supported. The bridge's wording (…prefill tail is not supported) does not match, so magpie never retries without prefill. The client gets a non-retryable 400, and the turn ends. - The agy session is dropped on timeout (
timeoutTurn→dropFromPool+stop()). On the user's next "continue", a fresh agy session replays roughly 156k tokens of history through the 112 KiB budget ([... earlier conversation history omitted to fit ACP 112 KiB prompt budget ...]). The model loses the earlier agreed spec and starts re-reading its own thread through tools. It also tries to parse agy's spilled.system_generated/steps/N/output.txtfile, which is not plain JSON because it starts with aCreated At / Wall time / Output:header.
Environment
- agy-cli plugin 0.1.95 (same code on
maintoday) - agy 1.5.7, magpie 0.1.3960 (as printed by
magpie --version), macOS 27.0.1 - Client: Codex Desktop (Responses API) via magpie; model
agy-cli/gemini-4-carbon, effort high
Timeline (local time, from magpie app.log and the Codex rollout)
| time | event |
|---|---|
| 17:06:44 | tool result returned; Codex sends the next request (tail = tool_result) |
| 17:08:41 | reasoning done; assistant text starts streaming ("…now writing manifest + build code…") |
| 17:08:41 → 17:13:43 | no further output for ~302 s; agy's transcript shows that step with output_tokens: 7332 but no tool_calls recorded |
| 17:13:43 | gateway: codex's call to agy-cli/gemini-4-carbon on agy-cli failed: 200 after 6m58.612s: agy-cli: agy bridge: a request combining a live tool_result with an assistant prefill tail is not supported |
| 17:13:44 | client retry: failed: 400 after 667ms: same message → turn ends |
| 17:17:54 | user says "continue" → new agy conversation; first prompt contains the 112 KiB "history omitted" notes |
Expected
- A long silent generation (big tool call / file write) should not be killed at 5 min. Or, if it is, the error should be one the client or magpie can retry cleanly, and the conversation should not lose its agy session.
- A continuation after a cut-off reply should not dead-end when the tail contains a tool_result.
Suggestions
- In
validateBody, either support "tool_result + prefill tail": render the tool results, then the prefill carrier note, in the same turn. Or reject it with wording or a code that magpie's prefill-refusal detection recognizes (e.g.does not support assistant message prefill …). Then magpie falls back to a non-prefill retry instead of surfacing a 400. - Consider a larger default
AGY_BRIDGE_TURN_TIMEOUT_MSfor turns that have already streamed text. Or reset the timer on agy-side progress signals during long tool-call generation, if ACP exposes any. - On
timeoutTurn, keep the session resumable (session/resume) rather than forcing a 112 KiB history replay on the next turn.
Workaround on my side: launchctl setenv AGY_BRIDGE_TURN_TIMEOUT_MS 900000, then restart magpie.
Happy to provide more log excerpts if useful.
Metadata
Metadata
Assignees
Labels
No labels
Activity