LLM retry has no max attempts: stream errors cause infinite retry loop, UI stuck on Thinking

Summary

LLM retry has no maximum attempt limit. RETRY_MAX_DELAY is set to 2147483647ms (~24 days). When DeepSeek returns a stream error, the retry loop continues indefinitely — the UI shows "Thinking..." forever with no error feedback.

Evidence

5 processes died on llm runtime selected immediately after stream error
RETRY_MAX_DELAY = 2147483647 (should be ~30 seconds for typical API retries)
No RETRY_MAX_ATTEMPTS constant exists
retryable() matches stream errors against patterns like /stream/ and /error/ — always retryable
RETRY_MAX_DELAY_NO_HEADERS = 30000 exists but only applies when no response headers present
With response headers, cap is full RETRY_MAX_DELAY (24 days)

Code location

File: retry.ts (bundled in main chunk)

RETRY_MAX_DELAY = 2147483647  // ~24 days — effectively unlimited
RETRY_MAX_DELAY_NO_HEADERS = 30000  // only used without headers
// RETRY_MAX_ATTEMPTS is MISSING

function policy2(opts) {
  return fromStepWithMetadata(succeed((meta3) => {
    const error50 = opts.parse(meta3.input);
    const retry6 = retryable(error50, opts.provider);
    if (!retry6) return done(meta3.attempt);  // only stops if NOT retryable
    // continues forever otherwise
  }));
}

Expected behavior

After N attempts (e.g., 3-5), emit a providerError event and stop — let the UI show an error instead of eternal "Thinking..."

Version: 1.18.16, Electron 42.3.3

This issue might be a duplicate of existing issues. Please check:

fix(session): SessionRetry.policy() retries forever with no max attempt count #21960: Nearly identical — SessionRetry.policy() retries forever with no max attempt count, same root cause in retry.ts, filed same day
[Bug]: Session processor retries indefinitely with unbounded exponential backoff — no max retries or circuit breaker #17648: Session processor retries indefinitely with unbounded exponential backoff — no max retries or circuit breaker (same underlying issue)
Retry backoff is uncapped when a retryable error has response headers but no retry-after #33728: Retry backoff is uncapped when a retryable error has response headers but no retry-after (same RETRY_MAX_DELAY concern)

Please review the above issues and consider adding any new information there rather than opening a separate report.

Corroboration — MCP error triggers infinite retry, Windows 1.18.16\n\n### Evidence from incident 2026-08-05 (PID 39476)\n\nWe observed a zombie process that spun for 3 days at ~136% CPU (≈39 CPU-hours, 4.5GB RAM) with zero network, zero disk writes, zero log output. The trigger was an MCP server error:\n\n`\nnotebooklm MCP error: -32601: Method not found\n→ session boot + dispose within 14s\n→ server agent loop stayed "active" but unsettled\n→ compute-bound spin with no I/O for 3 days\n`\n\n### The `-32601` error is NOT transient — but retry treats it as retryable\n\nA JSON-RPC `-32601: Method not found` is a protocol-level permanent error (the method does not exist). A correct client MUST NOT retry it — it\u0027s not transient and retrying will never succeed. However, given the broad `retryable()` pattern matching (`/stream/`, `/error/`), this error likely matches and enters the infinite retry loop.\n\nThis confirms the original report\u0027s finding: `retryable()` is too permissive and with no `RETRY_MAX_ATTEMPTS`, any error coded as retryable spins forever.\n\n### Context-size amplifier\n\nWe also found that larger opening context increases the probability of triggering this bug: with ~43K tokens of opening context (vs ~18K before), API streams are 2.5x longer, increasing the window for mid-stream errors (timeouts, rate limits, connection drops) that trigger the unbounded retry.\n\n### WAL bloat as post-mortem indicator\n\nDuring the zombie\u0027s 3-day spin, the opencode SQLite WAL file grew to 596MB (normal is \u003c10MB). This is a fingerprint of the runaway DB write loop inside the retry spin. WAL size could serve as a secondary detection signal — a process with WAL \u003e 100MB is almost certainly in a retry spin.\n\nVersion: 1.18.16, Windows 11

Addendum from a Windows environment (opencode v1.18.16 / Electron, native Windows Terminal via ConPTY) — 4 incidents over 6 days, supporting the unbounded-retry diagnosis and linking it to the sibling #41806.

Corroborating evidence:

5 server processes died at llm runtime selected immediately after a stream error, exactly matching the retry-loop mechanism described here (retryable() → /stream///error/ → always retryable → no RETRY_MAX_ATTEMPTS).
Timing correlation: the incidents began precisely when our boot context window grew (~18K→46K tokens). Larger turns → longer API streams → higher probability of stream error → the unbounded retry loop engages. Our internal context reduction (~−34%) is our current stopgap.
Cascade with Instance bootstrap hangs forever (Linux): spawned git child exits but is never reaped — TUI renders, but Enter can never start a session #41806: the runaway writes from the retry loop increase git snapshot activity, which feeds the exit-handler race in Instance bootstrap hangs forever (Linux): spawned git child exits but is never reaped — TUI renders, but Enter can never start a session #41806. The two compound; they look like separate symptoms but share the retry trigger.

Windows-specific symptom (new): the hung/dead server leaves a torrent of ANSI SGR codes printed as literal text in the terminal scrollback (e.g. [555;40;30M strings with the ESC byte stripped — a Mode-2 TUI render loop residue, confirmed by OCR of a capture and a clean router-stdout audit — not MCP output). This is a strong post-mortem fingerprint on Windows/ConPTY.

Storage collateral: the zombie server leaves a runaway SQLite WAL — our opencode.db reached 13.3 GB (checkpoint starvation). Worth confirming the bundled SQLite version is ≥ 3.51.3 (the 2026 WAL-reset fix); ours could not be verified from the shipped bundle.

Request: (1) a RETRY_MAX_ATTEMPTS cap (3–5) so a persistent stream error surfaces as a providerError instead of eternal "Thinking…"; (2) confirmation that the exit-handler race (#41806) settles on fast/immediate git-child exits — on Windows we see the same un-reaped-child pattern under index.lock contention from concurrent instances.

Happy to provide an instrumented build pull, captured raw bytes from a live repro (to disambiguate the ESC-strip: console ENABLE_VIRTUAL_TERMINAL_PROCESSING flag vs. a zombie duplicate writer), or a full opencode.log + ps snapshot on request.

Continuation of our prior corroboration (4 Windows incidents, context trigger, WAL fingerprint, ANSI flood symptom).

Independent confirmation on 1.18.32 (Linux, npm opencode-ai@1.18.32).

Setup: OpenCode pointed at a self-hosted OpenAI-compatible gateway (multi-provider failover chain). When every upstream provider refused the content (provider-side moderation), the gateway initially returned a generic overload-style error and OpenCode re-issued the identical request indefinitely — no max attempts, turn stuck with no terminal error.

Gateway-side workaround we had to add: return HTTP 400 with a deliberately non-transient body (code: content_filter_rejected, avoiding any exhausted/overload substring) so the client would stop. The fact that a gateway must signal "do not retry" via body fingerprinting is the symptom of the missing client-side bound.

Suggestion (matches the report): a RETRY_MAX_ATTEMPTS (e.g. 3-5) plus emitting a terminal providerError/finish_reason after N attempts, with RETRY_MAX_DELAY capped to a human timescale (the current 2147483647 ≈ 24 days is effectively unlimited). Also treat 4xx content-policy/terminal errors as non-retryable. Related: #17648.

Fixed in #41939, available in v1.18.17.

Also recommend updating to: https://opencode.ai/v2/docs

Navigation Menu

Repository navigation

LLM retry has no max attempts: stream errors cause infinite retry loop, UI stuck on Thinking #41848

Summary

Evidence

Code location

Expected behavior

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

LLM retry has no max attempts: stream errors cause infinite retry loop, UI stuck on Thinking #41848

Description

Summary

Evidence

Code location

Expected behavior

Activity

github-actions commented on Aug 11, 2026

martialebeau commented on Aug 12, 2026

martialebeau commented on Aug 12, 2026

ImperioCivek commented on Sep 23, 2026

rekram1-node commented on Oct 2, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions