fix: prevent context blowup by preserving previous_response_id on bridge recovery - IMPORTANT FIX FOR CONTEXT BLOWING UP

[balakumardev]

Problem

When an upstream websocket session is lost (session.closed) or a request fails mid-flight, the HTTP bridge recovery paths introduced in v1.12.0 return 400 previous_response_not_found errors. This tells the Codex CLI that the server-side conversation state is gone, causing it to drop previous_response_id and resend the entire conversation history as input.

The result is catastrophic for context window usage:

Metric	Before (v1.11.x)	After (v1.12.0)
Per-turn growth	~2-3K tokens	~70K tokens
Turns to fill 876K window	200+	17
Compactions per session	0	2+

Root Cause

Three recovery paths in the bridge return 400 previous_response_not_found instead of 502 upstream_unavailable:

1. Mid-request websocket failure - when the upstream websocket dies during send() and previous_response_id is present, the error handler returns 400 previous_response_not_found instead of 502 upstream_unavailable.

2. _retry_http_bridge_request_on_fresh_upstream - when called with send_request=True and previous_response_id is present, immediately marks the request as previous_response_not_found without attempting recovery.

3. _retry_http_bridge_precreated_request - same pattern.

When the CLI receives 400 previous_response_not_found, it interprets this as "the server doesn't have this conversation anymore" and resends the full conversation history without previous_response_id, inflating per-turn context by ~20x.

Fix

• When session.closed and previous_response_id is present, raise 502 immediately (matching pre-v1.12.0 behaviour) so the client retries with previous_response_id intact.
• In error handler, always raise 502 instead of 400 previous_response_not_found.
• In retry helpers, return False without marking previous_response_not_found so callers raise a retriable 502.

Test Results

7 regression tests added in test_bridge_context_blowup.py. Ran against both broken and fixed code:

Test	Broken (pre-fix)	Fixed
test_mid_request_failure_with_previous_response_id_raises_502	FAIL (returns 400)	PASS (returns 502)
test_retry_with_previous_response_id_returns_false_without_marking_error	FAIL (sets previous_response_not_found)	PASS
test_closed_session_with_previous_response_id_raises_502	PASS	PASS
test_closed_session_without_previous_response_id_attempts_recovery	PASS	PASS
test_healthy_growth_rate_stays_within_budget	PASS	PASS
test_broken_growth_rate_fills_window_fast	PASS	PASS
test_error_code_determines_growth_rate	PASS	PASS

All 44 existing test_proxy_http_bridge.py tests pass with zero regressions.

Live Verification

Tested on a single-instance deployment with gpt-5.4 (876K effective context window):

• v1.12.0 (broken): 4 user messages -> 853K tokens in 17 turns -> compaction triggered, avg 70K tok/turn
• v1.11.x (reverted): 4 user messages -> ~80K tokens in 17 turns -> no compaction, avg 518 tok/turn

[Soju06]

Thank you for this important fix @balakumardev! 🎉

The context blowup issue is a critical regression — great catch with the session log analysis showing 70K vs 2.3K tokens/turn.

I've rebased your changes on main and applied Codex adversarial review fixes (preserving the reconnect recovery path for between-turn disconnects). The reviewed version is in #401.

Closing this in favor of #401 which includes your original fix + review improvements.

[Soju06]

Thank you for catching this critical context blowup regression @balakumardev! 🎉

The session log analysis (2.3K vs 70K tokens/turn) was incredibly helpful in diagnosing the root cause. The Codex adversarial review identified some edge cases around multi-replica recovery and reconnect paths which were resolved together.

@all-contributors please add @balakumardev for code and test

[allcontributors]

@Soju06

I've put up a pull request to add @balakumardev! 🎉