agent/engine: guard against stale sdk_session_id when .jsonl is missing#65
Merged
pufit merged 1 commit intoMay 13, 2026
Merged
Conversation
When a Nerve container is restarted without persisting /root/.claude, the .jsonl conversation files are wiped but the DB still holds the old sdk_session_id. On the next turn the CLI is invoked with --resume and immediately exits 1 with "No conversation found with session ID", which hangs the session mid-turn. Two fixes: 1. Engine guard (_sdk_resume_file_exists): before passing --resume to the CLI, check that the .jsonl exists. If it is missing, clear the stale sdk_session_id from the DB and start a fresh conversation instead of crashing. Forks are exempt because their context is owned by the source session. 2. Bootstrap (.claude mount): add ~/.nerve/claude:/root/.claude to the generated docker-compose so conversation history survives restarts. The mount is siloed under ~/.nerve to keep the agent's CLI state separate from the host user's personal ~/.claude directory. Tests cover file-present, file-absent, fail-open on exception, and workspace-slash encoding. Co-Authored-By: Claude Opus <model> <noreply@anthropic.com>
3 tasks
Contributor
Author
|
Bumping for review. This is one of five small reliability fixes (#63, #64, #65, #66, #67) that compound: each closes one cause of chats that wedge mid-turn, lose context after restart, or get routed into the wrong session. Happy to split, rebase, or land any of them in any order; they don't depend on each other. In daily use the chat surface is the single biggest friction right now, so any of these merging would help. Let me know if you want changes. |
This was referenced May 13, 2026
pufit
approved these changes
May 13, 2026
pufit
pushed a commit
that referenced
this pull request
May 13, 2026
User-reported glitch: "I ask for something and there is no response, then I ask again and it answers to the previous question." Five reliability PRs (#63 shorthand-schema, #64 synthetic done, #65 stale sdk_session_id, #66 idle timeout, #67 sticky session) each close one underlying cause. Two gaps remain that none of those PRs cover. Gap 1: client-side send silently drops payloads. web/src/api/websocket.ts checked readyState === OPEN and no-op'd otherwise. The 3s reconnect window leaves a hole: send() returns to the caller and chatStore.sendMessage has already optimistically appended the user message and set isStreaming=true. The user thinks the agent is thinking but the message never reached the server, so the next reply lands against a stale prompt. Track readyState explicitly. CONNECTING or reconnect-scheduled now queues the payload (bounded to 5 entries; oldest evicted) and flushes from onopen. CLOSED-without-reconnect and CLOSING return 'dropped' so the caller can revert. chatStore.sendMessage pops the optimistic user message on 'dropped' and surfaces an inline assistant error so the user can retry. Gap 2: gateway initial-bind never replayed the broadcaster buffer. The switch_session handler already shipped session_status with buffered_events on session switch, but the initial-connect handshake at server.py:286-311 didn't. Reload mid-turn (or a transient 3s WS drop) and the in-flight stream was lost from the client's view even though the events sat in broadcaster._session_buffers waiting to be replayed. Lift the duplicated send-status construction into _send_session_status and call it from both branches. Initial-bind gates on broadcaster.is_buffering so idle sessions stay silent; switch_session calls unconditionally so the client refreshes is_running/status on every selection. The frontend handleSessionStatus already restores streamingBlocks, panels, todos, and interaction state from the buffer (handled by #69), so this is purely additive at the gateway. Tests: - 9 new asserts in tests/test_gateway_ws.py covering the helper output, the initial-bind gate, the switch_session regression path, and a load-fidelity check for buffer ordering. - Full pytest run: 444 pass, 2 skip, 2 pre-existing failures unrelated (test_bootstrap docker-env detection and test_cli_upgrade docker mode, both noted in notes/repo-conventions/nerve.md). - web/ tsc --noEmit clean, npm run build clean. Out of scope (followups, not blocking): - Stale-listener cleanup on swallowed send_json exceptions (server.py:298-301). - Application-level message_received ack from engine after sessions.add_message. - _session_locks TTL on session archive.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What and why
When a Nerve container restarts without a persisted
/root/.claudebind mount, the Claude Code.jsonlconversation files are wiped. The Nerve DB still holds the oldsdk_session_id, so the next turn invokes the CLI with--resume <stale-id>and the CLI exits 1 immediately: "No conversation found with session ID." The session hangs mid-turn.This PR has two fixes:
1. Engine guard (
nerve/agent/engine.py)Added
_sdk_resume_file_exists(sdk_session_id)which checks whether~/.claude/projects/<encoded-cwd>/<sdk_session_id>.jsonlis on disk before passing--resumeto the CLI. If the file is missing, the stalesdk_session_idis cleared in the DB and the turn starts a fresh conversation instead of crashing. On any unexpected error the method returnsTrue(fail-open) so genuine bugs still surface via the CLI.Fork-source sessions are exempt: a fork's resume target is the source session's .jsonl, which lives on its own row.
2. Bootstrap mount (
nerve/bootstrap.py)Added
~/.nerve/claude:/root/.claudeto the generateddocker-composeso conversation history survivesdocker compose down/up. The mount lives under~/.nerveto keep the agent's CLI state isolated from the host user's personal~/.claude(macOS stores OAuth tokens there).The entrypoint now runs
mkdir -p /root/.claude && chmod 700 /root/.claudeon startup so the bind-mounted directory is always accessible before any tool touches it.Tests
Five new cases in
tests/test_engine.py:TrueFalseos.path.isfileraises →True(fail-open)/encoded to-in the projects subdir path<session_id>.jsonlAll 23 tests pass.
Related
Companion to #64 (synthetic done events for stuck sessions). That PR closes the streaming-state bug; this one closes the resume-target bug.