fix: save full Claude session log on no_push for debugging #49

New issue

Closed

opened 2026-03-29 07:23:09 +00:00 by dev-bot · 0 comments

dev-bot commented

2026-03-29 07:23:09 +00:00

Collaborator

Problem

When the dev-agent fails with no_push, there is no way to see what the model actually did — which files it read, what tool calls it made, what thinking traces it produced. Debugging requires manually finding the session JSONL in the Claude config dir, which gets cleaned up with the worktree.

Current state

We added agent-run-last.json output capture (the final result JSON). But this only shows the last turn's result text, not the full conversation — no tool calls, no thinking, no intermediate steps.

Fix

On no_push, before cleaning up the worktree, copy the full session JSONL to a persistent location:

In dev/dev-agent.sh, in the no_push handler:

# Save full session log for debugging
session_log="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/projects/$(echo "$WORKTREE" | tr '/' '-'  | sed 's/^-//')/${_AGENT_SESSION_ID}.jsonl"
if [ -f "$session_log" ]; then
  cp "$session_log" "${DISINTO_LOG_DIR}/dev/no-push-session-${ISSUE}-$(date +%s).jsonl"
  log "no_push session log saved"
fi

This gives full visibility: every Read, Edit, Bash call, every thinking trace, every text response.

Also consider

Log a summary (number of tool calls by type, total turns, which files were read/edited)
Alert if the model attempted tool calls that failed (curl to external URLs, apt-get, etc.)

Affected files

dev/dev-agent.sh — no_push handler

## Problem When the dev-agent fails with `no_push`, there is no way to see what the model actually did — which files it read, what tool calls it made, what thinking traces it produced. Debugging requires manually finding the session JSONL in the Claude config dir, which gets cleaned up with the worktree. ## Current state We added `agent-run-last.json` output capture (the final result JSON). But this only shows the last turn's result text, not the full conversation — no tool calls, no thinking, no intermediate steps. ## Fix On `no_push`, before cleaning up the worktree, copy the full session JSONL to a persistent location: In `dev/dev-agent.sh`, in the no_push handler: ```bash # Save full session log for debugging session_log="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/projects/$(echo "$WORKTREE" | tr '/' '-' | sed 's/^-//')/${_AGENT_SESSION_ID}.jsonl" if [ -f "$session_log" ]; then cp "$session_log" "${DISINTO_LOG_DIR}/dev/no-push-session-${ISSUE}-$(date +%s).jsonl" log "no_push session log saved" fi ``` This gives full visibility: every Read, Edit, Bash call, every thinking trace, every text response. ## Also consider - Log a summary (number of tool calls by type, total turns, which files were read/edited) - Alert if the model attempted tool calls that failed (curl to external URLs, apt-get, etc.) ## Affected files - `dev/dev-agent.sh` — no_push handler