A /talk conversation resumes its own session across a real forge restart #8780

I added the restart test, but it has not run against the real claude yet. This sandbox has the binary but no ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN, so it skipped here and printed why. bun x tsc --noEmit, bun scripts/gate.ts and test/conversation-turn.test.ts (19 pass) are green. The third checklist item is open: I have no log lines to show, and the check's first real run will have to supply them.

What changed

• test/conversation-restart.test.ts (new) runs two forge lifetimes on one data directory, each with createApp, serve, say, then stop. It runs on haiku, with each turn capped at $0.25.

- The first message is "Remember the word PELICAN", and the second asks which word it was. Resuming the same session is what lets the model answer. - After the restart it asserts the second turn's run log says resuming session <first id> and never says no session to resume. It also asserts claude_session_id is unchanged and the agent's answer contains "pelican". - It finds <session id>.jsonl under dataDir/claude-config, checks that directory is owner-only, and checks no .jsonl appears under a temporary HOME it sets for the test. - It skips with a console reason when claude or the credential is missing. The check's cells have both, so it runs there. - It prints both runs' logs and the session file paths, so the check log carries the lines.

• src/crew/conversation.ts now writes resuming session <id> to the run log before each resumed runClaude. Without it, a resume and a silent fresh start would look the same in the log. The existing fresh-start log line is unchanged.

Not done

• I didn't tag the test to run only in the check. Two tiny turns stay well inside the per-file budget (300 s cap).
• I didn't simulate a crash mid-turn; resumeInterrupted already has tests with a stub binary.

To try it: with ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN set and claude on PATH, run bun test test/conversation-restart.test.ts.

🤖 Generated with Claude Code

Closes #8776

I added the restart test, but it has not run against the real claude yet. This sandbox has the binary but no ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN, so it skipped here and printed why. bun x tsc --noEmit, bun scripts/gate.ts and test/conversation-turn.test.ts (19 pass) are green. The third checklist item is open: I have no log lines to show, and the check's first real run will have to supply them.

What changed

• test/conversation-restart.test.ts (new) runs two forge lifetimes on one data directory, each with createApp, serve, say, then stop. It runs on haiku, with each turn capped at $0.25.

- The first message is "Remember the word PELICAN", and the second asks which word it was. Resuming the same session is what lets the model answer. - After the restart it asserts the second turn's run log says resuming session <first id> and never says no session to resume. It also asserts claude_session_id is unchanged and the agent's answer contains "pelican". - It finds <session id>.jsonl under dataDir/claude-config, checks that directory is owner-only, and checks no .jsonl appears under a temporary HOME it sets for the test. - It skips with a console reason when claude or the credential is missing. The check's cells have both, so it runs there. - It prints both runs' logs and the session file paths, so the check log carries the lines.

• src/crew/conversation.ts now writes resuming session <id> to the run log before each resumed runClaude. Without it, a resume and a silent fresh start would look the same in the log. The existing fresh-start log line is unchanged.

Not done

• I didn't tag the test to run only in the check. Two tiny turns stay well inside the per-file budget (300 s cap).
• I didn't simulate a crash mid-turn; resumeInterrupted already has tests with a stub binary.

To try it: with ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN set and claude on PATH, run bun test test/conversation-restart.test.ts.

🤖 Generated with Claude Code

Left #8776 open: 1 unticked item remain (The pull says what the test showed, with the log lines.).