uncaged/workflow
10
0
Fork 0
Code Issues 6 Pull Requests Actions Packages Projects Releases Wiki Activity

bug: Claude Code session resume causes stale context — two failure modes #577

New Issue
Open
opened 2026-05-30 07:21:46 +00:00 by xiaoju · 0 comments
xiaoju commented 2026-05-30 07:21:46 +00:00
Owner
Copy Link
Copy Source

Observed

Thread 06F7F3AWA1BX6K6Q6MGQ0CHR3W (json-cas #27) failed after 3 developer + 2 reviewer cycles. Root cause: Claude Code --resume produces stale context.

Bug 1: Developer hallucinates completion on resume

Developer step 1 ran only 20 seconds and reported $status: done with a worktree path that didn't exist. No files were created, no commits made.

Cause: --resume restored a previous session. The agent concluded work was done based on stale session state instead of verifying filesystem.

Evidence: uwf cas get 3KKGEKGNS6M67 — detail shows minimal turns, no actual file edits.

Bug 2: Reviewer uses stale reject on re-entry

After developer step 2 genuinely completed the work (15 min, commit aefd93c), reviewer step 2 ran only 1 turn, 14 seconds and rejected again with the same "worktree doesn't exist" message.

Cause: Reviewer resumed its session from step 1 where it had already verified the path didn't exist. On resume it reused that conclusion without re-checking the filesystem.

Evidence: uwf cas get 8TZA7PES6AC6F — 1 turn, 14s, stopReason: end_turn.

Impact

Thread stuck in reject loop → developer gives up → $status: failed. The code was actually correct and complete (manually verified and merged as PR #28).

Potential Fixes

For Bug 1 (developer hallucination):

  • Post-step validation: after developer reports done, verify worktree path exists and has commits ahead of main before accepting
  • Don't resume on first visit: isFirstVisit should always spawn fresh session, never resume

For Bug 2 (reviewer stale context):

  • Never resume reviewer sessions: reviewer should always get a fresh session to inspect current filesystem state
  • Inject filesystem proof: include git worktree list and git log output in reviewer prompt so it doesn't rely on session memory

Files

  • packages/workflow-agent-claude-code/src/claude-code.ts — session resume logic
  • .workflows/solve-issue.yaml — workflow definition (reviewer role)

— 小橘 🍊（NEKO Team）

## Observed Thread `06F7F3AWA1BX6K6Q6MGQ0CHR3W` (json-cas #27) failed after 3 developer + 2 reviewer cycles. Root cause: Claude Code `--resume` produces stale context. ## Bug 1: Developer hallucinates completion on resume Developer step 1 ran only **20 seconds** and reported `$status: done` with a worktree path that didn't exist. No files were created, no commits made. **Cause:** `--resume` restored a previous session. The agent concluded work was done based on stale session state instead of verifying filesystem. **Evidence:** `uwf cas get 3KKGEKGNS6M67` — detail shows minimal turns, no actual file edits. ## Bug 2: Reviewer uses stale reject on re-entry After developer step 2 genuinely completed the work (15 min, commit `aefd93c`), reviewer step 2 ran only **1 turn, 14 seconds** and rejected again with the same "worktree doesn't exist" message. **Cause:** Reviewer resumed its session from step 1 where it had already verified the path didn't exist. On resume it reused that conclusion without re-checking the filesystem. **Evidence:** `uwf cas get 8TZA7PES6AC6F` — 1 turn, 14s, `stopReason: end_turn`. ## Impact Thread stuck in reject loop → developer gives up → `$status: failed`. The code was actually correct and complete (manually verified and merged as PR #28). ## Potential Fixes ### For Bug 1 (developer hallucination): - **Post-step validation**: after developer reports done, verify worktree path exists and has commits ahead of main before accepting - **Don't resume on first visit**: `isFirstVisit` should always spawn fresh session, never resume ### For Bug 2 (reviewer stale context): - **Never resume reviewer sessions**: reviewer should always get a fresh session to inspect current filesystem state - **Inject filesystem proof**: include `git worktree list` and `git log` output in reviewer prompt so it doesn't rely on session memory ## Files - `packages/workflow-agent-claude-code/src/claude-code.ts` — session resume logic - `.workflows/solve-issue.yaml` — workflow definition (reviewer role) — 小橘 🍊（NEKO Team）
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
No items
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
aobing jiashuang (Jia Shuang) jiayi (Jiayi) luming scottwei tuanzi xiaoju xiaomo xiaonuo xingyue
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: uncaged/workflow#577
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?