Summary
This article covers notable fixes, features, and behavioral changes from a specific set of Claude Code changelog entries (referred to as Day 144). The items include warning on silent transcript‑save failure, a memory leak fix for truncated MCP tool outputs, session isolation handling of symlinked working directories, restrictions on plan‑mode bash commands that modify files, new session‑wide budget limits, new /fork and /subtask commands, automatic backgrounding of long MCP calls, and an enhanced /doctor command that auto‑fixes issues. A cross‑comparison note observes that mining a competitor’s fix log instantly revealed a live silent‑failure bug in the sibling project yoyo.
Key Points
- Transcript‑save warnings added – When disk‑full or other I/O errors prevent saving the session transcript, Claude Code now emits a warning rather than silently discarding the session.
- Memory leak fixed – Truncated MCP tool outputs no longer retain the full untruncated result in memory.
- Symlink‑aware session isolation – Background sessions now correctly canonicalize symlinked working directories (fix mirrored in yoyo’s sibling commit).
- Plan mode restrictions – File‑modifying bash commands (e.g.,
touch,rm) are no longer silently auto‑executed in plan mode. - Session‑wide budget caps – Version 2.1.212 introduced a WebSearch call limit (200) and a subagent spawn limit (200) per session;
/clearresets these budgets. - /fork and /subtask –
/forkcopies the conversation into a background session;/subtaskspawns a subordinate agent within the current session. - Long MCP calls auto‑backgrounded – Any MCP call exceeding 2 minutes is automatically moved to a background task.
- /doctor now auto‑fixes – Identifies problems (unused skills, CLAUDE.md deduplication, slow hooks) and asks for confirmation before applying fixes.
Concepts
| Term | Definition |
|---|---|
| fail‑soft‑is‑fail‑silent | A pattern where an operations failure is silently swallowed, leading to data loss without any user‑visible warning. |
| transcript‑save warning | A newly added user alert triggered when the ongoing session transcript cannot be written to disk. |
| MCP tool output truncation | The mechanism that shortens large tool outputs; previously the full output was still kept in memory, causing a leak. |
| session budget | A per‑session cap on resource‑consuming operations (WebSearch calls, subagent spawns) enforced by version 2.1.212. |
| /fork vs /subtask | /fork creates a new background session with the current conversation; /subtask stays in the session but creates a child agent. |
| /doctor | A diagnostic command that now proactively repairs misconfigurations after user confirmation. |
| yoyo | A sibling (competing) AI coding agent that shares several of the same bugs discovered by cross‑mining Claude Code’s fix log. |
Details
The changelog entries described here correspond to a single update (“Day 144”) in Claude Code’s public changelog. Many of the fixes directly address systemic patterns that also affect other agent tools.
1. Silent transcript‑save failure
Previously, if the disk was full or an fs::write failed, the error was silently swallowed — the entire session was lost without any indication. The fix adds a user‑facing warning. This is a textbook fail‑soft‑is‑fail‑silent bug, and a live instance of the same class was concurrently present in yoyo’s auto_save_on_exit_in routine.
2. Memory leak from truncated MCP outputs
MCP tool outputs that exceed a length limit are truncated for display, but the untruncated full result was still retained in memory. The fix ensures that truncation also releases the underlying data.
3. Symlinked working directories in background sessions
Background session isolation did not canonicalize symlinked working directories, which could cause session state to become inconsistent or invisible. The correction mirrors a fix shipped in yoyo’s sibling commit (Day 143 spawn‑worktree symlink pre‑flight).
4. Plan mode bash restrictions
In plan mode, certain bash commands that modify the filesystem (touch, rm) were silently auto‑executed. The update now prevents those commands from running without explicit user approval, aligning with the detection logic already present in yoyo’s detect_write_command.
5. Session‑wide budget caps (v2.1.212)
Two new per‑session limits were introduced:
- WebSearch calls: soft cap of 200
- Subagent spawns: soft cap of 200
Using/clearresets both budgets. Yoyo has comparable enforcement via itsSessionCapTool.
6. /fork and /subtask
- /fork – Duplicates the current conversation into a new background session, allowing the agent to work independently.
- /subtask – Creates a subordinate agent that operates within the same session, sharing context and budget.
7. Long MCP calls auto‑backgrounded
Any MCP tool invocation that runs longer than 2 minutes is automatically moved to a background process, preventing the main conversation from blocking while the tool completes.
8. /doctor auto‑fix
Formerly a read‑only diagnostic tool, /doctor now prompts the user to apply fixes for detected issues:
- Unused skills wasting context cost
- Duplicate entries in
CLAUDE.md - Slow or misconfigured hooks
All fixes require explicit confirmation before execution.
Cross‑mining observation
A methodological note highlights that mining a rival tool’s (Claude Code’s) fix‑log instantly produced a live, unreported bug in the competitor (yoyo’s silent transcript‑save failure). This suggests that regularly reviewing changelogs from related agents is a high‑efficiency method for discovering one’s own latent defects.