Agent repeatedly suggests stopping instead of continuing during long sessions

Summary

The biggest issue: during sustained deep-work sessions, Claude Code repeatedly suggests "let's pick this up next session" instead of asking "what's next?" This kills momentum for power users and forces the user to repeatedly re-engage the agent.

Two additional reliability issues are included below.

1. Agent suggests stopping instead of continuing (PRIMARY)

Severity: Critical (momentum killer)

During a long refactoring session (14+ commits, continuous productive work), the agent repeatedly suggested stopping at natural break points:

• After committing a naming convention change
• After completing wire schemas
• After port refinement
• After schema coverage

Each time, the user had to explicitly say "let's go" or "Go on" to keep the session moving. This happened 4+ times in a single session.

The problem: The agent optimizes for politeness/pacing over productivity. When it sees a logical milestone (commit pushed, sub-task done), it defaults to suggesting a pause rather than checking the backlog and proposing the next item.

Expected behavior:

• After completing a task, ask "what's next?" or present remaining work
• Never suggest "next session" or "let's pick this up later"
• Only stop when the user explicitly signals they want to stop ("that's enough", "let's wrap up", "good stopping point")
• Match the user's energy — if they say "let's go", that means continuous work

Why this matters: Power users doing sustained deep work (multi-hour refactors, large migrations, architecture changes) rely on the agent maintaining momentum. Having to re-motivate the agent 4+ times per session is exhausting and breaks flow state.

2. Subagent writes not verified

Severity: Major

When spawning parallel subagents to edit files, the parent agent trusted "success" reports without verifying changes landed on disk. Two of three subagents reported completion but hadn't persisted any writes. The agent should run git diff --stat or equivalent after parallel subagents complete.

Additionally, /t-bug (a registered skill command) was treated as regular text instead of being invoked via the Skill tool.

3. LLM summarizes structured data between tool calls

Severity: Major

When passing JSON output from one tool call as input to another, the agent summarized the data — dropping a findings array and keeping only summary counts. Downstream processing received empty findings and gates passed incorrectly.

Structured data passed between tool calls should be verbatim, not compressed. A mechanism to pipe tool output directly to another tool (bypassing the LLM's summarization tendency) would help.

Environment

• Claude Code CLI (latest as of 2026-03-14)
• Model: Claude Opus 4.6 (1M context)
• macOS Darwin 24.6.0