Subagent recovery after MAX_TURNS is reported as GOAL success, hiding interruption

[matei-anghel]

What happened?

Summary: codebase_investigator subagent reports status: "success" and Termination Reason: "GOAL" even though its own result says it hit the maximum turn limit before doing any analysis.

While using Gemini CLI to investigate two local repositories (t3code and superset), the agent invoked the codebase_investigator subagent and it failed immediately without doing any analysis.

I ran the Gemini CLI on these two local repositories with an objective asking it to understand their architectures and AI agent integrations. The saved session log shows:

• The main agent called codebase_investigator with a valid, concrete objective.
• The tool result was reported with status: "success" and Termination Reason: "GOAL".
• The embedded result text simultaneously says the investigation was interrupted because the "maximum number of turns was exceeded" before any specific task was identified or initiated.
• RelevantLocations was empty and no actual codebase analysis was performed.

This looks incorrect because:

1. The subagent appears to have exhausted its turn budget before performing any actual analysis.
2. The termination metadata is internally inconsistent: the parent tool call is surfaced as status: "success" with Termination Reason: "GOAL", but the embedded payload describes a limit-induced early termination with no findings.

Relevant evidence from the session log:

session file: session-2026-03-09T18-34-10479ef5.json

line 50: "name": "codebase_investigator"
line 53: objective: "Understand the architecture of t3code and superset..."
line 61: "Termination Reason: GOAL"
line 61: "The investigation was interrupted by a system-mandated termination (maximum turns exceeded)..."
line 61: "\"RelevantLocations\": []"
line 66: "status": "success"
line 79: "The `codebase_investigator` was unable to complete its analysis due to a turn limit..."

Full snippet from the saved session:

{
  "id": "codebase_investigator_1773081420202_0",
  "name": "codebase_investigator",
  "args": {
    "objective": "Understand the architecture of t3code and superset, focusing on AI agent integrations.\nSpecifically:\n1. In t3code, how is 'Codex' integrated? Where is the core logic for agent interaction?\n2. In superset, what are the different agent/harness integrations (e.g., Mastra, MCP, Claude, etc.) and how are they implemented?\n3. What are the common patterns or interfaces used in superset that could be adapted for t3code to support multiple harnesses?\n\nIdentify key files, symbols, and architectural patterns in both repositories."
  },
  "result": [
    {
      "functionResponse": {
        "response": {
          "output": "Subagent 'codebase_investigator' finished.\nTermination Reason: GOAL\nResult:\n{\n  \"SummaryOfFindings\": \"The investigation was interrupted by a system-mandated termination (maximum turns exceeded) before any specific task could be identified or initiated. As a result, no codebase analysis was performed, and no relevant locations or insights could be gathered.\",\n  \"ExplorationTrace\": [\n    \"Received system notification that the maximum number of turns was exceeded.\",\n    \"Followed emergency termination protocol as instructed.\"\n  ],\n  \"RelevantLocations\": []\n}"
        }
      }
    }
  ],
  "status": "success"
}

What did you expect to happen?

I expected one of these outcomes:

• codebase_investigator should actually explore the repositories and return findings.
• If it truly hits a turn limit, it should do so after some visible work, and the tool invocation should be surfaced as a failure or explicit limit condition, not as status: "success" with Termination Reason: "GOAL".

At minimum, the CLI should not report a contradictory success state for a subagent that says it exceeded its turn budget before beginning analysis. If the subagent hits a system-imposed turn limit before doing useful work, the tool result should be reported as a failure or explicit limit condition (e.g. status: "error" with a clear error code), not as a successful GOAL completion.

Client information

Client Information

Run gemini to enter the interactive CLI, then run the /about command.

> /about
Gemini CLI package version: 0.32.1
Platform: Windows
OS: Microsoft Windows NT 10.0.26200.0
Node.js: v20.20.0
PowerShell: 5.1.26100.7920

Login information

Google Account, Pro tier from students promotion

Anything else we need to know?

• After the failed codebase_investigator call, the main agent continued with manual file exploration and eventually produced an implementation plan. So the overall conversation did not fully crash, but the specialized investigator tool clearly failed.
• My current inference is that this may be either:
  • turn-budget state leaking into a fresh subagent invocation, or
  • an incorrect mapping of a subagent limit/error state into a successful GOAL termination in the parent session.
• If useful, I can provide the full saved session JSON or a redacted version of it.

[gemini-cli]

👋 @student-ankitpandit, thanks for your interest in this issue! We're reserving self-assignment for issues that have been marked with the help wanted label. Feel free to check out our list of issues that need attention.