bug: make check in formula steps exceeds default Bash timeout; override instructions unreliable

[rileywhite]

Gas City version

0.13.5 (commit: 98a429e, built: 2026-04-08T22:44:10Z)

Environment

Ubuntu, bash, tmux provider. Claude Code agent runtime (Claude Opus / Sonnet).

Reproduction

1. Create a formula with a step that instructs the agent to run make check:

[[steps]]
id = "quality-gates"
title = "Run quality gates"
description = """
Run make check with an explicit 10-minute timeout:
\```
Bash(command="make check 2>&1", timeout=600000)
\```
"""

2. Dispatch to a pool agent via gc sling.

3. Observe the agent's behavior when it reaches the quality gate step. Two failure modes:

   Mode A — default timeout: The agent runs make check in a normal Bash call without setting timeout=600000. The default 2-minute timeout kills the process with SIGTERM (exit code 144). The agent retries, hits the same timeout, and crash-loops.

   Mode B — commented out: The agent echoes the tool invocation syntax as a bash comment (# Bash(command="make check 2>&1", timeout=600000)) rather than translating it into an actual tool call with the timeout parameter.

4. make check for this repo runs go test ./... which takes 3+ minutes — reliably exceeds the 2-minute default.

Expected behavior

Formula step descriptions should be able to reliably instruct agents to run long-running commands with appropriate timeouts. make check is the standard quality gate per CONTRIBUTING.md and should be runnable from formula-driven workflows.

Actual behavior

There is no reliable way to override the Bash tool timeout from within a formula step description. The Bash(command=..., timeout=...) syntax in step descriptions is either:

• Ignored (agent uses default 2-minute timeout)
• Treated as literal text / commented out

This makes it impossible to create formula-driven workflows that follow the project's own contributing guidelines requiring make check before push.

Logs, screenshots, or traces

From polecat session logs, the agent's interpretation of the formula instruction:

# 1. Run quality gates — MUST use timeout: 600000 or SIGTERM will crash you
# Bash(command="make check 2>&1", timeout=600000)
# Bash(command="make check-docs 2>&1", timeout=600000)   # if docs changed

The tool invocation syntax is rendered as bash comments, not executed as tool calls.

Additional context

The formula (mol-polecat-work) included increasingly emphatic warnings about the timeout:

• "CRITICAL — Bash timeout"
• "NEVER use default timeout"
• "you will crash-loop"
• Bold text, tables, repeated instructions

None of these reliably caused the agent to use the correct timeout. The fundamental issue is that formula step descriptions are natural language processed by an LLM, so tool invocation parameters like timeout cannot be guaranteed.

Possible directions:

• A formula-level mechanism to set default Bash timeout for all commands in a step (e.g., [step.tool_defaults] in the formula schema)
• A project-level config that raises the default Bash timeout for agents working in this repo
• Making make check faster (under 2 minutes) so the default timeout suffices

[rileywhite]

Related upstream: anthropics/claude-code#45717 — the Bash tool timeout sends SIGTERM to the Claude Code process itself, not just the child command. This is the root cause of the crash-loop described here: when make check exceeds the 2-minute default timeout, the agent process dies rather than receiving a recoverable timeout error.

If that upstream bug is fixed, agents would survive the timeout and could retry with a longer timeout or skip the step — eliminating the crash-loop even when the agent doesn't follow the timeout: 600000 instruction.

[rileywhite]

Additional note: make test (which runs go test ./... directly) also exceeds the 2-minute default Bash timeout. This matters for workarounds — splitting make check into individual gates (make fmt-check, make lint, make vet, make test) only avoids the timeout for the first three. The test step still hits the same wall.

My current workaround runs make fmt-check, make lint, and make vet individually (all complete within the timeout) but skips make test entirely. Any solution needs to cover both make check and make test.

[julianknutsen]

Investigation Summary

We've investigated and confirmed this bug. Here's what we found:

Reproduction (confirmed on current main at 894861ac):

Command	Wall clock	Exceeds 120s default?
go test ./... -count=1 -short	10m 03s (602.91s)	Yes (5x)
make fmt-check	6m 03s (362.60s)	Yes (3x)

Your report of "3+ minutes" is conservative — the full suite is well over 10 minutes now.

Root cause — two independent issues:

1. No structured formula-to-tool parameter binding. Formula step descriptions are natural language that the LLM interprets. Tool call parameters (like timeout) are structured API parameters. There is no mechanism in Gas City to guarantee that natural language instructions ("use timeout: 600000") translate into actual tool parameters. This is why bold text, CRITICAL markers, and repeated instructions all fail — they're the wrong abstraction layer for the problem.

2. Upstream: Claude Code SIGTERM propagation (anthropics/claude-code#45717). The Bash tool timeout sends SIGTERM to the agent process itself rather than just the child command, making timeout fatal instead of recoverable. This causes the crash-loop you described.

Fix path:

We'll add a CLAUDE.md-level instruction for long-running commands in this repo. CLAUDE.md content is system-level prompt context, which is more reliably followed than step description text. We also confirmed that Claude Code's settings.json schema does not currently support tool parameter defaults, so the settings path is blocked upstream — but Gas City's overlay infrastructure is ready if that changes.

Classification: confirmed-bug · priority/p2 · status/accepted

A fix PR will follow.