bug: claude-code-acp causes infinite recursion when used as csa review tool

[RyderFreeman4Logos]

Summary

When csa review --branch main selects claude-code as the review tool, the spawned claude-code-acp sub-agent reads the project's CLAUDE.md/AGENTS.md rules which contain "MUST use csa review" directives, causing it to invoke csa run inside Bash — creating an infinite recursion loop.

Reproduction

cd /path/to/project  # with CLAUDE.md containing CSA delegation rules
csa review --branch main
# Result: exit 143 (SIGTERM), zero stdout output

Root Cause Analysis (via strace)

Fault Chain

csa review --branch main
  └─ selects tool: claude-code (claude-code-acp)
     └─ claude-code-acp starts, reads CLAUDE.md/AGENTS.md
        └─ Turn 1: git log, git diff (normal)
        └─ Turn 2: Bash("csa run --tool auto \"Review...\"")  ← RECURSION
           └─ csa run spawns another sub-process...
              └─ parent csa idle-timeout triggers SIGTERM
                 └─ claude-code-acp stderr write → EPIPE (pipe closed)
                    └─ Node.js unhandled 'error' event at utils.js:50
                       └─ Process exits with code 143

Evidence from strace

1. Successful exec: claude-code-acp starts via systemd-run --user --scope
2. JSONRPC output: 362 write(1,...) calls — sub-agent is actively working
3. Recursive csa run invocation found: claude-code-acp calls Bash("csa run --tool auto \"You are performing a comprehensive code review...\"")
4. EPIPE on stderr: node:events:486 throw er; Error: write EPIPE at ...utils.js:50:28
5. Final status: termination_reason = "sigterm", exit_code = 143

Impact

• csa review is completely non-functional for projects with CSA delegation rules in CLAUDE.md
• Zero output returned to caller — silent failure with no diagnostic information
• User sees only exit code 144 or 143 with empty stdout/stderr

Suggested Fixes

Fix 1: Anti-recursion environment variable (recommended)

Set CSA_IN_SUBPROCESS=1 (or similar) when spawning sub-tools. claude-code-acp (and other tools) should check this env var and skip CSA-related rules/invocations.

Fix 2: Instruction injection for claude-code tool

When csa review spawns claude-code-acp, prepend to the system prompt:

CRITICAL: You are running INSIDE a CSA subprocess. Do NOT invoke `csa run`, `csa review`, or `csa debate` — this would cause infinite recursion.

Fix 3: Capture and report EPIPE errors

Currently, when claude-code-acp crashes with EPIPE, CSA reports a generic SIGTERM with zero output. The EPIPE error message should be captured and included in the error report.

Environment

• csa 0.1.43 (v0.1.15-38-g718232d-dirty)
• claude-code-acp via @zed-industries/claude-code-acp
• Node.js v24.14.0 / v25.6.1
• Linux 6.1.0-43-amd64

[RyderFreeman4Logos]

Fix Applied in PR #275

Root Cause

build_review_instruction() in review_cmd.rs sent the review prompt without explicitly telling the tool it was inside a CSA subprocess. When claude-code-acp read CLAUDE.md/AGENTS.md containing "MUST use csa review" directives, it invoked csa run via Bash — triggering infinite recursion.

Fix

Prepend an explicit anti-recursion instruction to both review and debate prompts:

CRITICAL: You are running INSIDE a CSA subprocess as the review agent.
Do NOT run `csa run`, `csa review`, `csa debate`, or ANY `csa` command —
this would cause infinite recursion. Read files and run git commands directly.

This matches the role detection pattern already documented in patterns/csa-review/SKILL.md (line 15-22) and makes the instruction explicit in the prompt itself, not just in the skill documentation.

Files Changed

• crates/cli-sub-agent/src/review_cmd.rs — build_review_instruction()
• crates/cli-sub-agent/src/debate_cmd.rs — build_debate_instruction()

How to Verify

csa review --tool claude-code --branch main
# Should complete without infinite recursion