--print / -p non-interactive mode: TTY hang + 400 developer role rejected by third-party APIs

[naaika50]

Bug: --print / -p non-interactive mode hangs with TTY and fails with third-party APIs

Summary

The --print (-p) non-interactive flag has two critical bugs that make it unusable:

1. Hangs forever when stdin is a TTY (normal terminal) — never exits
2. 400 error with third-party APIs — sends role: "developer" which most OpenAI-compatible APIs reject

---

Environment

Kimchi version	0.1.72
OS	Windows 11
Shell	PowerShell / Git Bash
Install method	kimchi update (binary)
Config	Multi-model with 7 modelRoles configured

---

Reproduction: Bug #1 — TTY Hang

# This hangs FOREVER (no exit, no output beyond config warnings):
kimchi -p --no-session "Say hello"

# Expected: process prompt, print response, exit
# Actual: hangs indefinitely, must Ctrl+C

Workaround (forces non-TTY):

kimchi -p --no-session "Say hello" < /dev/null 2>&1
# This EXITS, revealing the underlying API error

Root cause: When stdin is a TTY, kimchi initializes the full interactive TUI even in -p mode. The --mode rpc output shows it gets stuck at:

{"method":"setStatus","statusKey":"permissions-mode","statusText":"default"}
{"method":"setStatus","statusKey":"active-tags","statusText":"phase: explore"}

After this — no further output. Agent is waiting for user interaction that never comes.

Flags tested that DON'T fix it:

• --yolo → still hangs
• --auto → still hangs
• --plan → still hangs
• --no-session → still hangs
• --mode json → still hangs
• --mode rpc → shows init messages then hangs

---

Reproduction: Bug #2 — 400 developer Role Error

# Non-TTY mode reveals the underlying API error:
kimchi -p --no-session "Say hello" < /dev/null 2>&1

# Output:
# Error: 400 developer is not one of ['system', 'assistant', 'user', 'tool', 'function']
# EXIT: 1

Session log evidence (~/.config/kimchi/harness/sessions/.../...jsonl):

"errorMessage":"400 developer is not one of ['system', 'assistant', 'user', 'tool', 'function']"

This error repeats across ALL retries (3 retries × multiple turns) before the exploration guard triggers.

Root cause: Kimchi's harness sends system-level messages with role: "developer" (OpenAI's newer role for o1/gpt-4o+ models). However, most third-party OpenAI-compatible APIs only accept ['system', 'assistant', 'user', 'tool', 'function'] and reject developer.

Providers tested — ALL failed with developer role:

Provider	Model	Error
freetokenfaucet2	qwen3.7-plus	400 developer is not one of [...]
aiand	zai-org/glm-5.2	403 Forbidden (or 400)
aiand2	deepseek-v4-pro	403 Forbidden
hcnsec	DeepSeek-V4-Flash	Timeout
opnegateway	tencent/hy3:free	Error
bynara1	tencent-hy3	Error
cccctoken	gpt-5.5	Error

---

What DOES Work

Only kimchi-dev provider succeeds:

kimchi -p --provider kimchi-dev --model deepseek-v4-flash --no-session "Say hello" < /dev/null

# Output: "Hello! How can I help you today?"
# EXIT: 0

Confirmed working models on kimchi-dev: deepseek-v4-flash, minimax-m3, kimi-k2.7

---

Current Multi-Model Config (failing)

{
  "modelRoles": {
    "orchestrator": "freetokenfaucet2/qwen3.7-plus",
    "planner":     "opnegateway/tencent/hy3:free",
    "builder":     "aiand2/zai-org/glm-5.2",
    "reviewer":    "aiand/zai-org/glm-5.2",
    "explorer":    "aiand2/deepseek-ai/deepseek-v4-pro",
    "researcher":  "bynara1/tencent-hy3",
    "compactor":   "aiand/zai-org/glm-5.2"
  }
}

All 7 roles use third-party APIs → all fail with 400 developer error.

---

Expected Behavior

1. -p should NOT initialize TUI when stdin is a TTY — it should run in true non-interactive/headless mode and exit after processing
2. -p should work with third-party APIs — either use role: "system" instead of role: "developer", or auto-detect and fall back

---

Suggested Fixes

Fix 1: TTY detection in -p mode
When -p flag is set, force non-interactive/headless code path regardless of TTY detection.

Fix 2: developer → system role fallback
Either change the harness to use role: "system" by default (all APIs support this), or detect the 400 developer error and retry with role: "system".

Fix 3: Better error surface in non-interactive mode
When -p mode encounters an API error after all retries, print the error to stderr and exit with non-zero code, rather than entering the interactive retry/session loop.

---

Note on developer role

The developer role is an OpenAI-specific feature for newer models (o1, gpt-4o, etc.). Most open-source model APIs and third-party proxies use pydantic validation that strictly checks for ['system', 'assistant', 'user', 'tool', 'function']. Using role: "system" would be universally compatible while still functioning identically on OpenAI's API (which accepts both).