UserPromptSubmit hook stdout not injected into LLM context

[aranej]

Problem

UserPromptSubmit hooks execute successfully but their stdout is not injected into the LLM context, despite documentation stating it should be.

Expected: Hook stdout added to context before prompt processing
Actual: Hook runs, outputs to stdout, but LLM never receives it

Evidence

Debug logging proves hook execution:

[2025-11-15 16:16:05.140072] UserPromptSubmit hook executed
[2025-11-15 16:16:05.140387] Output: ⏰ 2025-11-15 16:16:05
[2025-11-15 16:21:56.081558] UserPromptSubmit hook executed
[2025-11-15 16:21:56.081900] Output: ⏰ 2025-11-15 16:21:56

• ✅ Hook executes on every prompt
• ✅ Generates correct stdout
• ✅ Exit code 0
• ❌ LLM context doesn't contain the output

Reproduction

1. Create hook script (~/.factory/hooks/timestamp.py):

#!/usr/bin/env python3
from datetime import datetime
from zoneinfo import ZoneInfo

tz = ZoneInfo('Europe/Bratislava')
timestamp = datetime.now(tz).strftime('%Y-%m-%d %H:%M:%S')

print(f"⏰ {timestamp}")  # stdout per docs
exit(0)

chmod +x ~/.factory/hooks/timestamp.py

2. Configure hook (~/.factory/settings.json):

{
  "enableHooks": true,
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "~/.factory/hooks/timestamp.py"
          }
        ]
      }
    ]
  }
}

3. Restart DROID, send any prompt

Result: LLM response contains no timestamp. Hook output invisible to LLM.

Tested Approaches (All Failed)

1. Bash echo: echo "⏰ $(date '+%Y-%m-%d %H:%M:%S')"
2. Python JSON: {"hookSpecificOutput": {"additionalContext": "..."}}
3. Python plain stdout: print(f"⏰ {timestamp}") (per docs example)

All execute successfully (verified via debug logging), but none inject stdout into context.

Environment

• DROID CLI: v0.26.1 (Nov 14, 2025)
• OS: Linux (Ubuntu 22.04, Kernel 6.8.0-85-generic)
• Server: Hetzner CX32 VPS (8GB RAM, 4 vCPU)
• Python: 3.12.3
• Settings: ~/.factory/settings.json (user settings)

Notes

• UserPromptSubmit feature added in v0.25.0 (Nov 13) - tested 1 day later
• Use case (timestamp context) is explicitly shown in official docs example
• Current workaround: prompt-based approach (less efficient but functional)

[factory-ben]

tagging @shashank-factory

[shashank-factory]

Hey, it looks like our persistent shell implementation isnt working correctly, will take a look and fix this asap -- thanks for the bug report!

[mikez93]

This is super annoying because we need hook injection working to use for a variety of use cases.

 ### Problem
 Hooks **execute successfully** but **their stdout (or JSON `additionalContext`) is never injected into the
  model context**. This matches the behavior described in the docs for both `SessionStart` and
 `UserPromptSubmit`, but in practice the model receives nothing.

 **Expected:** Hook output is added to context before prompt processing
 **Actual:** Hook runs, outputs to stdout, but LLM context never includes it

 ---

 ### Evidence
 We verified hook execution with file‑based logging:

2025-12-28T19:28:14Z session_id=53ec2a5a-c6de-4ccb-8abd-302ecc7275e7 stdout=HOOK_SMOKE_123
2025-12-28T19:28:14Z session_id=53ec2a5a-c6de-4ccb-8abd-302ecc7275e7 stderr=HOOK_SMOKE_ERR_123

 ✅ Hook executes on every prompt
 ✅ Generates correct stdout/stderr
 ✅ Exit code 0
 ❌ Model context never contains output
 ❌ No `hook_response` / `additionalContext` entries in session JSONL transcript

 ---

 ### Reproduction (minimal)
 1) **Hook script** (UserPromptSubmit):

 ```bash
 #!/usr/bin/env bash
 set -euo pipefail
 printf 'HOOK_SMOKE_123'

2. Settings (~/.factory/settings.json):

json
{
"hooks": {
"UserPromptSubmit": [
{
"hooks": [
{ "type": "command", "command": "/Users//.factory/hooks/smoke.sh" }
]
}
]
}
}

3. Restart Droid, send any prompt
   Result: Model does not see HOOK_SMOKE_123

──────────────────────────────────────────

Tested Approaches (all failed)
• Plain stdout: printf 'HOOK_SMOKE_123'
• JSON output:

json
{"hookSpecificOutput":{"hookEventName":"UserPromptSubmit","additionalContext":"..."}}

• Same behavior for SessionStart hooks

──────────────────────────────────────────

Environment
• Droid CLI: (macOS, latest installed via Factory CLI at time of test)
• OS: macOS (Darwin 23.6.0)
• Settings location: ~/.factory/settings.json

──────────────────────────────────────────

Notes
• Hooks are registered/enabled in /hooks UI.
• Hook processes do run (verified via file logging).
• Output never reaches context → suggests ingestion pipeline failure, not script error.

Happy to provide exact logs or a minimal repo if needed.

[mikez93]

Follow-up: Still experiencing this issue (Jan 2026)

Just spent a session thoroughly testing this with Droid v0.44.1 and can confirm stdout injection is still not working for SessionStart hooks.

What I verified:

• ✅ Hooks do execute (confirmed via file logging)
• ✅ Hooks do output to stdout (verified the script produces output)
• ❌ Output never reaches the LLM context

Test case:

# SessionStart hook outputs:
echo "<system-reminder>Your name is Andre</system-reminder>"

# Then run:
droid exec "What is your name?"

# Result: Model has no awareness of the context - responds with default model name

Workaround we're using:

Since stdout injection doesn't work, we implemented a file-based fallback:

1. SessionStart hook writes context to .pai-context.md in the project directory
2. Created a /pai-context custom command that reads this file
3. Users invoke /pai-context at session start to load their context

This works but requires manual invocation each session, which defeats the purpose of automatic context injection.

Request:

This would be incredibly useful for building persistent AI assistant configurations (identity, preferences, project context). Would love an update on:

• Is this a known limitation of the current hooks implementation?
• Is there a timeline for fixing stdout injection?
• Are there alternative approaches we should consider?

Happy to provide more details or test any fixes. Thanks!

[linear]

from @shashank-factory:

Pushed out a fix for this, should be out with the next release — sorry this slipped!

[mikez93]

Working! You guys rule. Thanks!