debugging-rook

name	debugging-rook
description	Debugging patterns, CLI commands, scripts, and workflows for the Rook monorepo. Use when investigating bugs, testing server/client behavior, stepping through session replay, or inspecting environment state.

Debugging Rook

Use these tools and patterns to investigate and fix bugs in Rook.

Priority order

1. rook CLI + mock agent — fastest iteration, no native rebuilds
2. rook CLI + real runtime — when you need real AI behavior
3. Codex + computer use on the mac client — only when the bug is in native UI rendering

rook CLI

All commands need --auth-token "$ROOK_AUTH_TOKEN" (source .env first).

One-shot exec

rook exec --runtime MyPiOpenAiAgent --auth-token "$ROOK_AUTH_TOKEN" "tell me a joke"
rook exec --sessionId <id> --auth-token "$ROOK_AUTH_TOKEN" "what did you just say?"
rook exec --last-message-only --runtime MockAcpAgent --auth-token "$ROOK_AUTH_TOKEN" "12+34"

Session management

rook sessions --auth-token "$ROOK_AUTH_TOKEN"                 # list sessions
rook sessions --limit 5 --auth-token "$ROOK_AUTH_TOKEN"      # limit output
rook --transcript --sessionId <id> --auth-token "$ROOK_AUTH_TOKEN"  # dump replay

Named sessions (for mac client testing)

rook exec --runtime MyPiOpenAiAgent --auth-token "$ROOK_AUTH_TOKEN" --title "my-test" "ls"

--title only works with --runtime (new session), not --sessionId.

Environment inspection

rook environments --auth-token "$ROOK_AUTH_TOKEN"             # list known environments
rook environments --limit 5 --auth-token "$ROOK_AUTH_TOKEN"   # with bundle counts

Join/leave environments on session create or resume

rook exec --runtime MyPiOpenAiAgent --auth-token "$ROOK_AUTH_TOKEN" --join location:office "hi"
rook exec --sessionId <id> --auth-token "$ROOK_AUTH_TOKEN" --leave web:example.com "done"

--join and --leave are repeatable. Works with both --runtime and --sessionId.

Interactive mode

rook --runtime MyPiOpenAiAgent --auth-token "$ROOK_AUTH_TOKEN"   # new session, chat
rook --sessionId <id> --auth-token "$ROOK_AUTH_TOKEN"            # resume, transcript replays

Ctrl+C prints sessionId and exits.

Mock agent

File: server/src/server/agents/test-fixtures/mockAcpServer.mjs

• stores a transcript and replays on session/load
• streams thoughts, tool calls, tool outputs, assistant text
• handles common prompt patterns: jokes, ls, arithmetic, prime checking
• serialized via enqueue so load replay and prompt processing don't interleave
• edit it to add new test scenarios

Debug scripts

scripts/run-rook.sh

Launch the server and/or clients:

./scripts/run-rook.sh mac server     # mac client + server (most common)
./scripts/run-rook.sh server         # server only
./scripts/run-rook.sh stop           # kill everything

scripts/interact-with-remote-agent.sh

Exercise the remote-agent bridge without any UI. Needs server/ deps installed.

scripts/print-environments.sh

Dump active/recent environment state from the server:

./scripts/print-environments.sh
./scripts/print-environments.sh --raw    # full JSON, no jq formatting

Uses GET /api/diagnostics/environments. Useful for inspecting what environments the server knows about, their status (active/recent), and bundles.

Environment paths

Environment bundles live in two places:

• Repo: environment-repository/<kind>/<path>/.bundles/<bundle-id>/ — checked-in bundles (skills, AGENTS.md, tools)
• User-local: ~/.rook/environment-repository/<kind>/<path>/.bundles/<bundle-id>/ — bundles Rook writes at runtime (e.g. agent-authored skills, memories)

For the full filesystem shape and authoring model, see:

• PRODUCT/environment-repository.md
• PRODUCT/environment-local-authoring.md

When debugging missing skills or instructions, check both locations. The user-local path is where --join will pick up agent-authored artifacts.

scripts/dump-environment-decisions.sh

Dump the SQLite environment_decisions table:

./scripts/dump-environment-decisions.sh

Shows all approve/reject/accept/ignore decisions keyed by bundle hash.

scripts/tail-logs.sh

Tail Pi provider-payload traces:

./scripts/tail-logs.sh
./scripts/tail-logs.sh --instructions    # pretty-print system instructions
./scripts/tail-logs.sh --tools           # YAML-style tool definitions

Reads from .var/pi-traces.jsonl. Shows the raw prompts sent to the LLM — useful for debugging prompt construction and environment context injection.

Inspecting agent instructions

To see exactly what system prompt, skills, and context the agent is receiving:

# Start tailing Pi traces in one terminal:
./scripts/tail-logs.sh --instructions

# In another terminal, trigger a prompt:
rook exec --runtime MyPiOpenAiAgent --auth-token "$ROOK_AUTH_TOKEN" "hi"

The trace logs show the full provider payload including system instructions, skill content, tool definitions, and environment context — everything the agent sees.

scripts/run-tests.sh

Run the known server, Swift package, iPhone, and macOS test/build checks all at once.

Mac client debugging with Codex

The mac client's sessions list does not auto-refresh. Restart to see new sessions:

./scripts/run-rook.sh mac server

Then use Codex — always specify the full app path (multiple builds share the same bundle ID):

codex exec "Use computer use. Interact with the Rook app at /Users/johnberryman/projects/github/rookkeeper/rook/.var/run-rook/build/Rook/Build/Products/Debug/Rook.app. [instruction]" 2>/dev/null

Examples:

codex exec "Use computer use. Interact with the Rook app at .../.var/run-rook/build/Rook/...Rook.app. Tell me what screen it's on." 2>/dev/null
codex exec "Use computer use. Interact with the Rook app at .../.var/run-rook/build/Rook/...Rook.app. Click the session named 'my-test' and report what you see." 2>/dev/null
codex exec "Use computer use. Interact with the Rook app at .../.var/run-rook/build/Rook/...Rook.app. Type 'hi' into the chat input and press enter. Report what happens." 2>/dev/null

Key bits: codex exec for one-shot, "Use computer use." first, always full app path, click sessions by name not position, 2>/dev/null hides banner.

Full replay debug workflow

1. Create a named session:

   rook exec --runtime MyPiOpenAiAgent --auth-token "$ROOK_AUTH_TOKEN" --title "replay-test" "ls the directory"
   

2. Verify the transcript:

   rook --transcript --sessionId <id> --auth-token "$ROOK_AUTH_TOKEN"
   

3. Restart the mac app:

   ./scripts/run-rook.sh mac server
   

4. Codex clicks the session by name:

   codex exec "Use computer use. Interact with the Rook app at .../.var/run-rook/build/Rook/...Rook.app. Click the session named 'replay-test'. Describe every message in order." 2>/dev/null
   

5. Compare CLI transcript with Codex's report — they should match.

Common replay bugs

• blocks must be cleared before session/load, not after — otherwise the runtime's replay events get wiped
• user/assistant/thinking/tool events during replay need separate buffering from active-turn streaming
• isRunning must stay false during replay so the status dot doesn't glow