| name | openagent-dev-chat |
| description | Drive and diagnose the currently running OpenAgent Tauri debug instance through its authenticated loopback API. Use for agent-driven chat or goal runs, checkpoint/compaction diagnosis, tool approval, ask-user recovery, and low-noise debug traces from `bun tauri dev`. |
OpenAgent Dev Chat
Use scripts/openagent_dev.py; it discovers the local API and forces UTF-8
terminal output, including on a Windows GBK console. It defaults to compact JSON
and never returns complete prompts, tool schemas, or historical snapshots unless
you explicitly call the raw traces or checkpoints commands.
Minimal recovery loop
python scripts/openagent_dev.py health
python scripts/openagent_dev.py chat "Inspect the failing checkpoint flow"
python scripts/openagent_dev.py interrupts <conv-id>
python scripts/openagent_dev.py approve <conv-id> <interrupt-id>
python scripts/openagent_dev.py conversation <conv-id>
python scripts/openagent_dev.py diagnose <conv-id>
Trigger pre-flight context compaction
Use the dedicated diagnostic instead of manually fabricating a large
checkpoint. It reads the running application's configured compaction threshold,
creates an isolated conversation containing random user and assistant text at
that size, and then submits a normal user message through the product chat path:
python scripts/openagent_dev.py context-compaction
The result reports configured_threshold_tokens, seeded_estimated_tokens,
and the conversation ID. A successful request should contain a durable
compaction boundary; verify it with renderability <conv-id> and diagnose <conv-id>. The command fails clearly when context compaction is disabled in
the running configuration.
An HTTP 200 means only that the API request completed. Inspect phase: a value
of interrupted requires recovery, while final_completed is the normal
terminal state.
Use approve and deny for a tool approval. They always submit the correct
payload ({"values":{"approved":true|false}}). For an ask_user interrupt,
first inspect its listed fields, then answer exactly once:
python scripts/openagent_dev.py answer <conv-id> <interrupt-id> --json '{"values":{"choice":"A"}}'
Do not use a bare {"approved":true} value: it is invalid and rejected.
Replay a checkpoint snapshot
For an intentional recovery/replay experiment, the development API can replace
one checkpoint's complete CheckpointData payload. It is a durable write, so
first save the original with checkpoints <conv-id> --full, edit only the
target payload, and repeat the checkpoint ID as an explicit confirmation:
python scripts/openagent_dev.py replace-checkpoint <conv-id> <checkpoint-id> --data-file .\checkpoint-data.json --confirm-checkpoint <checkpoint-id>
The result includes renderability.errors and renderability.warnings; do not
claim the replay is valid if errors are present. If the edited checkpoint is
the active tip, the runtime snapshot cache is refreshed, so interrupts,
answer, approve, deny, and the next chat request exercise the replacement
state. An interrupted ask_user snapshot must retain its unresolved tool call;
inspect it with interrupts and answer it once with answer before assessing
the resumed trace.
Output and sensitive context
Use --compact (the default) for one-line agent output, --json for formatted
JSON, or --text for a short human view. traces and checkpoints are raw,
high-volume Inspector data; call them only after conversation or diagnose
identifies a concrete need. Never copy their complete prompts, tool schemas, or
trace payloads to an issue, commit message, or final response.
