# Find a conversation

Every task leaves a read-only conversation: each message, model call, and tool call, kept after the sandbox is gone.

A **conversation** is a task's transcript, and its ID is the task ID. Every task appears under **Conversations** in the dashboard, whatever started it.

**You need:** `introspection login` and a task ID, or a time window to search.

## Read one conversation

```bash
introspection conversations get <task-id>                  # full transcript and feedback
introspection conversations get <task-id> --summary-only   # totals only
```

A task that failed before the agent ran has no conversation: read `introspection tasks get <task-id>` instead.

## Filter the list

```bash
introspection conversations list --lookback 24h \
  --filter environment=production --filter status=Error --sort cost
```

Add `-o table` for one row per conversation. Common filters:

| Filter | Matches |
| --- | --- |
| `environment` | `development`, `staging`, or `production` |
| `status` | `Ok` or `Error` (a model call failed; failed tool calls alone stay `Ok`) |
| `recipe_git_commit_sha` | One release: the full commit SHA (`git rev-parse HEAD` prints it) |
| `trace_id` | The conversation containing that turn (the `?trace=` in a turn's dashboard URL) |
| `metadata` | `key:value` from the task's `conversation_metadata`; repeat to AND |

All filters and sort keys

| Filter | Matches |
| --- | --- |
| `runtime_group_id` | One runtime (`runtimes get <version-id>` shows it) |
| `runtime_id` | One version |
| `experiment_id` | One experiment |
| `model`, `agent_name` | A model or recipe agent used in the conversation |
| `owner_key` | One owner: `user:<id>`, `anon:<id>`, or `member:<id>` |
| `resolution` | `resolved`, `blocked`, `unresolved` (latest task-resolution observation is `abandoned`, or none exists), or `pending` (not yet observed) |
| `sentiment` | Platform-assessed user sentiment: `positive`, `negative`, `mixed`, `neutral` |

`--sort` takes `created`, `duration`, `turns`, `tool_calls`, `llm_calls`, `tokens`, or `cost`. Window flags: `--lookback 7d`, or `--start`/`--end`.

## See what one release did

Filter by the full commit SHA:

```bash
introspection conversations list --lookback 7d --filter recipe_git_commit_sha=<commit-sha>
```

To compare releases, group by `recipe_git_commit_sha` in [Metrics](/observe/metrics).

## Export

One conversation, in `json`, `arrow`, or `trajectory` format:

```bash
introspection conversations export <task-id> --format json --output-file conv.json
```

Several at once, one JSON line each. List the IDs first, then pass them:

```bash
introspection conversations list --lookback 30d --filter environment=production --query "[].id" -o table
introspection conversations get <id-1> <id-2> <id-3> --output-file conversations.jsonl
```

Or put up to 20 IDs, one per line, in a file and pass `--ids-file ids.txt`.

## Verify

`conversations get <task-id>` prints the turns you expect. A `completed` task is not necessarily a success: it can complete on an inactivity timeout. Check that `tasks get` shows `completion_reason: task_completion`.

## Limits

| Not supported | Do this instead |
| --- | --- |
| Searching message text on the server | Export, then search locally. |
| More than 20 IDs in one `--ids-file` | Pass the rest as arguments, or run `conversations get` again. |
| Filtering by end-user feedback | `sentiment` is the platform's own assessment. See [Read feedback](/observe/patterns#read-feedback). |
| Misspelled filter keys | The server ignores unknown keys, so an empty list can mean a typo. |

Or ask your agent: "Show production conversations with failed tool calls in the last 24 hours."

Find a conversation

Every task leaves a read-only conversation: each message, model call, and tool call, kept after the sandbox is gone.

A conversation is a task’s transcript, and its ID is the task ID. Every task appears under Conversations in the dashboard, whatever started it.

You need: introspection login and a task ID, or a time window to search.

Read one conversation


introspection conversations get <task-id>                  # full transcript and feedback
introspection conversations get <task-id> --summary-only   # totals only

A task that failed before the agent ran has no conversation: read introspection tasks get <task-id> instead.

Filter the list


introspection conversations list --lookback 24h \
  --filter environment=production --filter status=Error --sort cost

Add -o table for one row per conversation. Common filters:

Filter	Matches
`environment`	`development`, `staging`, or `production`
`status`	`Ok` or `Error` (a model call failed; failed tool calls alone stay `Ok`)
`recipe_git_commit_sha`	One release: the full commit SHA (`git rev-parse HEAD` prints it)
`trace_id`	The conversation containing that turn (the `?trace=` in a turn’s dashboard URL)
`metadata`	`key:value` from the task’s `conversation_metadata`; repeat to AND

All filters and sort keys

Filter	Matches
`runtime_group_id`	One runtime (`runtimes get <version-id>` shows it)
`runtime_id`	One version
`experiment_id`	One experiment
`model`, `agent_name`	A model or recipe agent used in the conversation
`owner_key`	One owner: `user:<id>`, `anon:<id>`, or `member:<id>`
`resolution`	`resolved`, `blocked`, `unresolved` (latest task-resolution observation is `abandoned`, or none exists), or `pending` (not yet observed)
`sentiment`	Platform-assessed user sentiment: `positive`, `negative`, `mixed`, `neutral`

--sort takes created, duration, turns, tool_calls, llm_calls, tokens, or cost. Window flags: --lookback 7d, or --start/--end.

See what one release did

Filter by the full commit SHA:


introspection conversations list --lookback 7d --filter recipe_git_commit_sha=<commit-sha>

To compare releases, group by recipe_git_commit_sha in Metrics.

Export

One conversation, in json, arrow, or trajectory format:


introspection conversations export <task-id> --format json --output-file conv.json

Several at once, one JSON line each. List the IDs first, then pass them:


introspection conversations list --lookback 30d --filter environment=production --query "[].id" -o table
introspection conversations get <id-1> <id-2> <id-3> --output-file conversations.jsonl

Or put up to 20 IDs, one per line, in a file and pass --ids-file ids.txt.

Verify

conversations get <task-id> prints the turns you expect. A completed task is not necessarily a success: it can complete on an inactivity timeout. Check that tasks get shows completion_reason: task_completion.

Limits

Not supported	Do this instead
Searching message text on the server	Export, then search locally.
More than 20 IDs in one `--ids-file`	Pass the rest as arguments, or run `conversations get` again.
Filtering by end-user feedback	`sentiment` is the platform’s own assessment. See Read feedback.
Misspelled filter keys	The server ignores unknown keys, so an empty list can mean a typo.

Or ask your agent: “Show production conversations with failed tool calls in the last 24 hours.”