Skip to main content

​
crewship inspect

Print the journal entries for a run as a one-line-per-event timeline. Each line shows timestamp, severity chip, entry type, and summary. Costs and tool calls are pulled out into a footer block when present. Sibling design with crewship explain:
  • crewship inspect <run-id> — structured timeline (this command, fast, no LLM)
  • crewship explain <run-id> — LLM-summarised narrative (slower, costs tokens)
Both share fetchRun plus the journal endpoint; explain feeds the entries to an agent, inspect formats them directly. 
crewship inspect <run-id> [flags]
Requires an authenticated session and a selected workspace (requireAuthAndWorkspace).

​
Flags

FlagDefaultPurpose
--types <list>"" (all)Comma-separated journal entry types to include — e.g. error,exec.error,keeper.decision.
--filter <jq-expr>""Pipe JSON output through jq <expr> (requires jq in $PATH).
--watch <duration>""Auto-refresh every duration (e.g. 5s, 1m). Minimum 1s.
The global --format flag (table / json / yaml / ndjson / quiet) is honoured. JSON / YAML / --filter paths bypass the table formatter and emit the raw entry list with run_id, agent_id, and entries.

​
Examples

​
Default (table view)

crewship inspect r_abc
# Run r_abc · agent code-review  (42 entries)
# ────────────────────────────────────────────────────────────────────────────────
# 14:02:11  [        ]  exec.start              spawning claude --print
# 14:02:12  [        ]  llm.request             prompt 1842 tokens
# 14:02:14  [        ]  tool_call               read_file: src/api/router.go
# 14:02:15  [notice  ]  cost.incurred           $0.0142 (1842 prompt + 280 output)
# 14:02:18  [error   ]  exec.error              tool exit 1: linter failed
# ────────────────────────────────────────────────────────────────────────────────
# total cost: $0.0142   tool calls: 3   errors: 1

​
Narrow to errors and keeper decisions

crewship inspect r_abc --types error,exec.error,keeper.decision
--types is forwarded to the server as the entry_type query param on /api/v1/journal, so filtering is applied DB-side before the entries are reversed for chronological display.

​
Pull a single field via jq

crewship inspect r_abc --filter '.entries[] | select(.severity=="error")'
--filter short-circuits to JSON output; the expression is piped through jq.

​
Raw JSON for piping

crewship inspect r_abc --format json | jq '.entries[] | .entry_type' | sort | uniq -c

​
Live tail

crewship inspect r_abc --watch 5s
Re-runs the inspect every 5 seconds. Useful for runs that are still streaming events.

​
How the run is scoped

The CLI fetches journal entries scoped to the run’s agent and the run’s time window, then filters down to entries whose trace_id matches the run ID. By convention trace_id == run_id for journal entries (see internal/journal/types.go). Older entries that pre-date the trace rollout, or non-run-scoped events that the user explicitly --types-included, have no trace_id and are kept on the assumption the user wants to see them.

​
Common errors

  • run <id> has no agent_id — the run record is malformed (or you passed something that isn’t a run ID). Run IDs start with r_.
  • no entries in journal window — run may be older than recall horizon — the run is older than the journal retention window (typically 7d). Printed dimmed under the header instead of a hard error.

​
See also