Trajectory bundles

Trajectory capture is OpenClaw's per-session flight recorder. It records a structured timeline for each agent run, then /export-trajectory packages the current session into a redacted support bundle covering:

• The prompt, system prompt, and tools sent to the model
• Which transcript messages and tool calls led to an answer
• Whether the run timed out, aborted, compacted, or hit a provider error
• Which model, plugins, skills, and runtime settings were active
• Usage and prompt-cache metadata the provider returned

For a broad Gateway support report, start with /diagnostics instead; it collects the sanitized Gateway bundle and, for OpenAI Codex harness sessions, can send Codex feedback to OpenAI after approval. Use /export-trajectory when you need the detailed per-session prompt, tool, and transcript timeline.

Quick start

Send in the active session (alias /trajectory):

/export-trajectory

OpenClaw writes the bundle under the workspace:

.openclaw/trajectory-exports/openclaw-trajectory-<session>-<timestamp>/

Pass a relative output directory name to override it:

/export-trajectory bug-1234

The name resolves inside .openclaw/trajectory-exports/. Absolute paths and ~ paths are rejected.

Trajectory bundles can contain prompts, model messages, tool schemas, tool results, runtime events, and local paths, so the chat command always runs through exec approval. Approve the export once when you intend to create the bundle; do not use allow-all. In group chats, OpenClaw sends the approval prompt and export result to the owner privately instead of posting trajectory details back to the shared room. The room receives only a status notice that distinguishes confirmed, pending, and suppressed private delivery.

For local inspection or support workflows, run the underlying CLI command directly:

openclaw sessions export-trajectory --session-key "agent:main:telegram:direct:123" --workspace .

Other flags: --output <path> (directory name inside .openclaw/trajectory-exports), --store <path> (session store override), --agent <id> (agent id for store resolution), --json (structured output).

Access

Trajectory export is an owner command. The sender must pass the normal command authorization checks plus the owner check for the channel.

What gets recorded

Trajectory capture is on by default for OpenClaw agent runs.

Runtime events include:

• session.started
• trace.metadata
• context.compiled
• prompt.submitted
• tool.call, with the tool identity and sanitized arguments
• tool.result, with the sanitized result and execution outcome
• model.fallback_step, including the source model, next model, failure reason/detail, chain position, and whether the chain advanced, succeeded, or was exhausted
• model.completed
• trace.artifacts
• session.ended

Native tool results include success only for a settled outcome. A tool that is still running or has an unknown outcome omits success; it is not recorded as a successful completion.

Transcript events are reconstructed from the active session branch: user messages, assistant messages, tool calls, tool results, compactions, model changes, labels, and custom session entries.

Events are written as JSON Lines with this schema marker:

{  "traceSchema": "openclaw-trajectory",  "schemaVersion": 1}

Bundle files

manifest.json lists the files present in a given bundle; some files are omitted when the session did not capture the corresponding runtime data.

Capture storage

Runtime trajectory events are stored with the session in the per-agent SQLite database. Runtime batches carry the captured JSON bytes to the database worker for persistence. Normal turns keep their events together until cleanup drains the final batch. Long turns also flush in the background at 32 pending events or when the pending queue overflows the session's rolling byte limit. A process crash before a flush can lose the pending diagnostic events. Queued events follow the same rolling byte limit while the database writer is busy. Exporting a trajectory materializes a redacted JSONL support bundle; the live runtime capture is not a session-adjacent JSONL sidecar.

Each session keeps a rolling 10 MiB runtime window. Global retention removes complete runs older than 14 days, then the oldest remaining runs to target a 512 MiB total. The session triggering cleanup is protected by its own rolling window. Global cleanup begins on first use and hourly thereafter, separately from appends. It selects bounded batches without holding the database writer and rechecks the selection before each deletion. Cleanup continues through bounded batches; concurrent changes defer a stale selection until a later append.

Runtime export reads the SQLite capture. Pre-July 2026 .trajectory.jsonl and .trajectory-path.json sidecars are no longer read or followed. Existing files can be removed after retaining any needed diagnostic data.

Disable capture

export OPENCLAW_TRAJECTORY=0

This disables runtime trajectory capture before starting OpenClaw. /export-trajectory can still export the transcript branch, but runtime-only data such as compiled context, provider artifacts, and prompt metadata may be missing.

Tune flush timeout

OpenClaw drains any pending runtime trajectory rows during agent cleanup. The default cleanup timeout is 10,000 ms. On slow disks or large stores, set OPENCLAW_TRAJECTORY_FLUSH_TIMEOUT_MS before starting OpenClaw:

export OPENCLAW_TRAJECTORY_FLUSH_TIMEOUT_MS=30000

This controls when OpenClaw logs an openclaw-trajectory-flush timeout and continues; it does not change the trajectory size caps. To tune all agent cleanup steps that do not pass an explicit timeout, set OPENCLAW_AGENT_CLEANUP_TIMEOUT_MS.

Privacy and limits

Trajectory bundles are for support and debugging, not public posting. OpenClaw redacts sensitive values before writing export files:

• credentials and known secret-like payload fields
• image data
• local state paths
• workspace paths, replaced with $WORKSPACE_DIR
• home directory paths, where detected

The exporter also bounds input size:

• runtime capture: the live capture is a rolling window capped at 10 MiB, dropping the oldest events to make room for new ones; export accepts the SQLite runtime store up to 50 MiB
• session (transcript) sources: 50 MiB, whether read from the SQLite transcript store or a session file
• runtime events per export: 200,000
• total exported events: 250,000
• individual runtime event lines are truncated above 256 KiB

Runtime capture preserves structured results within its field limits, including explicit truncation markers. If a tool result exceeds the event size limit, capture keeps its tool identity and outcome while dropping oversized payload fields.

The source byte limits are checked before parsing events. SQLite sources use UTF-8 JSONL size, including separators between rows, regardless of the database encoding. SQLite runtime event counts are also checked before loading rows. Oversized sources fail with a size error before an export bundle is created.

These are per-source limits, not a bound on total process memory. Parsing, projection, redaction, and output serialization can retain additional copies; large exports can still require more memory than their source size.

Review bundles before sharing them outside your team. Redaction is best-effort and cannot know every application-specific secret.

Troubleshooting

If the export has no runtime events:

• confirm OpenClaw was started without OPENCLAW_TRAJECTORY=0
• run another message in the session, then export again
• inspect manifest.json for runtimeEventCount

If the command rejects the output path:

• use a relative name like bug-1234
• do not pass /tmp/... or ~/...
• keep the export inside .openclaw/trajectory-exports/

If the export fails with a size error, a transcript or runtime source — the SQLite store or a file — exceeded the export safety limits above. Start a new session or export a smaller reproduction.

Related

• Diffs
• Session management
• Exec tool
• Codex harness runtime