LLP 0265: the OpenClaw sweep reads the run trajectory for system_text and tools

[bgmcmullen]

OpenClaw rows were the thinnest in ai_gateway_messages. On the machine this was developed against: 73 rows, cwd on all of them, model on half, and system_text / tools null on every one. The session transcript records the conversation and nothing about the run that produced it; the run trajectory OpenClaw already writes beside every session does.

What changed

• New reader openclaw/src/trajectory_file.js, the one reader of <sessionId>.trajectory.jsonl (resolved through OpenClaw's .trajectory-path.json pointer when present). It takes exactly two facts from context.compiled, the compiled system prompt and the tool definitions, and leaves the rest of the stream on disk.
• Per-message stamping. AiGatewayProjectedMessage gains system_text and tools, and the row builder prefers them over the exchange-level pair, the precedence model (LLP 0026) and provider (LLP 0194) already had. An OpenClaw session recompiles per run: one observed session ran with 23, then 21, then 26 tools under three different system prompts, so an exchange-level value would present one turn's answer as every turn's. No live projector sets either field, so live rows are unchanged.
• Run windows. A run's window opens at its context.compiled and closes at its session.ended; a message outside every window is stamped with nothing, because a run that failed before compiling (or one an embedded CLI harness owned) is not described by whatever the previous run compiled.
• Recording time, not message time. OpenclawSessionMessage now surfaces recordedAtMs (the record line's timestamp) alongside timestampMs (the envelope's, still what message_created_at uses). A webchat prompt carries the moment the user sent it, seconds before its own run compiled.

The truncation problem

OpenClaw truncates a recorded system prompt at two caps and announces only one:

• past 32768 characters (TRAJECTORY_RUNTIME_DATA_STRING_MAX_CHARS, hard-coded) the value is replaced by a { truncated: true, originalChars, limitChars } stub;
• below that, a long prompt is still clipped to 20000 characters plus an ellipsis and written as an ordinary string, with no marker.

Truncation is therefore detected rather than assumed absent: the stub, a recorded string shorter than the chars the run's own trace.metadata reports, or the trailing ellipsis. Clipped text is kept (a 20000-char prefix is useful) and flagged; attributes.openclaw.system_prompt carries chars, hash, recorded_chars, and truncated.

Across the 33 verified runs, no real agent run recorded a complete system prompt (24 stubbed at 36k-41k, 9 clipped at exactly 20001 chars, 2 complete two-line probes). Complete prompt text still depends on the wire, so LLP 0175 remains the path to it.

Effect on real data (73 rows)

	before	after
tools	0	71
system_text, complete	0	1
system_text, clipped and flagged	0	20
digest only (prompt over 32k)	0	50

Enrichment applies from the next import onward: backfill dedupe skips rows already committed under the same part_id, the same boundary LLP 0194 accepted.

Docs

LLP 0265 is new. LLP 0171 gains an Extended-by and a narrowed non-goal: the trajectory-file half is lifted for these two columns only; the probe-session stream and the live tail stay out.

Testing

• test/plugins/openclaw-trajectory.test.js (new, 12 cases): pointer resolution and its refusals, per-run contexts, both truncation caps, the report cross-check, window matching.
• test/plugins/openclaw-backfill.test.js (+8 cases): end-to-end stamping through the real materializer, two runs in one session, the recording-time rule, both truncation states, untraced sessions, and the run's log counters.
• npm run smoke -- backfill_openclaw_session_shape now stages a trajectory and asserts both columns through the query path. Its existing two-level timestamp disagreement doubles as the recording-time proof: the assistant envelope's 10:00:09 falls outside the run window, its record line's 10:00:02 falls inside.
• npm test at the 28 pre-existing failures (unchanged), npm run typecheck at its 3 pre-existing errors, backfill_openclaw_fixture green.

🤖 Generated with Claude Code