fix(cli): mark synthetic SDK user messages as meta

[hqhq1025]

Problem

In remote mode, invoking a skill renders the entire skill document as a user message bubble in the web UI, as if the human had pasted it into the chat. The tool card itself renders correctly; the injected body shows up right after it.

The cause is a field-name mismatch between the two transcript sources, and a converter that drops the flag.

Claude Code injects its own user-role turns (skill bodies, compact continuation summaries). The on-disk JSONL flags them isMeta:

{"type":"user","isMeta":true,"sourceToolUseID":"toolu_01…",
 "message":{"role":"user","content":[{"type":"text","text":"Base directory for this skill: …"}]}}

The stream-json the SDK path consumes flags the same event isSynthetic instead — captured from a real session:

{"type":"user","message":{"role":"user","content":[{"type":"text",
  "text":"Base directory for this skill: /home/…/skills/brainstorming"}]},
 "parent_tool_use_id":null,"isSynthetic":true}

sdkToLogConverter's case 'user' copied only message, so the flag was silently discarded and the emitted line had neither isMeta nor isSynthetic. Every downstream guard then let it through:

• OutgoingMessageQueue.ts:124 — checks !logMessage.isMeta, which is undefined, so it forwards.
• apiSession.ts:88 — isExternalUserMessage checks isMeta, then falls back to a prefix allowlist (<system-reminder>, <task-notification>, …). A skill body is bare markdown starting with Base directory for this skill:, so it matches nothing and is classified as genuine human input.
• The web UI receives role:'user' and renders a user bubble.

This is why local mode is unaffected: claudeLocalLauncher.ts:29 sees isMeta on the JSONL and drops the message, covered by claudeLocalLauncher.test.ts:167 ('filters out isMeta messages (e.g. skill injections)'). The SDK path simply never had an equivalent.

Fix

Normalize isSynthetic (and a passthrough isMeta) onto the converted line, restoring parity with the transcript path. No new filtering logic is introduced — the three existing isMeta guards start working as they were already written to.

This also covers compact continuation summaries, which carry the same isSynthetic flag over stream-json and were leaking through remote mode for the same reason.

Tests

Three cases added to sdkToLogConverter.test.ts: isSynthetic maps to isMeta, an explicit isMeta survives, and a genuine user message stays unflagged.

• bun run typecheck in cli/ passes.
• bun test in cli/: 1054 pass / 110 fail. The same 110 fail on main at this commit (1051 pass there — the delta is exactly the three new tests), so no regression.

Follow-up (not in this PR)

web/src/chat/normalizeAgent.ts:360-376 turns any non-sidechain user message whose content is entirely text blocks into role:'user'. That branch is the backstop this bug slipped past, and its stated premise — "isExternalUserMessage rejects array content" — is not accurate, since extractRawUserTextContent handles arrays. It has no test coverage. Worth tightening separately; with this fix the message no longer reaches the web layer at all.

[github-actions]

Review mode: initial

No findings. The change is narrowly scoped and the new tests cover the intended flag normalization. Residual risk: I could not run bun test here because bun is not available in this environment, so the added unit tests were not executed locally.

HAPI Bot