Deliver launch context out-of-band via the system-prompt channel

[jamesrochabrun]

Summary

Curated launch context (the Context tab / context sets feature from #448) was prepended to the first user message — pasted into the Claude TUI, or passed as Codex's argv prompt. That polluted the visible first message, the transcript, and session previews.

Context now rides each provider's purpose-built background channel instead, folded into the same value as the existing simulator guidance in EmbeddedTerminalLaunchBuilder:

• Claude: --append-system-prompt "<preamble>\n<context>…"
• Codex: -c developer_instructions="<preamble>\n<context>…"

The first message carries only what the user actually typed; a context-only launch types nothing into the terminal. A one-line preamble gives the agent provenance ("the user attached this curated context when launching this session"). ContextAssembler / ContextPayloadStore are unchanged — blocks over 48 KB still spill to a payload file with a short reference prompt, which is what rides the flag.

Resume persistence

System-prompt text is per-invocation, not conversation history, so a resume that drops the flag silently loses the context. Both resume branches (claude -r, codex resume) now re-pass it:

• New table session_launch_context (migration v17_create_session_launch_context) records the exact context text when a pending session resolves to its real session id; rows are pruned after 30 days on save.
• cliLaunch reads the persisted text via a sync accessor (getSessionLaunchContextTextSync, mirroring getAIConfigSync) on resume launches.
• Side effect: simulator guidance now also re-applies on resume, which is consistent — the XcodeBuildMCP config it depends on was already passed on resume.

Monitoring stays clean by construction

• The Codex parser only surfaces event_msg/user_message events as user content and ignores message-type response items, so the injected developer-role instructions never appear in monitoring. Verified against the parser, not assumed.
• Claude's appended system prompt is not a JSONL message, so nothing new shows up there either.
• Session card previews and auto-naming now see the user's real first message instead of the context blob.

Also fixed

AgentHubGhosttyTerminalSurface.restart() re-configured with metadataStore: nil and unconditionally overwrote the stored reference, dropping AI-config (and now context) lookups after a restart. It now keeps the stored reference when nil is passed, mirroring TerminalContainerView.

Test plan

• New: LaunchContextRoutingTests (context never merges into the prompt, both providers; whitespace-only context is byte-identical to no context), CombinedAppendSystemPromptTests, EmbeddedTerminalLaunchBuilderLaunchContextTests (flag placement + resume round-trip through a real temp SQLite store), SessionLaunchContextStoreTests (round-trip, upsert, retention pruning, and a v17 migration-preservation test seeded from the v16 baseline).
• Updated: CLICommandConfigurationArgumentHandlingTests resume assertions (resume now re-passes the system prompt on both providers).
• All targeted suites pass; Ghostty builds clean via xcodebuild.
• Full ./scripts/test.sh gate: the AgentHubCore suite fails ~20 known load-dependent flaky tests (workspace detection, worktree progress, CI notifier, Monitor queue — the documented clean-tree flake set). Verified by running the identical suite on a clean worktree at the parent commit: same failure families, and every test that failed only on this branch passes in isolation here (38/38). Real app session_workspace_state verified byte-identical before/after the gate.

🤖 Generated with Claude Code