fix(server): Claude /compact and Claude-initiated turns no longer end the user's turn; subagents show their own model

[rynfar]

Problem

Two Claude adapter bugs from upstream apply to Pylon:

• After a resume, Claude can first run a turn of its own, such as a task notification for a background task the previous process left behind, and only then run the user's queued prompt. Pylon completed the user's turn on that first result. A queued /compact then ran with no turn open and the thread looked busy afterwards. Pylon's zero-turn handshake guard hides this only when the notification turn makes no model request. A notification that Claude answers ends the user's turn early.
• A subagent with its own model showed the parent session's model in the Agents panel. The correction waited for the next task row, which could be minutes away. A task launched from inside a subagent never found its owner.

Fix

• handleResultMessage ignores a result that answers a different Claude turn. A result for another turn echoes prompt uuids that do not include the active turn id, or it echoes nothing and has a non-human origin. Results from older CLIs with neither field still complete the turn. Synthetic turns are unchanged. Pylon steers send no uuid, but the SDK's user_message_uuids keeps the turn's own prompt uuid alongside any folded-in steer, so steered turns still complete.
• When a subagent's assistant snapshot changes its model, Pylon emits a metadata-only task.updated at once. A bounded tool_use_id → parent_tool_use_id map lets task_started find the owning subagent and seed a nested agent with its owner's model. Pylon's resume carry-over (previousAgent.model) still ranks above the owner's model.

The Claude adapter edits stay out of startSessionInternal, except for two lines that add the new map to the session context. They merge cleanly with fix/claude-resume-unsaved-first-turn.

Sources (upstream cycle #967)

Source	Subject	Disposition
9da066dbe906310671b86290b09df0b56a88ffa5 (#14497)	Claude /compact no longer ends early and leaves the thread busy	Adopted. The upstream test was adapted so the notification turn makes a model request, because Pylon's handshake guard already covered the num_turns: 0 form. Added a steered-turn regression test.
7ab800a43c770b2c48a2d9306c89b6177f61ad4a (#14540)	Claude subagents with their own model no longer show the parent's model	Adopted. Pylon's offerRuntimeEvent(sessionIncarnationId, …) signature and resume model carry-over were kept.
bd89c1302026255c62cc09278207bfaf2664da4a (#14486)	Grok CLIs older than 1.0.13 marked broken	Skipped. The change edits only the manifest compatibility ranges. Pylon's manifest has no compatibility section because the version-compatibility pipeline (96c4bfa0a2, d4cd7d5c33) is still held, and Pylon has no Grok version floor. Pylon's only CLI floor is the provider-local MINIMUM_CODEX_CLI_VERSION in CodexProvider.ts. Follow-up candidate: a matching provider-local Grok floor (>=1.0.13, because xAI returns 426 Upgrade Required to older CLIs) in GrokProvider.ts.

Verification

• vp test run src/provider/Layers/ClaudeAdapter.test.ts: 184 passed. Run with CLAUDE_CODE_ENABLE_TODO_TOOLS unset, since the harness host leaks it into one existing env test.
• Fail-before checks:
  • With the other-turn guard disabled, keeps a turn open past the result of a Claude-initiated turn and completes a steered turn from a result that echoes its folded-in prompts fail.
  • With the pre-#14540 adapter, task.started carries model/effort; subagent snapshots refine the model fails.
• vp run -F t3 typecheck passes. vp check on the changed files passes.
• git merge-tree with fix/claude-resume-unsaved-first-turn: clean.

Backend only, no UI evidence.

Claude Opus 5.5 in Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	14.0 KiB	14.0 KiB	+7 B (+0.0%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.3 KiB	7.3 KiB	−13 B (−0.2%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.7 KiB	6.7 KiB	+20 B (+0.3%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.0 KiB	58.0 KiB	+44 B (+0.1%)	66.4 KiB	✅
Codex	Live turn messages	9	10	+1 (+11.1%)	21	✅
Claude	Total thread wire	14.0 KiB	14.0 KiB	+17 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.3 KiB	7.3 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.7 KiB	6.8 KiB	+23 B (+0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.9 KiB	58.9 KiB	+44 B (+0.1%)	66.4 KiB	✅
Claude	Live turn messages	9	10	+1 (+11.1%)	21	✅

Baseline: f677ed4 · PR result: 7087ffa · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 115.7 KiB
• Claude decoded thread snapshot: 116.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.