docs(lineage): live Claude Code 2.1.259 orchestration capture (parallel, sequential, nested spawns; resumed turns; /compact)

.changeset/claude-orchestration-lineage-evidence.md

@@ -0,0 +1,5 @@
+---
+"agent-bundle": patch
+---
+
+Extend the Claude host capability table (`claude-2.1.250.json`, rendered on the hosts reference page) with live Claude Code 2.1.259 evidence on the `lineage.subagent-events`, `lineage.root`, `lineage.parent`, `lineage.depth`, and `lineage.mcp-correlation` rows: two `Agent` spawns issued in one message bind to their own spawn calls, `Explore` subagents carry the same lineage fields as `general-purpose` ones, `PostToolUseFailure` carries the subagent's `agent_id`, the root `session_id` survives `--resume` and `/compact`, and `claudecode/toolUseId` correlates MCP calls at depth 0, 1, and 2. Emitted host output and `adapterRevision` are unchanged. (#455)

docs/audits/2026-09-03-claude-live-session-proofs.md

@@ -102,6 +102,25 @@ is reproduced here beyond counts, codes, and field names.
    the `native-host-smoke` workflow's Claude source leg will report
    `harness-failure` on every run against 2.1.257.
 
+## Re-run later the same day: unmodified harness, Claude Code 2.1.259
+
+After the operator signed in again, `probe:install claude` — the checked-in
+harness, byte-for-byte copy of `~/.claude/.credentials.json`, no local edits —
+produced a signed-in isolated home at `/tmp/host-test/claude-home`
+(`claude auth status` → `loggedIn: true`, `authMethod: claude.ai`). The three
+proofs were re-run with `HOME=/tmp/host-test/claude-home` (raw logs under
+`/tmp/claude-live/orch/`), right after that home had hosted the four-turn
+orchestration capture (`docs/audits/2026-09-03-host-lineage-matrix.md`,
+"orchestration run"):
+
+| Proof | Result | Assertions that held |
+| --- | --- | --- |
+| `pnpm test:host-install:session:claude` | **pass** — 2/2, 0 skipped | the same assertions as above against Claude Code 2.1.259 |
+| `pnpm test:packed:native:claude` | **pass** — 8/8, 0 skipped, no fix needed this time (defects 1–2 are merged) | `hosts = ['claude']`, `status: 'passed'`, Eval `pass: 1, fail: 0` |
+| `AGENT_BUNDLE_NATIVE_CLAUDE_SMOKE=1 rstest …/native-claude-contract.test.ts` | **fail** — 17/18, `status: 'harness-failure'` | everything measured held again (`authentication.status: 'subscription-session'`, `version: '2.1.259'`, `validation.exitCode: 0`, `stream.activationEvidence: 'observed'`, `stream.authSource: 'non-environment'`, `stream.plugins: ['agent-bundle-native-smoke']`, `errorEnvelopes: []`, one `result`/`success`, `stderr.present: false`); sole diagnostic `claude-native.normal-home.changed` — defect 3, still open as [#439](https://github.com/ScriptedAlchemy/agent-bundle/issues/439), reproduced on 2.1.259 |
+
+The real `~/.claude` was again never written by any of these runs.
+
 ## Cost and footprint
 
 The two lineage captures reported `total_cost_usd` of roughly $0.44 and
@@ -112,4 +131,6 @@ isolated home used here was under `/tmp/claude-recapture/` or
 claude` removed the probe's home (`/tmp/host-test/claude-home` and
 `/tmp/host-test/claude-workspace` are gone; the captures under
 `/tmp/host-test/claude/` remain) and `rm -rf /tmp/claude-recapture/proof-home`
-removed the proofs' home.
+removed the proofs' home. The 2.1.259 orchestration capture cost about $1.23
+across its four turns (`total_cost_usd` 0.79 + 0.14 + 0.24 + 0.07); its
+isolated home was removed by `probe:uninstall claude` after the re-run above.

examples/host-test/README.md

@@ -36,6 +36,9 @@ Two MCP servers ship in the plugin:
 - `host-test` (generated routes, `src/mcp/host-test/tools/`): `dump` (filter by
   any conversation/session/subagent id, `full` for raw lines) and `reset`. Each
   `dump` call records the request context the generated server mounted for it.
+  A bare `dump` returns the newest 50 matching records — a whole log of a few
+  hundred records overflows the tool-result document — so pass `limit` (up to
+  5000) for more; `matched` and `total` always count the whole log.
 - `host-test-raw` (hand-rolled stdio factory, `src/mcp/host-test-raw.ts`):
   `probe` records the raw SDK request context — session id, JSON-RPC id,
   `_meta`, lifted envelope, negotiated client info — so hook↔MCP correlation is
@@ -77,6 +80,29 @@ pnpm --filter @agent-bundle-example/host-test probe:uninstall claude
   land in `/tmp/host-test/<host>/`, followed by a rendered `host-test dump`.
   Earlier runs' records are never re-copied, and the command exits non-zero
   when the host fails or when the run produced no hook record or no MCP record.
+- `probe:capture <host> --scenario <file.json>` replaces the default prompt with
+  an ordered list of turns (`{ "turns": ["...", "..."] }`; a turn may also be
+  `{ "prompt": "..." }`). For Claude every turn after the first runs
+  `claude -p --resume <session_id>` against the session the first turn's
+  `system/init` envelope reported, so one capture holds a multi-turn session
+  with `SessionStart source: resume` and one `SessionEnd` per invocation.
+  `Stop` follows the prompt path rather than the invocation: a turn whose
+  background subagents finish re-prompts itself with a `<task-notification>`
+  and stops twice, while a `/compact` turn submits no prompt and never stops
+  (`fixtures/host-lineage/claude-2.1.259-orchestration.ndjson`). If the first
+  turn reports no `session_id`, the capture fails instead of running the
+  remaining turns as fresh sessions; `--prompt` and `--scenario` refuse each
+  other, and `--scripted-model` plays a fixed transcript and refuses both.
+  Codex and Cursor drivers take
+  the first turn only and refuse longer scenarios. `scenarios/claude-orchestration.json` is
+  the checked-in orchestration scenario (two parallel `Agent` spawns, one
+  sequential spawn that nests another, the `host-test:host-test` skill, a
+  plugin-command probe, a manual `/compact`, a final stop).
+- Claude turns run with `--output-format stream-json --verbose`, so the model's
+  own tool-use stream is saved next to the hook payloads as
+  `session-<stamp>[.turn-N].stream.ndjson`; envelopes produced inside a
+  subagent carry `parent_tool_use_id` (the parent's `Agent` `tool_use_id`),
+  which is how the transcript and the hook log are cross-checked.
 - Host processes get an allowlisted environment (PATH, locale, display, proxy,
   TLS plumbing) plus the isolated `HOME`; nothing else from your shell is
   inherited, and even allowlisted values are dropped when they carry a

examples/host-test/scenarios/claude-orchestration.json

@@ -0,0 +1,10 @@
+{
+  "name": "claude-orchestration",
+  "description": "Multi-turn Claude Code session that exercises the host's own orchestration surfaces: two parallel Agent spawns, one sequential spawn that nests another, the installed skill, a plugin-command probe, a manual /compact, and a final stop. Each turn after the first resumes the same session.",
+  "turns": [
+    "You are exercising the host-test probe plugin inside an isolated harness. Do exactly these steps in order, without asking questions and without stopping early.\n1. Run the shell command `pwd`.\n2. Call the `dump` tool of the host-test MCP server with {} as arguments and remember its `log.path`.\n3. In ONE assistant message, spawn TWO subagents with the Agent tool IN PARALLEL (two Agent tool calls in the same message):\n   (a) subagent_type \"Explore\", prompt: \"Run the shell command `ls -la`. Then call the host-test MCP server's `dump` tool with {} and the host-test-raw MCP server's `probe` tool with {\\\"note\\\":\\\"parallel-explore\\\"} if those tools are available to you. Reply with every id you saw and state explicitly whether the MCP tools were available to you.\"\n   (b) subagent_type \"general-purpose\", prompt: \"Run the shell command `pwd`, call the host-test MCP server's `dump` tool with {}, call the host-test-raw MCP server's `probe` tool with {\\\"note\\\":\\\"parallel-general\\\"}, then reply with every id you saw.\"\n4. After BOTH have returned, spawn ONE more subagent sequentially with the Agent tool, subagent_type \"general-purpose\", prompt: \"Run the shell command `pwd`, call the host-test MCP server's `dump` tool with {}, call the host-test-raw MCP server's `probe` tool with {\\\"note\\\":\\\"sequential\\\"}. Then spawn a nested subagent with the Agent tool (subagent_type \\\"general-purpose\\\") and instruct it to run `pwd`, call the host-test `dump` tool with {}, call the host-test-raw `probe` tool with {\\\"note\\\":\\\"nested\\\"}, and reply with every id it saw. Finally reply with every id you and the nested subagent saw.\"\n5. Reply with exactly one final line: HOST_TEST_TURN1_DONE <the log.path from step 2>",
+    "Same session, next turn. Do exactly these steps in order, without asking questions.\n1. Invoke the installed host-test skill through the Skill tool (its name is `host-test:host-test`) and follow the steps it lists, using {\"note\":\"root-skill\"} for the probe call; do NOT spawn any subagent this turn.\n2. If the host-test plugin provides any slash command or custom command other than that skill, run it once; if it provides none, write the single word NO_PLUGIN_COMMAND.\n3. Reply with exactly one final line: HOST_TEST_TURN2_DONE",
+    "/compact",
+    "Same session, final turn. Run the shell command `pwd` once more, then reply with exactly one final line: HOST_TEST_DONE <the log.path you learned earlier in this session>"
+  ]
+}