test(scenarios): reproducible test scenarios harness

.github/workflows/ci.yml

@@ -50,3 +50,21 @@ jobs:
           cache: 'pnpm'
       - run: pnpm install --frozen-lockfile
       - run: bash scripts/e2e-publish.sh
+
+  # Reproducible multi-agent scenarios under tests/scenarios/. Kept out
+  # of the `pnpm test` aggregate so a scenario failure stays attributable
+  # to the harness rather than blending into the per-package test job.
+  scenarios:
+    if: github.event_name != 'pull_request' || github.event.pull_request.draft == false
+    runs-on: ubuntu-latest
+    needs: build
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+      - uses: pnpm/action-setup@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+          cache: 'pnpm'
+      - run: pnpm install --frozen-lockfile
+      - run: pnpm scenarios

README.md

@@ -1015,7 +1015,7 @@ validate.
 - 🟡 Cursor and Gemini CLI installers exist but have less smoke coverage
 - 🔵 Per-runtime smoke for claim-before-edit emission
 - 🔵 Cross-runtime handoff smoke (Codex hands off to Claude, both run)
-- ⏳ Reproducible test fixture set under `tests/scenarios/`
+- ✅ Reproducible test fixture set under `tests/scenarios/` (5 scenarios, harness self-tests, `pnpm scenarios`)
 
 > **`time-to-healthy`: still hours**, but the time the human spends _deciding what to run_ drops sharply because every signal carries its `cmd:` and `tool:` already.
 

openspec/changes/scenarios-harness-2026-05-16/CHANGE.md

@@ -0,0 +1,31 @@
+---
+slug: scenarios-harness-2026-05-16
+---
+
+# CHANGE · scenarios-harness-2026-05-16
+
+## §P proposal
+
+### Problem
+
+README §v0.x "Multi-runtime confidence" lists "Reproducible test fixture set under `tests/scenarios/`" as the last open item. Today, multi-agent situations (claim-before-edit, cross-runtime handoff, stale-claim sweep, plan claim adoption, pre/post path mismatch) live as ad-hoc smoke tests scattered across `packages/hooks/test/` and `apps/cli/test/`. Each rebuilds its own tempdir + git repo + fake-timer scaffolding inline. Reproducing a regression means hand-porting that scaffolding into a fresh file.
+
+### Proposal
+
+Add a reproducible test-scenarios harness under `tests/scenarios/`. Each scenario is a directory of plaintext artifacts (no binary snapshots):
+
+- `seed.sql` — applied after schema migrations against a fresh tempdir SQLite DB.
+- `inputs.jsonl` — one envelope per line: `{kind, at_ms, payload}` where `kind` is `lifecycle | mcp | tick`. Lifecycle flows through the same `runOmxLifecycleEnvelope` that production hooks call.
+- `expected.json` — normalized substrate snapshot with subset matchers (`toMatchObject` style), not full-row equality. Paths normalized to `<REPO_ROOT>`.
+- Optional `meta.yaml` — runtimes, tags, description.
+
+A shared `_harness/` drives all scenarios via `vi.useFakeTimers` + `vi.setSystemTime(BASE_TS + at_ms)` per envelope so timing is deterministic. Embeddings forced to `provider: none` to remove network. Five canonical scenarios ship in this PR: claim-before-edit, cross-runtime handoff, stale-claim sweep, plan claim adoption, pre/post path mismatch. Two harness self-tests prove the runner fails closed on missing expected and reports a clear diff on mismatch. A separate CI job runs `pnpm scenarios` on Node 20 after `build`, kept out of `pnpm test` so failure attribution stays clean.
+
+### Acceptance criteria
+
+- `pnpm scenarios` runs all five scenarios plus two harness self-tests, all green.
+- `pnpm scenarios:filter <slug>` runs a single scenario by name.
+- `pnpm scenarios:explain <slug>` prints a human-readable timeline.
+- `pnpm scenarios:record <slug>` regenerates `expected.json` from a live run (manual trim still required for subset matcher discipline).
+- `.github/workflows/ci.yml` gains a `scenarios` job after `build` running on Node 20 only.
+- `pnpm typecheck` and `pnpm build` clean.

package.json

@@ -33,13 +33,18 @@
     "p": "pnpm run release",
     "publish:cli": "bash scripts/publish-cli.sh",
     "publish:cli:dry-run": "bash scripts/publish-cli.sh --dry-run",
-    "release": "changeset publish"
+    "release": "changeset publish",
+    "scenarios": "vitest run --config tests/scenarios/_harness/vitest.config.ts",
+    "scenarios:filter": "vitest run --config tests/scenarios/_harness/vitest.config.ts -t",
+    "scenarios:explain": "tsx tests/scenarios/_harness/explain.mts",
+    "scenarios:record": "tsx tests/scenarios/_harness/record.mts"
   },
   "devDependencies": {
     "@biomejs/biome": "^1.9.4",
     "@changesets/cli": "^2.27.9",
     "@types/node": "^22.9.0",
     "tsup": "^8.3.5",
+    "tsx": "^4.19.2",
     "typescript": "^5.6.3",
     "vitest": "^2.1.5"
   }

tests/scenarios/01-claim-before-edit/expected.json

@@ -0,0 +1,112 @@
+{
+  "observations": [
+    {
+      "kind": "lifecycle_event",
+      "ts_offset": 10,
+      "metadata_subset": {
+        "event_name": "session_start",
+        "session_id": "codex@scenario-01",
+        "agent": "codex",
+        "branch": "agent/scenario/default",
+        "binding_status": "bound_task"
+      }
+    },
+    {
+      "kind": "omx-lifecycle",
+      "metadata_subset": {
+        "event_id": "evt_01_session",
+        "event_type": "session_start",
+        "ok": true
+      }
+    },
+    {
+      "kind": "omx-lifecycle",
+      "metadata_subset": {
+        "event_id": "evt_01_bind",
+        "event_type": "task_bind",
+        "ok": true
+      }
+    },
+    {
+      "kind": "claim",
+      "ts_offset": 40,
+      "metadata_subset": {
+        "kind": "claim",
+        "source": "pre-tool-use",
+        "file_path": "src/target.ts",
+        "auto_claimed_before_edit": true,
+        "tool": "Edit"
+      }
+    },
+    {
+      "kind": "claim-before-edit",
+      "ts_offset": 40,
+      "metadata_subset": {
+        "kind": "claim-before-edit",
+        "source": "pre-tool-use",
+        "outcome": "auto_claimed_before_edit",
+        "file_path": "src/target.ts",
+        "tool": "Edit",
+        "conflict": false
+      }
+    },
+    {
+      "kind": "omx-lifecycle",
+      "ts_offset": 40,
+      "metadata_subset": {
+        "event_id": "evt_01_pre",
+        "event_type": "pre_tool_use",
+        "tool_name": "Edit",
+        "extracted_paths": ["src/target.ts"]
+      }
+    },
+    {
+      "kind": "tool_use",
+      "ts_offset": 60,
+      "metadata_subset": {
+        "tool": "Edit",
+        "lifecycle_event_id": "evt_01_post",
+        "parent_event_id": "evt_01_pre",
+        "file_path": "src/target.ts"
+      }
+    },
+    {
+      "kind": "omx-lifecycle",
+      "ts_offset": 60,
+      "metadata_subset": {
+        "event_id": "evt_01_post",
+        "event_type": "post_tool_use",
+        "parent_event_id": "evt_01_pre",
+        "tool_name": "Edit"
+      }
+    }
+  ],
+  "claims": [
+    {
+      "task_id": 1,
+      "file_path": "src/target.ts",
+      "session_id": "codex@scenario-01",
+      "state": "active"
+    }
+  ],
+  "mcp_metrics": [],
+  "lifecycle_events": [
+    {
+      "event_type": "session_start",
+      "event_id": "evt_01_session"
+    },
+    {
+      "event_type": "task_bind",
+      "event_id": "evt_01_bind"
+    },
+    {
+      "event_type": "pre_tool_use",
+      "event_id": "evt_01_pre"
+    },
+    {
+      "event_type": "post_tool_use",
+      "event_id": "evt_01_post",
+      "parent_event_id": "evt_01_pre"
+    }
+  ]
+}

tests/scenarios/01-claim-before-edit/inputs.jsonl

@@ -0,0 +1,7 @@
+# Scenario 01 — claim-before-edit
+# A codex session starts, binds to a task, then claims src/target.ts
+# before issuing a pre_tool_use + post_tool_use Edit on it.
+{"kind":"lifecycle","at_ms":10,"payload":{"event_id":"evt_01_session","event_name":"session_start","session_id":"codex@scenario-01","agent":"codex","branch":"agent/scenario/default"}}
+{"kind":"lifecycle","at_ms":20,"payload":{"event_id":"evt_01_bind","event_name":"task_bind","session_id":"codex@scenario-01","agent":"codex","branch":"agent/scenario/default"}}
+{"kind":"lifecycle","at_ms":40,"payload":{"event_id":"evt_01_pre","event_name":"pre_tool_use","session_id":"codex@scenario-01","agent":"codex","branch":"agent/scenario/default","tool_name":"Edit","tool_input":{"operation":"replace","paths":[{"path":"<REPO_ROOT>/src/target.ts","role":"target","kind":"file"}]}}}
+{"kind":"lifecycle","at_ms":60,"payload":{"event_id":"evt_01_post","event_name":"post_tool_use","parent_event_id":"evt_01_pre","session_id":"codex@scenario-01","agent":"codex","branch":"agent/scenario/default","tool_name":"Edit","tool_input":{"operation":"replace","paths":[{"path":"<REPO_ROOT>/src/target.ts","role":"target","kind":"file"}]},"tool_response":{"success":true}}}

tests/scenarios/01-claim-before-edit/meta.yaml

@@ -0,0 +1,10 @@
+description: |
+  Codex session binds to a task on agent/scenario/default and edits
+  src/target.ts. Proves pre_tool_use synthesizes a claim-before-edit
+  signal and that the post_tool_use observation lands with the claim
+  already in place.
+runtimes:
+  - codex
+tags:
+  - claim-before-edit
+  - lifecycle

tests/scenarios/01-claim-before-edit/seed.sql

@@ -0,0 +1,2 @@
+-- No seed needed: the lifecycle session_start + task_bind events drive
+-- session creation and task binding by themselves.

tests/scenarios/02-cross-runtime-handoff/expected.json

@@ -0,0 +1,92 @@
+{
+  "observations": [
+    {
+      "kind": "lifecycle_event",
+      "metadata_subset": {
+        "event_name": "session_start",
+        "session_id": "codex@scenario-02",
+        "agent": "codex"
+      }
+    },
+    {
+      "kind": "omx-lifecycle",
+      "metadata_subset": {
+        "event_id": "evt_02_codex_session",
+        "event_type": "session_start"
+      }
+    },
+    {
+      "kind": "omx-lifecycle",
+      "metadata_subset": {
+        "event_id": "evt_02_codex_bind",
+        "event_type": "task_bind"
+      }
+    },
+    {
+      "kind": "claim",
+      "ts_offset": 30,
+      "metadata_subset": {
+        "kind": "claim",
+        "file_path": "src/target.ts"
+      }
+    },
+    {
+      "kind": "relay",
+      "ts_offset": 50,
+      "metadata_subset": {
+        "kind": "relay",
+        "from_session_id": "codex@scenario-02",
+        "from_agent": "codex",
+        "reason": "quota"
+      }
+    },
+    {
+      "kind": "claim-weakened",
+      "ts_offset": 50,
+      "metadata_subset": {
+        "kind": "claim-weakened",
+        "file_path": "src/target.ts",
+        "state": "handoff_pending"
+      }
+    },
+    {
+      "kind": "lifecycle_event",
+      "ts_offset": 300000,
+      "metadata_subset": {
+        "event_name": "session_start",
+        "session_id": "claude@scenario-02",
+        "agent": "claude"
+      }
+    },
+    {
+      "kind": "omx-lifecycle",
+      "metadata_subset": {
+        "event_id": "evt_02_claude_session",
+        "event_type": "session_start"
+      }
+    }
+  ],
+  "claims": [
+    {
+      "task_id": 1,
+      "file_path": "src/target.ts",
+      "session_id": "claude@scenario-02",
+      "state": "active"
+    }
+  ],
+  "mcp_metrics": [],
+  "lifecycle_events": [
+    {
+      "event_type": "session_start",
+      "event_id": "evt_02_codex_session"
+    },
+    {
+      "event_type": "task_bind",
+      "event_id": "evt_02_codex_bind"
+    },
+    {
+      "event_type": "session_start",
+      "event_id": "evt_02_claude_session"
+    }
+  ]
+}

tests/scenarios/02-cross-runtime-handoff/inputs.jsonl

@@ -0,0 +1,11 @@
+# Scenario 02 — cross-runtime handoff
+# Codex binds to a task, claims src/target.ts, then relays out. Claude
+# session adopts via accept_relay. Proves the baton pass survives
+# across runtimes and the claim ends up owned by claude's session.
+{"kind":"lifecycle","at_ms":10,"payload":{"event_id":"evt_02_codex_session","event_name":"session_start","session_id":"codex@scenario-02","agent":"codex","branch":"agent/scenario/default"}}
+{"kind":"lifecycle","at_ms":20,"payload":{"event_id":"evt_02_codex_bind","event_name":"task_bind","session_id":"codex@scenario-02","agent":"codex","branch":"agent/scenario/default"}}
+{"kind":"task","at_ms":30,"payload":{"action":"claim_file","task_id":1,"session_id":"codex@scenario-02","file_path":"src/target.ts","note":"codex claims before relay"}}
+{"kind":"task","at_ms":50,"payload":{"action":"relay","task_id":1,"from_session_id":"codex@scenario-02","from_agent":"codex","reason":"quota","one_line":"codex hit quota, handing off","base_branch":"main","expires_in_ms":600000}}
+{"kind":"lifecycle","at_ms":300000,"payload":{"event_id":"evt_02_claude_session","event_name":"session_start","session_id":"claude@scenario-02","agent":"claude","branch":"agent/scenario/default"}}
+{"kind":"task","at_ms":300100,"payload":{"action":"join","task_id":1,"session_id":"claude@scenario-02","agent":"claude"}}
+{"kind":"task","at_ms":300200,"payload":{"action":"accept_relay","task_id":1,"session_id":"claude@scenario-02"}}

tests/scenarios/02-cross-runtime-handoff/meta.yaml

@@ -0,0 +1,13 @@
+description: |
+  Codex binds to a task on agent/scenario/default and claims
+  src/target.ts. Codex then relays out (quota reason). 5 minutes later
+  a claude session starts, joins the task, and accepts the relay.
+  Proves the cross-runtime baton pass and that the claim ends up
+  owned by claude's session.
+runtimes:
+  - codex
+  - claude
+tags:
+  - handoff
+  - relay
+  - cross-runtime

tests/scenarios/02-cross-runtime-handoff/seed.sql

@@ -0,0 +1,2 @@
+-- No seed needed: lifecycle session_start + task_bind create the task,
+-- task envelopes drive the relay and adoption.