feat(staging): deterministic staging-runner replacing LLM deployer on deploy-staging (ORCH-115)

Replace the LLM `deployer` agent on the `deploy-staging` stage (self-hosting
orchestrator) with a deterministic staging-runner intercepted in launch_job
BEFORE _spawn (the deploy-finalizer / post-deploy-monitor reserved-agent
precedent). The runner executes the SAME staging suite, maps the exit-code to
`staging_status:` via the existing self_deploy.map_exit_code_to_status contract,
writes 15-staging-log.md, and initiates the UNCHANGED check_staging_status gate
exactly as a finished LLM-deployer would.

Invariant (NFR-1): this replaces only the *producer* of the artifact — the
artifact contract, the gate / _parse_staging_status / check_staging_status name,
STAGE_TRANSITIONS, the machine-verdict key `staging_status:` and the DB schema are
byte-for-byte unchanged. Additive, under a kill-switch + repo-scope CSV,
never-raise, fail-safe back to the LLM path.

Two-level outcome (D5, anti ORCH-110): suite executed -> verdict -> advance
(FAILED -> the existing deploy-staging -> development rollback + developer-retry,
same as a FAILED LLM verdict); tool-error (suite did not execute) -> bounded DEFER
-> fail-closed FAILED + alert on exhaustion (infra != code fault; never a silent
advance / false green).

First implemented slice of the LLM determinization roadmap (ORCH-118 A6,
replace-deterministic-now).

- New leaf src/staging_runner.py (never-raise; proc_group tree-kill + timeout)
- launch_job intercept + _run_staging_runner_job (mirror _run_deploy_finalizer_job)
- config: ORCH_STAGING_RUNNER_* keys (enabled/repos/timeout/infra-retry budget)
- GET /queue staging_runner observability block
- docs: llm-call-sites/roadmap/usage-policy (A6 implemented; machine blocks +
  single-transport invariant intact), deployer.md (LLM branch -> fallback),
  CLAUDE.md, CHANGELOG.md, overview (tech-pipeline/tech-agents/tech-quality-security),
  .env.example
- tests/test_orch115_staging_runner.py (TC-01..TC-13); LLM anti-drift green (TC-14)

Refs: ORCH-115

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.env.example

@@ -557,6 +557,27 @@ ORCH_COVERAGE_EPSILON=0.5
 ORCH_COVERAGE_TOOL_FAIL_CLOSED=false
 ORCH_COVERAGE_RUN_TIMEOUT_S=900
 
+# ORCH-115: deterministic staging-runner replacing the LLM `deployer` on the
+# `deploy-staging` stage (self-hosting orchestrator). Intercepted in launch_job
+# BEFORE _spawn (deploy-finalizer / post-deploy-monitor precedent): runs the same
+# staging suite, maps exit-code -> staging_status:, writes 15-staging-log.md and
+# initiates the UNCHANGED check_staging_status gate. Replaces only the producer of
+# the artifact; the gate / STAGE_TRANSITIONS / DB schema are byte-for-byte unchanged.
+# See ADR-001-deterministic-staging-runner.md / adr-0048.
+#   STAGING_RUNNER_ENABLED            -> kill-switch; false -> the prior LLM deployer
+#                                        runs on deploy-staging via _spawn 1:1.
+#   STAGING_RUNNER_REPOS              -> CSV scope; empty -> self-hosting only.
+#   STAGING_RUNNER_TIMEOUT_S          -> wall-clock budget for the docker-exec suite
+#                                        (malformed/non-positive -> default 600 + WARNING).
+#   STAGING_RUNNER_INFRA_MAX_RETRIES  -> tool-error (suite did NOT execute) bounded DEFER
+#                                        budget before a fail-closed FAILED (anti ORCH-110).
+#   STAGING_RUNNER_INFRA_RETRY_DELAY_S-> delay before the re-queued deployer job.
+ORCH_STAGING_RUNNER_ENABLED=true
+ORCH_STAGING_RUNNER_REPOS=
+ORCH_STAGING_RUNNER_TIMEOUT_S=600
+ORCH_STAGING_RUNNER_INFRA_MAX_RETRIES=2
+ORCH_STAGING_RUNNER_INFRA_RETRY_DELAY_S=30
+
 # ORCH-057 (follow-up ORCH-040): legacy root-owned ownership detect + actionable
 # worktree error. After the uid migration (user: "1000:1000") legacy root:root files
 # in /repos broke worktree creation under uid 1000 with a raw "Permission denied".

.openclaw/agents/deployer.md

@@ -45,6 +45,16 @@ then emit `staging_status:` / `deploy_status:`.
 
 Run the staging test suite against the live staging environment and write the verdict.
 
+> **ORCH-115 — deterministic runner leads this stage for in-scope repos.** On `deploy-staging` for
+> the self-hosting `orchestrator` repo, this stage is now driven by **deterministic code**
+> (`src/staging_runner.py`, intercepted in `launch_job` BEFORE `_spawn`, mirroring the prod Phase
+> A/B/C pattern) — it runs the SAME canonical staging suite below, maps the exit code to
+> `staging_status:` via the same `0 → SUCCESS / non-zero → FAILED` contract, writes
+> `15-staging-log.md`, and initiates the unchanged `check_staging_status` gate. The LLM steps below
+> remain the **fallback** under a disabled kill-switch (`ORCH_STAGING_RUNNER_ENABLED=false`) or for
+> non-self repos. The artifact contract / gate / machine key `staging_status:` are unchanged. Details:
+> `docs/work-items/ORCH-115/06-adr/ADR-001-deterministic-staging-runner.md`.
+
 **Steps:**
 
 1. Run the staging suite. **CANONICAL: run INSIDE the `orchestrator-staging` container via

.task-dev.md

@@ -1,4 +1,4 @@
-Work item: ORCH-112
+Work item: ORCH-115
 Repo: orchestrator
-Branch: feature/ORCH-112-bug-failed-cancelled-task-arti
+Branch: feature/ORCH-115-orch-replace-llm-deployer-with
 Stage: development

CHANGELOG.md

@@ -3,6 +3,12 @@
 Формат: [Keep a Changelog](https://keepachangelog.com/). Записи — на смысловой PR/задачу.
 
 ## [Unreleased]
+- **Детерминированный staging-раннер вместо LLM-деплойера на `deploy-staging`** (ORCH-115, `feat`): первый реализованный срез determinization-roadmap (ORCH-118 A6, `replace-deterministic-now`) — на стадии `deploy-staging` для self-hosting `orchestrator` **LLM-агент `deployer` заменён детерминированным кодом** (`src/staging_runner.py`). Работа агента на этой стадии была чисто механической (запуск staging-сюиты + маппинг exit-кода `staging_check.py` → `staging_status:`); каждый прогон жёг токены/время opus-агента (~40–120k / 3–15 мин) и встраивал недетерминизм LLM в точку ветвления гейта. **Инвариант (NFR-1):** это замена *продюсера* артефакта, **не** гейта — контракт `15-staging-log.md`, гейт `check_staging_status`/`_parse_staging_status`, `STAGE_TRANSITIONS`, machine-verdict `staging_status:`, схема БД — **байт-в-байт не тронуты**. Аддитивно, под kill-switch, never-raise, fail-closed, скоуп self-hosting. ADR: `docs/work-items/ORCH-115/06-adr/ADR-001-deterministic-staging-runner.md`, сквозной `docs/architecture/adr/adr-0048-deterministic-staging-runner.md`.
+  - **Перехват в `launch_job` до `_spawn` (D1):** `if job.agent=="deployer" and staging_runner.should_intercept(job)` → `_run_staging_runner_job` (зеркало `_run_deploy_finalizer_job`, прецедент `deploy-finalizer`/`post-deploy-monitor` `launcher.py:389/394`): синхронно ведёт `jobs`-строку через `mark_job`, возвращает `None` (нет `agent_runs`-строки, нет токенов). Дискриминатор «staging vs prod» — **стадия задачи** `deploy-staging` (роль `deployer` общая для `deploy-staging`/`deploy`), не имя роли; `should_intercept` never-raise → `False` → штатный `_spawn` (fail-safe к LLM-пути). Для не-self репо `applies==False` → прод-`deployer` никогда не перехватывается.
+  - **Leaf `src/staging_runner.py` (новый, чистый never-raise):** по образцу `self_deploy`/`proc_group`/`staging_verdict` (на импорте только `config`/`proc_group`; `db`/`git_worktree`/`stage_engine`/`qg.checks`/`notifications` — лениво). Исполняет ту же сюиту (`docker exec orchestrator-staging python3 <repos_dir>/<self-repo>/scripts/staging_check.py --base-url http://localhost:<staging_port> --mode stub`, арги из config — ORCH-101, без host-хардкодов) через `proc_group.run_in_process_group` (tree-kill, таймаут `staging_runner_timeout_s=600`, малформ/непозитив → дефолт + WARNING); маппит exit-код **единым** контрактом `self_deploy.map_exit_code_to_status` (`0→SUCCESS`/иначе/None→`FAILED`, fail-closed; ORCH-061 infra-tolerance остаётся внутри `staging_check.py`, раннер не пересуживает — BR-4); пишет `15-staging-log.md` (тот же machine-key `staging_status:` UPPERCASE + 52c-схема, `author_agent: staging-runner`/`model_used: n/a`) + best-effort commit/push в **фичеветку** (не в `main` — гейт читает worktree первым, усиливает AC-8/BR-7); вызывает **существующий** `advance_stage(current_stage="deploy-staging", finished_agent="deployer")` — без новых рёбер/исходов (transition-lease ORCH-114 берётся внутри `advance_stage`, раннер не трогает — граница O1).
+  - **Двухуровневый исход (D5, анти-ORCH-110):** сюита **исполнилась** (реальный exit-код) → verdict→advance (FAILED → тот же откат `deploy-staging → development` + developer-retry, что у FAILED-вердикта LLM, `stage_engine.py:932`); сюита **не исполнилась** (tool-error: spawn-error/таймаут/`returncode None`) → инфра-сбой ≠ код-фейл → bounded **DEFER** (re-queue `deployer`-джоба с задержкой + restart-safe маркер `staging-runner infra-retry` в `task_content`, счётчик подсчётом маркера — без правки схемы БД), на исчерпании `staging_runner_infra_max_retries=2` → fail-closed `FAILED` + advance + инфра-alert. Раннер **никогда** не делает тихий advance/ложный green, **никогда** не клинит очередь и **не** жжёт developer-retry на транзиентной инфре.
+  - **Self-hosting safety (BR-7/AC-8):** в командной строке раннера нет рестарта 8500 / `docker compose up orchestrator` / `--build` / force-push в `main` / правок `.env` — раннер только читает/исполняет staging-сюиту (8501) и пишет лог. Наблюдаемость (D10) — in-process счётчики (`runs`/`success`/`failed`/`tool_error`/`deferred`) + read-only блок `staging_runner` в `GET /queue` + один структурный лог-вердикт на прогон (различает код-фейл и tool-error). Флаги (`config.py`, дефолт = боевое): `staging_runner_enabled` (env `ORCH_STAGING_RUNNER_ENABLED`), `staging_runner_repos` (CSV; пусто → self-hosting only), `staging_runner_timeout_s`, `staging_runner_infra_max_retries`, `staging_runner_infra_retry_delay_s`. Откат = `ORCH_STAGING_RUNNER_ENABLED=false` → на `deploy-staging` снова LLM-`deployer` через `_spawn` **байт-в-байт**.
+  - **Норматив сопровождения ORCH-118 (NFR-6):** обновлены `docs/architecture/llm-call-sites.md` (A6 — реализован; машинный `ORCH-118-INVENTORY-BLOCK` сохраняет deployer как `avoidable=yes`/`axis=C` — LLM-ветвь жива как fallback), `llm-determinization-roadmap.md` (rank 1 deployer — ✅ реализован; машинный блок/инвариант «ровно один `first_slice = yes`» целы), `llm-usage-policy.md` (§5 — единственный транспорт S0 не нарушен, раннер LLM не зовёт), `.openclaw/agents/deployer.md` (LLM-ветвь `deploy-staging` — fallback), витрина `docs/overview/tech-pipeline.md`/`tech-agents.md`. Покрытие — `tests/test_orch115_staging_runner.py` (TC-01…TC-13) + зелёные `tests/test_llm_call_site_inventory.py`/`test_llm_determinization_docs.py` (TC-14).
 - **Карта LLM-консультаций + control-path-ось «avoidable» + roadmap + нормативная политика** (ORCH-118, `docs`+`test`, inventory-first, docs+tests only): зонтичный follow-up RCA-трека ORCH-114/117 — у оркестратора не было ни нормативного критерия «где LLM нужен, а где это avoidable control path», ни карты мест вызова LLM, прибитой к коду. Выпущена **доказательная карта** каждого места, где control-path потребляет (или способен потребить) суждение LLM, с control-path-разметкой и классификацией; **упорядоченный roadmap** детерминированных замен; **нормативная политика** использования LLM; набор **структурных анти-дрейф тестов**. Это **docs + tests only**: `src/**`-рантайм не меняется → `STAGE_TRANSITIONS` / реестр и имена `QG_CHECKS`/`check_*` / machine-verdict-ключи / схема БД — **байт-в-байт не тронуты**; раннеры замен **не** реализуются (FR-7); конкретные follow-up Plane-ID **не** фиксируются (R3/NFR-6 — кандидаты по роли). kill-switch не нужен (нет рантайм-поведения), как ORCH-077/079/101/102/103/011. ADR: `docs/work-items/ORCH-118/06-adr/ADR-001-llm-call-site-map-and-determinization-roadmap.md`, сквозной `docs/architecture/adr/adr-0047-llm-usage-policy-and-call-site-map.md`.
   - **Единица инвентаря — LLM-консультация** (control-path потребляет суждение LLM), а **не** «спавн процесса / существование Claude CLI» (R4, capability ≠ consultation). Карта разводит **три ортогональных факта**: (1) consultation ≠ transport/slot (единственный транспорт LLM-консультации в `src/**` — `launcher._spawn`, `launcher.py:472`/CLI-сборка `610-614`; иного транспорта нет; job-роли `deploy-finalizer`/`post-deploy-monitor` занимают слот, но перехватываются в `launch_job` **до** `_spawn`, `launcher.py:389/394` — консультации нет); (2) **control-path (C) ≠ artifact-producer (P)** по коду-потребителю в `src/qg/checks.py` (C: гейт ветвится на LLM-вердикте; P: детерминированный гейт судит артефакт независимо — файлы/CI); (3) деривируемость вердикта из tool-сигналов.
   - **Нормативное определение** «avoidable LLM control path» = двухбитный предикат (C-консультация **И** вердикт деривируем из exit-кодов `pytest`/smoke/`staging_check.py`/деплоя). Поимённый целевой набор (доказательно, прибит тестами): **avoidable = `{tester, deployer}`**; control-path-но-keep = `{reviewer}` (вердикт «приемлемость кода/решения» НЕ деривируем); не-control-path (P, keep) = `{analyst, architect, developer}`; уже детерминированы (эталон) = `{deploy-finalizer, post-deploy-monitor}`.

CLAUDE.md

@@ -411,6 +411,52 @@ Plane, **не** Quality Gate и **не** стадия).
   `docs/work-items/ORCH-117/06-adr/ADR-001-sandbox-only-plane-write-guard.md`, сквозной
   `docs/architecture/adr/adr-0046-sandbox-only-plane-write-guard.md`.
 
+## Детерминированный staging-раннер вместо LLM-деплойера (ORCH-115)
+Первый реализованный срез determinization-roadmap (ORCH-118 A6, `replace-deterministic-now`): на
+стадии `deploy-staging` для self-hosting `orchestrator` **LLM-агент `deployer` заменён
+детерминированным кодом** (`src/staging_runner.py`). Работа агента на этой стадии была чисто
+механической (запуск staging-сюиты + маппинг exit-кода) — теперь её делает leaf, перехватываемый в
+`launch_job` **до `_spawn`** (прецедент `deploy-finalizer`/`post-deploy-monitor`,
+`launcher.py:389/394`). **Инвариант (NFR-1):** замена *продюсера* артефакта, **не** гейта —
+контракт `15-staging-log.md`, гейт/`_parse_staging_status`/имя `check_staging_status`,
+`STAGE_TRANSITIONS`, machine-verdict `staging_status:`, схема БД — **байт-в-байт не тронуты**.
+Аддитивно, под kill-switch, never-raise, fail-closed.
+- **Перехват (D1):** `launch_job` — `if job.agent=="deployer" and staging_runner.should_intercept(job)`
+  → `_run_staging_runner_job` (зеркало `_run_deploy_finalizer_job`): синхронно ведёт `jobs`-строку
+  через `mark_job`, возвращает `None` (нет `agent_runs`). Дискриминатор «staging vs prod» —
+  **стадия задачи** `deploy-staging` (роль `deployer` общая для `deploy-staging`/`deploy`), не имя
+  роли; `should_intercept` never-raise → `False` → штатный `_spawn` (fail-safe к LLM-пути).
+- **Раннер (D2-D7):** leaf по образцу `self_deploy`/`proc_group`/`staging_verdict` (на импорте только
+  `config`/`proc_group`; `db`/`git_worktree`/`stage_engine`/`qg.checks`/`notifications` — лениво).
+  Исполняет ту же сюиту (`docker exec orchestrator-staging python3 .../staging_check.py --base-url
+  http://localhost:8501 --mode stub`, арги из config — ORCH-101) через `proc_group` (tree-kill,
+  таймаут `staging_runner_timeout_s=600`); маппит exit-код **единым** контрактом
+  `self_deploy.map_exit_code_to_status` (`0→SUCCESS`/иначе/None→`FAILED`; ORCH-061 infra-tolerance
+  остаётся внутри `staging_check.py`, раннер не пересуживает); пишет `15-staging-log.md`
+  (`author_agent: staging-runner`/`model_used: n/a`, 52c-схема) + best-effort push в **фичеветку**
+  (не в `main` — гейт читает worktree первым, усиливает AC-8); вызывает **существующий**
+  `advance_stage(current_stage="deploy-staging", finished_agent="deployer")` — без новых
+  рёбер/исходов (lease ORCH-114 берётся внутри `advance_stage`, раннер не трогает).
+- **Двухуровневый исход (D5, анти-ORCH-110):** сюита **исполнилась** (реальный exit-код) → verdict→
+  advance (FAILED → тот же откат `deploy-staging → development` + developer-retry, что у LLM); сюита
+  **не исполнилась** (tool-error: spawn-error/таймаут/`returncode None`) → инфра-сбой ≠ код-фейл →
+  bounded **DEFER** (re-queue `deployer`-джоба + restart-safe маркер `staging-runner infra-retry` в
+  `task_content`, без правки схемы БД), на исчерпании `staging_runner_infra_max_retries` → fail-closed
+  `FAILED` + advance + alert. Никогда тихий advance/ложный green; не клинит очередь; не жжёт
+  developer-retry на транзиентной инфре.
+- **Флаги** (`config.py`, дефолт = боевое): `staging_runner_enabled` (kill-switch, env
+  `ORCH_STAGING_RUNNER_ENABLED`), `staging_runner_repos` (CSV; **пусто → self-hosting only**),
+  `staging_runner_timeout_s=600`, `staging_runner_infra_max_retries=2`,
+  `staging_runner_infra_retry_delay_s=30`. Откат = `ORCH_STAGING_RUNNER_ENABLED=false` → на
+  `deploy-staging` снова LLM-`deployer` через `_spawn` **байт-в-байт**. Наблюдаемость — in-process
+  счётчики + read-only блок `staging_runner` в `GET /queue` + один структурный лог-вердикт на прогон
+  (различает код-фейл и tool-error). Норматив сопровождения ORCH-118: обновлены
+  `llm-call-sites.md`/`llm-determinization-roadmap.md`/`llm-usage-policy.md` (A6 — реализован, машинные
+  блоки/инвариант «единственный транспорт S0» целы) + `deployer.md` (LLM-ветвь — fallback). Покрытие —
+  `tests/test_orch115_staging_runner.py` (TC-01…TC-13) + зелёные LLM-анти-дрейф тесты (TC-14). Детали —
+  `docs/work-items/ORCH-115/06-adr/ADR-001-deterministic-staging-runner.md`, сквозной
+  `docs/architecture/adr/adr-0048-deterministic-staging-runner.md`.
+
 ## Машинный журнал уроков (ORCH-098)
 Шаг 1 («Фундамент», F2) эпика саморазвития: формализует свободнотекстовые «уроки» из `memory/` в
 **машинную структурированную таблицу отклонений конвейера** `lessons`, фундамент для будущих

docs/architecture/llm-call-sites.md

@@ -58,7 +58,7 @@
 | **A3** | `.openclaw/agents/developer.md` | стадия `development` | developer | код + PR | — | `check_ci_green:82` (+ `check_branch_mergeable:657`) — CI/merge | ~150–400k / 10–40 мин | да (через S0) | **P** | `keep-LLM` | написание кода — настоящее суждение; гейт судит CI/merge, не самоотчёт |
 | **A4** | `.openclaw/agents/reviewer.md` | стадия `review` | reviewer | `12-review.md` | `verdict:` | `check_reviewer_verdict:336` (`verdict:`) | ~100–300k / 5–25 мин | да (через S0) | **C** | `keep-LLM` | control path, но вердикт «приемлемость кода/решения» **НЕ деривируем** из exit-кода — настоящее суждение |
 | **A5** | `.openclaw/agents/tester.md` | стадия `testing` | tester | `13-test-report.md` | `result:` | `check_tests_passed:182` → `_parse_tests_verdict:226` (`result:`) | ~60–150k / 5–20 мин | да (через S0) | **C** | `needs-hybrid-fallback` | **avoidable**: PASS/FAIL = exit-code `pytest`+smoke (деривируем); LLM нужен лишь на триаж падений / маппинг TC↔критерии |
-| **A6** | `.openclaw/agents/deployer.md` | стадии `deploy-staging` / `deploy` | deployer | `15-staging-log.md` / `14-deploy-log.md` | `staging_status:` / `deploy_status:` | `check_staging_status:599` → `_parse_staging_status:538` (`staging_status:`); `check_deploy_status:473` → `_parse_deploy_status:413` (`deploy_status:`) | ~40–120k / 3–15 мин | да (через S0) | **C** | `replace-deterministic-now` | **avoidable**: staging-вердикт = маппинг exit-кода `staging_check.py`; прод уже детерминирован Phase A/B/C (ORCH-036) |
+| **A6** | `.openclaw/agents/deployer.md` | стадии `deploy-staging` / `deploy` | deployer | `15-staging-log.md` / `14-deploy-log.md` | `staging_status:` / `deploy_status:` | `check_staging_status:599` → `_parse_staging_status:538` (`staging_status:`); `check_deploy_status:473` → `_parse_deploy_status:413` (`deploy_status:`) | ~40–120k / 3–15 мин | да (через S0; для in-scope репо на `deploy-staging` — **нет**, перехват до `_spawn`) | **C** | `replace-deterministic-now` | **avoidable, СРЕЗ РЕАЛИЗОВАН (ORCH-115):** на `deploy-staging` для self-hosting `orchestrator` вердикт `staging_status:` производит детерминированный `src/staging_runner.py` (перехват в `launch_job` до `_spawn`, как D1/D2) — маппинг exit-кода `staging_check.py`; прод уже детерминирован Phase A/B/C (ORCH-036). LLM-ветвь остаётся fallback'ом под выключенным флагом / для не-self репо |
 | **D1** | `src/agents/launcher.py:389` (перехват в `launch_job` **до** `_spawn`; «Not an LLM spawn» `407`) | post-deploy edge | deploy-finalizer | jobs-row | — | — | — (детерминированный) | **нет** (слот, перехват до `_spawn`) | — | `already-deterministic` (эталон) | Занимает слот агента, но LLM не консультируется — рабочий прецедент замены |
 | **D2** | `src/agents/launcher.py:394` (перехват в `launch_job` **до** `_spawn`; «Not an LLM spawn» `428`) | post-deploy observation | post-deploy-monitor | jobs-row | — | — | — (детерминированный) | **нет** (слот, перехват до `_spawn`) | — | `already-deterministic` (эталон) | Тик наблюдения; LLM не консультируется |
 
@@ -147,8 +147,14 @@
 
 - **Docs + tests only (ORCH-118).** `STAGE_TRANSITIONS` / реестр и имена `QG_CHECKS`/`check_*` /
   machine-verdict-ключи (`verdict:`/`result:`/`staging_status:`/`deploy_status:`/`security_status:`/
-  `coverage_status:`) / схема БД — **байт-в-байт не тронуты**. Раннеры замен **не** реализованы
-  (см. roadmap — follow-up'ы по роли, без Plane-ID).
+  `coverage_status:`) / схема БД — **байт-в-байт не тронуты** (это инвариант самой карты).
+- **Реализованные срезы.** Первый срез roadmap'а — **deployer (staging-status)** — реализован
+  **ORCH-115** (`src/staging_runner.py`, перехват в `launch_job` до `_spawn`): на `deploy-staging`
+  для in-scope репо вердикт `staging_status:` производит детерминированный код, не LLM. Это
+  `replace-deterministic-now` без ввода второго транспорта (раннер LLM не зовёт) — карта/инвариант
+  «единственный транспорт S0» соблюдены. Машинный `ORCH-118-INVENTORY-BLOCK` сохраняет deployer как
+  `avoidable=yes`/`axis=C` (LLM-ветвь жива как fallback под выключенным флагом / для не-self репо).
+  Второй кандидат (`tester`) остаётся follow-up'ом по роли.
 - **Анти-дрейф тесты:** `tests/test_llm_call_site_inventory.py` (TC-01…TC-06, TC-09, TC-12, TC-13,
   TC-14) и `tests/test_llm_determinization_docs.py` (TC-07/08/11). Дискриминатор всех проверок —
   **«консультирует LLM», а не «спавнит subprocess»**.

docs/architecture/llm-determinization-roadmap.md

@@ -15,7 +15,17 @@
 
 ---
 
-## 1. Рекомендованный первый срез — **deployer (staging-status)**
+## 1. Рекомендованный первый срез — **deployer (staging-status)** — ✅ РЕАЛИЗОВАН (ORCH-115)
+
+> **Статус: реализовано.** Срез выполнен в **ORCH-115** — `src/staging_runner.py` (перехват в
+> `launch_job` до `_spawn`, как `D1`/`D2`): на стадии `deploy-staging` для self-hosting `orchestrator`
+> вердикт `staging_status:` производит детерминированный код (маппинг exit-кода `staging_check.py`
+> через `self_deploy.map_exit_code_to_status`), а не LLM. Под kill-switch `staging_runner_enabled` +
+> скоуп `staging_runner_repos` (пусто → self-hosting only); LLM-ветвь остаётся fallback'ом.
+> Контракт артефакта/гейта `check_staging_status`/`STAGE_TRANSITIONS`/схема БД — не тронуты. Детали —
+> `docs/work-items/ORCH-115/06-adr/ADR-001-deterministic-staging-runner.md`, сквозной
+> `docs/architecture/adr/adr-0048-deterministic-staging-runner.md`. Запись `rank 1` в машинном блоке
+> §4 сохраняется (`first_slice = yes`, инвариант карты) как историческая фиксация первого среза.
 
 Обоснование (самый низкорисковый «чисто деривируемый» control path):
 

docs/architecture/llm-usage-policy.md

@@ -94,3 +94,8 @@ Call-site является **avoidable LLM control path** тогда и толь
   `launcher._spawn` (S0). Ввод второго транспорта (новый `_spawn`, импорт `anthropic`/`openai`/иного
   LLM-SDK, прямой HTTP Anthropic/Claude, второй model-invoking subprocess) запрещён без явного ADR;
   прибито тестами TC-01/TC-12.
+- **Реализованный срез (ORCH-115).** Снятие C-консультации с деривируемым вердиктом — это разрешённое
+  `replace-deterministic-now`, а не ввод новой LLM-консультации. ORCH-115 снял A6/staging-status:
+  детерминированный `src/staging_runner.py` производит `staging_status:` без `_spawn` (перехват до
+  него, как `D1`/`D2`) — раннер **LLM не зовёт** и **второй транспорт не вводит**, поэтому инвариант
+  «единственный транспорт S0» соблюдён (TC-12 зелёный). Это образец для последующих срезов roadmap'а.

docs/overview/tech-agents.md

@@ -48,6 +48,12 @@ Machine-verdict ключи читаются гейтами **только из Y
 Особенность: промпт `deployer` сознательно на английском (самый safety-critical — несёт
 запреты self-hosting в видной рамке); остальные пять — на русском.
 
+Особенность (ORCH-115): на стадии `deploy-staging` для self-hosting `orchestrator` LLM-`deployer`
+заменён **детерминированным staging-раннером** (`src/staging_runner.py`) — его работа была чисто
+механической (запуск staging-сюиты + маппинг exit-кода). LLM-промпт `deployer` остаётся fallback'ом
+под выключенным флагом / для не-self репо и продолжает вести прод-стадию `deploy`. Подробнее —
+[конвейер](tech-pipeline.md) и [карта LLM-консультаций](../architecture/llm-call-sites.md).
+
 ## Человек как седьмая роль
 
 Человек не пишет артефакты конвейера, но принимает два решения, которые не делегированы

docs/overview/tech-pipeline.md

@@ -34,6 +34,16 @@ created → analysis → architecture → development → review → testing →
 | `done` | — | — | терминал |
 | `cancelled` | — | — | терминал (сток отмены) |
 
+> **Детерминированный staging-раннер (ORCH-115).** На стадии `deploy-staging` для self-hosting
+> `orchestrator` работу ведёт **детерминированный код** (`src/staging_runner.py`), а не LLM-агент
+> `deployer`: он перехватывается в `launch_job` до запуска агента (как Phase A/B/C прод-деплоя),
+> исполняет ту же staging-сюиту, маппит exit-код в `staging_status:` и инициирует **тот же** гейт
+> `check_staging_status`. Это замена *продюсера* артефакта, а не гейта: контракт `15-staging-log.md`,
+> имя/семантика `check_staging_status`, `STAGE_TRANSITIONS` — не изменились. Под kill-switch
+> `staging_runner_enabled` (скоуп `staging_runner_repos`, пусто → self-hosting only); при выключении
+> на стадии снова работает LLM-`deployer` байт-в-байт. Это первый реализованный срез
+> determinization-roadmap (см. `docs/architecture/llm-determinization-roadmap.md`).
+
 ## Под-гейты деплойного ребра — врезки, не стадии
 
 На переходе `deploy-staging → deploy` исполняются четыре под-гейта в нормативном порядке

docs/overview/tech-quality-security.md

@@ -50,7 +50,10 @@ LLM, и **нормативную политику** «LLM — только та
 control-path и его вердикт деривируем из exit-кодов (тогда консультацию можно заменить
 детерминированным раннером). Поимённо: avoidable = `{tester, deployer}`; настоящее суждение
 сохраняется у `{analyst, architect, developer, reviewer}`. Карта — снимок, прибитый структурными
-анти-дрейф тестами; реализация замен — отдельные follow-up'ы по роли.
+анти-дрейф тестами. **Первый срез реализован (ORCH-115):** на `deploy-staging` для self-hosting
+`orchestrator` LLM-`deployer` заменён детерминированным `src/staging_runner.py` (вердикт
+`staging_status:` = маппинг exit-кода staging-сюиты); LLM-ветвь остаётся fallback'ом, гейт
+`check_staging_status` не тронут. Замена второго кандидата (`tester`) — follow-up по роли.
 
 - Карта вызовов LLM: [`../architecture/llm-call-sites.md`](../architecture/llm-call-sites.md)
 - Нормативная политика: [`../architecture/llm-usage-policy.md`](../architecture/llm-usage-policy.md)

src/agents/launcher.py

@@ -385,6 +385,14 @@ class AgentLauncher:
         (no-LLM) job — intercept it BEFORE _spawn (which would raise
         "Unknown agent", R-6) and run the deploy finalizer synchronously, driving
         the jobs row status itself. Returns None (no agent_run row).
+
+        ORCH-115: the LLM ``deployer`` on the ``deploy-staging`` stage (self-hosting
+        scope) is replaced by a DETERMINISTIC staging-runner — intercepted here
+        BEFORE _spawn (same precedent as deploy-finalizer / post-deploy-monitor). The
+        discriminator is the TASK STAGE (deploy-staging), not the role name, so the
+        prod ``deploy`` deployer is never caught (staging_runner.should_intercept).
+        Kill-switch off / out of scope -> should_intercept False -> the prior LLM
+        deployer runs via _spawn byte-for-byte.
         """
         if job.get("agent") == "deploy-finalizer":
             return self._run_deploy_finalizer_job(job)
@@ -393,6 +401,11 @@ class AgentLauncher:
         # observation tick synchronously. Returns None (no agent_run row).
         if job.get("agent") == "post-deploy-monitor":
             return self._run_post_deploy_monitor_job(job)
+        # ORCH-115: deterministic staging-runner intercept (BEFORE _spawn).
+        if job.get("agent") == "deployer":
+            from .. import staging_runner
+            if staging_runner.should_intercept(job):
+                return self._run_staging_runner_job(job)
         return self._spawn(
             job["agent"],
             job["repo"],
@@ -422,6 +435,28 @@ class AgentLauncher:
                 pass
         return None
 
+    def _run_staging_runner_job(self, job: dict):
+        """ORCH-115: run the deterministic staging gate for a deployer job.
+
+        Not an LLM spawn — there is no subprocess/monitor of an agent, so we mark the
+        jobs row done/failed here (mirror of _run_deploy_finalizer_job). The runner
+        never-raises, but we guard anyway so a runner fault can't wedge the worker.
+        Returns None (no agent_run row, _spawn not called).
+        """
+        from ..db import mark_job
+        from .. import staging_runner
+        try:
+            staging_runner.run_staging_gate(job)
+            mark_job(job["id"], "done")
+            logger.info(f"staging-runner job {job['id']} done")
+        except Exception as e:
+            logger.error(f"staging-runner job {job['id']} failed: {e}")
+            try:
+                mark_job(job["id"], "failed", error=f"staging-runner error: {e}")
+            except Exception:
+                pass
+        return None
+
     def _run_post_deploy_monitor_job(self, job: dict):
         """ORCH-021: run one deterministic post-deploy monitor tick for a job.
 

src/config.py

@@ -413,6 +413,51 @@ class Settings(BaseSettings):
     coverage_tool_fail_closed: bool = False
     coverage_run_timeout_s: int = 900
 
+    # ORCH-115: deterministic staging-runner replacing the LLM `deployer` agent on
+    # the `deploy-staging` stage for the self-hosting orchestrator. A new leaf
+    # src/staging_runner.py (never-raise) is intercepted in launch_job BEFORE _spawn
+    # (mirroring the deploy-finalizer / post-deploy-monitor reserved-agent
+    # precedent, launcher.py:389/394): it runs the SAME staging suite the LLM ran
+    # (`docker exec orchestrator-staging python3 .../staging_check.py`), maps the
+    # exit-code -> staging_status: via the existing self_deploy.map_exit_code_to_status
+    # contract, writes 15-staging-log.md, and initiates the EXISTING check_staging_status
+    # gate exactly as a finished LLM-deployer would. The artifact contract, the gate,
+    # STAGE_TRANSITIONS and the DB schema are byte-for-byte UNCHANGED — this only
+    # replaces the *producer* of the artifact. Pattern = coverage_gate_* / self_deploy_*.
+    # See docs/work-items/ORCH-115/06-adr/ADR-001-deterministic-staging-runner.md and
+    # docs/architecture/adr/adr-0048-deterministic-staging-runner.md.
+    #   staging_runner_enabled            -> SINGLE kill-switch (env
+    #                                        ORCH_STAGING_RUNNER_ENABLED). False -> the
+    #                                        intercept never fires -> the prior LLM
+    #                                        deployer runs on deploy-staging via _spawn
+    #                                        byte-for-byte as before ORCH-115 (D8/AC-6).
+    #   staging_runner_repos              -> CSV scope (env ORCH_STAGING_RUNNER_REPOS).
+    #                                        Empty -> self-hosting only (orchestrator)
+    #                                        via is_self_hosting_repo; non-empty ->
+    #                                        membership. Mirrors coverage_gate_repos.
+    #   staging_runner_timeout_s          -> wall-clock budget for the docker-exec
+    #                                        staging suite (env ORCH_STAGING_RUNNER_TIMEOUT_S).
+    #                                        Malformed/non-positive -> default + WARNING
+    #                                        (never-break). Aligned with the cross-cutting
+    #                                        budget invariant ORCH-065/109/110 WITHOUT
+    #                                        touching reaper_max_running_s (D9): it replaces
+    #                                        the up-to-900s LLM staging window with a bounded
+    #                                        <=600s deterministic one (Σ on the edge does not grow).
+    #   staging_runner_infra_max_retries  -> tool-error (suite did NOT execute: spawn-error /
+    #                                        timeout / returncode None) bounded DEFER budget
+    #                                        before a fail-closed FAILED (env
+    #                                        ORCH_STAGING_RUNNER_INFRA_MAX_RETRIES). Mirrors
+    #                                        merge_retest_infra_max_retries — infra hiccup is
+    #                                        NOT a code-fault, so it never burns a developer-retry
+    #                                        until the budget is exhausted (D5, anti ORCH-110).
+    #   staging_runner_infra_retry_delay_s-> delay before the re-queued deployer job
+    #                                        (env ORCH_STAGING_RUNNER_INFRA_RETRY_DELAY_S).
+    staging_runner_enabled: bool = True
+    staging_runner_repos: str = ""
+    staging_runner_timeout_s: int = 600
+    staging_runner_infra_max_retries: int = 2
+    staging_runner_infra_retry_delay_s: int = 30
+
     # ORCH-098 (FND/F2): machine lessons-journal — additive `lessons` table + leaf
     # src/lessons.py (never-raise observer, by образцу serial_gate/coverage_gate/
     # metrics). The journal is an OBSERVER, never a Quality Gate: writing a lesson

src/main.py

@@ -235,6 +235,7 @@ async def queue():
     from . import lessons
     from . import checkout_hygiene
     from . import transition_lease
+    from . import staging_runner
     from .disk_watchdog import disk_watchdog
     from .build_cache_pruner import build_cache_pruner
     return {
@@ -283,6 +284,11 @@ async def queue():
         # (owner/stage/age/live) + defer/reclaim/CAS-lost counters. Additive block;
         # never-raise.
         "transition_lease": transition_lease.snapshot(),
+        # ORCH-115 (FR-7 / AC-10): deterministic staging-runner observability
+        # (read-only) — kill-switch, scope, timeout/infra budget + run/success/
+        # failed/tool_error/deferred counters, so a code-fail FAILED is distinguishable
+        # from an infra tool-error. Additive block; never-raise.
+        "staging_runner": staging_runner.snapshot(),
         # ORCH-098 (FR-4 / AC-4): lessons-journal observability (read-only) —
         # kill-switch + counts by type/status + last N lessons. Additive block;
         # never-raise (snapshot() returns {"enabled": ...} minimum on error).

src/staging_runner.py

@@ -0,0 +1,581 @@
+"""Deterministic staging-runner (ORCH-115).
+
+The `deploy-staging` stage for the self-hosting ``orchestrator`` repo was driven by
+the LLM ``deployer`` agent, but the actual work is purely deterministic
+(``.openclaw/agents/deployer.md`` steps 1-4): run the staging suite, map its
+**exit-code** to a verdict (``0 -> SUCCESS``, else ``FAILED``), write
+``15-staging-log.md`` and commit it. This leaf replaces that LLM consultation with
+deterministic code, intercepted in ``launcher.launch_job`` BEFORE ``_spawn`` (the
+``deploy-finalizer`` / ``post-deploy-monitor`` reserved-agent precedent,
+``launcher.py:389/394``).
+
+What is and is NOT changed (NFR-1, the critical invariant):
+  * UNCHANGED — the artifact contract (``15-staging-log.md`` with
+    ``staging_status: SUCCESS|FAILED``), the gate ``check_staging_status`` /
+    ``_parse_staging_status``, ``STAGE_TRANSITIONS``, the DB schema. This module
+    replaces only the *producer* of the artifact, never the gate that reads it.
+  * NEW — a deterministic producer + a launch_job intercept. Under a kill-switch +
+    repo-scope CSV; fail-safe back to the LLM path when off / out of scope.
+
+This module is a **leaf** (mirror of ``self_deploy`` / ``proc_group`` /
+``staging_verdict``): it imports only ``config`` / ``logging`` / ``proc_group`` at
+module load; ``db`` / ``git_worktree`` / ``self_deploy.map_exit_code_to_status`` /
+``qg.checks`` / ``stage_engine.advance_stage`` / ``notifications`` are imported
+LAZILY inside functions so the heavy ``stage_engine`` is never pulled at import and
+no import cycle forms (pattern: ``transition_lease`` / ``merge_gate``). Every public
+function honours a **never-raise** contract (AC-7): a staging-infra hiccup can never
+crash the worker / wedge the queue.
+
+Two-level outcome (D5 — the key safety decision, anti ORCH-110):
+  * the suite EXECUTED (a real exit-code, 0 or non-zero) -> trust the code:
+    ``0 -> SUCCESS -> advance``; ``!=0 -> FAILED -> the existing rollback
+    deploy-staging -> development`` (same developer-retry path as a FAILED LLM
+    verdict). ORCH-061 infra-tolerance is already INSIDE ``staging_check.py`` — the
+    runner never re-judges it (BR-4).
+  * the suite did NOT execute (tool-error: spawn-error / timeout / ``returncode is
+    None``) -> an infra fault, NOT a code fault -> a bounded DEFER (re-queue a fresh
+    ``deployer`` job with a delay + a restart-safe marker). On budget exhaustion ->
+    fail-closed ``FAILED`` + advance + alert. So the runner NEVER does a silent
+    advance / false green, and NEVER wedges the queue, but does NOT burn a
+    developer-retry on transient infra.
+"""
+
+import logging
+import time
+
+from .config import settings
+from . import proc_group
+
+logger = logging.getLogger("orchestrator.staging_runner")
+
+# Platform service literal — the staging compose service the suite runs inside.
+# Already an accepted platform literal (mirror image_freshness._STAGING_SERVICE);
+# NOT a host hardcode (test_no_host_hardcodes forbids host IP/home/hostname only).
+STAGING_SERVICE = "orchestrator-staging"
+
+# Default wall-clock budget for the docker-exec staging suite (D9). Kept <= the LLM
+# staging window it replaces so Σ(work on the deploy-staging edge) does not grow and
+# the cross-cutting reaper invariant (ORCH-065/109/110) holds WITHOUT touching
+# reaper_max_running_s.
+_DEFAULT_TIMEOUT_S = 600
+
+_GIT_TIMEOUT = 60
+
+# Restart-safe DEFER marker (counted from the persisted jobs queue, mirror of
+# stage_engine._merge_infra_retry_count). Embedded verbatim in the re-queued job's
+# task_content so a service restart never resets the infra-retry budget.
+_INFRA_RETRY_MARKER = "staging-runner infra-retry"
+
+# In-process observability counters (mirror merge_gate._MERGE_GATE_COUNTERS, ORCH-110).
+_STAGING_RUNNER_COUNTERS: dict = {
+    "runs": 0,          # run_staging_gate entered
+    "success": 0,       # suite ran, exit 0 -> SUCCESS
+    "failed": 0,        # suite ran non-zero, OR infra budget exhausted -> FAILED
+    "tool_error": 0,    # suite did NOT execute (spawn-error / timeout / None)
+    "deferred": 0,      # bounded infra DEFER (re-queued)
+}
+
+
+def _bump(key: str) -> None:
+    """Increment an observability counter. Never raises."""
+    try:
+        _STAGING_RUNNER_COUNTERS[key] += 1
+    except Exception:  # noqa: BLE001 - observability must never break a decision
+        pass
+
+
+# ---------------------------------------------------------------------------
+# Conditionality (D8 / FR-6 / AC-6)
+# ---------------------------------------------------------------------------
+def applies(repo: str) -> bool:
+    """Whether the deterministic staging-runner is REAL for ``repo``.
+
+    Mirrors ``self_deploy_applies`` / ``coverage_gate``:
+      * ``staging_runner_enabled=False`` -> always False (global kill-switch); the
+        legacy LLM-deployer path runs on deploy-staging via ``_spawn``.
+      * ``staging_runner_repos`` (CSV) non-empty -> only the listed repos.
+      * empty CSV -> only the self-hosting repo (``orchestrator``), which is the only
+        one with a staging instance.
+    Checked FIRST in ``should_intercept`` (local, no network, no DB) so a disabled
+    flag costs nothing. never-raise -> False (fail-safe to the prior LLM path).
+    """
+    try:
+        if not settings.staging_runner_enabled:
+            return False
+        raw = (settings.staging_runner_repos or "").strip()
+        if raw:
+            allowed = {r.strip().lower() for r in raw.split(",") if r.strip()}
+            return (repo or "").strip().lower() in allowed
+        # Lazy import keeps this module a leaf (no qg import at module load).
+        from .qg.checks import is_self_hosting_repo
+        return is_self_hosting_repo(repo)
+    except Exception as e:  # noqa: BLE001 - never-raise contract
+        logger.warning("staging_runner.applies error for %s: %s", repo, e)
+        return False
+
+
+def should_intercept(job: dict) -> bool:
+    """True iff this ``deployer`` job is the deterministic staging-suite job (D1).
+
+    The discriminator of "staging vs prod" is the TASK STAGE, not the role name
+    (``deployer`` owns both ``deploy-staging`` and ``deploy``): intercept iff
+    ``agent == "deployer"`` AND the task's ``tasks.stage == "deploy-staging"`` AND
+    ``applies(repo)``. For self-hosting the prod ``deploy`` edge runs Phase A (no
+    deployer spawn); for non-self repos ``applies`` is False, so a non-self prod
+    ``deployer`` job is never intercepted (R-1 / TC-06).
+
+    never-raise -> False (a DB-lookup failure falls through to ``_spawn``, fail-safe
+    to the prior LLM path).
+    """
+    try:
+        if (job.get("agent") or "") != "deployer":
+            return False
+        # applies() FIRST (local, no DB): disabled flag -> zero DB overhead.
+        if not applies(job.get("repo")):
+            return False
+        task_id = job.get("task_id")
+        if task_id is None:
+            return False
+        from .db import get_db
+        conn = get_db()
+        row = conn.execute("SELECT stage FROM tasks WHERE id=?", (task_id,)).fetchone()
+        conn.close()
+        if not row:
+            return False
+        return (row[0] or "") == "deploy-staging"
+    except Exception as e:  # noqa: BLE001 - never-raise contract
+        logger.warning("staging_runner.should_intercept error: %s", e)
+        return False
+
+
+# ---------------------------------------------------------------------------
+# Suite execution (D3 / FR-2 / NFR-3 / AC-8 / AC-9)
+# ---------------------------------------------------------------------------
+def build_staging_command() -> list[str]:
+    """Build the canonical staging-suite argv (same command the LLM-deployer ran).
+
+    ``docker exec <STAGING_SERVICE> python3 <repos_dir>/<self-repo>/scripts/staging_check.py
+    --base-url http://localhost:<staging_port> --mode stub``. Host-specifics come from
+    config (ORCH-101, no host hardcodes). Self-hosting safety (BR-7 / AC-8 / TC-12):
+    NO restart of 8500, NO ``docker compose up orchestrator`` / ``--build``, NO
+    force-push, NO ``.env`` edit — the runner only reads/executes the staging suite
+    (8501) and writes a log.
+    """
+    from .qg.checks import SELF_HOSTING_REPO
+    repos_dir = (settings.repos_dir or "/repos").rstrip("/")
+    script = f"{repos_dir}/{SELF_HOSTING_REPO}/scripts/staging_check.py"
+    base_url = f"http://localhost:{int(settings.staging_port)}"
+    return [
+        "docker", "exec", STAGING_SERVICE,
+        "python3", script,
+        "--base-url", base_url,
+        "--mode", "stub",
+    ]
+
+
+def _resolve_timeout() -> int:
+    """Resolve ``staging_runner_timeout_s`` (malformed/non-positive -> default +
+    WARNING, never-break — mirror of ``merge_gate._resolve_retest_timeout``)."""
+    raw = getattr(settings, "staging_runner_timeout_s", _DEFAULT_TIMEOUT_S)
+    try:
+        t = int(raw)
+        if t > 0:
+            return t
+        logger.warning(
+            "staging_runner_timeout_s non-positive (%r) -> default %ds",
+            raw, _DEFAULT_TIMEOUT_S,
+        )
+    except (TypeError, ValueError):
+        logger.warning(
+            "staging_runner_timeout_s malformed (%r) -> default %ds",
+            raw, _DEFAULT_TIMEOUT_S,
+        )
+    return _DEFAULT_TIMEOUT_S
+
+
+def run_staging_suite() -> proc_group.ProcResult:
+    """Execute the staging suite in its own process group, tree-killed on timeout.
+
+    Reuses ``proc_group.run_in_process_group`` (ORCH-110) so a hung docker-exec /
+    pytest subtree is killed whole (no orphans, AC-9). Never raises (proc_group
+    degrades any OS error to a safe ``ProcResult``).
+    """
+    cmd = build_staging_command()
+    timeout = _resolve_timeout()
+    try:
+        grace = float(getattr(settings, "agent_kill_grace_seconds", 5) or 5)
+    except (TypeError, ValueError):
+        grace = 5.0
+    return proc_group.run_in_process_group(
+        cmd,
+        cwd=None,
+        timeout=timeout,
+        grace_s=grace,
+        tree_kill=bool(getattr(settings, "subprocess_tree_kill_enabled", True)),
+    )
+
+
+# ---------------------------------------------------------------------------
+# exit-code -> verdict (D4 / FR-3 / AC-3) — reuse the single contract, no 2nd mapping
+# ---------------------------------------------------------------------------
+def map_exit_code_to_status(exit_code) -> str:
+    """``0 -> SUCCESS``; non-zero / None / non-int -> ``FAILED`` (fail-closed).
+
+    Delegates to ``self_deploy.map_exit_code_to_status`` — the SAME pure, unit-tested
+    contract the deploy-finalizer uses (BR-4: no second, drifting mapping).
+    """
+    from .self_deploy import map_exit_code_to_status as _map
+    return _map(exit_code)
+
+
+# ---------------------------------------------------------------------------
+# Artifact 15-staging-log.md (D6 / FR-4 / AC-2 / AC-8) — mirror write_deploy_log
+# ---------------------------------------------------------------------------
+def _extract_infra_waived(stdout: str) -> list[str]:
+    """Lines from the suite stdout carrying the ORCH-061 ``INFRA-WAIVED:`` marker
+    (copied into the log body for observability, as the prompt required)."""
+    if not stdout:
+        return []
+    return [ln.strip() for ln in stdout.splitlines() if "INFRA-WAIVED" in ln]
+
+
+def build_staging_log(
+    work_item_id: str, exit_code, status: str, stdout: str = "", *, tool_error: bool = False
+) -> str:
+    """Render a ``15-staging-log.md`` body whose ``staging_status:`` frontmatter is the
+    verdict ``check_staging_status`` / ``_parse_staging_status`` reads (contract
+    UNCHANGED, AC-2). Carries the mandatory 52c schema (``work_item`` / ``stage`` /
+    ``author_agent`` / ``status`` / ``created_at`` / ``model_used``); ``author_agent:
+    staging-runner`` / ``model_used: n/a`` honestly reflect the DETERMINISTIC producer.
+    The machine key ``staging_status:`` and its UPPERCASE value are NOT changed.
+
+    Written as a literal block (mirror of ``self_deploy.build_deploy_log``) so the
+    machine-read frontmatter is byte-exact; only the frontmatter is machine-read, the
+    body is informational.
+    """
+    import datetime
+    created = datetime.date.today().isoformat()
+    sub_status = "success" if status == "SUCCESS" else "failed"
+    base_url = f"http://localhost:{int(settings.staging_port)}"
+
+    waived = _extract_infra_waived(stdout)
+    waived_body = ""
+    if waived:
+        waived_body = "\nINFRA-WAIVED lines (ORCH-061, copied for observability):\n" + "\n".join(
+            f"- {ln}" for ln in waived
+        ) + "\n"
+
+    tail = ""
+    if stdout:
+        tail_text = stdout.strip()[-1500:]
+        if tail_text:
+            tail = f"\nStaging suite stdout (tail):\n```\n{tail_text}\n```\n"
+
+    note = (
+        "Staging suite did NOT execute (tool-error) and the infra-retry budget was "
+        "exhausted -> fail-closed FAILED."
+        if tool_error
+        else f"Staging suite exit-code `{exit_code}` -> `staging_status: {status}`."
+    )
+
+    return (
+        "---\n"
+        f"staging_status: {status}\n"
+        f"work_item: {work_item_id}\n"
+        "stage: deploy-staging\n"
+        "author_agent: staging-runner\n"
+        f"status: {sub_status}\n"
+        f"created_at: {created}\n"
+        "model_used: n/a\n"
+        f"exit_code: {exit_code}\n"
+        f"base_url: {base_url}\n"
+        "---\n\n"
+        "# Staging Gate Log (deterministic runner, ORCH-115)\n\n"
+        f"{note}\n\n"
+        "Вердикт зафиксирован детерминированным staging-раннером (ORCH-115), не LLM. "
+        "infra-tolerance (ORCH-061) уже учтена внутри `staging_check.py` — раннер её не "
+        "пересуживает.\n"
+        f"{waived_body}"
+        f"{tail}"
+    )
+
+
+def write_staging_log(
+    repo: str, work_item_id: str, branch: str, exit_code, status: str,
+    stdout: str = "", *, tool_error: bool = False,
+) -> bool:
+    """Write ``15-staging-log.md`` into the task worktree (so ``check_staging_status``
+    reads it) and best-effort commit+push it to the FEATURE BRANCH. Returns True iff
+    the file was written. Never raises.
+
+    Mirror of ``self_deploy.write_deploy_log`` EXCEPT: the actor name is
+    ``staging-runner`` and the log is pushed only to the feature branch — there is NO
+    separate PR-merge of the log into ``main`` (the gate reads the worktree first;
+    excluding any direct work on ``main`` strengthens AC-8 / BR-7). The feature branch
+    is merged into ``main`` later by the normal merge-gate path.
+    """
+    import os
+    import subprocess
+    from .git_worktree import get_worktree_path
+
+    rel = f"docs/work-items/{work_item_id}/15-staging-log.md"
+    try:
+        wt = get_worktree_path(repo, branch)
+    except Exception as e:  # noqa: BLE001 - never-raise
+        logger.error("write_staging_log: worktree error for %s/%s: %s", repo, branch, e)
+        return False
+
+    path = os.path.join(wt, rel)
+    content = build_staging_log(work_item_id, exit_code, status, stdout, tool_error=tool_error)
+    try:
+        os.makedirs(os.path.dirname(path), exist_ok=True)
+        with open(path, "w", encoding="utf-8") as f:
+            f.write(content)
+    except OSError as e:
+        logger.error("write_staging_log: write error at %s: %s", path, e)
+        return False
+
+    # Best-effort commit + push to the feature branch (the gate also falls back to
+    # origin/main). ORCH-101: HOME + email domain from Settings; the actor NAME stays
+    # the platform literal `staging-runner` (deterministic system-actor commits).
+    _email = f"staging-runner@{settings.git_email_domain}"
+    git_env = {
+        **os.environ,
+        "HOME": settings.agent_home_dir,
+        "GIT_AUTHOR_NAME": "staging-runner",
+        "GIT_AUTHOR_EMAIL": _email,
+        "GIT_COMMITTER_NAME": "staging-runner",
+        "GIT_COMMITTER_EMAIL": _email,
+    }
+    try:
+        subprocess.run(["git", "-C", wt, "add", rel],
+                       capture_output=True, timeout=_GIT_TIMEOUT, env=git_env)
+        commit = subprocess.run(
+            ["git", "-C", wt, "commit", "-m",
+             f"staging(ORCH-115): staging gate {status} for {work_item_id}"],
+            capture_output=True, text=True, timeout=_GIT_TIMEOUT, env=git_env,
+        )
+        if commit.returncode == 0:
+            subprocess.run(["git", "-C", wt, "push", "origin", branch],
+                           capture_output=True, timeout=_GIT_TIMEOUT, env=git_env)
+    except (subprocess.SubprocessError, OSError) as e:
+        logger.warning("write_staging_log: git commit/push best-effort failed: %s", e)
+    return True
+
+
+# ---------------------------------------------------------------------------
+# Existing-gate initiation (D7 / FR-5 / AC-4) — no new routing branch
+# ---------------------------------------------------------------------------
+def _advance(task_id, repo: str, work_item_id: str, branch: str) -> None:
+    """Initiate the SAME ``advance_stage`` evaluation a finished LLM-deployer would
+    (``finished_agent="deployer"`` on ``deploy-staging``): SUCCESS -> sub-gates
+    security->merge->coverage->image-freshness (ORCH-022/043/027/058) + Phase A
+    (ORCH-036); FAILED -> the existing rollback deploy-staging -> development
+    (``stage_engine.py:932``). No new routing branch. The transition-lease (ORCH-114)
+    is taken INSIDE advance_stage on the side-effectful edge — the runner never
+    touches it (task boundary O1). never-raise."""
+    try:
+        from . import stage_engine
+        stage_engine.advance_stage(
+            task_id=task_id,
+            current_stage="deploy-staging",
+            repo=repo,
+            work_item_id=work_item_id,
+            branch=branch,
+            finished_agent="deployer",
+        )
+    except Exception as e:  # noqa: BLE001 - never-raise into the worker
+        logger.error(
+            "staging_runner._advance: advance_stage failed for task %s (%s): %s",
+            task_id, work_item_id, e,
+        )
+
+
+# ---------------------------------------------------------------------------
+# Two-level outcome (D5) — tool-error DEFER bookkeeping
+# ---------------------------------------------------------------------------
+def _infra_retry_count(task_id) -> int:
+    """How many times this task was re-queued by the tool-error DEFER path
+    (restart-safe; counted from the persisted jobs queue by the marker — mirror of
+    ``stage_engine._merge_infra_retry_count``). Never raises -> 0 on error."""
+    try:
+        from .db import get_db
+        conn = get_db()
+        n = conn.execute(
+            "SELECT COUNT(*) FROM jobs WHERE task_id=? AND task_content LIKE ?",
+            (task_id, f"%{_INFRA_RETRY_MARKER}%"),
+        ).fetchone()[0]
+        conn.close()
+        return int(n)
+    except Exception as e:  # noqa: BLE001 - never-raise
+        logger.warning("staging_runner._infra_retry_count error for %s: %s", task_id, e)
+        return 0
+
+
+def _handle_tool_error(
+    task_id, repo: str, work_item_id: str, branch: str, result: proc_group.ProcResult
+) -> None:
+    """Suite did NOT execute (tool-error) -> bounded DEFER, then fail-closed (D5).
+
+    Anti ORCH-110: an infra fault is NOT a code fault, so we re-queue a fresh
+    ``deployer`` job (which re-enters this runner) with a delay instead of an
+    immediate FAILED-rollback that would burn a developer-retry. On budget exhaustion
+    -> write ``staging_status: FAILED`` + advance (the existing rollback) + an
+    INFRA-specific alert (explicitly "not a code defect"). Never a silent advance /
+    false green; never wedges the queue. never-raise."""
+    retries = _infra_retry_count(task_id)
+    try:
+        max_retries = int(settings.staging_runner_infra_max_retries)
+    except (TypeError, ValueError):
+        max_retries = 2
+    try:
+        delay = int(settings.staging_runner_infra_retry_delay_s)
+    except (TypeError, ValueError):
+        delay = 30
+
+    if retries < max_retries:
+        _bump("deferred")
+        reason = "timeout" if result.timed_out else "suite did not execute (tool-error)"
+        task_desc = (
+            f"Work item: {work_item_id}\nRepo: {repo}\nBranch: {branch}\n"
+            f"Stage: deploy-staging\nNote: {_INFRA_RETRY_MARKER} "
+            f"(attempt {retries + 1}/{max_retries}) — {reason}, retrying after {delay}s."
+        )
+        try:
+            from .db import enqueue_job
+            new_job = enqueue_job(
+                "deployer", repo, task_desc, task_id=task_id, available_at_delay_s=delay,
+            )
+            logger.warning(
+                "Task %s (%s): staging suite did not execute (%s) -> infra-DEFER "
+                "(job_id=%s, attempt %d/%d)",
+                task_id, work_item_id, reason, new_job, retries + 1, max_retries,
+            )
+        except Exception as e:  # noqa: BLE001 - never-raise
+            logger.error("staging_runner: infra-DEFER enqueue failed for %s: %s", task_id, e)
+        return
+
+    # Budget exhausted -> fail-closed FAILED (terminal, never a false green).
+    _bump("failed")
+    logger.error(
+        "Task %s (%s): staging tool-error DEFER budget exhausted (%d) -> fail-closed FAILED",
+        task_id, work_item_id, max_retries,
+    )
+    write_staging_log(repo, work_item_id, branch, result.returncode, "FAILED",
+                      result.stdout, tool_error=True)
+    _alert_infra_exhausted(work_item_id, max_retries)
+    _advance(task_id, repo, work_item_id, branch)
+
+
+def _alert_infra_exhausted(work_item_id: str, max_retries: int) -> None:
+    """Best-effort Telegram alert that the staging suite never executed (infra, NOT a
+    code defect) after the retry budget. never-raise."""
+    try:
+        from .notifications import send_telegram, link_for
+        send_telegram(
+            f"\U0001f6a8 {link_for(work_item_id)}: staging suite не запустилась "
+            f"(инфра, НЕ дефект кода) после {max_retries} попыток — fail-closed FAILED, "
+            f"откат на development. Нужно проверить staging-стенд."
+        )
+    except Exception as e:  # noqa: BLE001 - never-raise
+        logger.warning("staging_runner: infra-exhausted alert failed for %s: %s", work_item_id, e)
+
+
+# ---------------------------------------------------------------------------
+# Entry point (D2) — owns the full deterministic flow, mirror run_deploy_finalizer
+# ---------------------------------------------------------------------------
+def run_staging_gate(job: dict) -> None:
+    """Deterministic staging gate for a ``deployer`` job on ``deploy-staging``.
+
+    Flow (mirror of ``stage_engine.run_deploy_finalizer``):
+      1. resolve ``work_item_id`` / ``branch`` by ``task_id``;
+      2. execute the staging suite (D3) -> ProcResult;
+      3. suite EXECUTED -> map exit-code -> ``staging_status:``, write
+         ``15-staging-log.md``, initiate the existing gate via ``advance_stage`` (D7);
+      4. suite did NOT execute (tool-error) -> bounded DEFER / fail-closed (D5);
+      5. observability counters + one structured verdict log (D10).
+    Never raises into the caller (the launcher marks the job done/failed)."""
+    started = time.time()
+    _bump("runs")
+    task_id = job.get("task_id")
+    repo = job.get("repo")
+
+    # 1. resolve task fields.
+    work_item_id, branch = None, None
+    try:
+        from .db import get_db
+        conn = get_db()
+        row = conn.execute(
+            "SELECT work_item_id, branch FROM tasks WHERE id=?", (task_id,)
+        ).fetchone()
+        conn.close()
+        if row:
+            work_item_id, branch = row[0], row[1]
+    except Exception as e:  # noqa: BLE001 - never-raise
+        logger.error("staging_runner: task lookup failed for task_id=%s: %s", task_id, e)
+    if not work_item_id or not branch:
+        logger.error(
+            "staging_runner: missing work_item_id/branch for task_id=%s — aborting", task_id
+        )
+        return
+
+    # 2-4. execute + classify + route — guarded so AC-7 (never-raise) holds even if
+    # an unexpected error escapes a sub-step (the worker must never crash on staging
+    # infra; the task is left on deploy-staging for the reconciler/reaper to re-drive).
+    try:
+        result = run_staging_suite()
+        duration_s = round(time.time() - started, 1)
+        suite_ran = (result.returncode is not None) and (not result.timed_out)
+
+        if suite_ran:
+            # 3. trust the exit-code (ORCH-061 already inside staging_check.py).
+            status = map_exit_code_to_status(result.returncode)
+            _bump("success" if status == "SUCCESS" else "failed")
+            logger.info(
+                "staging_runner verdict: work_item=%s repo=%s exit_code=%s status=%s "
+                "duration_s=%s outcome=%s",
+                work_item_id, repo, result.returncode, status, duration_s,
+                "code-pass" if status == "SUCCESS" else "code-fail",
+            )
+            write_staging_log(repo, work_item_id, branch, result.returncode, status, result.stdout)
+            _advance(task_id, repo, work_item_id, branch)
+            return
+
+        # 4. tool-error (suite did not execute) -> DEFER / fail-closed (D5).
+        _bump("tool_error")
+        logger.warning(
+            "staging_runner verdict: work_item=%s repo=%s exit_code=%s status=%s "
+            "duration_s=%s outcome=tool-error (timed_out=%s)",
+            work_item_id, repo, result.returncode, "TOOL-ERROR", duration_s, result.timed_out,
+        )
+        _handle_tool_error(task_id, repo, work_item_id, branch, result)
+    except Exception as e:  # noqa: BLE001 - never-raise into the worker (AC-7)
+        logger.error(
+            "staging_runner.run_staging_gate: unexpected error for task %s (%s): %s",
+            task_id, work_item_id, e,
+        )
+
+
+# ---------------------------------------------------------------------------
+# Observability (D10 / FR-7 / AC-10)
+# ---------------------------------------------------------------------------
+def snapshot() -> dict:
+    """Read-only staging-runner summary for ``GET /queue`` (FR-7 / AC-10).
+
+    Additive block; existing ``/queue`` keys are untouched. never-raise: any error ->
+    a minimal dict with the kill-switch state."""
+    try:
+        return {
+            "enabled": bool(settings.staging_runner_enabled),
+            "repos": getattr(settings, "staging_runner_repos", "") or "",
+            "timeout_s": getattr(settings, "staging_runner_timeout_s", _DEFAULT_TIMEOUT_S),
+            "infra_max_retries": getattr(settings, "staging_runner_infra_max_retries", 2),
+            "runs": _STAGING_RUNNER_COUNTERS["runs"],
+            "success": _STAGING_RUNNER_COUNTERS["success"],
+            "failed": _STAGING_RUNNER_COUNTERS["failed"],
+            "tool_error": _STAGING_RUNNER_COUNTERS["tool_error"],
+            "deferred": _STAGING_RUNNER_COUNTERS["deferred"],
+        }
+    except Exception as e:  # noqa: BLE001 - never-raise -> minimal dict
+        logger.warning("staging_runner.snapshot error: %s", e)
+        return {"enabled": False}

tests/test_orch115_staging_runner.py

@@ -0,0 +1,438 @@
+"""ORCH-115 — deterministic staging-runner replacing the LLM deployer on
+`deploy-staging` (self-hosting orchestrator).
+
+Covers Phase 1 (04-test-plan.yaml TC-01…TC-13): the launch_job intercept BEFORE
+_spawn, the exit-code -> staging_status: mapping, 15-staging-log.md render + the
+UNCHANGED gate contract, advance_stage initiation, kill-switch/scope, never-raise /
+fail-safe, the bounded tool-error DEFER, process timeout, self-hosting safety, the
+anti-drift invariant (STAGE_TRANSITIONS / QG_CHECKS / DB schema untouched), and the
+/queue observability block.
+
+No live Claude CLI, docker or network: the staging subprocess and advance_stage are
+mocked; the pure mapping/render is tested directly. (TC-14 — the LLM call-site map
+anti-drift — lives in tests/test_llm_call_site_inventory.py / test_llm_determinization_docs.py.)
+"""
+import os
+import tempfile
+
+import pytest
+
+_test_db = os.path.join(tempfile.gettempdir(), "test_orch115_staging_runner.db")
+os.environ["ORCH_DB_PATH"] = _test_db
+os.environ["ORCH_REPOS_DIR"] = tempfile.gettempdir()
+os.environ.setdefault("ORCH_GITEA_TOKEN", "test-token")
+os.environ.setdefault("ORCH_PLANE_API_TOKEN", "test-token")
+
+from unittest.mock import MagicMock  # noqa: E402
+
+import src.db as _db  # noqa: E402
+from src.db import init_db, get_db, enqueue_job  # noqa: E402
+from src import staging_runner  # noqa: E402
+from src import stage_engine  # noqa: E402
+from src import config as cfg  # noqa: E402
+from src.proc_group import ProcResult  # noqa: E402
+from src.agents.launcher import AgentLauncher  # noqa: E402
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+@pytest.fixture(autouse=True)
+def fresh_db(monkeypatch, tmp_path):
+    monkeypatch.setattr(_db.settings, "db_path", _test_db)
+    if os.path.exists(_test_db):
+        os.unlink(_test_db)
+    init_db()
+    # Worktree artefacts land in tmp; git commit/push degrade gracefully (no repo).
+    monkeypatch.setattr("src.git_worktree.settings.worktrees_dir", str(tmp_path), raising=False)
+    # Runner ON, self-hosting scope (default).
+    monkeypatch.setattr(cfg.settings, "staging_runner_enabled", True, raising=False)
+    monkeypatch.setattr(cfg.settings, "staging_runner_repos", "", raising=False)
+    monkeypatch.setattr(cfg.settings, "staging_runner_infra_max_retries", 2, raising=False)
+    # Reset in-process counters between tests.
+    for k in staging_runner._STAGING_RUNNER_COUNTERS:
+        staging_runner._STAGING_RUNNER_COUNTERS[k] = 0
+    yield
+
+
+def _make_task(stage, repo="orchestrator", wi="ORCH-115", branch="feature/ORCH-115-x"):
+    conn = get_db()
+    cur = conn.execute(
+        "INSERT INTO tasks (plane_id, work_item_id, repo, branch, stage) VALUES (?, ?, ?, ?, ?)",
+        (f"plane-{wi}", wi, repo, branch, stage),
+    )
+    tid = cur.lastrowid
+    conn.commit()
+    conn.close()
+    return tid
+
+
+def _make_job(agent, repo, task_id, content="x"):
+    return enqueue_job(agent, repo, content, task_id=task_id)
+
+
+def _job_dict(job_id, agent, repo, task_id):
+    return {"id": job_id, "agent": agent, "repo": repo, "task_id": task_id, "task_content": "x"}
+
+
+def _read_log(tmp_path, repo, branch, wi):
+    from src.git_worktree import get_worktree_path
+    p = os.path.join(get_worktree_path(repo, branch), f"docs/work-items/{wi}/15-staging-log.md")
+    with open(p, "r", encoding="utf-8") as f:
+        return f.read()
+
+
+# ---------------------------------------------------------------------------
+# TC-01 — applies(): kill-switch / scope / fail-safe (FR-6 / AC-6)
+# ---------------------------------------------------------------------------
+def test_tc01_applies_killswitch_and_scope(monkeypatch):
+    # enabled=False -> False (fall back to the LLM path) regardless of repo.
+    monkeypatch.setattr(cfg.settings, "staging_runner_enabled", False)
+    assert staging_runner.applies("orchestrator") is False
+
+    # enabled, empty CSV -> self-hosting only.
+    monkeypatch.setattr(cfg.settings, "staging_runner_enabled", True)
+    monkeypatch.setattr(cfg.settings, "staging_runner_repos", "")
+    assert staging_runner.applies("orchestrator") is True
+    assert staging_runner.applies("enduro-trails") is False
+    assert staging_runner.applies("") is False
+
+    # non-empty CSV -> membership (case-insensitive).
+    monkeypatch.setattr(cfg.settings, "staging_runner_repos", "enduro-trails, orchestrator")
+    assert staging_runner.applies("ENDURO-TRAILS") is True
+    assert staging_runner.applies("other-repo") is False
+
+
+def test_tc01_applies_never_raises(monkeypatch):
+    # A settings attribute that explodes on read -> applies() degrades to False.
+    class Boom:
+        @property
+        def staging_runner_enabled(self):
+            raise RuntimeError("boom")
+    monkeypatch.setattr(staging_runner, "settings", Boom())
+    assert staging_runner.applies("orchestrator") is False
+
+
+# ---------------------------------------------------------------------------
+# TC-02 — exit-code -> verdict, single shared contract (FR-3 / AC-3)
+# ---------------------------------------------------------------------------
+def test_tc02_map_exit_code():
+    from src import self_deploy
+    assert staging_runner.map_exit_code_to_status(0) == "SUCCESS"
+    for bad in (1, 2, 137, -9):
+        assert staging_runner.map_exit_code_to_status(bad) == "FAILED"
+    for noncode in (None, "x", object()):
+        assert staging_runner.map_exit_code_to_status(noncode) == "FAILED"
+    # Same contract as the deploy-finalizer (no second, drifting mapping).
+    for v in (0, 1, None, "garbage"):
+        assert staging_runner.map_exit_code_to_status(v) == self_deploy.map_exit_code_to_status(v)
+
+
+# ---------------------------------------------------------------------------
+# TC-03 — 15-staging-log.md render: machine key + 52c schema + INFRA-WAIVED (FR-4 / AC-2)
+# ---------------------------------------------------------------------------
+def test_tc03_log_render_schema_and_infra_waived():
+    stdout = "B6 OK\nINFRA-WAIVED: B9 image-build skipped (sandbox)\nDONE"
+    body = staging_runner.build_staging_log("ORCH-115", 0, "SUCCESS", stdout)
+    assert "staging_status: SUCCESS" in body  # UPPERCASE machine key
+    for field in ("work_item: ORCH-115", "stage: deploy-staging",
+                  "author_agent: staging-runner", "status: success",
+                  "created_at:", "model_used: n/a"):
+        assert field in body, f"missing 52c field: {field}"
+    # INFRA-WAIVED line copied into the body for observability.
+    assert "INFRA-WAIVED: B9 image-build skipped (sandbox)" in body
+
+    failed = staging_runner.build_staging_log("ORCH-115", 1, "FAILED", "boom")
+    assert "staging_status: FAILED" in failed
+    assert "status: failed" in failed
+
+
+# ---------------------------------------------------------------------------
+# TC-04 — generated log read by the UNCHANGED _parse_staging_status (AC-2)
+# ---------------------------------------------------------------------------
+def test_tc04_gate_parser_unchanged():
+    from src.qg.checks import _parse_staging_status
+    ok, reason = _parse_staging_status(staging_runner.build_staging_log("ORCH-115", 0, "SUCCESS"))
+    assert ok is True and "SUCCESS" in reason
+    bad, reason2 = _parse_staging_status(staging_runner.build_staging_log("ORCH-115", 2, "FAILED"))
+    assert bad is False and "FAILED" in reason2
+
+
+# ---------------------------------------------------------------------------
+# TC-05 — launch_job intercepts deployer on deploy-staging BEFORE _spawn (AC-1)
+# ---------------------------------------------------------------------------
+def test_tc05_launch_job_intercepts_before_spawn(monkeypatch):
+    tid = _make_task("deploy-staging")
+    jid = _make_job("deployer", "orchestrator", tid)
+    lr = AgentLauncher()
+    spawn = MagicMock(return_value=999)
+    monkeypatch.setattr(lr, "_spawn", spawn)
+    run_gate = MagicMock()
+    monkeypatch.setattr(staging_runner, "run_staging_gate", run_gate)
+
+    ret = lr.launch_job(_job_dict(jid, "deployer", "orchestrator", tid))
+
+    assert ret is None                      # no agent_runs row
+    spawn.assert_not_called()               # LLM path NOT taken
+    run_gate.assert_called_once()           # deterministic runner ran
+    # jobs row driven to done by the launcher (no monitor/agent).
+    conn = get_db()
+    status = conn.execute("SELECT status FROM jobs WHERE id=?", (jid,)).fetchone()[0]
+    conn.close()
+    assert status == "done"
+
+
+# ---------------------------------------------------------------------------
+# TC-06 — stage discriminator: deployer on `deploy` is NOT intercepted (AC-1 / R-1)
+# ---------------------------------------------------------------------------
+def test_tc06_stage_discriminator_prod_not_intercepted(monkeypatch):
+    tid = _make_task("deploy")  # prod edge, not deploy-staging
+    jid = _make_job("deployer", "orchestrator", tid)
+    lr = AgentLauncher()
+    spawn = MagicMock(return_value=42)
+    monkeypatch.setattr(lr, "_spawn", spawn)
+    run_gate = MagicMock()
+    monkeypatch.setattr(staging_runner, "run_staging_gate", run_gate)
+
+    ret = lr.launch_job(_job_dict(jid, "deployer", "orchestrator", tid))
+
+    assert ret == 42                        # _spawn path taken
+    spawn.assert_called_once()
+    run_gate.assert_not_called()
+    assert staging_runner.should_intercept(_job_dict(jid, "deployer", "orchestrator", tid)) is False
+
+
+def test_tc06_non_self_repo_not_intercepted(monkeypatch):
+    tid = _make_task("deploy-staging", repo="enduro-trails", wi="ET-9", branch="feature/ET-9-x")
+    jid = _make_job("deployer", "enduro-trails", tid)
+    assert staging_runner.should_intercept(_job_dict(jid, "deployer", "enduro-trails", tid)) is False
+
+
+# ---------------------------------------------------------------------------
+# TC-07 — routing equivalence: SUCCESS -> advance; FAILED -> same path (AC-4)
+# ---------------------------------------------------------------------------
+@pytest.mark.parametrize("rc,expected", [(0, "SUCCESS"), (1, "FAILED")])
+def test_tc07_advance_initiated_like_llm(monkeypatch, tmp_path, rc, expected):
+    tid = _make_task("deploy-staging")
+    monkeypatch.setattr(staging_runner, "run_staging_suite",
+                        lambda: ProcResult(returncode=rc, stdout="out", stderr="", timed_out=False))
+    advance = MagicMock()
+    monkeypatch.setattr(stage_engine, "advance_stage", advance)
+
+    staging_runner.run_staging_gate(_job_dict(1, "deployer", "orchestrator", tid))
+
+    advance.assert_called_once()
+    kwargs = advance.call_args.kwargs
+    assert kwargs["current_stage"] == "deploy-staging"
+    assert kwargs["finished_agent"] == "deployer"
+    assert kwargs["work_item_id"] == "ORCH-115"
+    # The log written for the gate carries the matching verdict.
+    assert f"staging_status: {expected}" in _read_log(tmp_path, "orchestrator", "feature/ORCH-115-x", "ORCH-115")
+
+
+# ---------------------------------------------------------------------------
+# TC-08 — kill-switch: enabled=False -> LLM path via _spawn (AC-6)
+# ---------------------------------------------------------------------------
+def test_tc08_killswitch_falls_back_to_spawn(monkeypatch):
+    monkeypatch.setattr(cfg.settings, "staging_runner_enabled", False)
+    tid = _make_task("deploy-staging")
+    jid = _make_job("deployer", "orchestrator", tid)
+    lr = AgentLauncher()
+    spawn = MagicMock(return_value=7)
+    monkeypatch.setattr(lr, "_spawn", spawn)
+    run_gate = MagicMock()
+    monkeypatch.setattr(staging_runner, "run_staging_gate", run_gate)
+
+    ret = lr.launch_job(_job_dict(jid, "deployer", "orchestrator", tid))
+
+    assert ret == 7
+    spawn.assert_called_once()              # byte-for-byte the prior LLM-deployer path
+    run_gate.assert_not_called()
+
+
+# ---------------------------------------------------------------------------
+# TC-09 — anti-drift NFR-1: STAGE_TRANSITIONS / QG_CHECKS / schema untouched (AC-5)
+# ---------------------------------------------------------------------------
+def test_tc09_pipeline_contract_unchanged():
+    from src.stages import STAGE_TRANSITIONS
+    from src.qg.checks import QG_CHECKS
+    # The deploy-staging edge + its gate are byte-for-byte the prior contract.
+    assert STAGE_TRANSITIONS["deploy-staging"] == {
+        "next": "deploy", "agent": "deployer", "qg": "check_staging_status"
+    }
+    assert "check_staging_status" in QG_CHECKS
+    # No new ORCH-115 table/column: only the existing tables exist.
+    conn = get_db()
+    tables = {r[0] for r in conn.execute(
+        "SELECT name FROM sqlite_master WHERE type='table'").fetchall()}
+    conn.close()
+    assert not any("staging_runner" in t for t in tables)
+    # The machine key is not renamed: the runner emits `staging_status:`.
+    assert "staging_status:" in staging_runner.build_staging_log("ORCH-115", 0, "SUCCESS")
+
+
+# ---------------------------------------------------------------------------
+# TC-10 — never-raise / fail-safe: tool-error never a silent advance/false green (AC-7)
+# ---------------------------------------------------------------------------
+def test_tc10_nonzero_exit_is_failed_and_advances(monkeypatch, tmp_path):
+    tid = _make_task("deploy-staging")
+    monkeypatch.setattr(staging_runner, "run_staging_suite",
+                        lambda: ProcResult(returncode=3, stdout="fail", stderr="", timed_out=False))
+    advance = MagicMock()
+    monkeypatch.setattr(stage_engine, "advance_stage", advance)
+    staging_runner.run_staging_gate(_job_dict(1, "deployer", "orchestrator", tid))
+    advance.assert_called_once()
+    assert "staging_status: FAILED" in _read_log(tmp_path, "orchestrator", "feature/ORCH-115-x", "ORCH-115")
+
+
+def test_tc10_timeout_defers_without_advance(monkeypatch):
+    tid = _make_task("deploy-staging")
+    monkeypatch.setattr(cfg.settings, "staging_runner_infra_max_retries", 2)
+    monkeypatch.setattr(staging_runner, "run_staging_suite",
+                        lambda: ProcResult(returncode=None, stdout="", stderr="", timed_out=True))
+    advance = MagicMock()
+    monkeypatch.setattr(stage_engine, "advance_stage", advance)
+
+    staging_runner.run_staging_gate(_job_dict(1, "deployer", "orchestrator", tid))
+
+    advance.assert_not_called()             # NO silent advance / false green on infra hiccup
+    # A fresh deployer job was re-queued with the restart-safe marker (bounded DEFER).
+    conn = get_db()
+    n = conn.execute(
+        "SELECT COUNT(*) FROM jobs WHERE task_id=? AND task_content LIKE ?",
+        (tid, f"%{staging_runner._INFRA_RETRY_MARKER}%"),
+    ).fetchone()[0]
+    conn.close()
+    assert n == 1
+    assert staging_runner._STAGING_RUNNER_COUNTERS["deferred"] == 1
+    assert staging_runner._STAGING_RUNNER_COUNTERS["tool_error"] == 1
+
+
+def test_tc10_tool_error_budget_exhausted_fails_closed(monkeypatch, tmp_path):
+    tid = _make_task("deploy-staging")
+    monkeypatch.setattr(cfg.settings, "staging_runner_infra_max_retries", 0)  # exhausted immediately
+    monkeypatch.setattr(staging_runner, "run_staging_suite",
+                        lambda: ProcResult(returncode=None, stdout="", stderr="", timed_out=True))
+    advance = MagicMock()
+    monkeypatch.setattr(stage_engine, "advance_stage", advance)
+
+    staging_runner.run_staging_gate(_job_dict(1, "deployer", "orchestrator", tid))
+
+    advance.assert_called_once()            # fail-closed: write FAILED + advance (existing rollback)
+    assert "staging_status: FAILED" in _read_log(tmp_path, "orchestrator", "feature/ORCH-115-x", "ORCH-115")
+
+
+def test_tc10_run_gate_never_raises(monkeypatch):
+    tid = _make_task("deploy-staging")
+
+    def boom():
+        raise RuntimeError("docker exec exploded")
+    monkeypatch.setattr(staging_runner, "run_staging_suite", boom)
+    # Must NOT raise (AC-7): the worker is protected even on an unexpected error.
+    staging_runner.run_staging_gate(_job_dict(1, "deployer", "orchestrator", tid))
+
+
+def test_tc10_launcher_contains_runner_fault(monkeypatch):
+    tid = _make_task("deploy-staging")
+    jid = _make_job("deployer", "orchestrator", tid)
+    lr = AgentLauncher()
+    monkeypatch.setattr(lr, "_spawn", MagicMock())
+    monkeypatch.setattr(staging_runner, "run_staging_gate",
+                        MagicMock(side_effect=RuntimeError("kaboom")))
+    # The launcher wrapper contains the fault -> job marked failed, never raises.
+    ret = lr.launch_job(_job_dict(jid, "deployer", "orchestrator", tid))
+    assert ret is None
+    conn = get_db()
+    status = conn.execute("SELECT status FROM jobs WHERE id=?", (jid,)).fetchone()[0]
+    conn.close()
+    assert status == "failed"
+
+
+# ---------------------------------------------------------------------------
+# TC-11 — timeout resolution + propagation (NFR-4 / AC-9)
+# ---------------------------------------------------------------------------
+def test_tc11_resolve_timeout_default_on_bad_value(monkeypatch):
+    monkeypatch.setattr(cfg.settings, "staging_runner_timeout_s", 1234)
+    assert staging_runner._resolve_timeout() == 1234
+    for bad in (0, -5, "abc", None):
+        monkeypatch.setattr(cfg.settings, "staging_runner_timeout_s", bad)
+        assert staging_runner._resolve_timeout() == staging_runner._DEFAULT_TIMEOUT_S
+
+
+def test_tc11_timeout_passed_to_proc_group(monkeypatch):
+    monkeypatch.setattr(cfg.settings, "staging_runner_timeout_s", 321)
+    monkeypatch.setattr(cfg.settings, "subprocess_tree_kill_enabled", True)
+    captured = {}
+
+    def fake_run(cmd, *, cwd, timeout, grace_s, tree_kill):
+        captured.update(cmd=cmd, timeout=timeout, tree_kill=tree_kill)
+        return ProcResult(returncode=0, stdout="", stderr="", timed_out=False)
+    monkeypatch.setattr(staging_runner.proc_group, "run_in_process_group", fake_run)
+
+    staging_runner.run_staging_suite()
+    assert captured["timeout"] == 321
+    assert captured["tree_kill"] is True
+    assert captured["cmd"][:2] == ["docker", "exec"]
+
+
+# ---------------------------------------------------------------------------
+# TC-12 — self-hosting safety: no forbidden literals in the runner command (AC-8)
+# ---------------------------------------------------------------------------
+def test_tc12_command_has_no_dangerous_literals():
+    cmd = " ".join(staging_runner.build_staging_command())
+    forbidden = ("compose", "up -d", "--build", "8500", "force", "push", ".env",
+                 "rm ", "restart")
+    for token in forbidden:
+        assert token not in cmd, f"forbidden literal {token!r} in runner command: {cmd}"
+    # It IS the staging suite against 8501 via docker exec.
+    assert "docker exec" in cmd
+    assert "staging_check.py" in cmd
+    assert "8501" in cmd
+    assert "--mode stub" in cmd
+
+
+# ---------------------------------------------------------------------------
+# TC-13 — observability: /queue block + structured verdict log (AC-10)
+# ---------------------------------------------------------------------------
+def test_tc13_snapshot_shape():
+    snap = staging_runner.snapshot()
+    for k in ("enabled", "repos", "timeout_s", "infra_max_retries",
+              "runs", "success", "failed", "tool_error", "deferred"):
+        assert k in snap, f"snapshot missing key {k}"
+
+
+def test_tc13_queue_endpoint_includes_block(monkeypatch):
+    import asyncio
+    from src import main
+    payload = asyncio.run(main.queue())
+    assert "staging_runner" in payload
+    assert "enabled" in payload["staging_runner"]
+
+
+def test_tc13_structured_verdict_log_distinguishes_outcomes(monkeypatch, caplog, tmp_path):
+    tid = _make_task("deploy-staging")
+    monkeypatch.setattr(stage_engine, "advance_stage", MagicMock())
+    # code-pass
+    monkeypatch.setattr(staging_runner, "run_staging_suite",
+                        lambda: ProcResult(returncode=0, stdout="", stderr="", timed_out=False))
+    with caplog.at_level("INFO", logger="orchestrator.staging_runner"):
+        staging_runner.run_staging_gate(_job_dict(1, "deployer", "orchestrator", tid))
+    assert any("outcome=code-pass" in r.message for r in caplog.records)
+
+    caplog.clear()
+    # tool-error
+    monkeypatch.setattr(cfg.settings, "staging_runner_infra_max_retries", 2)
+    monkeypatch.setattr(staging_runner, "run_staging_suite",
+                        lambda: ProcResult(returncode=None, stdout="", stderr="", timed_out=True))
+    with caplog.at_level("WARNING", logger="orchestrator.staging_runner"):
+        staging_runner.run_staging_gate(_job_dict(1, "deployer", "orchestrator", tid))
+    assert any("outcome=tool-error" in r.message for r in caplog.records)
+
+
+def test_tc13_snapshot_never_raises(monkeypatch):
+    class Boom:
+        def __getattr__(self, name):
+            raise RuntimeError("boom")
+    monkeypatch.setattr(staging_runner, "settings", Boom())
+    snap = staging_runner.snapshot()
+    assert snap["enabled"] is False