Retire live model workflow E2E harnesses

.github/pull_request_template.md

@@ -40,7 +40,7 @@ Maintainer live check: no | yes
 
 Surface: N/A | provider | browser | gateway | channel | release
 
-Maintainer-only note: contributors are not expected to provide secrets or run credentialed live checks. Maintainers may run `Live Release E2E` for provider, browser, gateway, channel, or release smoke coverage.
+Maintainer-only note: contributors are not expected to provide secrets or run credentialed live checks. Maintainers may manually run `LLM Smoke`, `Live Search E2E`, or `Live Telegram Smoke` for model API connectivity, search providers, or Telegram delivery. These workflows do not provide full Agent or Gateway model E2E coverage.
 
 ## Safety
 

.github/scripts/windows_test_assignments.json

@@ -136,7 +136,6 @@
       "tests/test_execution_status_contract.py",
       "tests/test_git_runtime.py",
       "tests/test_github_issue_link_sync.py",
-      "tests/test_long_task_fault_proxy.py",
       "tests/test_lossless_toml.py",
       "tests/test_memory/test_profile_import.py",
       "tests/test_meta_skill_openclaw_lifestyle_comparison.py",
@@ -225,7 +224,6 @@
       "tests/test_sandbox/test_windows_default_runner.py",
       "tests/test_scripts/test_check_treatment_delivery.py",
       "tests/test_scripts/test_exp_quarantine.py",
-      "tests/test_scripts/test_live_meta_skill_creator_e2e.py",
       "tests/test_skill_pdf_toolkit.py",
       "tests/test_skills/test_awesome_webpage_meta_skill.py",
       "tests/test_skills/test_clarify_schema_protocol.py",
@@ -722,10 +720,6 @@
       "tests/test_identity/test_user_profile_prompt.py",
       "tests/test_identity/test_workspace_injection_scan.py",
       "tests/test_inference_postprocess_rules.py",
-      "tests/test_live_long_task_case_driver.py",
-      "tests/test_live_mixed_provider_gateway.py",
-      "tests/test_live_multi_provider_matrix.py",
-      "tests/test_live_provider_profile_gateway_e2e.py",
       "tests/test_mcp_server/test_bridge.py",
       "tests/test_mcp_server/test_gateway_client.py",
       "tests/test_memory/test_store_vec_extension_cleanup.py",
@@ -838,7 +832,6 @@
       "tests/test_scheduler/test_webhook_delivery.py",
       "tests/test_scripts/test_analyze_dashscope_payload_risk.py",
       "tests/test_scripts/test_check_docker_image_lock.py",
-      "tests/test_scripts/test_live_meta_soft_activation_e2e.py",
       "tests/test_scripts/test_meta_trigger_accuracy_script.py",
       "tests/test_scripts/test_refresh_models_dev_snapshot.py",
       "tests/test_scripts/test_replay_finalize_gate.py",
@@ -1317,10 +1310,9 @@
       "tests/test_engine/turn_runner/test_stream_consumer_stage_unit.py",
       "tests/test_engine/turn_runner/test_turn_finalizer_paused_im_layout.py",
       "tests/test_envelope_policy_deny_cap.py",
+      "tests/test_gateway/test_offline_document_workbench_e2e.py",
       "tests/test_gemini_thought_signature.py",
       "tests/test_install_scripts.py",
-      "tests/test_live_artifact_prompt_annotations_e2e.py",
-      "tests/test_live_long_task_release_gate.py",
       "tests/test_llm_ensemble_config.py",
       "tests/test_mcp/test_sse_client.py",
       "tests/test_mcp_server/test_cli.py",
@@ -1496,30 +1488,25 @@
       "reason": "Reviewed affinity exception for an environment-independent test; rebalance the measured Windows critical path using three comparable successful runs."
     },
     {
-      "path": "tests/test_live_long_task_case_driver.py",
-      "from": "gateway-sqlite",
-      "to": "recovery-migration",
-      "reason": "Rebalance the measured Windows critical path using three comparable successful runs."
-    },
-    {
-      "path": "tests/test_live_multi_provider_matrix.py",
+      "path": "tests/test_observability/test_bundle.py",
       "from": "gateway-sqlite",
       "to": "recovery-migration",
-      "reason": "Rebalance the measured Windows critical path using three comparable successful runs."
+      "affinity_exception": true,
+      "reason": "Retain the reviewed environment-independent affinity exception and balance the remaining historical Windows weights after retiring model E2E harnesses."
     },
     {
-      "path": "tests/test_observability/test_bundle.py",
+      "path": "tests/test_persistence/test_router_decision_writer.py",
       "from": "gateway-sqlite",
-      "to": "desktop-installer-contracts",
+      "to": "core",
       "affinity_exception": true,
       "reason": "Reviewed affinity exception for an environment-independent test; rebalance the measured Windows critical path using three comparable successful runs."
     },
     {
-      "path": "tests/test_persistence/test_router_decision_writer.py",
+      "path": "tests/test_persistence/test_turn_error_writer.py",
       "from": "gateway-sqlite",
-      "to": "core",
+      "to": "recovery-migration",
       "affinity_exception": true,
-      "reason": "Reviewed affinity exception for an environment-independent test; rebalance the measured Windows critical path using three comparable successful runs."
+      "reason": "Balance the remaining historical Windows weights after retiring model E2E harnesses; these synthetic SQLite tests need no shard-specific setup."
     },
     {
       "path": "tests/test_sandbox/test_windows_default_capability.py",

.github/scripts/windows_test_durations.json

@@ -791,16 +791,10 @@
     "tests/test_identity/test_workspace_injection_scan.py": 0.067,
     "tests/test_inference_postprocess_rules.py": 0.076,
     "tests/test_install_scripts.py": 6.582,
-    "tests/test_live_artifact_prompt_annotations_e2e.py": 0.01,
-    "tests/test_live_long_task_case_driver.py": 54.424,
-    "tests/test_live_long_task_release_gate.py": 0.23,
-    "tests/test_live_mixed_provider_gateway.py": 0.2,
-    "tests/test_live_multi_provider_matrix.py": 13.55,
-    "tests/test_live_provider_profile_gateway_e2e.py": 0.131,
+    "tests/test_gateway/test_offline_document_workbench_e2e.py": 0.01,
     "tests/test_live_provider_profile_smoke.py": 0.065,
     "tests/test_live_tokenrhythm_billing_audit.py": 0.146,
     "tests/test_llm_ensemble_config.py": 1.845,
-    "tests/test_long_task_fault_proxy.py": 4.571,
     "tests/test_lossless_toml.py": 0.094,
     "tests/test_mcp/test_discovery_lifecycle.py": 0.203,
     "tests/test_mcp/test_sse_client.py": 0.832,
@@ -1177,8 +1171,6 @@
     "tests/test_scripts/test_check_treatment_delivery.py": 0.073,
     "tests/test_scripts/test_exp_ledger.py": 0.01,
     "tests/test_scripts/test_exp_quarantine.py": 0.01,
-    "tests/test_scripts/test_live_meta_skill_creator_e2e.py": 5.755,
-    "tests/test_scripts/test_live_meta_soft_activation_e2e.py": 1.914,
     "tests/test_scripts/test_meta_skill_validation_matrix.py": 0.03,
     "tests/test_scripts/test_meta_trigger_accuracy_script.py": 0.016,
     "tests/test_scripts/test_prestage_release_to_oss.py": 12.831,

.github/workflows/live-release-e2e.yml

@@ -1,4 +1,4 @@
-name: Live Release E2E
+name: Live Telegram Smoke
 
 "on":
   workflow_dispatch:
@@ -18,16 +18,14 @@ concurrency:
 
 jobs:
   live-release-e2e:
-    name: Maintainer live release gates
+    name: Maintainer Telegram channel smoke
+    if: ${{ inputs.run_telegram_channel }}
     runs-on: ubuntu-latest
     timeout-minutes: 30
 
     env:
       PYTHONPATH: ${{ github.workspace }}
       OPENSQUILLA_TURN_CALL_LOG: "0"
-      OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
-      OPENROUTER_BASE_URL: ${{ vars.OPENROUTER_BASE_URL || 'https://openrouter.ai/api/v1' }}
-      LLM_TEST_MODEL: ${{ vars.LLM_TEST_MODEL || 'openai/gpt-4o-mini' }}
       OPENSQUILLA_LIVE_TELEGRAM_BOT_TOKEN: ${{ secrets.OPENSQUILLA_LIVE_TELEGRAM_BOT_TOKEN }}
       OPENSQUILLA_LIVE_TELEGRAM_CHAT_ID: ${{ secrets.OPENSQUILLA_LIVE_TELEGRAM_CHAT_ID }}
 
@@ -52,16 +50,7 @@ jobs:
       - name: Install dependencies
         run: uv sync --extra dev --extra recommended --frozen
 
-      - name: Fail if OpenRouter secret is missing
-        shell: bash
-        run: |
-          if [ -z "$OPENROUTER_API_KEY" ]; then
-            echo "OPENROUTER_API_KEY GitHub secret is required"
-            exit 1
-          fi
-
       - name: Fail if Telegram secrets are missing when channel smoke is enabled
-        if: ${{ inputs.run_telegram_channel }}
         shell: bash
         run: |
           if [ -z "$OPENSQUILLA_LIVE_TELEGRAM_BOT_TOKEN" ]; then
@@ -73,13 +62,7 @@ jobs:
             exit 1
           fi
 
-      - name: Run gateway LLM e2e
-        env:
-          OPENSQUILLA_GATEWAY_LLM_E2E: "1"
-        run: uv run pytest tests/functional/test_gateway_llm_e2e.py -q -s
-
       - name: Run live Telegram channel smoke
-        if: ${{ inputs.run_telegram_channel }}
         env:
           OPENSQUILLA_LIVE_TELEGRAM_E2E: "1"
         run: uv run pytest tests/functional/test_live_channel_telegram_smoke.py -q -s

.github/workflows/live-search-e2e.yml

@@ -18,17 +18,14 @@ concurrency:
 
 jobs:
   live-search-e2e:
-    name: Live search provider and agent gates
+    name: Live search provider gates
     runs-on: ubuntu-latest
     timeout-minutes: 30
 
     env:
       PYTHONPATH: ${{ github.workspace }}
       OPENSQUILLA_TURN_CALL_LOG: "0"
       OPENSQUILLA_LIVE_SEARCH: "1"
-      OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
-      OPENROUTER_BASE_URL: ${{ vars.OPENROUTER_BASE_URL || 'https://openrouter.ai/api/v1' }}
-      LLM_TEST_MODEL: ${{ vars.LLM_TEST_MODEL || 'openai/gpt-4o-mini' }}
       TAVILY_API_KEY: ${{ secrets.TAVILY_API_KEY }}
       BRAVE_SEARCH_API_KEY: ${{ secrets.BRAVE_SEARCH_API_KEY }}
       EXA_API_KEY: ${{ secrets.EXA_API_KEY }}
@@ -58,10 +55,6 @@ jobs:
       - name: Fail if required live search secrets are missing
         shell: bash
         run: |
-          if [ -z "$OPENROUTER_API_KEY" ]; then
-            echo "OPENROUTER_API_KEY GitHub secret is required"
-            exit 1
-          fi
           if [ -z "$TAVILY_API_KEY" ]; then
             echo "TAVILY_API_KEY GitHub secret is required"
             exit 1
@@ -86,7 +79,6 @@ jobs:
         run: |
           timeout 180s uv run pytest \
             tests/live/test_search_retrieval_live.py \
-            tests/live/test_web_search_agent_e2e.py \
             -m live_search \
             -q -s
 

CONTRIBUTING.md

@@ -85,7 +85,11 @@ Default tests must be offline, deterministic, credential-free, and safe for fork
 
 Add or update public regression tests for behavior changes and bug fixes. Prefer focused unit or integration tests unless the behavior crosses the gateway, browser UI, provider, or channel boundary.
 
-Live checks are maintainer-only gates. The `Live Release E2E` workflow covers real provider, browser, and optional channel smoke tests with GitHub secrets and explicit opt-in inputs.
+Live checks are maintainer-only gates. The manually triggered `LLM Smoke`,
+`Live Search E2E`, and `Live Telegram Smoke` workflows cover model API
+connectivity, search providers, and Telegram delivery respectively. They use
+GitHub secrets and explicit opt-in inputs; they do not provide full Agent or
+Gateway model E2E coverage.
 
 ## Private Materials
 

desktop/electron/scripts/test-offline-document-workbench-e2e.mjs

@@ -50,7 +50,7 @@ await run(
     'run',
     'pytest',
     '-q',
-    'tests/test_live_artifact_prompt_annotations_e2e.py::test_owned_gateway_html_workbench_lifecycle_is_offline_and_immutable',
+    'tests/test_gateway/test_offline_document_workbench_e2e.py::test_owned_gateway_html_workbench_lifecycle_is_offline_and_immutable',
   ],
 )
 

docs/authoring/meta-skills.md

@@ -438,10 +438,7 @@ Before sharing or enabling a MetaSkill:
    reflect the workflow's true side effects.
 7. If supporting `meta_skill.auto_trigger = true`, run deterministic trigger
    checks with `scripts/meta_trigger_accuracy.py`.
-8. If supporting `meta_skill.auto_trigger = true`, run model-decision soft
-   activation checks with
-   `scripts/live_meta_soft_activation_e2e.py --env-file /path/to/.env`.
-9. For generated skills, inspect the Web UI proposal detail and its auto-enable
+8. For generated skills, inspect the Web UI proposal detail and its auto-enable
    audit before accepting or enabling.
 
 ## Troubleshooting

docs/features/prompt-annotation-editing.md

@@ -377,82 +377,15 @@ write. A discarded or interrupted candidate creates no revision. It is
 credential-free and requires Electron foreground focus;
 a locked or background-only macOS session fails the gate.
 
-### Live provider certification
-
-Keep provider credentials only in the process environment or an ignored local
-environment file. Never put a key in a command, fixture, report, or committed
-configuration. A sanitized provider/Gateway prerequisite can be run with:
-
-```sh
-uv run python scripts/live_provider_profile_gateway_e2e.py \
-  --providers tokenrhythm \
-  --output "${TMPDIR:?}/opensquilla-provider-gateway.json"
-```
-
-That script certifies provider transport and accounting; it is not by itself
-PromptAnnotation certification. For every release that changes this path, an
-isolated owned Gateway and Desktop build must additionally pass this live matrix:
-
-The dedicated certification boundary can be checked with:
-
-```sh
-uv run python scripts/live_artifact_prompt_annotations_e2e.py \
-  --output "${TMPDIR:?}/opensquilla-prompt-annotations.json" \
-  --confirm-live-cost \
-  --confirm-rotated-key \
-  --execute-live-matrix
-```
-
-Without `--execute-live-matrix`, the command is an intentional zero-call dry
-run that writes `certification=incomplete`. With the flag, the isolated worker
-runs the owned Gateway/provider path and requires the ten-tool document-agent surface.
-It does not replace the separate real-Electron selection gate.
-The fixed Direct `glm-5.2` source-fallback case must exercise a repair loop:
-`document_inspect → document_read → document_patch →
-document_browser_inspect → document_browser_screenshot` (or a bounded browser
-action) `→ document_patch → document_browser_inspect → document_finish(commit)
-→ tools=[]` finalization. The model chooses whether to take each observation,
-repair, or discard step; the harness checks this representative sequence but
-does not impose a PromptAnnotation-specific iteration count. The live harness
-therefore uses a 120-second per-case deadline; this is a task deadline, not a
-provider retry. A B5 Ensemble case uses the configured proposer/Aggregator
-rounds while keeping proposer and finalizer tools empty. The outcome-finalization
-round is required Agent work; it must not be removed or treated as a free call.
-The approved autonomous-loop matrix therefore reserves 42 baseline physical
-provider calls, allows a bounded worst-case reservation of 63, and hard-stops
-raw provider traffic at 64. Preview, browser action, `document_finish`, and
-tools-disabled finalizer requests are all included in that physical count;
-none is deducted or treated as free work.
-A build is not release-ready until the following matrix is
-verified end to end:
-
-- Direct: one source-fallback insertion and a two-annotation semantic batch
-  using fixed model `glm-5.2`;
-- Router: a single selection at `c2`, a structural batch at `c3`, and no
-  fallback below the effective floor;
-- Ensemble: the full configured lineup, zero executable proposer tool calls,
-  an Aggregator-owned candidate loop, and one final `document_finish(commit)`;
-- rejection cases for stale head, cross-session draft, DOM mismatch, and visual
-  selection, each with zero provider calls;
-- exactly one revision and one change set for every successful batch, plus a
-  successful whole-turn revert.
-
-Store only case name, mode/tier/model, tool name and count, content hashes, and
-boolean results. Scan the report and temporary directory for credentials and
-delete the temporary data after review. If this feature-specific live matrix
-has not been completed, set the release override for
-`artifactPromptAnnotations` to `false`.
-
 ## Release, rollback, and maintenance
 
 For each release:
 
 1. Run the offline, packaged-wheel, Web UI, and real Electron suites against
    the release candidate.
-2. Complete the live Direct/Router/Ensemble matrix with an isolated profile.
-3. Canary the exact Desktop build and watch sanitized audit events,
+2. Canary the exact Desktop build and watch sanitized audit events,
    annotation remap outcomes, validation failures, and orphan cleanup.
-4. Keep the default enabled only while the one-turn/one-change-set, zero-call
+3. Keep the default enabled only while the one-turn/one-change-set, zero-call
    rejection, and Aggregator-only mutation invariants remain true.
 
 For an incident, apply the explicit `artifactPromptAnnotations: false` override