clean the unused part

Author: ZFancy

prototype/agingbench/telemetry/README.md

@@ -89,7 +89,7 @@ extractors push it to L3 without extra integration.
 ```python
 from agingbench.telemetry import list_supported_formats
 list_supported_formats()
-# ['claude_code', 'generic', 'langfuse', 'langsmith',
+# ['claude_code', 'generic', 'langfuse',
 #  'openai_assistants', 'openhands', 'otlp']
 ```
 
@@ -108,9 +108,13 @@ list_supported_formats()
 | `openai_assistants` | `thread.message` / `thread.run` / `thread.run.step` objects |
 | `openhands` | OpenHands SDK event log (`source`, `action`, `observation`, `llm_metrics`) |
 | `langfuse` | Langfuse SDK exports or REST-API JSON (camelCase or snake_case) |
-| `langsmith` | LangSmith run JSON — routed through the generic adapter |
 | `otlp` | OTLP JSON spans (`gen_ai.*` semconv + legacy `llm.*` namespace) |
 
+LangSmith run JSON works today via `trace_format="generic"` — its field
+shape is covered by the generic adapter's aliasing. A dedicated
+`langsmith` format will be added once we ship a fixture + adapter-level
+test for it.
+
 Each adapter normalises into the canonical `TelemetryRecord`; all
 downstream inference is format-agnostic, so the parse-tested adapters
 work today if you already have a JSONL in the expected shape — what's
@@ -145,15 +149,14 @@ empty list, not a crash.
 ## Deployment profiles
 
 A profile encodes domain conventions (outcome-extraction rules,
-subject-linkage, mechanism weights, privacy patterns, session
-detection):
+default privacy patterns, session-detection defaults):
 
 ```python
 from agingbench.telemetry import list_profiles, load_profile
 list_profiles()                     # ['code_assistant', 'generic']
 p = load_profile("code_assistant")
 p.outcome_rules                     # {'pr_merged': 'success', ...}
-p.mechanism_weights                 # {'compression': 0.8, 'revision': 1.5, ...}
+p.privacy_patterns                  # [{'pattern': 'AKIA...', 'replacement': '[AWS_ACCESS_KEY]'}, ...]
 ```
 
 Override per call:
@@ -362,25 +365,29 @@ extractors, deployment profiles, synthetic-probe orchestrator, ASCII
 card renderer, `prepare_trace` preprocessor) are shipped and covered
 by the test suite (~85 telemetry-specific tests across
 `test_telemetry_adapters.py`, `test_telemetry_stub.py`,
-`test_telemetry_v11.py`; 143+ total `prototype/tests/`).
+`test_telemetry_v11.py`; 228+ tests total in `prototype/tests/`).
 
 **Trace-format coverage in this release**: Claude Code is verified
 end-to-end on real production traces; the `generic` adapter is
-verified against fixture data. The remaining five adapters
-(`openai_assistants`, `openhands`, `langfuse`, `langsmith`, `otlp`)
-pass adapter-level tests against shipped fixtures, but their
-*extraction recipes* — the steps needed to dump a JSONL of the right
-shape from each live third-party SDK — have not been validated against
-current SDK versions and are tracked as future work.
+verified against fixture data. The remaining four adapters
+(`openai_assistants`, `openhands`, `langfuse`, `otlp`) pass
+adapter-level tests against shipped fixtures, but their *extraction
+recipes* — the steps needed to dump a JSONL of the right shape from
+each live third-party SDK — have not been validated against current
+SDK versions and are tracked as future work. LangSmith run JSON is
+routable today via `trace_format="generic"`.
 
 ## Roadmap
 
 | Milestone | Scope |
 |---|---|
-| **Next** | End-to-end validation of the five parse-tested adapters against current SDKs (`openai_assistants`, `openhands`, `langfuse`, `langsmith`, `otlp`), promoting each to "verified" as it lands. More outcome extractors (GitHub Actions CI status, Langfuse score events, Slack reactions). Validation correlation study against scenario-derived metrics. |
+| **Next** | End-to-end validation of the four parse-tested adapters against current SDKs (`openai_assistants`, `openhands`, `langfuse`, `otlp`), promoting each to "verified" as it lands. Add a dedicated `langsmith` format (currently routes through `generic`). More outcome extractors (GitHub Actions CI status, Langfuse score events, Slack reactions). Validation correlation study against scenario-derived metrics. |
 | **Later** | Cross-tenant aggregation with differential privacy. Streaming ingestion. Native protobuf OTLP. |
 | **v2** | Multilingual user-correction detection. Workspace-fidelity inference for self-planning agents (S5). |
 
 Contributing a validated recipe for one of the parse-tested adapters
-is the highest-leverage way to widen format coverage — see the
-top-level `docs/CONTRIBUTING.md`.
+is the highest-leverage way to widen format coverage. The top-level
+`docs/CONTRIBUTING.md` covers SUT YAMLs and integration adapters;
+telemetry-adapter recipes can be contributed by adding a fixture under
+`example_traces/`, a `normalize()` implementation under `adapters/`,
+and registering the format in `adapters/__init__.py`.

prototype/agingbench/telemetry/adapters/__init__.py

@@ -16,6 +16,10 @@
 
 
 # format name → normaliser
+# NOTE: `langsmith` routes through the generic adapter (no dedicated
+# normaliser). Kept in the registry for backward compat, but NOT
+# advertised as a first-class format in the README until we ship a
+# dedicated fixture + adapter-level test for it.
 ADAPTERS: dict[str, Callable[[dict], Optional[TelemetryRecord]]] = {
     "generic":            generic.normalize,
     "claude_code":        claude_code.normalize,

prototype/agingbench/telemetry/profiles/__init__.py

@@ -3,8 +3,6 @@
 
 A profile encodes (per deployment type):
   - default outcome-extraction rules
-  - subject-linkage rules
-  - mechanism-inference weights
   - default privacy patterns
   - session-detection defaults
 
@@ -25,8 +23,6 @@
 class Profile:
     deployment_type:        str
     outcome_rules:          dict = field(default_factory=dict)
-    subject_linkage:        dict = field(default_factory=dict)
-    mechanism_weights:      dict = field(default_factory=dict)
     session_detection:      dict = field(default_factory=dict)
     privacy_patterns:       list = field(default_factory=list)
     raw:                    dict = field(default_factory=dict)
@@ -44,8 +40,6 @@ def load_profile(name: str = "generic") -> Profile:
     return Profile(
         deployment_type=doc.get("deployment_type", name),
         outcome_rules=doc.get("outcome_rules", {}) or {},
-        subject_linkage=doc.get("subject_linkage", {}) or {},
-        mechanism_weights=doc.get("mechanism_weights", {}) or {},
         session_detection=doc.get("session_detection", {}) or {},
         privacy_patterns=doc.get("privacy_patterns", []) or [],
         raw=doc,
@@ -63,8 +57,6 @@ def merge_overrides(profile: Profile, overrides: Optional[dict]) -> Profile:
     return Profile(
         deployment_type=profile.deployment_type,
         outcome_rules={**profile.outcome_rules, **overrides.get("outcome_rules", {})},
-        subject_linkage={**profile.subject_linkage, **overrides.get("subject_linkage", {})},
-        mechanism_weights={**profile.mechanism_weights, **overrides.get("mechanism_weights", {})},
         session_detection={**profile.session_detection, **overrides.get("session_detection", {})},
         privacy_patterns=profile.privacy_patterns + (overrides.get("privacy_patterns") or []),
         raw=profile.raw,

prototype/agingbench/telemetry/profiles/code_assistant.yaml

@@ -15,20 +15,6 @@ outcome_rules:
   completion_rejected:          user_rejected
   abort:                        abandoned
 
-subject_linkage:
-  # In a code workflow, the natural "subject" is the file or function.
-  primary: file_path
-  secondary: [function_name, class_name, pr_id]
-
-mechanism_weights:
-  # In code-assistant deployments revision matters more (consistency
-  # across edits to the same file); pure-compression matters less since
-  # the file system absorbs state.
-  compression:   0.8
-  interference:  1.0
-  revision:      1.5
-  maintenance:   1.0
-
 session_detection:
   primary: explicit_session_id    # Claude Code, OpenHands provide this
   fallback_idle_gap_minutes: 60   # coding sessions can have long pauses

prototype/agingbench/telemetry/profiles/generic.yaml

@@ -14,16 +14,6 @@ outcome_rules:
   user_thumbs_up:    success
   user_thumbs_down:  fail
 
-subject_linkage:
-  # Default: no cross-session subject linkage.
-  primary: null
-
-mechanism_weights:
-  compression:   1.0
-  interference:  1.0
-  revision:      1.0
-  maintenance:   1.0
-
 session_detection:
   primary: explicit_session_id
   fallback_idle_gap_minutes: 30

prototype/agingbench/telemetry/schema.py

@@ -126,7 +126,7 @@ class TraceAuditBlock:
     deployment_type:        str = "generic"
     n_sessions_detected:    int = 0
     n_outcome_events:       int = 0
-    session_detection_mode: str = "idle_gap"   # 'explicit_id'|'idle_gap'|'user_id_split'
+    session_detection_mode: str = "idle_gap"   # 'explicit_id'|'idle_gap'
     outcome_rules_hash:     Optional[str] = None
 
     # Per-mechanism sub-dicts — populated by inference modules.

prototype/agingbench/telemetry/session_detection.py

@@ -4,13 +4,11 @@
 Strategy (in priority order):
   1. Explicit session_id field (Langfuse, OpenAI threads, Claude Code)
   2. Explicit reset markers in user messages (/clear, /reset)
-  3. user_id grouping + idle-gap split
-  4. Pure idle-gap split (last resort)
+  3. Pure idle-gap split (last resort)
 """
 from __future__ import annotations
 
 from collections import defaultdict
-from typing import Optional
 
 from .schema import TelemetryRecord
 
@@ -21,7 +19,6 @@
 def detect_sessions(
     records: list[TelemetryRecord],
     idle_gap_minutes: float = 30.0,
-    user_id_field: Optional[str] = None,
 ) -> tuple[list[list[TelemetryRecord]], str]:
     """Return (sessions, mode) where sessions is list of lists and mode is the
     detection strategy that fired.
@@ -38,17 +35,6 @@ def detect_sessions(
         sessions = [sorted(g, key=lambda r: r.timestamp) for g in groups.values()]
         return _stable_session_order(sessions), "explicit_id"
 
-    # Strategy 2 + 4: idle-gap split (with optional user-id pre-grouping).
-    if user_id_field:
-        by_user = defaultdict(list)
-        for r in records:
-            uid = (r.user_id_hash or r.raw.get(user_id_field) or "_anon")
-            by_user[uid].append(r)
-        sessions = []
-        for uid, recs in by_user.items():
-            sessions.extend(_split_by_idle_gap(recs, idle_gap_minutes))
-        return _stable_session_order(sessions), "user_id_split"
-
     sessions = _split_by_idle_gap(records, idle_gap_minutes)
     return sessions, "idle_gap"
 

prototype/agingbench/telemetry/trace_to_card.py

@@ -24,7 +24,12 @@
 from typing import Any, Optional
 
 
-# Trace formats the v1 stub knows about.
+# Trace formats the v1 stub knows about. NOTE: `langsmith` is included
+# (silently routes through the `generic` adapter) for backward compat,
+# but is intentionally NOT advertised as a first-class format in the
+# README — it ships without a dedicated fixture/test. Users with
+# LangSmith run JSON should prefer `trace_format="generic"` going
+# forward; a dedicated langsmith adapter + fixture is on the roadmap.
 SUPPORTED_TRACE_FORMATS = (
     "langfuse", "langsmith", "otlp", "generic", "claude_code",
     "openai_assistants", "openhands",

prototype/tests/test_dry_run.py prototype/tests/dry_run.pyprototype/tests/test_dry_run.py renamed to prototype/tests/dry_run.py

@@ -97,7 +97,7 @@ def test_validate_card_path_missing_file(tmp_path):
 
 def test_sample_card_validates():
     """The fixture sample card (committed in examples/sample_cards/) must validate."""
-    sample = Path(__file__).parent.parent / "examples" / "sample_cards" / "s1_haiku45_lossy_compress.json"
+    sample = Path(__file__).parent.parent / "examples" / "sample_cards" / "s1_research_literature_haiku45_lossy_compress.json"
     if not sample.is_file():
         pytest.skip("sample card not present")
     errors = validate_card_path(sample)