coderouter-cli 2.1.0__py3-none-any.whl → 2.2.0__py3-none-any.whl

coderouter/cli.py

@@ -196,6 +196,104 @@ def _build_parser() -> argparse.ArgumentParser:
         help="Print one snapshot as plain text and exit (scripts / non-tty).",
     )
 
+    # v2.0-K: `coderouter audit` — read structured JSONL audit log.
+    audit = sub.add_parser(
+        "audit",
+        help="Read the structured audit log (v2.0-K).",
+        description=(
+            "Read and filter the JSONL audit log written by `coderouter serve` "
+            "when state_dir and audit_log are configured. Shows guard activations, "
+            "chain fallbacks, budget warnings, self-healing events, and drift "
+            "transitions in chronological order."
+        ),
+    )
+    audit.add_argument(
+        "--state-dir",
+        default=None,
+        help=(
+            "Path to the state directory containing audit.jsonl. "
+            "Defaults to ~/.coderouter/state/."
+        ),
+    )
+    audit.add_argument(
+        "--tail",
+        type=int,
+        default=None,
+        metavar="N",
+        help="Show only the last N entries.",
+    )
+    audit.add_argument(
+        "--filter",
+        default=None,
+        metavar="EVENT",
+        help="Only entries whose event name contains this substring (case-insensitive).",
+    )
+    audit.add_argument(
+        "--since",
+        default=None,
+        metavar="DATETIME",
+        help="Only entries with ts >= this ISO 8601 prefix (e.g. '2026-05-06').",
+    )
+    audit.add_argument(
+        "--summary",
+        action="store_true",
+        help="Print event type → count summary instead of individual entries.",
+    )
+
+    # v2.0-K (Replay): `coderouter replay` — statistical A/B analysis
+    # of request journal metadata across providers.
+    replay = sub.add_parser(
+        "replay",
+        help="Statistical replay analysis of request journal (v2.0-K).",
+        description=(
+            "Read the request metadata journal and display per-provider "
+            "statistics (token counts, cost, cache hit ratios). Optionally "
+            "compare two providers side-by-side. Request/response bodies "
+            "are not recorded, so this is statistical analysis — not "
+            "literal re-execution."
+        ),
+    )
+    replay.add_argument(
+        "--state-dir",
+        default=None,
+        help=(
+            "Path to the state directory containing requests.jsonl. "
+            "Defaults to ~/.coderouter/state/."
+        ),
+    )
+    replay.add_argument(
+        "--log",
+        default=None,
+        metavar="PATH",
+        help="Direct path to the request journal JSONL file (overrides --state-dir).",
+    )
+    replay.add_argument(
+        "--provider",
+        default=None,
+        metavar="NAME",
+        help="Filter entries to this provider only.",
+    )
+    replay.add_argument(
+        "--compare",
+        nargs=2,
+        metavar=("A", "B"),
+        default=None,
+        help="Compare two providers side-by-side (e.g. --compare anthropic-api openrouter-free).",
+    )
+    replay.add_argument(
+        "--since",
+        default=None,
+        metavar="DATETIME",
+        help="Only entries with ts >= this ISO 8601 prefix (e.g. '2026-05-06').",
+    )
+    replay.add_argument(
+        "--limit",
+        type=int,
+        default=None,
+        metavar="N",
+        help="Use only the last N entries (applied after --since and --provider filters).",
+    )
+
     return parser
 
 
@@ -277,6 +375,12 @@ def main(argv: list[str] | None = None) -> int:
 
         return stats_main(args.url, interval=args.interval, once=args.once)
 
+    if args.command == "audit":
+        return _run_audit(args)
+
+    if args.command == "replay":
+        return _run_replay(args)
+
     print(f"unknown command: {args.command}", file=sys.stderr)
     return 2
 
@@ -476,6 +580,121 @@ def _run_apply_or_dry_run(
 
 
 
+def _run_audit(args: argparse.Namespace) -> int:
+    """v2.0-K: read and display the structured audit log.
+
+    Resolves the audit log path from --state-dir (or default
+    ~/.coderouter/state/) and renders entries with optional filtering.
+    """
+    import json
+
+    from coderouter.state.audit_log import read_audit_log, summarize_audit_log
+
+    state_dir = Path(args.state_dir).expanduser() if args.state_dir else (
+        Path.home() / ".coderouter" / "state"
+    )
+    log_path = state_dir / "audit.jsonl"
+
+    if not log_path.exists():
+        print(f"audit: no audit log found at {log_path}", file=sys.stderr)
+        print(
+            "  Ensure state_dir and audit_log are configured in providers.yaml.",
+            file=sys.stderr,
+        )
+        return 1
+
+    entries = read_audit_log(
+        log_path,
+        tail=args.tail,
+        event_filter=args.filter,
+        since=args.since,
+    )
+
+    if not entries:
+        print("audit: no matching entries found.")
+        return 0
+
+    if args.summary:
+        summary = summarize_audit_log(entries)
+        print(f"Audit log summary ({len(entries)} entries):\n")
+        for event, count in summary.items():
+            print(f"  {event:<40s} {count:>6d}")
+        return 0
+
+    for entry in entries:
+        ts = entry.get("ts", "")
+        event = entry.get("event", "")
+        level = entry.get("level", "")
+        # Build a compact one-line display.
+        extras = {
+            k: v
+            for k, v in entry.items()
+            if k not in ("ts", "event", "level")
+        }
+        extra_str = ""
+        if extras:
+            extra_str = " " + json.dumps(extras, default=str, ensure_ascii=False)
+        print(f"[{ts}] {level:<7s} {event}{extra_str}")
+
+    return 0
+
+
+def _run_replay(args: argparse.Namespace) -> int:
+    """v2.0-K (Replay): statistical A/B analysis of request journal.
+
+    Reads the request journal (requests.jsonl) and either displays a
+    per-provider summary table or a side-by-side comparison of two
+    providers.
+    """
+    from coderouter.state.replay import (
+        compare_providers,
+        format_comparison_table,
+        format_summary_table,
+        summarize_window,
+    )
+    from coderouter.state.request_log import read_request_log
+
+    # Resolve the journal file path.
+    if args.log:
+        log_path = Path(args.log).expanduser()
+    else:
+        state_dir = Path(args.state_dir).expanduser() if args.state_dir else (
+            Path.home() / ".coderouter" / "state"
+        )
+        log_path = state_dir / "requests.jsonl"
+
+    if not log_path.exists():
+        print(f"replay: no request journal found at {log_path}", file=sys.stderr)
+        print(
+            "  Ensure state_dir and request_log are configured in providers.yaml.",
+            file=sys.stderr,
+        )
+        return 1
+
+    entries = read_request_log(
+        log_path,
+        provider_filter=args.provider,
+        since=args.since,
+    )
+
+    if args.limit is not None and args.limit > 0:
+        entries = entries[-args.limit:]
+
+    if not entries:
+        print("replay: no matching entries found.")
+        return 0
+
+    if args.compare:
+        provider_a, provider_b = args.compare
+        comparison = compare_providers(entries, provider_a, provider_b)
+        print(format_comparison_table(comparison))
+    else:
+        summary = summarize_window(entries)
+        print(format_summary_table(summary))
+
+    return 0
+
+
 def _run_check_env(arg_value: str) -> int:
     """v1.6.3: filesystem / git security checks for `.env`.
 

coderouter/config/schemas.py

@@ -242,6 +242,21 @@ class ProviderConfig(BaseModel):
             "LM Studio Qwen3.5 128K → 131072, Anthropic Claude → 200000."
         ),
     )
+    # v2.0-J: optional shell command to restart this provider's backend
+    # process when it becomes UNHEALTHY. Executed via subprocess when
+    # self-healing is enabled and the provider crosses the UNHEALTHY
+    # threshold. Security: opt-in only — unset means no restart attempt.
+    restart_command: str | None = Field(
+        default=None,
+        description=(
+            "v2.0-J (Self-healing): shell command to restart this "
+            "provider's backend process. Examples: 'ollama serve', "
+            "'open -a LM\\ Studio'. Only executed when the profile's "
+            "backend_health_action is 'exclude' and the provider "
+            "transitions to UNHEALTHY. Unset = no automatic restart "
+            "(recovery probe still runs, waiting for manual restart)."
+        ),
+    )
 
     @model_validator(mode="after")
     def _check_output_filters_known(self) -> ProviderConfig:
@@ -350,6 +365,24 @@ class FallbackChain(BaseModel):
             "error response. See FallbackChain comment for trade-offs."
         ),
     )
+    # v2.2: total tool-call count hard cap. A safety valve against
+    # runaway agents that call many *different* tools without looping
+    # (which the streak-based L3 detector misses). Set to 0 to
+    # disable the cap entirely.
+    max_tool_calls: int = Field(
+        default=50,
+        ge=0,
+        le=1000,
+        description=(
+            "v2.2: maximum total tool_use blocks allowed in the "
+            "conversation. When exceeded, the request is rejected with "
+            "a ``tool_count_exceeded`` error (if tool_loop_action is "
+            "``break``) or logged (if ``warn``). Set to 0 to disable. "
+            "Default 50 is deliberately more permissive than Unsloth "
+            "Studio's 25 — Claude Code agent sessions routinely reach "
+            "25+ calls in normal operation."
+        ),
+    )
     # v1.9-E phase 2 (L2): memory-pressure detection + cooldown.
     #
     # Local backends (Ollama / LM Studio / llama.cpp) report VRAM
@@ -410,7 +443,7 @@ class FallbackChain(BaseModel):
     # Distinct from the v1.9-C ``adaptive`` gradient (continuous
     # latency / error-rate buffer with debounce) which handles the
     # "slow but alive" case; L5 handles the "hard crash" case.
-    backend_health_action: Literal["off", "warn", "demote"] = Field(
+    backend_health_action: Literal["off", "warn", "demote", "exclude"] = Field(
         default="warn",
         description=(
             "v1.9-E (L5 phase 2): action when a provider transitions "
@@ -420,7 +453,12 @@ class FallbackChain(BaseModel):
             "moves the UNHEALTHY provider to the back of the chain "
             "for the next ``_resolve_chain`` (similar to v1.9-C "
             "adaptive demotion but state-machine-based, not "
-            "rolling-window-based). ``off`` disables the monitor "
+            "rolling-window-based). ``exclude`` (v2.0-J) removes the "
+            "UNHEALTHY provider from the chain entirely + triggers "
+            "self-healing (restart helper if configured, recovery "
+            "probe with exponential backoff). On recovery, the "
+            "provider is automatically restored to its original "
+            "chain position. ``off`` disables the monitor "
             "entirely (zero observation overhead, identical to "
             "v1.9.x behavior)."
         ),
@@ -439,6 +477,41 @@ class FallbackChain(BaseModel):
             "blips that the v1.9-C adaptive adjuster already handles."
         ),
     )
+    # v2.0-J: self-healing recovery probe configuration.
+    recovery_probe_initial_s: float = Field(
+        default=30.0,
+        ge=5.0,
+        le=600.0,
+        description=(
+            "v2.0-J: initial interval (seconds) for recovery probes "
+            "sent to an UNHEALTHY-excluded provider. Each failed probe "
+            "doubles the interval up to ``recovery_probe_max_s``. "
+            "A successful probe restores the provider to its original "
+            "chain position immediately."
+        ),
+    )
+    recovery_probe_max_s: float = Field(
+        default=300.0,
+        ge=30.0,
+        le=3600.0,
+        description=(
+            "v2.0-J: maximum interval (seconds) for recovery probe "
+            "exponential backoff. Default 300 s (5 min) means a dead "
+            "backend is probed at most every 5 minutes indefinitely "
+            "until it recovers or the server shuts down."
+        ),
+    )
+    restart_timeout_s: float = Field(
+        default=30.0,
+        ge=5.0,
+        le=120.0,
+        description=(
+            "v2.0-J: timeout (seconds) for the restart_command "
+            "subprocess. If the command doesn't complete within this "
+            "window, it is killed. Prevents hung restart commands from "
+            "blocking recovery."
+        ),
+    )
     adaptive: bool = Field(
         default=False,
         description=(
@@ -871,6 +944,63 @@ class CodeRouterConfig(BaseModel):
         ),
     )
 
+    # v2.0-K: Persistent state — survive restarts.
+    state_dir: str | None = Field(
+        default=None,
+        description=(
+            "v2.0-K: directory for persistent state (sqlite3 KV store + "
+            "audit log). None = in-memory only (no persistence, backward-"
+            "compatible). Set to a path like '~/.coderouter/state/' to "
+            "enable cross-restart durability for budget totals, health "
+            "state, and self-healing exclusions. The directory is created "
+            "automatically if it doesn't exist."
+        ),
+    )
+    audit_log: Literal["off", "active"] = Field(
+        default="off",
+        description=(
+            "v2.0-K: structured audit log. 'active' writes guard "
+            "activations, chain fallbacks, budget warnings, self-healing "
+            "events, and drift transitions to a JSONL file in state_dir. "
+            "'off' = no audit log (backward-compatible default). Requires "
+            "state_dir to be set."
+        ),
+    )
+    audit_log_max_bytes: int = Field(
+        default=10_485_760,
+        ge=1_048_576,
+        le=1_073_741_824,
+        description=(
+            "v2.0-K: maximum audit log file size before rotation (bytes). "
+            "Default 10 MiB. When exceeded, the current file is renamed "
+            "to audit.jsonl.1 and a fresh file is started. Only one "
+            "backup is kept."
+        ),
+    )
+    request_log: Literal["off", "active"] = Field(
+        default="off",
+        description=(
+            "v2.0-K (Replay): request metadata journal. 'active' records "
+            "per-request metadata (provider, token counts, cost, streaming "
+            "flag) to a JSONL file in state_dir on every successful "
+            "response. Request/response bodies are NOT recorded (privacy "
+            "+ size). Used by ``coderouter replay`` for statistical A/B "
+            "analysis. 'off' = no journal (backward-compatible default). "
+            "Requires state_dir to be set."
+        ),
+    )
+    request_log_max_bytes: int = Field(
+        default=52_428_800,
+        ge=1_048_576,
+        le=1_073_741_824,
+        description=(
+            "v2.0-K (Replay): maximum request journal file size before "
+            "rotation (bytes). Default 50 MiB. Same single-backup "
+            "rotation as audit_log — when exceeded, the current file is "
+            "renamed to requests.jsonl.1 and a fresh file is started."
+        ),
+    )
+
     @model_validator(mode="after")
     def _check_default_profile_exists(self) -> CodeRouterConfig:
         """v0.6-A: surface a typo'd ``default_profile`` at load time.

coderouter/guards/__init__.py

@@ -6,10 +6,12 @@ failure modes that a continuously-running local-LLM agent loop tends
 to hit:
 
   * :mod:`coderouter.guards.tool_loop`      — L3 stuck-tool detection
-  * :mod:`coderouter.guards.memory_pressure` — L2 backend OOM
-                                                 awareness (planned)
-  * :mod:`coderouter.guards.backend_health`  — L5 continuous probe +
-                                                 chain reorder (planned)
+  * :mod:`coderouter.guards.memory_pressure` — L2 backend OOM awareness
+  * :mod:`coderouter.guards.backend_health`  — L5 health state machine +
+                                                 chain reorder
+  * :mod:`coderouter.guards.self_healing`    — v2.0-J auto-exclude +
+                                                 restart + recovery probe
+  * :mod:`coderouter.guards.continuous_probe` — v2.0-I background probing
 
 Each guard is a pure-functional / single-class module that the engine
 consults at the appropriate dispatch point. Guards never block the

coderouter/guards/backend_health.py

@@ -200,6 +200,40 @@ class BackendHealthMonitor:
         """True iff ``provider``'s current state is ``UNHEALTHY``."""
         return self.state_for(provider) == "UNHEALTHY"
 
+    # ------------------------------------------------------------------
+    # v2.0-K: Persistence
+    # ------------------------------------------------------------------
+
+    def save_state(self) -> dict[str, object]:
+        """Export the current per-provider health state for persistence."""
+        with self._lock:
+            return {
+                name: {
+                    "state": entry.state,
+                    "consecutive_failures": entry.consecutive_failures,
+                }
+                for name, entry in self._state.items()
+            }
+
+    def load_state(self, state: dict[str, object]) -> None:
+        """Restore health state from a previously saved dict."""
+        if not isinstance(state, dict):
+            return
+        with self._lock:
+            for name, data in state.items():
+                if not isinstance(data, dict):
+                    continue
+                saved_state = data.get("state", "HEALTHY")
+                if saved_state not in ("HEALTHY", "DEGRADED", "UNHEALTHY"):
+                    saved_state = "HEALTHY"
+                failures = data.get("consecutive_failures", 0)
+                if not isinstance(failures, int) or failures < 0:
+                    failures = 0
+                self._state[name] = _ProviderHealth(
+                    state=saved_state,  # type: ignore[arg-type]
+                    consecutive_failures=failures,
+                )
+
 
 __all__ = [
     "BackendHealthMonitor",