coderouter-cli 2.15.1__py3-none-any.whl → 2.17.0__py3-none-any.whl

coderouter/routing/fallback.py

@@ -47,6 +47,7 @@ from coderouter.adapters.base import (
     ProviderCallOverrides,
     StreamChunk,
     StreamTruncatedError,
+    TTFTTimeoutError,
 )
 from coderouter.adapters.registry import build_adapter
 from coderouter.config.schemas import CodeRouterConfig, ProviderConfig
@@ -111,6 +112,7 @@ from coderouter.routing.fallback_trace import (
     REASON_PAID_GATE,
     REASON_SELF_HEALING_EXCLUDED,
     REASON_STREAM_TRUNCATED,
+    REASON_TTFT_TIMEOUT,
     REASON_UNKNOWN_PROVIDER,
     FallbackTrace,
     begin_fallback_trace,
@@ -436,6 +438,16 @@ def _apply_context_budget_guard(
     if profile.context_budget_action == "warn" or not estimate.over_trim_threshold:
         return request, "warning"
 
+    # v2.16.0: reject → let the client compact instead of trimming here.
+    if profile.context_budget_action == "reject":
+        from coderouter.guards.context_budget import ContextBudgetRejectError
+
+        raise ContextBudgetRejectError(
+            estimated_tokens=estimate.estimated_tokens,
+            max_context_tokens=estimate.max_context_tokens,
+            profile=profile.name,
+        )
+
     # Over trim threshold + action is trim → trim messages
     trimmed_request, trim_result = trim_to_budget(
         request,
@@ -888,6 +900,38 @@ def _emit_cache_observed_streaming(
     )
 
 
+async def _await_within_ttft(
+    awaitable: Any,
+    *,
+    deadline: float | None,
+    provider: str,
+    ttft_timeout_s: float | None,
+) -> Any:
+    """v2.17.0: await ``awaitable`` but give up at the TTFT ``deadline``.
+
+    ``deadline`` is a ``time.monotonic()`` instant (``None`` = no limit). On
+    expiry raises :class:`TTFTTimeoutError` — a retryable ``AdapterError``, so
+    every existing pre-content fallback branch handles it unchanged. The
+    cancelled await unwinds the adapter's generator, which releases its
+    upstream connection.
+    """
+    if deadline is None:
+        return await awaitable
+    remaining = deadline - time.monotonic()
+    if remaining <= 0:
+        close = getattr(awaitable, "close", None)
+        if close is not None:
+            close()
+        raise TTFTTimeoutError(provider=provider, ttft_timeout_s=ttft_timeout_s or 0.0)
+    try:
+        async with asyncio.timeout(remaining):
+            return await awaitable
+    except TimeoutError as exc:
+        raise TTFTTimeoutError(
+            provider=provider, ttft_timeout_s=ttft_timeout_s or 0.0
+        ) from exc
+
+
 class NoProvidersAvailableError(CodeRouterError):
     """Raised when every provider in the chain has failed (or was filtered out)."""
 
@@ -3344,7 +3388,7 @@ class FallbackEngine:
         # keeps the default byte-for-byte identical to v2.14.0.
         last_empty_stream_truncated = False
 
-        for adapter, will_degrade in chain:
+        for chain_idx, (adapter, will_degrade) in enumerate(chain):
             is_native = isinstance(adapter, AnthropicAdapter)
             downgrading = (not is_native) and bool(request.tools)
             effective_request = request
@@ -3404,6 +3448,14 @@ class FallbackEngine:
             # only the non-streaming generate_anthropic recorded, so the
             # busiest path produced no adaptive signal.
             attempt_started = time.monotonic()
+            # v2.17.0: TTFT cut-off. Never on the last provider — there is
+            # nothing left to fall back to, so waiting is the better answer.
+            ttft_s = getattr(adapter.config, "ttft_timeout_s", None)
+            ttft_deadline = (
+                attempt_started + ttft_s
+                if ttft_s is not None and chain_idx < len(chain) - 1
+                else None
+            )
             event_iter: AsyncIterator[AnthropicStreamEvent]
             first: AnthropicStreamEvent
             try:
@@ -3412,11 +3464,21 @@ class FallbackEngine:
                 # mypy narrows for stream_anthropic (not on BaseAdapter).
                 if isinstance(adapter, AnthropicAdapter):
                     event_iter = adapter.stream_anthropic(effective_request, overrides=overrides)
-                    first = await anext(event_iter)
+                    first = await _await_within_ttft(
+                        anext(event_iter),
+                        deadline=ttft_deadline,
+                        provider=adapter.name,
+                        ttft_timeout_s=ttft_s,
+                    )
                 elif downgrading:
                     # v0.3-D downgrade: run non-streaming, repair, replay.
                     chat_req = to_chat_request(effective_request)
                     chat_req.stream = False
+                    # v2.17.0: no TTFT deadline here. This branch is a
+                    # *non-streaming* call (the tool-turn downgrade), so its
+                    # "first output" is the whole answer — a TTFT cut would
+                    # become a total-generation limit and throw away slow
+                    # but healthy local replies. ``timeout_s`` bounds it.
                     chat_resp = await adapter.generate(chat_req, overrides=overrides)
                     # H-4: same as the non-streaming branch — keep the
                     # CPU-bound repair scan off the event loop.
@@ -3431,7 +3493,12 @@ class FallbackEngine:
                     event_iter = stream_chat_to_anthropic_events(
                         adapter.stream(chat_req, overrides=overrides)
                     )
-                    first = await anext(event_iter)
+                    first = await _await_within_ttft(
+                        anext(event_iter),
+                        deadline=ttft_deadline,
+                        provider=adapter.name,
+                        ttft_timeout_s=ttft_s,
+                    )
             except StopAsyncIteration:
                 # M2: record the empty-stream failure.
                 self._adaptive.record_attempt(
@@ -3547,7 +3614,18 @@ class FallbackEngine:
                                 yield buffered_ev
                             buffer = []
                     try:
-                        pending = await iterator.__anext__()
+                        if content_started:
+                            pending = await iterator.__anext__()
+                        else:
+                            # v2.17.0: the TTFT deadline also covers the
+                            # withheld preamble — first *real* output is
+                            # what counts, not the first event.
+                            pending = await _await_within_ttft(
+                                iterator.__anext__(),
+                                deadline=ttft_deadline,
+                                provider=adapter.name,
+                                ttft_timeout_s=ttft_s,
+                            )
                     except StopAsyncIteration:
                         break
             except AdapterError as exc:
@@ -3599,8 +3677,28 @@ class FallbackEngine:
                     # message — so the hop is recorded as ``stream-truncated``
                     # and the empty-response log is skipped. The adapter has
                     # already emitted ``stream-truncation-detected``.
-                    truncated = isinstance(exc, StreamTruncatedError)
-                    if truncated:
+                    #
+                    # v2.17.0: a TTFT cut-off lands here too, and its buffer
+                    # is just as unterminated (opening events, no
+                    # ``message_stop``). Treat it like a truncation so the
+                    # chain-exhausted flush below never replays it as a
+                    # "complete empty answer" — that would hand the client a
+                    # 200 stream that never ends. Close the abandoned upstream
+                    # stream now rather than waiting for GC.
+                    ttft_cut = isinstance(exc, TTFTTimeoutError)
+                    truncated = isinstance(exc, StreamTruncatedError) or ttft_cut
+                    if ttft_cut:
+                        aclose_iter = getattr(event_iter, "aclose", None)
+                        if aclose_iter is not None:
+                            with contextlib.suppress(Exception):
+                                await aclose_iter()
+                        trace.record_failure(
+                            adapter.name,
+                            REASON_TTFT_TIMEOUT,
+                            detail=describe_adapter_error(exc),
+                            stream=True,
+                        )
+                    elif truncated:
                         trace.record_failure(
                             adapter.name,
                             REASON_STREAM_TRUNCATED,

coderouter/routing/fallback_trace.py

@@ -111,6 +111,8 @@ REASON_EMPTY_STREAM: Final = "empty-stream"
 # ``connection`` (the transport itself broke). Produced only when the profile
 # sets ``stream_truncation_action: error``.
 REASON_STREAM_TRUNCATED: Final = "stream-truncated"
+# v2.17.0: no first real output within the provider's ``ttft_timeout_s``.
+REASON_TTFT_TIMEOUT: Final = "ttft-timeout"
 
 # -- degrade reasons (pre-existing strings; never produce a hop) -----------
 # Kept here so the whole ``reason`` vocabulary is readable in one place.
@@ -141,6 +143,7 @@ ATTEMPT_FAILURE_REASONS: Final[frozenset[str]] = frozenset(
         REASON_EMPTY_RESPONSE,
         REASON_EMPTY_STREAM,
         REASON_STREAM_TRUNCATED,
+        REASON_TTFT_TIMEOUT,
     }
 )
 
@@ -207,6 +210,8 @@ def describe_adapter_error(exc: AdapterError) -> str:
         # Not a transport fault — the HTTP body ended cleanly; it was the
         # LLM protocol inside it that never closed.
         return "no-terminator"
+    if reason == REASON_TTFT_TIMEOUT:
+        return "no-first-token"
     return f"transport={reason}"
 
 
@@ -227,10 +232,12 @@ def classify_adapter_error(exc: AdapterError) -> str:
     ``adapters.base`` is a TYPE_CHECKING-only import at module scope here
     (the runtime import order would otherwise cycle through the registry).
     """
-    from coderouter.adapters.base import StreamTruncatedError
+    from coderouter.adapters.base import StreamTruncatedError, TTFTTimeoutError
 
     if isinstance(exc, StreamTruncatedError):
         return REASON_STREAM_TRUNCATED
+    if isinstance(exc, TTFTTimeoutError):
+        return REASON_TTFT_TIMEOUT
     status = exc.status_code
     if status is None:
         message = str(exc).lower()

coderouter/vscode_init.py

@@ -112,6 +112,37 @@ _MANAGED_ENV_KEYS: tuple[str, ...] = (
     "CODEROUTER_MODE",
 )
 
+# v2.16.0: Claude Code settings for a gateway that is not api.anthropic.com.
+# Written by default (``--no-claude-code-env`` opts out):
+#
+# * ``ENABLE_TOOL_SEARCH=false`` — 2.1.304 turned tool search on behind a
+#   custom ANTHROPIC_BASE_URL. Only api.anthropic.com runs the search; the
+#   gateway would answer 400 on every session's first request and Claude
+#   Code would retry without it. Saying so up front skips that round trip.
+# * ``CLAUDE_CODE_AUTO_MODE_SERVER=0`` — on gateways auto mode defaults to
+#   the server-side permission classifier, which a local backend does not
+#   provide; ``0`` keeps the classifier client-side (its requests then go
+#   through CodeRouter like any other request).
+# * ``CLAUDE_CODE_MAX_CONTEXT_TOKENS`` — only when ``--max-context-tokens``
+#   is given. Since 2.1.285 Claude Code assumes a 1M window behind a
+#   gateway for Opus 4.7+ / Sonnet 5+ / Fable ids, so auto-compaction would
+#   wait far past what a local model holds.
+CLAUDE_CODE_ENV_KEYS: tuple[str, ...] = (
+    "ENABLE_TOOL_SEARCH",
+    "CLAUDE_CODE_AUTO_MODE_SERVER",
+    "CLAUDE_CODE_MAX_CONTEXT_TOKENS",
+)
+
+
+def claude_code_env(*, max_context_tokens: int | None = None) -> dict[str, str]:
+    """v2.16.0: the Claude Code env vars ``vscode-init`` writes by default."""
+    env = {"ENABLE_TOOL_SEARCH": "false", "CLAUDE_CODE_AUTO_MODE_SERVER": "0"}
+    if max_context_tokens is not None:
+        if max_context_tokens <= 0:
+            raise ValueError("max_context_tokens must be positive")
+        env["CLAUDE_CODE_MAX_CONTEXT_TOKENS"] = str(max_context_tokens)
+    return env
+
 
 @dataclass(frozen=True)
 class FileOutcome:
@@ -155,6 +186,7 @@ def run_vscode_init(
     with_envrc: bool = False,
     dry_run: bool = False,
     force: bool = False,
+    extra_env: dict[str, str] | None = None,
 ) -> VSCodeInitResult:
     """Scaffold VSCode workspace settings for CodeRouter.
 
@@ -182,6 +214,9 @@ def run_vscode_init(
         When true, existing conflicting values are overwritten. Without
         this flag, a conflict is reported with a diff and the file is
         left untouched.
+    extra_env:
+        v2.16.0: additional variables to write next to the base ones
+        (the CLI passes :func:`claude_code_env`). Same conflict rules.
 
     Raises
     ------
@@ -205,6 +240,7 @@ def run_vscode_init(
             profile=profile,
             force=force,
             dry_run=dry_run,
+            extra_env=extra_env,
         )
     )
 
@@ -217,6 +253,7 @@ def run_vscode_init(
                 profile=profile,
                 force=force,
                 dry_run=dry_run,
+                extra_env=extra_env,
             )
         )
 
@@ -235,6 +272,7 @@ def _apply_settings_json(
     profile: str | None,
     force: bool,
     dry_run: bool,
+    extra_env: dict[str, str] | None = None,
 ) -> FileOutcome:
     desired_env: dict[str, str] = {
         "ANTHROPIC_BASE_URL": base_url,
@@ -242,6 +280,8 @@ def _apply_settings_json(
     }
     if profile:
         desired_env["CODEROUTER_MODE"] = profile
+    if extra_env:
+        desired_env.update(extra_env)
 
     old_text = ""
     if path.exists():
@@ -401,6 +441,7 @@ def _apply_envrc(
     profile: str | None,
     force: bool,
     dry_run: bool,
+    extra_env: dict[str, str] | None = None,
 ) -> FileOutcome:
     """Write (or refresh) the managed block in ``.envrc``.
 
@@ -413,7 +454,7 @@ def _apply_envrc(
       which now *appends* the block rather than replacing the file. The
       operator's lines survive either way.
     """
-    block = _render_envrc_block(base_url=base_url, profile=profile)
+    block = _render_envrc_block(base_url=base_url, profile=profile, extra_env=extra_env)
 
     old_text = ""
     if path.exists():
@@ -467,7 +508,7 @@ def _apply_envrc(
                 before.rstrip("\n") + ("\n\n" if before.strip() else "")
             ) + block + ("\n\n" + after.lstrip("\n") if after.strip() else "\n")
 
-        conflicts = _unmanaged_conflicts(outside)
+        conflicts = _unmanaged_conflicts(outside, extra_keys=tuple(extra_env or ()))
         if conflicts:
             # direnv applies exports in order, so a duplicate outside the
             # fence would silently win or lose depending on placement.
@@ -568,7 +609,9 @@ def _split_managed_block(text: str) -> tuple[str, str] | None:
     return text[:start], text[end + len(_ENVRC_END) :].lstrip("\n")
 
 
-def _render_envrc_block(*, base_url: str, profile: str | None) -> str:
+def _render_envrc_block(
+    *, base_url: str, profile: str | None, extra_env: dict[str, str] | None = None
+) -> str:
     """The fenced block CodeRouter owns, markers included."""
     lines = [
         _ENVRC_BEGIN,
@@ -579,18 +622,25 @@ def _render_envrc_block(*, base_url: str, profile: str | None) -> str:
     ]
     if profile:
         lines.append(f'export CODEROUTER_MODE="{profile}"')
+    for key, value in (extra_env or {}).items():
+        lines.append(f'export {key}="{value}"')
     lines.append(_ENVRC_END)
     return "\n".join(lines)
 
 
-def _unmanaged_conflicts(outside_text: str) -> list[str]:
-    """Managed variables the operator exports outside our block."""
+def _unmanaged_conflicts(outside_text: str, *, extra_keys: tuple[str, ...] = ()) -> list[str]:
+    """Managed variables the operator exports outside our block.
+
+    v2.16.0: ``extra_keys`` (the Claude Code variables this run writes) are
+    checked too — but only when this run writes them, so an operator who
+    passed ``--no-claude-code-env`` can keep their own exports.
+    """
     found: list[str] = []
     for line in outside_text.splitlines():
         stripped = line.strip()
         if stripped.startswith("#"):
             continue
-        for var in _MANAGED_ENVRC_VARS:
+        for var in (*_MANAGED_ENVRC_VARS, *extra_keys):
             if stripped.startswith((f"export {var}=", f"{var}=")) and var not in found:
                 found.append(var)
     return found

{coderouter_cli-2.15.1.dist-info → coderouter_cli-2.17.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: coderouter-cli
-Version: 2.15.1
+Version: 2.17.0
 Summary: Local-first, free-first, fallback-built-in LLM router. Claude Code / OpenAI compatible.
 Project-URL: Homepage, https://github.com/zephel01/CodeRouter
 Project-URL: Repository, https://github.com/zephel01/CodeRouter

{coderouter_cli-2.15.1.dist-info → coderouter_cli-2.17.0.dist-info}/RECORD

@@ -1,11 +1,12 @@
 coderouter/__init__.py,sha256=ghdjPrLtnRzY8fyQ4CJZI1UJKADyNTLtA3G7se8H7Ns,696
 coderouter/__main__.py,sha256=-LCgxJnvgUV240HjQKv7ly-mn2NuKHpC4nCpvTHjeSU,130
-coderouter/cli.py,sha256=Nbx1ur_HnssFtuM48yPELwhY40veVNnTAfK7OVF2uVc,42400
+coderouter/cli.py,sha256=-QPJ2wTJIpa4TvIFObNaV4hyv_OeDtoLebc692ZWYQA,44775
 coderouter/cli_stats.py,sha256=CCjzc1G4hTRHZ2gG1XhxhDpUkJnnl3NXbcbp1T18jpg,29894
 coderouter/cost.py,sha256=32h6uzb4nxh2eA5d2Hn3kD9yJbtis6CFDAbeIy5KRkM,7431
 coderouter/credentials.py,sha256=R_PdrcA941_jJX_U8hxrjPDTCVVzLd4dsw2c00KwAwE,12991
 coderouter/doctor.py,sha256=0EUqKaEw2GaqX6zGKjR-bVDdlmBznFIklSz3yIoQbE0,91969
 coderouter/doctor_apply.py,sha256=tZT3bEd1G6t67YRhf1TGY4NlSvZuJJbnEqN-3AKy8Hc,29972
+coderouter/doctor_claude_code.py,sha256=15ZdyUg0M_wZGEBN9Q3eb_D75lncqxQi1dNV6irWh2U,6621
 coderouter/env_security.py,sha256=FEBZnXfJ0xE39kmMMn39zk0W_DRRnmcB_REmP9f4xWo,14796
 coderouter/errors.py,sha256=Xmq67lheyw8iv3Ox39jh2c4tvNI5RcUR4QkoxVDN6l4,1130
 coderouter/gguf_introspect.py,sha256=Scu0JjUfLHj7T7TL0HQIm1AZcGYrXsjkaMMSN9HkTro,16121
@@ -20,23 +21,23 @@ coderouter/rollback.py,sha256=Bqxx1MALTDLYc0BtGMEMqRyiZIWFoAZJH_RBQw-M5ow,7675
 coderouter/secret_redaction.py,sha256=Rg71gqKN6jnybZ6UvCvXIvnThI9uRBvmOqC2grSUSUk,23159
 coderouter/token_estimation.py,sha256=7klQYA4FIg1vtjQGWAhwiBhiLwyz8KTetrLVXFJLGIc,16957
 coderouter/token_estimation_accurate.py,sha256=GTfzrBVnvAGjeVzmzAeUdOYZvWZKLAxcxPpFiJGlzjk,4609
-coderouter/vscode_init.py,sha256=ZEROtETsHjyBkLzZKPnF-zsBpLJx4ph0vmgyty6kNYQ,33433
+coderouter/vscode_init.py,sha256=rs3S0FOuPFZhhDvE547Zonhj98al66Jc8G1wvBnFq40,35915
 coderouter/adapters/__init__.py,sha256=7dIDSZ-FE_0iSqLSDc_lK1idRdLTKcM2hP9tCJipgPI,463
-coderouter/adapters/anthropic_native.py,sha256=JxOaU-ymU1K3Tnvhufi_li1qChW69XtUszKJ2DdFlDQ,29936
-coderouter/adapters/base.py,sha256=qW0QMbO08Dw2ZDruL9rCvCxA0FXhKnSodlK9g-8Y350,13535
-coderouter/adapters/openai_compat.py,sha256=Kl5AACqOHxTPUn8c-4IFu4ug9_E3loLKTm2yYmvalho,23872
+coderouter/adapters/anthropic_native.py,sha256=fYM4uXdYYvXfYBgz2dPjdPwv-yKY7XfnXp18YZNCsjg,30185
+coderouter/adapters/base.py,sha256=Jv_a7d1RtrX_Y6fouL__-DlK4c3cjrYZmkjVKEcqf2w,15297
+coderouter/adapters/openai_compat.py,sha256=-6jD8RLjlocmnGCkGWm6C0j5fLumadJb9J-wZW4a-GQ,24121
 coderouter/adapters/registry.py,sha256=XNc1wmFv_oZlqKS0rA693TRDdbq1sZ-OIcofXOVOtq0,3680
 coderouter/config/__init__.py,sha256=FODEn74fN-qZnt4INPSHswqhOlEgpL6-_onxsitSx8g,274
 coderouter/config/capability_registry.py,sha256=RZgzHF54dXy0CCDocRCz2ee21qXGbbgTSKTkRI9iLL4,16936
 coderouter/config/env_file.py,sha256=CoMK27fuAXm-NtoLzXb8yN2E-wDFjHQuFwiIlmgTBQw,10356
 coderouter/config/loader.py,sha256=7FoJXWaO_vN9FAjUStJDwI4aAYcUQQgMJtq5BNy8Pnc,12419
-coderouter/config/schemas.py,sha256=0NM5KtVGsjH8BPBeBfgM-ItfOYQKhyI-P_3EbJjIaps,146436
+coderouter/config/schemas.py,sha256=i3j1JCMhZTQm753vUhfmvGcf-k2LRlSYaxLugd3Wv34,149822
 coderouter/data/__init__.py,sha256=uNyfD9jaCvTWsBAWtaw1Fr25OSxzv3psGMfBjT1z0Cc,328
 coderouter/data/model-capabilities.yaml,sha256=S9jt6SC6-3s2-icZ_n-a14iEMnc2yB1C2R6q-N_tZWQ,19309
 coderouter/guards/__init__.py,sha256=5qliYBqygvVPneej7nx0uSjxDKsz7t8VzvrDgVBJlvU,1170
 coderouter/guards/_fingerprint.py,sha256=qsgNzIq9jv3FHrKL39nGJARp0cMenpN_QmWoJu87vU4,4835
 coderouter/guards/backend_health.py,sha256=GLUVQ5BhU6i8Cxz_FrgIMs5AZ-YSKAW8veFWgvn9GFI,12565
-coderouter/guards/context_budget.py,sha256=gw3MkiHDjpLmQTlTdJT6gRAxYEnoU0xmwSvPZk_miOQ,23059
+coderouter/guards/context_budget.py,sha256=zbplo_1G54kHITdr4kkgUxjQX20Y8MN2Z4XdxETAGwI,23834
 coderouter/guards/continuous_probe.py,sha256=WIfS-apVMWXGv7bPBxxkJssePa95I4fT0mi4PNqK5iE,12181
 coderouter/guards/drift_actions.py,sha256=A6pY5CR480Ct5rCVyjlBvjPFVc93eu_r5qcUpK9mWKc,3602
 coderouter/guards/drift_detection.py,sha256=lW45EZDGRleDUTstw_CHlfUiwLUrlDHX8g7ncuOYNEw,13962
@@ -45,12 +46,14 @@ coderouter/guards/memory_pressure.py,sha256=9Xm-b7rv9Td5kPNmAFoL3sNmaOR5zT8NLdn6
 coderouter/guards/self_healing.py,sha256=XeUlgFWGqMQ5E9YvvzR2CsHRJOAJEyDChqRN0xmZ0vo,16841
 coderouter/guards/tool_loop.py,sha256=EzeMcmU7BLeTW2jsRVevU81l5rhWcn1oUr7EpzgXjVM,15209
 coderouter/ingress/__init__.py,sha256=WQsCH2CGJCAhy0mS6GSEdeYZRkkQu2OHDsP4CJWTLug,155
-coderouter/ingress/anthropic_routes.py,sha256=wlOb7ChWHMvuERR0lj8TF4MfPis2ZKegceoak4eREes,31131
+coderouter/ingress/anthropic_errors.py,sha256=q8F4LmH57f8Jh2hoV4Mmnvf9sZH32uiXhQXPRBZPjCE,8122
+coderouter/ingress/anthropic_routes.py,sha256=0hJY_i2Zlj93jSGTTfrFETTY-_TRmEsllxWmiS-bSeU,40207
 coderouter/ingress/app.py,sha256=ZOx7dSvMxaWS4CzXOOEwgzFuycAVgAeWKCN6lFNF-F0,22433
 coderouter/ingress/dashboard_routes.py,sha256=nPW5_pJe27-hzqCs4sqVVn0BrgXB_7Wcd4qZ6cirvWo,26736
 coderouter/ingress/launcher_routes.py,sha256=uP_Otp9JE0fydUt1YP-wYcnS8Kj5RPLNIFf88YpfkPM,141535
 coderouter/ingress/metrics_routes.py,sha256=4s_f9CY8nVIdQBfXxyUMlN7r5_cXHOaAu7PvNbUuoxg,4609
 coderouter/ingress/openai_routes.py,sha256=9DHVtGe1mpcUBH3ziDL4RGpEA-LzjyANpBIN3Kfbpxc,14193
+coderouter/ingress/tool_search.py,sha256=JYnJ8SUqPmBfZmsl41TdBQNnux0mwbidKdsr5qn15Hk,3753
 coderouter/ingress/ui_auth.py,sha256=uCWB5IaMQRNFYdY39FQ5DDVhsoRQJoqimZNSAu_QK_Q,3742
 coderouter/metrics/__init__.py,sha256=7Es351DPS7yLM0yVF_F0eesmiD83n7Zzhie44chht38,1465
 coderouter/metrics/collector.py,sha256=-744UrUS6_aEY7hbwuvoX0Nn7wuOsQ57IIVr52nPe2M,63268
@@ -64,8 +67,8 @@ coderouter/routing/adaptive.py,sha256=G2o377twGSjbUh65wiIFx6klnpFGjsD_nI3oDvcBwh
 coderouter/routing/auto_router.py,sha256=y4v0c8u5F9f98Vmhx1vRcKPiOgAvpzbFqr6TIh058h0,13341
 coderouter/routing/budget.py,sha256=PblmVKJGs_BwNa9uDHAA8hmZ4XIVKv38mHAeU0V3OMs,8451
 coderouter/routing/capability.py,sha256=rRQhzTgQYFTFDNYcSLKvq3xjdw6B1w7T-3jS7PTMI14,27864
-coderouter/routing/fallback.py,sha256=Hzy-_2_zK64GUqGGcOZP-RgOq8wZ4PdBpVu4OBTx8dM,172282
-coderouter/routing/fallback_trace.py,sha256=0BysXY1CZqSvVTuePbgmHbu1AKkHLMX7hsuG6rKlDdE,19787
+coderouter/routing/fallback.py,sha256=vx1Vrl_vy7-dLxkKvipfba_pbhQPbuhRVFjmuMMWyaI,176825
+coderouter/routing/fallback_trace.py,sha256=LwUw56mhVIkP4Gu-fNvZmOjL8hXxY1ci6b1MuGjl5m0,20099
 coderouter/state/__init__.py,sha256=XoGcPmmBQSiZWML2S0juSveQ78xfhtdeCliNnVyzu7E,1088
 coderouter/state/audit_log.py,sha256=CEFNJht2WZ8mqesI8CzeuL63NfzFzAZ7wF88auyl6l0,11444
 coderouter/state/replay.py,sha256=Z_YHKroTKZdrL8qObFxcoLOAQWWXZvXFdLfxzvBhEJg,11230
@@ -77,8 +80,8 @@ coderouter/translation/anthropic.py,sha256=ZQQGS9itN_3o7IKD68lKr5gR_Kem32Jsd7OcG
 coderouter/translation/convert.py,sha256=VrstDAMsUpcQ1VgMTn_-ETp1OjGX8v0ZGhCkcd4icyU,53037
 coderouter/translation/tool_repair.py,sha256=1p8rmdyyscGX2g0DDOYitMnLv-nKfR1SLuLaRB4z7GM,64244
 coderouter/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-coderouter_cli-2.15.1.dist-info/METADATA,sha256=oJXnZHAWbbgbFy7wWTJW4LxBTHkA0NSDMWubhzUqo_0,18221
-coderouter_cli-2.15.1.dist-info/WHEEL,sha256=W3fkpkm7-wf9vBI5Z-7s0eWkeM-spu78I8Neb98DeEg,87
-coderouter_cli-2.15.1.dist-info/entry_points.txt,sha256=-dnLfD1YZ2WjH2zSdNCvlO65wYltM9bsHt9Fhg3yGss,51
-coderouter_cli-2.15.1.dist-info/licenses/LICENSE,sha256=wkEzoR86jFw33jvfOHjULqmkGEfxTFMgMaJnpR8mPRw,1065
-coderouter_cli-2.15.1.dist-info/RECORD,,
+coderouter_cli-2.17.0.dist-info/METADATA,sha256=LcxBbvilLSKMicAtVZLtm1tQSazmg52g_Mj21bTOo8k,18221
+coderouter_cli-2.17.0.dist-info/WHEEL,sha256=W3fkpkm7-wf9vBI5Z-7s0eWkeM-spu78I8Neb98DeEg,87
+coderouter_cli-2.17.0.dist-info/entry_points.txt,sha256=-dnLfD1YZ2WjH2zSdNCvlO65wYltM9bsHt9Fhg3yGss,51
+coderouter_cli-2.17.0.dist-info/licenses/LICENSE,sha256=wkEzoR86jFw33jvfOHjULqmkGEfxTFMgMaJnpR8mPRw,1065
+coderouter_cli-2.17.0.dist-info/RECORD,,