Fix repeated tool calls and upgrade github-copilot-sdk to 1.0.15

MIGRATION.md

@@ -1,3 +1,79 @@
+# Migration Guide: v2.7.x → v2.8.0
+
+## Overview
+
+v2.8.0 advances the pinned `github-copilot-sdk` from `1.0.7` to `1.0.15` (stable;
+bundled CLI 1.0.89) and changes the packaged default model to `auto`. The SDK move is
+**wire-compatible for tool forwarding**: no public API, config key, env var, or CLI
+flag changes.
+
+SDK surface changes absorbed by the provider:
+
+- **New `Tool.is_terminal` field.** SDK 1.0.15's `copilot.tools.Tool` dataclass gains
+  a trailing `is_terminal: bool = False` field, which `copilot/client.py` reads by
+  attribute in **both** `create_session` and `resume`
+  (`if tool.is_terminal: definition["isTerminal"] = True`). The provider hands the
+  SDK a duck-typed `SDKToolWrapper`, so the wrapper now mirrors that field with a
+  `False` default, which omits the `isTerminal` key and keeps the tool-forwarding
+  request byte-identical to 1.0.7. Same shape as the v1.0.2 `defer` and v1.0.7
+  `metadata` additions. Without it, every tool-forwarding turn would raise
+  `AttributeError`. The provider never marks a tool terminal: Amplifier's
+  orchestrator owns tool execution.
+- **Two new mode-gated session defaults pinned** (sdk-boundary MinimalMode:MUST:17-18):
+  `custom_agents_local_only=True` and `enable_experimental_mode=False`, mirroring
+  the SDK's own empty-mode defaults. Because `custom_agents_local_only` is non-None,
+  the SDK applies it with one extra `session.options.update` RPC after each session
+  is created.
+- **`ReasoningEffort` Literal gains `"max"`.** The provider's fallback allowlist
+  already accepted `"max"`, so behavior is unchanged; `"none"` is now the only
+  allowlist value outside the SDK Literal.
+- **40 new `SessionEventType` members**, all classified **DROP** (no kernel domain
+  mapping). 30 are listed explicitly in `config/data/events.yaml` (fusion,
+  auto-tier, workflow-run, model-call, MCP-server, sandbox, and UI telemetry, plus
+  `assistant.turn_retry`, `agent.interrupted`, `session.completion_receipt`); the
+  other 10 fall under the existing `permission.*`, `skill.*`, and `subagent.*`
+  wildcards.
+
+### Also in this release: development dependency on `amplifier-core>=2.0.0`
+
+This affects contributors, not users: the provider does not install `amplifier-core`
+as a runtime dependency. The `dev` extra now requires 2.x, and `uv.lock` resolves
+2.0.1 rather than 1.3.3, so routine tests use the kernel shipped by the current
+Amplifier CLI. Contributors with an existing development environment should run
+`uv sync --extra dev`. No user action is required.
+
+---
+
+## What Changed: default model is now `auto`
+
+- **What:** The packaged default model (used when neither the request nor the
+  `default_model` config sets one) changes from `claude-opus-4.5` to `auto`,
+  Copilot's server-side router and the Copilot CLI's own default. Each turn is
+  dispatched to a concrete model chosen by the service. Because `auto`
+  advertises no limits, `get_info()` now reports `context_window=128000` and
+  `max_output_tokens=16384` for the default (was `200000` / `32000`), matching
+  the policy fallbacks the provider already applied to `auto` at runtime.
+  `auto` advertises no reasoning levels, so a `reasoning_effort` (caller value
+  or provider default) is dropped with a log once the model list is cached.
+- **Action required:** none for configurations that set `default_model` or pass
+  a model per request. To use a fixed model, set `default_model` to an explicit
+  model ID that `list_models()` reports.
+
+---
+
+## What Changed: SDK requirement is now `github-copilot-sdk==1.0.15`
+
+- **What:** The provider now pins `github-copilot-sdk==1.0.15` (was `==1.0.7`). The
+  import-time support floor is unchanged (`>=1.0.0`; pre-GA beta SDKs remain
+  unsupported).
+- **Action required:** none. A clean `pip install` of `2.8.0` pulls `1.0.15`
+  automatically via the `pyproject.toml` pin; `amplifier provider install --force
+  github-copilot` reinstalls the pinned SDK for an existing environment.
+- **Rollback:** if the previous SDK pin is required, pin
+  `amplifier-module-provider-github-copilot<2.8.0`.
+
+---
+
 # Migration Guide: v2.6.x → v2.7.0
 
 ## Overview

README.md

@@ -109,7 +109,7 @@ providers:
   - module: provider-github-copilot
     source: git+https://github.com/microsoft/amplifier-module-provider-github-copilot@main
     config:
-      default_model: claude-opus-4.5
+      default_model: claude-opus-4.8
 ```
 
 ## Usage
@@ -127,7 +127,7 @@ amplifier provider models github-copilot
 
 ## Supported Models
 
-Models are discovered dynamically from the SDK at runtime — the list reflects your GitHub Copilot plan. The tables below show the current public set as of SDK 1.0.7; run `amplifier provider models github-copilot` for the live list. The Context column is the default-tier prompt window; models that advertise a larger long-context tier (shown as `long=` in the live list) are reached by enabling `enable_long_context`.
+Models are discovered dynamically from the SDK at runtime — the list reflects your GitHub Copilot plan. The tables below show the current public set as of SDK 1.0.15; run `amplifier provider models github-copilot` for the live list. The Context column is the default-tier prompt window; models that advertise a larger long-context tier (shown as `long=` in the live list) are reached by enabling `enable_long_context`.
 
 **Routing:**
 
@@ -187,12 +187,12 @@ providers:
   - module: provider-github-copilot
     name: github-copilot
     config:
-      default_model: claude-opus-4.5
+      default_model: auto
 ```
 
 | Key | Default | Description |
 | --- | --- | --- |
-| `default_model` | `"claude-opus-4.5"` | Model used when the caller does not specify one. Any ID from `list_models()` is valid. |
+| `default_model` | `"auto"` | Model used when the caller does not specify one. Any ID from `list_models()` is valid. `auto` is Copilot's server-side router (the Copilot CLI's default): each turn goes to a model the service picks. It advertises no limits or reasoning levels, so the provider reports a conservative 128k context / 16k output, and a `reasoning_effort` is dropped for it once the model list is cached. Set an explicit model ID for a fixed window, reasoning control, or reproducible routing. |
 | `reasoning_effort` | `"model default"` | Reasoning effort forwarded to models that support it (e.g. `low`, `medium`, `high`). Both this default and a caller-supplied value are best-effort: forwarded when the model supports it, dropped (field omitted, server uses its own default) when it does not — so a global or inherited effort never breaks delegation to a non-reasoning model. A malformed value (typo, mixed-case) still raises `ConfigurationError`. `"model default"` or unset defers to the model. Run `amplifier provider models github-copilot` for each model's levels. |
 | `enable_long_context` | `false` | Default to the model's long-context tier when the caller does not set a `context_tier`. Forwarded to the SDK as `long_context`; ignored by models without a long tier. |
 | `raw` | `false` | Include raw SDK payloads as a `"raw"` field in `llm:request` / `llm:response` events. See [Raw Payload Logging](#raw-payload-logging). |
@@ -388,7 +388,7 @@ For module structure, design decisions, and contract index see [docs/ARCHITECTUR
   metadata: the provider obtains the selected model's default-tier (or opted-in
   long-tier) prompt budget from SDK model discovery. They are not a per-request
   admission decision or a backend acceptance guarantee.
-- GitHub Copilot SDK 1.0.7 exposes experimental `context_info` and
+- GitHub Copilot SDK 1.0.15 exposes experimental `context_info` and
   `recompute_context_tokens` operations only for an already-created SDK
   session. They do not accept arbitrary next-request content, so this provider
   does not advertise a native `provider_count` preflight measurement.
@@ -548,7 +548,7 @@ Running `amplifier init` before authentication:
 ## Dependencies
 
 - `amplifier-core` (provided by Amplifier runtime, not installed separately)
-- `github-copilot-sdk==1.0.7`
+- `github-copilot-sdk==1.0.15`
 - `pyyaml>=6.0`
 
 > **Note:** `github-copilot-sdk` is installed automatically when you install or initialize

amplifier_module_provider_github_copilot/__init__.py

@@ -62,7 +62,7 @@ def _check_sdk_version(version_str: str) -> None:
     the technical symbol-availability minimum sits *below* this floor. The floor
     is deliberately set higher, at GA, to drop beta-SDK support entirely.
 
-    The exact pyproject pin (``==1.0.7``) is enforced separately by
+    The exact pyproject pin (``==1.0.15``) is enforced separately by
     ``tests/_sdk_version_gate.py``; this guard only enforces the GA floor.
 
     Extracted for testability — module-level code that runs under
@@ -75,8 +75,8 @@ def _check_sdk_version(version_str: str) -> None:
         raise ImportError(
             f"github-copilot-sdk=={version_str} is below the minimum supported "
             "version (>=1.0.0 required; pre-GA beta SDKs are no longer supported). "
-            "Pinned target: ==1.0.7 (pyproject.toml). "
-            "Reinstall with: pip install 'github-copilot-sdk==1.0.7' "
+            "Pinned target: ==1.0.15 (pyproject.toml). "
+            "Reinstall with: pip install 'github-copilot-sdk==1.0.15' "
             f"or: amplifier provider install --force {PROVIDER_ID}"
         )
 
@@ -105,7 +105,7 @@ def _parse_sdk_version(version_str: str) -> _Version:
         # SDK required; tests only run with SDK installed
         raise ImportError(
             "Required dependency 'github-copilot-sdk' is not installed. "
-            "Install with:  pip install 'github-copilot-sdk==1.0.7'"
+            "Install with:  pip install 'github-copilot-sdk==1.0.15'"
         ) from _e
     # Contract: sdk-boundary:Membrane:MUST:5 — fail at import time on wrong version.
     _check_sdk_version(_sdk_version)
@@ -127,7 +127,7 @@ def _parse_sdk_version(version_str: str) -> _Version:
 
 # Contract: provider-protocol:public_api:MUST:1 — must match pyproject.toml [project].version
 # Verified by tests/test_behaviors.py::TestPackageVersionConsistency
-__version__ = "2.7.0"
+__version__ = "2.8.0"
 
 # Amplifier module metadata
 __amplifier_module_type__ = "provider"

amplifier_module_provider_github_copilot/config/_models.py

@@ -22,15 +22,19 @@
     # NOTE: Per-model capabilities (vision, thinking) are set dynamically in models.py
     # based on SDK's supports_vision/supports_reasoning_effort flags
     "capabilities": ["streaming", "tools"],
-    # Updated defaults to use claude-opus-4.5 with SDK-verified limits
-    # SDK limits: max_context_window=200000, max_prompt_tokens=168000
-    # max_output_tokens = context_window - max_prompt_tokens = 32000
+    # Default model "auto" is Copilot's server-side router (the Copilot CLI's own
+    # default): each turn is dispatched to a concrete model chosen by the
+    # service. list_models() reports "auto" with empty limits and no reasoning
+    # efforts, so its window is unknown until routing happens. context_window /
+    # max_output_tokens therefore mirror FALLBACKS below, which is also exactly
+    # what model_translation derives for "auto" at runtime; get_info() reports
+    # the same values on a cold and a warm model cache.
     "defaults": {
-        "model": "claude-opus-4.5",
+        "model": "auto",
         "max_tokens": 4096,
         "timeout": 3600,
-        "context_window": 200000,
-        "max_output_tokens": 32000,
+        "context_window": 128000,
+        "max_output_tokens": 16384,
     },
 }
 

amplifier_module_provider_github_copilot/config/_sdk_protection.py

@@ -26,7 +26,7 @@
 class MinimalModeConfig:
     """Minimal mode session configuration policy.
 
-    Contract: sdk-boundary:MinimalMode:MUST:1-16
+    Contract: sdk-boundary:MinimalMode:MUST:1-18
 
     Disables SDK features that Amplifier handles, ensuring Amplifier is the sole
     orchestrator. Evidence: 57% wall-clock improvement (12.5s → 5.4s) confirmed
@@ -113,6 +113,26 @@ class MinimalModeConfig:
     # sdk-boundary:MinimalMode SDK-introspection closure).
     memory: dict[str, Any] = field(default_factory=lambda: {"enabled": False})
 
+    # MUST:17: Restrict custom-agent loading to local definitions only; Amplifier
+    # orchestrates agents (complements MUST:5 `custom_agents=[]`). Mode-gated
+    # create_session kwarg with empty-mode default helper
+    # `_custom_agents_local_only_default` (v1.0.15 `_mode.py:254-259`) that
+    # returns True ONLY when `mode == "empty"`. Under `mode="copilot-cli"` the
+    # helper returns None and the bundled-CLI default would otherwise apply.
+    # The SDK emits the value in the create payload and, because it is non-None,
+    # also applies it via one post-create `session.options.update` RPC
+    # (`_post_create_options_patch`, v1.0.15 `_mode.py:295-340`). Value mirrors
+    # the SDK empty-mode default.
+    custom_agents_local_only: bool = True
+
+    # MUST:18: Disable SDK experimental features; Amplifier needs a stable,
+    # reproducible runtime surface. Mode-gated create_session kwarg with
+    # empty-mode default helper `_enable_experimental_mode_default` (v1.0.15
+    # `_mode.py:262-267`) that returns False ONLY when `mode == "empty"`; under
+    # `mode="copilot-cli"` the runtime decides when omitted. Value mirrors the
+    # SDK empty-mode default.
+    enable_experimental_mode: bool = False
+
 
 @dataclass(frozen=True)
 class ToolCaptureConfig:

amplifier_module_provider_github_copilot/config/data/errors.yaml

@@ -91,8 +91,12 @@ error_mappings:
         field: conflict_type
 
   # P4: Model/provider configuration errors
+  # "is not supported for model": SDK v1.0.15 (bundled CLI 1.0.89) rejects an
+  # unsupported reasoning_effort at send time via a session.error whose message
+  # is "Reasoning effort '<value>' is not supported for model '<model>'."
+  # (earlier SDKs rejected at session.create with "does not support ...").
   - sdk_patterns: ["ConfigurationError"]
-    string_patterns: ["does not support", "configuration error", "not configured"]
+    string_patterns: ["does not support", "is not supported for model", "configuration error", "not configured"]
     kernel_error: ConfigurationError
     retryable: false
     # Context extraction for better debugging

amplifier_module_provider_github_copilot/config/data/events.yaml

@@ -245,6 +245,89 @@ event_classifications:
     - mcp.tools.list_changed
     - mcp.resources.list_changed
     - mcp.prompts.list_changed
+    # SDK v1.0.15 (bundled CLI 1.0.89): 40 new SessionEventType enum members
+    # since v1.0.7. All DROP: the provider's domain surface is single-completion
+    # + tool-call + usage + error + content-deltas; none of these carry a kernel
+    # domain mapping. Ten are already covered by the existing wildcards above:
+    # permission.assentDetected / .carriedForward / .contextualAuthorization /
+    # .messageAuthorization / .messageAuthorizationDegraded /
+    # .messageAuthorizationRead (permission.*), skill.context_delivered /
+    # .context_delivered_ref / .invoked_ref (skill.*), subagent.configured
+    # (subagent.*). The remaining 30 are listed explicitly:
+    #   assistant.turn_retry: extra model inference attempt inside an existing
+    #     turn (turn_id/model/reason). The retried output still arrives via the
+    #     bridged message/usage events; bridging this marker would add nothing.
+    #   agent.interrupted: user-interrupt telemetry (activity/elapsed/tool
+    #     names). The provider aborts sessions itself after tool capture; turn
+    #     completion and failures travel via session.idle / session.error.
+    #   model.call_start / model.call_finished: per-dispatch model lifecycle
+    #     telemetry; siblings of model.call_failure (DROP above). Usage rides
+    #     assistant.usage -> USAGE_UPDATE.
+    #   prompt_cache_break: prompt-cache loss diagnostics; no domain value.
+    #   session.completion_receipt: behavior-neutral record of runtime facts at
+    #     completion; TURN_COMPLETE already comes from assistant.message /
+    #     session.idle, so bridging would double-emit.
+    #   session.context_cleared: host-initiated history clear; the provider never
+    #     clears SDK history (ephemeral per-complete() sessions).
+    #   assistant.fusion_phase_started / _activity / _completed / _failed,
+    #     session.fusion_route_started / _route_failed / _resolved / _completed:
+    #     EXPERIMENTAL HydraFusion multi-model routing telemetry; the provider
+    #     selects models explicitly.
+    #   session.auto_tier_recommendation / session.auto_tier_switch_failed:
+    #     Auto-tier preference signals; siblings of auto_mode_switch.* and
+    #     session.auto_mode_resolved (DROP above).
+    #   session.model_deselected: explicit model withdrawn by its publishing host;
+    #     the provider passes the model on every create_session.
+    #   session.mode_notice_delivered: mode-transition notice bookkeeping for
+    #     resume; sibling of session.mode_changed (DROP above).
+    #   session.permission_recovery: Autopilot permission-recovery snapshot;
+    #     sibling of session.permissions_changed (DROP above). The provider
+    #     denies every permission request (deny-destroy).
+    #   sandbox.decision: process-sandbox governance record; the SDK executes no
+    #     tools under deny-destroy, so there is nothing to surface.
+    #   session.indexed_search: transient indexed-search diagnostics; Amplifier
+    #     owns retrieval (skip_embedding_retrieval=True).
+    #   session.managed_settings_enforced: enterprise-policy enforcement notice;
+    #     sibling of session.managed_settings_resolved (DROP above).
+    #   session.mcp_server_removed / session.mcp_server_needs_reconnect: MCP
+    #     server lifecycle; siblings of session.mcp_server_status_changed (DROP
+    #     above). mcp_servers={} (MinimalMode:MUST:3).
+    #   tool_search.activated: lazy tool-search activations restored on resume;
+    #     the provider sets defer=None on every tool and never resumes sessions.
+    #   ui.ephemeral_query: transient side-query output that does not modify
+    #     conversation history; no UI host.
+    #   workflow.run_started / _updated / _settled: ephemeral workflow-run
+    #     lifecycle signals; Amplifier orchestrates its own workflows.
+    - assistant.turn_retry
+    - agent.interrupted
+    - model.call_start
+    - model.call_finished
+    - prompt_cache_break
+    - session.completion_receipt
+    - session.context_cleared
+    - assistant.fusion_phase_started
+    - assistant.fusion_phase_activity
+    - assistant.fusion_phase_completed
+    - assistant.fusion_phase_failed
+    - session.fusion_route_started
+    - session.fusion_route_failed
+    - session.fusion_resolved
+    - session.fusion_completed
+    - session.auto_tier_recommendation
+    - session.auto_tier_switch_failed
+    - session.model_deselected
+    - session.mode_notice_delivered
+    - session.permission_recovery
+    - sandbox.decision
+    - session.indexed_search
+    - session.managed_settings_enforced
+    - session.mcp_server_removed
+    - session.mcp_server_needs_reconnect
+    - tool_search.activated
+    - ui.ephemeral_query
+    - workflow.run_started
+    - workflow.run_updated
+    - workflow.run_settled
 
 finish_reason_map:
   # Values MUST be lowercase to match amplifier-core proto: