routed runs fail when the router selects effort none or max #64831

Title: Model routing: Copilot harness rejects router-selected reasoning efforts none and max

Summary

With engine.model-routing enabled (#64593), the Copilot harness refuses to start Copilot
whenever AWF's routing selection uses reasoning effort none or max:

unexpected error: AWF model routing failed: AWF /reflect returned unsupported reasoning effort max; refusing to start Copilot

Both values are valid end to end:

• gh-aw-router and AWF's routing candidate pool use the effort set
  none | minimal | low | medium | high | xhigh | max.
• The pinned Copilot CLI (1.0.89) accepts all seven values:
  copilot --help → --reasoning-effort <level> [possible values: none, minimal, low, medium, high, xhigh, max].

Only gh-aw's allow-lists are narrower: they accept minimal | low | medium | high | xhigh.

Impact

This affects any routed workflow whose candidates include models that support max or none.
In the published gh-aw-router tables, the top-ranked Copilot choice uses max or none in:

When this happens, the agent step fails before any work is done. This is not a routing failure
(exit 78). A local AWF run selected gpt-6-astra with effort max, and the Copilot API served
that request successfully.

Analysis

Two allow-lists in actions/setup/js are the only mismatch:

1. copilot_harness.cjs, resolveAWFModelRoutingSelection (around line 509):

   if (effort && !["minimal", "low", "medium", "high", "xhigh"].includes(effort)) {
     return { selection: null, error: `AWF /reflect returned unsupported reasoning effort ${effort}` };
   }

2. copilot_sdk_driver.cjs (around line 113) only forwards reasoningEffort to the SDK session
   when GH_AW_COPILOT_ROUTING_EFFORT is one of minimal | low | medium | high | xhigh.
   Otherwise it silently drops the effort.

The other code paths already pass the value through unchanged:

• applyCopilotRoutingArgs (--reasoning-effort <effort>)
• applyCopilotRoutingSelection (GH_AW_COPILOT_ROUTING_EFFORT)
• buildCopilotSDKChildEnv (COPILOT_REASONING_EFFORT for sub-agent sessions)

Note: @github/copilot-sdk 1.0.13 types SessionConfig.reasoningEffort as
"low" | "medium" | "high" | "xhigh" | "max". The RPC layer uses string, and
copilot_sdk_session.cjs already passes the value as any. The headless server is the same CLI
that accepts all seven values.

Implementation plan

1. Add one shared constant in actions/setup/js (in copilot_harness.cjs, or a small shared
   module that the SDK driver also imports):

   const ROUTING_REASONING_EFFORTS = Object.freeze(["none", "minimal", "low", "medium", "high", "xhigh", "max"]);

   These are the values the pinned Copilot CLI lists for --reasoning-effort.
2. Use it in resolveAWFModelRoutingSelection in place of the inline list. Keep
   lower-casing and trimming. Keep the distinction between effort: null (no flag, no env) and
   effort: "none" (an explicit value that is passed through).
3. Use it in copilot_sdk_driver.cjs for the reasoningEffort forwarding check, so max and
   none reach runWithCopilotSDK.
4. Add tests:
   • copilot_harness.test.cjs:
     • resolveAWFModelRoutingSelection accepts max and none, and still rejects an unknown
       value such as "extreme".
     • applyCopilotRoutingArgs with effort: "max" produces --reasoning-effort max.
     • effort: "none" produces --reasoning-effort none.
     • effort: null produces no flag.
     • applyCopilotRoutingSelection sets GH_AW_COPILOT_ROUTING_EFFORT to max and none.
   • SDK driver: reasoningEffort is forwarded for max and none when GH_AW_MODEL_ROUTING=1.
5. Optional hardening: if the Copilot CLI version pin changes, a test or comment should tie
   ROUTING_REASONING_EFFORTS to the CLI's documented --reasoning-effort values.
6. Run make agent-finish.

Verification

• Unit tests above.
• End to end:
  • Compile a routed workflow with engine.model-routing, using mode: robust or balanced and
    allowed-models that include a model offering max (for example gpt-6-astra).
  • Confirm the agent step logs
    inference routing: mode=awf-routed model=<model> effort=max and Copilot starts.
  • Confirm the AWF api-proxy logs a model_routing request event with routed: "as_selected".

Out of scope

No AWF or gh-aw-router change is needed. Both already use the full effort set.