fix(providers): restore live provider pricing by bumping the Copilot SDK

[Jason Robert (jrob5756)]

Refs #386. Fixes the mechanism; the issue also asks about static-table entries, which this deliberately does not touch (see below).

What was actually wrong

CopilotProvider.get_model_pricing reads billing.token_prices. The pinned github-copilot-sdk 1.0.1 ships twoModelBilling classes:

file	declares	from_dict
copilot/generated/rpc.py	multiplier, token_prices	parses tokenPrices
copilot/client.py	multiplier only	reads multiplier, drops tokenPrices

client.list_models() returns the second one, which is what the hook reads. So the field never left the API and was still modelled in the generated types; the client dataclass just threw it away.

The effect is that the hook returned None for every model. The chain #265 built — workflow override, then hook, then static table, then unpriced — ran permanently on its fallback, and any model missing from DEFAULT_PRICING reported no cost at all. #388 made that visible; this makes it stop happening.

1.0.9 parses the field again. Verified against both wheels rather than taken on trust, and end to end: a model parsed by the real ModelInfo.from_dict yields $15.00 / $75.00 per million tokens for claude-opus-5, which is 1.5 credits per 1k tokens at 100 credits to the dollar.

Why the floor moves, not just the lock

>=1.0.0 is satisfied by 1.0.1, so an existing environment would keep dead pricing while reporting a perfectly healthy dependency. The feature genuinely requires 1.0.9, so the floor should say so.

Is the bump safe

I diffed every SDK method Conductor calls between the two versions by AST rather than by eye:

• create_session and resume_session gained 16 optional kwargs and lost nothing
• list_models, start, stop, abort, destroy, disconnect, on, rpc, send, RuntimeConnection.for_uri are unchanged
• approve_all's invocation narrowed from dict[str, str] to a PermissionInvocation TypedDict

That last one is the ty error Frank Li (@franklixuefei) reported on #388. PermissionInvocation is a TypedDict, so it was always dict-compatible at runtime; the annotation now matches.

One thing worth flagging that the type error hid: 1.0.9 also gave approve_all behaviour it did not have before.

ifinvocation.get("managed_settings_enabled", False):
raiseRuntimeError("approve_all cannot be used when managed settings are enabled")

That flag is populated from enable_managed_settings, an opt-in kwarg on create_session that Conductor never passes, so the raise is unreachable. I recorded it in the handler docstring rather than leave the next person to trace it. If we ever do opt in, the blanket approve-all handler becomes a hard failure and will need to handle it.

The static table

claude-opus-5 and gpt-5.6-sol are still absent, on purpose. With the hook alive they price from the SDK, so #386's practical complaint is answered without anyone inventing per-token rates — which is the trade Frank Li (@franklixuefei) declined to make on #388, correctly. Guessed rates are worse than no rates.

Tests

The existing hook tests build their models from SimpleNamespace, so they assert what Conductor does with a billing object, never whether the SDK still hands it one. That is why 21 models silently went unpriced against a fully green suite.

Two new tests build the model through the SDK's own ModelInfo.from_dict. I confirmed they fail on 1.0.1 and pass on 1.0.9, so the next SDK release that drops the field fails the build instead of quietly reverting every run to static pricing.

Validation

6091 passed, 47 skipped. ruff check, ruff format --check and ty check src all clean.

[Jason Robert (jrob5756)]

Merged main in and resolved two conflicts. One of them was load-bearing, so flagging it rather than burying it in a merge commit.

#390 hardened the "Remove bundled Copilot CLI binary" step to fail when find matches nothing:

removed=$(find .venv -path '*/copilot/bin/copilot*' -print -delete)if [ -z"$removed" ];thenecho"::error::No bundled Copilot CLI matched ..."exit 1
fi

The reasoning is right, and it is the same defect I found from the other end: a silent no-op reinstates the auth hang the step exists to prevent. But the assertion stops being true with this PR. SDK 1.0.1 ships per-platform wheels carrying copilot/bin/; 1.0.9 is a single pure-Python wheel that fetches the CLI on first use, so there is nothing to delete. Confirmed against both venvs.

So taking that side unchanged would have turned main red on the merge commit, for a condition that is correct rather than broken.

Resolution keeps everything #390 added — shell: bash, set -euo pipefail, and the pattern matching both lib/pythonX.Y/ and Windows Lib/ with the .exe suffix — and drops only the hard failure, replacing it with a message that says an empty match is expected. What now guarantees the original intent is COPILOT_SKIP_CLI_DOWNLOAD=1 in the test step's env, which blocks the on-demand fetch that replaced the bundled binary. The step is now only there for a venv restored from an older lock.

Exercised the resolved script against all three layouts before pushing: old POSIX (copilot/bin/copilot), old Windows (Lib/site-packages/copilot/bin/copilot.exe), and 1.0.9 with no binary. All exit 0, and the first two still delete.

The env var sits on the test step, so it covers the Windows matrix entry #390 added as well as the Linux ones.

CHANGELOG conflict was the ordinary both-sides-appended kind; kept both, no wording changed.

6129 passed locally after the merge, ruff and ty clean.