release: 0.7.7 — forward real model + tools to /gate pre-flight (T4)

[maltsev-dev]

0.7.7 — forward real model + tools to /gate pre-flight (T4)

Bug: Pre-0.7.7 every SDK /gate call for any workflow with a budget was hard-blocked with

"Tool 'llm' was blocked because policy 'Rule 1 (cost_limit)' (score 70.00) matched"

because the runtime hard-coded model="budget-precheck" as a sentinel placeholder. The backend's PolicyEvaluationGraph.evaluate() stub treated any synthetic cost_limit rule with score > 0.8 as Block (backend/src/policy/graph.rs:448-462, backend/src/proxy/http/gate/internal.rs:619-628), so the pricing lookup never landed on a real model and the rule fired with the wrong score.

Fix: Forward the real model name and tool list to /gate via a new per-call context API.

Added

• nullrun.set_call_context(model=..., tools=[...]) — per-call context the SDK forwards to /gate so the backend can enforce budget tiers and tool-block on real values.

  import nullrun
  
  with nullrun.workflow(name="support-bot"):
      nullrun.set_call_context(
          model="claude-sonnet-4-6",
          tools=["shell.run", "code.eval"],
      )
  
      @nullrun.protect
      def chat(message: str) -> str:
          return agent.run(message)

  • model (optional) — backend uses it to look up the per-model rate from tool_pricing (Postgres) so projected_cost matches what /track will compute from real token counts. None → backend falls back to the claude-sonnet-4 default rate.
  • tools (optional) — backend matches each against the workflow's effective blocked_tools aggregate and returns block on any match. None leaves whatever was previously set; [] clears.
  • nullrun.get_call_model() / nullrun.get_call_tools() are the read-side helpers (also reachable via nullrun.context.*).

Fixed

• /gate pre-flight no longer sends model="budget-precheck". check_workflow_budget now reads get_call_model() (or None when unset) instead of the placeholder. Default workflows with a budget now return allow instead of blanket-block.
• /gate pre-flight now forwards the per-call tools list. Transport.check previously dropped the tools key from the wire payload, so even set_call_context(tools=[...]) had no effect on /gate. The transport now propagates tools when set; [] vs missing-None are distinguished on the wire per gate/internal.rs::check_tool_block ("no tools will be called" is different from "I did not tell you what tools").

Tests

• New: tests/test_gate_real_path.py (226 lines) pins the fix:
  • TestGateRealPathRegression — default request allows a clean workflow (not the old blanket block), wire payload has no policy-N residue from the old graph plumbing, real decision="block" still raises WorkflowKilledInterrupt (so the fix didn't accidentally remove the real-block path).
  • TestSetCallContext — set_call_context(model=...) flows into the body, set_call_context(tools=[...]) flows into the body, no-context means no tools key at all (not []), and set_call_context(tools=[]) clears a previously-set tool list.
  • TestPackageExports — the new helpers are reachable from nullrun.*.
• tests/conftest.py — reset_runtime fixture now also clears _call_model_var and _call_tools_var so a test's set_call_context(...) doesn't leak into the next test's wire payload.

Migration

No breaking changes. New helpers default to None / empty, so existing call sites (and every test in the suite) keep working without modification.

Diff stats

 CHANGELOG.md                       |   97 ++++++
 pyproject.toml                     |    2 +-
 src/nullrun/__init__.py            |    8 +
 src/nullrun/context.py             |   62 +++++
 src/nullrun/runtime.py             |   28 ++-
 src/nullrun/transport.py           |   12 +
 tests/conftest.py                  |    8 +
 tests/test_gate_real_path.py       |  226 +++++++++++++++++ (new)
 8 files changed, 433 insertions(+), 3 deletions(-)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!