Objective
Make the actual LLM/runtime invocation consume the Mapper-selected task context instead of calling driver.execute(goal) with only the natural-language goal.
Problem
The distributed context pack contains goal, ACs and source references, but the routed runtime path currently derives a goal string and executes it directly. Exact spans, tests, graph evidence, fidelity and omissions are not delivered to the model, so the model must survey the repository again.
Implementation steps
- Define a provider-neutral runtime request carrying goal, ACs, exact Mapper spans, source refs, tests, graph evidence, constraints, omissions and remaining budget.
- Add a deterministic renderer for CLI-style drivers with explicit section boundaries and prompt-injection provenance labels.
- Keep raw repository instructions/source evidence separated from trusted system/operator constraints.
- Apply a final token budget after rendering; required evidence may not be silently truncated.
- Pass the structured request to drivers that support it and the rendered request to legacy string-only drivers.
- Preserve cwd/tool access while restricting mutation targets to the authorized per-task plan.
- Add redaction and secret scanning before dispatch.
- Capture a context-consumption receipt: envelope hash, rendered tokens, selected paths/spans, omitted fields, provider/model, cache status and observable token usage.
- When providers do not report usage, record
null plus reason; never estimate it as measured.
- Bind runtime-execution receipt to mapper envelope hash and plan/mutation authority.
- Reject stale/mismatched envelopes before calling a provider.
- Add compatibility tests for each real driver and malformed/hostile context.
Acceptance criteria
Non-goals
- Training/fine-tuning a model.
- Claiming token savings without observed evidence.
- Allowing the model to expand scope beyond mutation authority.
Objective
Make the actual LLM/runtime invocation consume the Mapper-selected task context instead of calling
driver.execute(goal)with only the natural-language goal.Problem
The distributed context pack contains goal, ACs and source references, but the routed runtime path currently derives a goal string and executes it directly. Exact spans, tests, graph evidence, fidelity and omissions are not delivered to the model, so the model must survey the repository again.
Implementation steps
nullplus reason; never estimate it as measured.Acceptance criteria
nullwith a machine-readable reason.Non-goals