Strengthen local agent execution

[jadercorrea]

Why

Local execution was inexpensive but not yet trustworthy enough to demonstrate the GPTCode thesis. Research could infer root causes without repository evidence, editors could stop before completing a plan, reviewers could keep requesting tools after deterministic checks passed, retry limits did not bound the complete workflow, and cancellation could leave child processes running.

The maintained CLI also carried disconnected hosted-product commands and reported local-model metrics with more precision than the provider supplied.

What changed

• ground deep research in repository files while rejecting path and symlink escapes;
• let the editor complete planned multi-file changes and use ranged file reads;
• force bounded, explicit review verdicts after deterministic verification;
• propagate --max-attempts into the internal conductor;
• terminate full subprocess trees when agent commands are cancelled;
• fall back to the embedded model catalog for clean local installations;
• report observed local-model calls and tool timing without false precision;
• isolate nested Git operations from the parent repository during commit hooks;
• remove the disconnected legacy monitor and imperative release commands.

User impact

GPTCode can now complete a controlled workflow with Ollama/Qwen without paid inference: inspect repository evidence, identify an incorrect Go implementation, plan the change, apply the patch, pass go test ./..., and finish model review within a caller-defined attempt budget.

This makes the local-first path useful for repeated evaluation and gives the project a stronger, auditable demonstration of its architecture.

Evidence

• complete Ollama/Qwen repair workflow: success in approximately 48 seconds;
• make verify;
• go test -race ./...;
• go vet ./...;
• staticcheck ./...;
• gosec -quiet ./...;
• public examples pass the race detector and enforce 100% statement coverage;
• regression coverage for repository evidence, model fallback, planned files, bounded review, process cancellation, telemetry, and Git hook isolation.

Follow-up

Instrument every model stage through one provider wrapper, then isolate the remaining Live, training, Supabase, and experimental surfaces behind explicit compatibility boundaries.