Summary

This change fixes a REPL prediction hang caused by entering multiline mode on incomplete input. Previously, that transition cancelled the visible prediction by calling OnInputChanged(""), which also started a hidden empty-input debounced prediction; if that prediction path stalled, it could keep the REPL tied up and interfere with foreground command execution because interpreter context was shared globally.

Risk Assessment: 🟡 Medium — Medium risk: tests are strong and the fix is reasonably bounded, but it changes core cancellation/timeout behavior and includes a real timeout-precedence flaw in prediction requests.

Architecture

flowchart TD
subgraph repl_flow["REPL Prediction Flow"]
  direction TB
  multiline_input["Multiline input transition (updated)"]
  prediction_state["PredictionState.Cancel() (updated)"]
  event_provider["EventPredictionProvider (updated)"]
  predict_middleware["repl.predict middleware / prediction agent (unchanged)"]

  multiline_input -->|"uses cancel instead of empty prediction request"| prediction_state
  prediction_state -->|"stops in-flight prediction work"| event_provider
  event_provider -->|"emits repl.predict"| predict_middleware
end

subgraph runtime_flow["Cancellation And Model Runtime"]
  direction TB
  command_exec["REPL command execution (updated)"]
  goroutine_contexts["Interpreter goroutine context store (added)"]
  lite_model["Default lite model config (updated)"]
  openai_provider["OpenAI provider timeout handling (updated)"]

  command_exec -->|"sets foreground command context"| goroutine_contexts
  lite_model -->|"supplies optional timeout"| openai_provider
end

event_provider -->|"sets prediction context with 30s fallback"| goroutine_contexts
predict_middleware -->|"uses gsh.models.lite for LLM predictions"| lite_model
predict_middleware -->|"invokes model requests"| openai_provider
goroutine_contexts -->|"keeps prediction and command cancellation isolated"| openai_provider

Key changes made

• Adds PredictionState.Cancel() and uses it for multiline input, so newline transitions invalidate in-flight predictions without launching a replacement empty prediction.
• Moves interpreter execution context storage to a per-goroutine map and switches callers to ClearContext(), so prediction work and foreground command execution no longer overwrite each other's cancellation state.
• Adds timeout protection around prediction and model requests: prediction events get a fallback deadline, OpenAI-compatible models now support a timeout field, and the default gsh.models.lite Ollama model is configured with 15000.
• Updates docs/tutorial/03-command-prediction.md so the command prediction tutorial documents the new timeout guidance.

How was this tested

• go test ./internal/script/interpreter ./internal/repl/input ./internal/repl/predict
• go test ./internal/repl/...
• go test ./cmd/gsh/...
• go test -run 'TestInterpreter_Context_|TestOpenAIProviderChatCompletion_UsesModelTimeout|TestMultiLineInputDoesNotStartHiddenPrediction' ./internal/script/interpreter ./internal/repl/input
• golangci-lint fmt
• golangci-lint run --fix
• golangci-lint run

Critique Comments

• 🔴 internal/repl/predict/event_provider.go:73 - This outer WithTimeout(..., 30s) caps every repl.predict run, so a user-configured model timeout above 30s can never take effect. Slow local models will still be cancelled early; derive the limit from model config or make this cap configurable.