fix(pi): use per-model context window for status bar display

[KorenKrita]

Problem

The Pi context bar always displays usage against a hardcoded 190k token budget (DEFAULT_PI_CONTEXT_WINDOW_TOKENS - CONTEXT_HEADROOM_TOKENS = 200000 - 10000), regardless of the actual model's context window. For example, a model with 128k context window would show incorrect percentages.

Pi reports each model's contextWindow via get_available_models, and this data is already stored in piAvailableModels metadata — but it was never used for the status bar.

Root Cause

Pi's turn_end usage events don't include context_window (unlike Claude/Codex which report it via ACP usage updates). So reduced.latestUsage?.contextWindow is always null for Pi sessions, causing the fallback to the hardcoded default.

Changes

In SessionChat.tsx:

• Compute piContextWindow by looking up the current model's contextWindow from piAvailableModels
• Pass it as a fallback: contextWindow={reduced.latestUsage?.contextWindow ?? piContextWindow}
• Extract the piModels variable to avoid duplicate computation

Behavior

• Pi sessions now show context usage against the correct model-specific context window
• Falls back to the 200k default only when model metadata is unavailable
• No changes to Claude, Codex, or other session types

[github-actions]

Findings

• [Minor] Pi context-window lookup ignores provider identity — session.model stores only the legacy model id, and Pi explicitly persists metadata.piSelectedModel so the web can disambiguate providers that expose the same modelId (shared/src/schemas.ts:73). The new lookup uses the first matching modelId, so the status bar can display the wrong context window for duplicate Pi models. Evidence: web/src/components/SessionChat.tsx:654.
  Suggested fix:

  const model = piSelectedModel
      ? piModels.find((m) => m.provider === piSelectedModel.provider && m.modelId === piSelectedModel.modelId)
      : piModels.find((m) => m.modelId === props.session.model)
  return model?.contextWindow ?? undefined

Summary

• Review mode: initial
• One issue found in the Pi status-bar fallback: it should use the provider-qualified selected model before falling back to the bare legacy model id.

Testing

• Not run (automation)

HAPI Bot

[github-actions]

Findings
No issues found.

Summary
Review mode: follow-up after new commits
No diff-introduced correctness, security, data-loss, performance, or maintainability issues found in the current changes. Residual risk: only the touched web unit test was attempted locally, and it could not run because bun is unavailable in this runner shell.

Testing
Not run (automation: bun not found)

HAPI Bot