fix(quota): provider-aware budget gating and per-provider usage tracking

[BabyKoan]

Problem

When using ollama-launch provider, Kōan displayed incorrect 100%/100% quota numbers from previous regular Claude API usage. The usage_state.json accumulated tokens across provider switches without resetting, causing stale data to persist.

Root Cause

• usage_estimator.py tracked tokens in a single shared state file regardless of which CLI provider was active.
• UsageTracker had no awareness of whether the current provider had metered API quota, so it gated on stale usage.md even for local/self-hosted providers.

Fix

1. Add CLIProvider.has_api_quota() — returns False for ollama-launch and LocalLLMProvider (no metered API quota). Base class defaults to True.

2. Make _get_budget_mode() provider-aware — when the active provider has no API quota, budget_mode is forced to 'disabled' so stale usage.md never gates the agent into wait/review mode.

3. Add per-provider tracking to usage_estimator.py — stores name in and resets session/weekly counters when the provider changes. This prevents stale Claude numbers from polluting ollama-launch display.

4. Update TUI dashboard — shows 'no API quota' messages for no-quota providers instead of misleading usage bars.

Verification

• Full test suite: 15877 passed, 1 warning
• New tests added for:
  • OllamaLaunchProvider.has_api_quota() returns False
  • LocalLLMProvider.has_api_quota() returns False
  • ClaudeProvider.has_api_quota() returns True
  • Provider-aware budget mode override
  • Usage counter reset on provider change
  • Usage counter preservation when provider unchanged

Related

• Fixes Improve local support using ollama command #319

---

Quality Report

Changes: 11 files changed, 228 insertions(+), 21 deletions(-)

Code scan: clean

Tests: passed (403
tests)

Branch hygiene: clean

Generated by Kōan

[BabyKoan]

PR Review — fix(quota): provider-aware budget gating and per-provider usage tracking

Clean, well-tested provider-aware quota implementation.

• Correctly adds has_api_quota() to provider hierarchy with appropriate overrides
• Per-provider tracking in usage_estimator prevents stale data accumulation
• TUI dashboard properly shows 'no API quota' for local providers
• Comprehensive test coverage for all provider types and edge cases
• Minor suggestion: call decide_mode() instead of hardcoding 'deep' in TUI

---

🟡 Important

1. Hardcoded mode display may be inaccurate (`koan/app/tui_dashboard.py`, L883)

When provider has no API quota, the TUI displays Mode deep (budget disabled for ...) with hardcoded 'deep'. This should call t.decide_mode() to show the actual mode the agent would operate in.

The UsageTracker.decide_mode() should return 'deep' when budget_mode is 'disabled', but this should be verified rather than assumed. If the logic changes, the TUI will show stale information.

Suggested fix:

try:
    mode = t.decide_mode()
    lines.append(f"Mode      [{_MINT}]{mode}[/]  [dim](budget disabled for {_provider_name()})[/]")
except Exception as exc:
    lines.append(f"Mode      [{_MINT}]deep[/]  [dim](budget disabled for {_provider_name()})[/]")

lines.append(f"Mode      [{_MINT}]deep[/]  [dim](budget disabled for {_provider_name()})[/]")

---

Checklist

• No hardcoded secrets
• Input validation at boundaries
• Error handling with proper cleanup
• No resource leaks
• No mutable default arguments
• Test coverage for new code paths
• Architecture and abstraction level appropriate

---

Silent Failure Analysis

🟡 **MEDIUM** — silent fallback on error (`koan/app/tui_dashboard.py:94-101`)

Risk: When provider quota check fails, returns True (has quota) which could incorrectly apply budget gating to local providers that should be exempt.

def _provider_has_api_quota() -> bool:
    try:
        from app.cli_provider import get_provider
        return get_provider().has_api_quota()
    except Exception as exc:
        _log.debug("provider_has_api_quota failed: %s", exc)
        return True

Fix: Consider logging at WARNING level since this affects budget gating behavior, or document why fail-closed (True) is the safe default.

🟡 **MEDIUM** — silent fallback on error (`koan/app/usage_tracker.py:285-291`)

Risk: If provider quota check throws, budget_mode falls back to config value instead of forcing 'disabled' - local providers could get incorrectly gated on stale usage.

if not get_provider().has_api_quota():
    return "disabled"
except Exception as exc:
    logger.debug("Provider quota check failed: %s", exc)
    # falls through to return config mode

Fix: Log at WARNING level since incorrect budget gating blocks agent iterations, or consider fail-safe behavior that assumes no quota when check fails.

---

Automated review by Kōan (Ollama-launch · model qwen3.5:cloud) HEAD=1b1fcbf 1 min 58s

[BabyKoan]

Rebase with requested adjustments

Branch koan.atoomic/fix-ollama-quota-display was rebased onto main and review feedback was applied.

Changes applied

• Applied all three review suggestions:
• Replace hardcoded deep mode in TUI Usage tab with dynamic t.decide_mode() call, falling back to deep on exception.
• Upgrade _provider_has_api_quota() failure log from debug to warning in tui_dashboard.py since incorrect fallback affects budget gating.
• Upgrade _get_budget_mode() failure log from debug to warning in usage_tracker.py since provider check failure can block agent iterations.

Stats

11 files changed, 233 insertions(+), 22 deletions(-)

Actions performed

• Already-solved check: negative (confidence=high, reasoning=None of the PR's core changes — has_api_quota(), _maybe_reset_provider(), or provider-aware budg)
• Resolved merge conflicts (1 round(s))
• Rebased koan.atoomic/fix-ollama-quota-display onto upstream/main
• Applied review feedback
• Pre-push CI check: previous run passed
• Force-pushed koan.atoomic/fix-ollama-quota-display to origin
• CI check enqueued in ## CI (async)

CI status

CI will be checked asynchronously.

---

Automated by Kōan