Skip to content

CHANGELOG.md

Latest commit

 

History

History
88 lines (64 loc) · 3.33 KB

CHANGELOG.md

File metadata and controls

88 lines (64 loc) · 3.33 KB

Changelog

v0.2.0 — 2026-03-16 — 1M Context Window & API Hardening

Critical Fixes

1. Context Window: 1M Auto-Detection

  • Root cause: context_limit hardcoded to 200k, Opus 4.6 [1m] has 1M window
  • Fix: detect [1m] in model.id (matches CLI's NO() function)
  • Before: 104k tokens showed 77% (against 200k). After: 10% (against 1M)

2. Context Calculation: Match CLI's /context

  • Formula: percentage = round(totalTokens / contextWindow * 100)
  • Removed erroneous system_overhead=24500 — cache_read + input already includes system prompt
  • Changed denominator from window - output_reserve to full window (matches CLI)
  • Fixed max_output_tokens: 16000 -> 32000

3. Cost Formatting Bug

  • $10.00 displayed as $1 due to sed 's/0$//' eating integer digits
  • Fix: sed 's/\.00$//' — only strip .00, keep $10.50 as $10.50

4. Model Detection Updated

  • Added opus-4-6, sonnet-4-6 pattern matching
  • Strip [1m] suffix from settings.json model before case-matching (prevents opus[1m][1m])

API Improvements

5. Adaptive Quota TTL

  • 5h utilization <20%: poll every 5min
  • 20-50%: every 2min
  • 50-80%: every 1min
  • 80%+: every 30s
  • Previously: fixed 60s regardless of utilization

6. Error Backoff

  • 120s cooldown after failed API call (was: retry every tick)
  • Error state tracked in ${session_id}.err file

7. OAuth Gate

  • Skip quota/user fetch when ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, or ANTHROPIC_BASE_URL set
  • Skip when credentials file exists but contains no OAuth token
  • Eliminates "no token found" log noise for API key users

8. Dynamic User-Agent

  • Uses CLI version from input JSON (was: hardcoded claude-code/2.1.22)

Code Quality

9. Atomic Cache Writes

  • Write to ${cache_file}.tmp.$$ then mv -f to prevent truncated JSON from race conditions

10. JSON Injection Prevention

  • detect_session_boundary() now uses jq -n --arg instead of raw string interpolation

11. Profile Parsing Consolidation

  • 10 separate echo | jq calls replaced with single eval "$(jq -r @sh ...)" call

12. Progress Bar Extraction

  • render_bar() function replaces 70+ lines of copy-paste across 7 bar styles
  • C-style for ((i=0; ...)) loops replace seq subshells

13. Deduplicated Logic

  • get_user_tier(): single function for tier detection (was: 3 copies)
  • get_adaptive_ttl(): single function for TTL calculation (was: 2 copies)

14. Portability

  • printf '%b' replaces echo -e for ANSI stripping (portable across shells)
  • sed 's/\.00$//' replaces sed -E 's/0+$//' (GNU/BSD compatible)

15. Input Parsing

  • Single eval "$(jq -r @sh ...)" call replaces 12 separate subshells

Environment Variables

Variable Default Description
CLAUDE_CONTEXT_LIMIT auto Context token limit override (auto-detected from model.id)
CLAUDE_CODE_MAX_OUTPUT_TOKENS 32000 Output token reserve
CLAUDE_DATA_DIR script dir Data directory (usage.jsonl)
CLAUDE_CACHE_DIR $CLAUDE_DATA_DIR/sessions Session cache directory

v0.1.0 — 2026-02-05 — API Correctness & macOS Support

Initial release with float-to-integer conversion fixes, macOS Keychain support, OrbStack home directory resolution, profile API alignment, per-model quotas, BSD compatibility, user/quota components, and dependency checking.