Claude Code 2.1.x catch-up + honest accounting

Correct totals. One API response split across several JSONL lines is now counted once (it was over-counted 2.4x to 3.4x, with Tools and Retrieved inflated most). Resume replays are deduplicated by uuid, and phantom system records no longer count as messages.

Correct window size. fable-5 and opus-5 sessions, the [1m] model in settings.json, the statusline payload's context_window_size, and pasted /context output all feed context-limit detection. A fresh 1M session no longer reads as 81% hot.

Correct compaction math. Auto-compaction is modelled as a 33K-token headroom buffer (83.5% of a 200K window, 96.7% of 1M) instead of a flat 80%; the window total includes the last response; compaction events use Claude Code's own postTokens and trigger (auto/manual).

System bucket restored. The fixed Claude Code context (system prompt, core tool schemas, memory, skills, hooks) is derived again instead of being discarded above 15K.

New record types. Attachment records (hook output, skill and tool listings, task reminders, nested CLAUDE.md, memory files) are classified; non-human user records (task notifications, slash-command wrappers, skill bodies) no longer masquerade as your prompt.

Tool catalogue rebuilt. Placeholder tool names removed; PowerShell, Workflow, ScheduleWakeup and ToolSearch-loaded deferred tools recognised; MCP servers discovered from ~/.claude.json, plugins, and the session itself.

Plus:

• calibrate parses the current /context table format.
• Buckets now reconcile to the API window total and expose raw contentTokens; trend records carry bucketBasis.
• Auto-inject advisory: same prefix, headroom-based default gate, corrected explanation of when Claude Code compacts.

Fully offline, zero API calls, as always.

Upgrade notes

• JSON consumers: bucket tokens now sum to total_tokens exactly; the raw classifier estimates moved to contentTokens per bucket. reconciliation.basis says which basis a view is on.
• COMPACTION_THRESHOLD is gone; the trigger is context_limit - 33,000 (override via autocompactBufferTokens in ~/.claude-crusts/config.json).
• Auto-inject without an explicit threshold in config now fires near the real auto-compaction trigger instead of at 70% usage. Set threshold explicitly to keep the old behaviour.

v0.7.1 — Auto-inject injection log + firmer advisory wording

• hooks auto-inject log — every fire of the auto-inject hook now appends a JSONL line to ~/.claude-crusts/auto-inject.log with timestamp, session ID, usage %, reclaimable tokens, and the full advisory text.
  View history with claude-crusts hooks auto-inject log (newest 20, table) or --limit 50 --verbose for full advisory text per entry.

  • Firmer advisory wording — the text the hook generates is now phrased as a directive rather than a passive suggestion, asking Claude to proactively surface the /compact focus command in its reply instead of
    silently acknowledging it as context.

  Install: npm install -g claude-crusts@latest

  Full details: https://github.com/Abinesh-L/claude-crusts#readme

v0.5.0 — Interactive TUI + integrations

• Interactive REPL shell (claude-crusts tui) with session picker, tab completion for commands and session IDs
• Clipboard copy for fix blocks (copy 1|2|3) — cross-platform
• Claude Code hook integration (claude-crusts hooks enable|disable|status) — one-line health summary after every response
• Fast-path status command for hook use

Full details: https://github.com/Abinesh-L/claude-crusts#readme

v0.4.0 — Measurement improvements

• Cross-session trend analysis (claude-crusts trend) — sparkline, direction, per-session averages
• Skills discovery from ~/.claude/skills/ and project-local configs
• Improved memory file detection across the project CLAUDE.md chain
• Derived overhead — internal system prompt + message framing now inferred per-session from API usage data instead of hardcoded
• fix command — three pasteable blocks (session prompt, CLAUDE.md snippet, /compact command)

v0.3.0

• Watch mode (claude-crusts watch) — live terminal dashboard that updates as your session runs. See compaction happen in real-time with token counts, category breakdown, and compaction alerts.
• Compaction loss detail (claude-crusts lost) — see exactly which files, conversations, and tool results were dropped during auto-compaction. Nobody else offers this.
• Slash command (/crusts) — run the analysis inside Claude Code. No terminal switching — Claude Code reads the data and gives personalized advice.
• Content-based compact hints — fix command generates meaningful /compact focus strings like "focus on the renderer.ts, classifier.ts changes" instead of message numbers.
• Markdown reports (claude-crusts report --format md) — preview in VS Code without opening a browser.
• Dashboard bar alignment fix (v0.3.1)

Install: npm install -g claude-crusts

v0.2.0

• Cross-session comparison (claude-crusts compare)
• HTML and Markdown report export (claude-crusts report --format html|md)
• Published to npm (npx claude-crusts analyze)
• Real model name and session duration in dashboard