agentacct 0.12.1

Kimi Code support end to end — usage import from ~/.kimi-code, pricing-table estimates for K3 and DeepSeek Flash, automatic user-level MCP registration (agentacct setup mcp --agent kimi-code --write / onboard --scope global --agent kimi-code), and client labels across the CLI, the local API, the TUI, and the app.

Also in this release:

• Usage filters: one row of independent Date (presets or custom range) / Agent / Model / Provider filters; the whole page follows the selection, options come from an unfiltered read, empty combinations show an explicit empty state, and unconfirmed filters withhold numbers instead of mislabeling them.
• Freshness stamp: the app top bar always shows the local-data stamp (last known build, saved copy, or read time) instead of disappearing.
• Pricing fixes: client routing model ids resolve against the LiteLLM-derived table, historical repricing prices a row the same way a fresh import does, and the cost basis of Claude Code, Codex, hermes, dsh, and Kimi Code usage rows is stated as the token-derived estimate it is.
• Evidence spool compaction: agentacct evidence compact-spool (dry run by default, --write --yes to run) reclaims already-dead shadow rows from the append-only spool with a blocking rebuild-and-compare before the atomic swap and a verified zstd archive first.
• Kimi Code index tolerance: deleted-session tombstones, stale rows, and a half-written final line no longer read as corruption.