agentacct

macOS app

Download agentacct 0.12.10 for macOS

Fixed

• Recovery notifications no longer accumulate behind identical banners. A recurring source issue replaces its previous recovery banner, and recovered notices disappear after ten seconds.
• Recovery titles identify the affected client and issue. Recent updates keeps the recovery history, while active faults remain visible until dismissed or confirmed recovered.

Includes #353. The DMG is rebuilt for this release because the Swift app payload changed; it contains the matching frozen 0.12.10 CLI and is signed, notarized, and stapled.

CLI installation after the PyPI publishing job completes: pip install --upgrade agentacct==0.12.10.

Codex's gpt-6.1-sol usage rows are priced instead of importing cost-unknown. Nothing else changed.

macOS app (no Python, pipx, or terminal needed): agentacct-0.12.9.dmg — signed, notarized, and stapled, built from this release's own commit. The app embeds the frozen agentacct CLI, so the DMG is rebuilt here even though no Swift file changed: the fix lives in the CLI, and reusing the 0.12.8 DMG would have shipped the old price table and silently dropped it.

CLI: pipx install agentacct==0.12.9 (or uv tool install agentacct==0.12.9).

What changed since 0.12.8:

• gpt-6.1-sol has a price row. Codex was reporting this model while neither the builtin price table nor the local LiteLLM snapshot carried it, so those rows imported cost_confidence: unknown — and a later scan only reprices what a catalog row now covers, so they stayed that way. The builtin table now carries the published list rates on both routes a codex row can report, the client's own codex key and OpenAI's openai key: $2/1M in, $10/1M out, $0.10/1M cache read, $2.50/1M cache write. No multiplier — the 2.5x ccusage fast-pricing convention stays scoped to gpt-5.5, so the two routes price identically rather than one of them silently carrying a convention.

• Both keys mean it prices immediately. Exact-key lookup answers for the codex route instead of waiting on the snapshot's TTL refresh to alias it onto OpenAI's row, and if a snapshot does carry the model the dollars agree.

• Coverage stays per model. A sibling id no row carries (gpt-6.1-sol-pro) remains cost-unknown rather than borrowing this row's price.

• Rows already in your ledger are repaired by a command, not by upgrading. Prices are estimated from local token counts (cost_basis: pricing_table), never provider billing:

  agentacct usage import-local --refresh --estimate-costs
  

  The watcher path does the same on its scans once it is running the new importer. Verified on a copy of a real ledger with no pricing snapshot at all (builtin table only, no network): 11/11 gpt-6.1-sol rows repriced, $6.528786 total, pricing_source: agent_sentinel_builtin.

Asset verification: agentacct-0.12.9.dmg — 37,059,269 bytes, sha256 6eb8202358b7b9adfcfae901a6b1afac8a42a39a05d35eebc22541d93d7b6025, built from 26c01a9dc089895821a74600100d897c9b920985; the mounted app reports agentacct 0.12.9 (26c01a9dc089895821a74600100d897c9b920985) and spctl accepts it as Notarized Developer ID from Developer ID Application: Yuxuan Hu (L7656X6G3Q).

A reboot no longer strands the macOS app with a dead recorder and no way to act: the recording-health surfaces, the window's recovery gate, the setup sheet, and the menu bar all offer a one-click Start recorder that runs agentacct start in the background — including on a machine whose agentacct came from pipx/uv rather than the app's own install.

macOS app (no Python, pipx, or terminal needed): agentacct-0.12.8.dmg — signed, notarized, and stapled, built from this release's own commit; the frozen agentacct CLI it embeds carries the same behavior, and the app installs that CLI on first setup. It opens with a normal double-click.

CLI: pipx install agentacct==0.12.8 (or uv tool install agentacct==0.12.8).

What changed since 0.12.7:

• The app can start the recorder after a reboot. The window's "The recorder needs recovery" gate previously offered only View saved work and Open recording setup; it now leads with the same one-click start the health popover, the notice stack, the menu-bar dropdown, and the setup sheet's reconnect button use. It runs agentacct start --no-sync-clients --store-dir <the store this window displays> --json and confirms the result with the same status --json readiness check the app-owned reconnect uses.
• It also works when the app owns no verified recorder — a development build, or a machine where pipx/uv installed agentacct over the app's wrapper. The app resolves ~/.local/bin/agentacct and then agentacct on PATH, and accepts a candidate only if its own --version banner says agentacct <version>. The button's caption names what will run: "Starts agentacct 0.12.4 at ~/.local/bin/agentacct — not installed or managed by this app."
• The recorder is genuinely in the background. The CLI's runtime manager spawns the watcher and dashboard detached, so they outlive the app; the app only waits for the start command and then re-reads the recorder's own state.
• The app-owned path is unchanged. reconnectRecorder's provenance and ownership checks, and its precedence whenever it is available, are exactly as before. The external path never installs, updates, replaces or uninstalls that CLI, never refreshes client integrations (--no-sync-clients), and never signals a process it did not start.
• Coverage: 13 new Swift tests pin the external path (candidate order, symlink following, banner rejection, exact argv, store gate, readiness failure and retry, no writes) and the routing rule that the app-owned recorder always wins; the full suite is 669 tests in release mode.

dsh (DeepSeek Harness) usage rows are priced again after a provider-name mismatch left every one cost-unknown.

DeepSeek Harness records its own route name, deepseek-official, as the provider on every usage row, while the pricing tables key DeepSeek's own models under provider deepseek. No lookup step matched — not the exact key, the provider aliases, the client model aliases, nor the generic <vendor>/<model> fallback (the id carries no slash) — so apply_pricing_estimate_to_event returned False and every imported dsh row stayed cost-unknown, which the Usage views render as unpriced.

macOS app (no Python, pipx, or terminal needed): agentacct-0.12.7.dmg — signed, notarized, and stapled, built from this release's own commit; the frozen agentacct CLI it embeds carries the fix, and the app installs that CLI on first setup. It opens with a normal double-click.

CLI: pipx install agentacct==0.12.7 (or uv tool install agentacct==0.12.7).

What changed since 0.12.6:

• dsh rows are priced again. A provider alias (deepseek-official → deepseek) resolves the client's route name to the catalog's DeepSeek list prices (deepseek-flash, deepseek-v4-flash, deepseek-v4-flash-vision-exp, deepseek-v4-pro). It is a name mapping only: the target row's list price applies as an ≈ estimate, never provider billing, and a dsh model id no catalog row covers still stays cost-unknown.
• Stored dsh rows repair themselves without rereading transcripts. agentacct usage import-local --client dsh --refresh --estimate-costs, and the watcher's equivalent scan, promote previously imported unknown-cost dsh rows to estimated_from_tokens estimates — 95 rows on the maintainer's own store, 75 of them outside the discovery window — while already-priced rows and recorded token counts stay untouched.
• The dsh import fixture now matches the client it claims to model (deepseek-official / deepseek-flash instead of the stale deepseek/deepseek-chat shape no current dsh version reports), alongside an alias test, an end-to-end pricing test, and a stored-row reprice test.