coding-agent-cost 0.1.0__tar.gz → 0.1.1__tar.gz

{coding_agent_cost-0.1.0 → coding_agent_cost-0.1.1}/PKG-INFO

@@ -1,10 +1,13 @@
 Metadata-Version: 2.4
 Name: coding-agent-cost
-Version: 0.1.0
+Version: 0.1.1
 Summary: Estimate AI coding agent (Claude Code / Codex CLI) token usage and cost from local logs
 Author: shiki-yusuke
 License: MIT
 Project-URL: Homepage, https://github.com/shiki-yusuke/agent-cost
+Project-URL: Repository, https://github.com/shiki-yusuke/agent-cost
+Project-URL: Issues, https://github.com/shiki-yusuke/agent-cost/issues
+Keywords: ai-agents,claude-code,codex-cli,cost-estimation,local-first,token-usage
 Classifier: Development Status :: 3 - Alpha
 Classifier: Environment :: Console
 Classifier: Intended Audience :: Developers
@@ -23,30 +26,159 @@ Dynamic: license-file
 
 # agent-cost
 
-> PyPI distribution: [`coding-agent-cost`](https://pypi.org/project/coding-agent-cost/) — the
-> command (`agent-cost`) and import name (`agent_cost`) are unchanged. The name `agent-cost`
-> itself is blocked by PyPI's name-similarity rule against `agentcost`, an unrelated project
-> by a different author; this tool is not affiliated with it.
-
-Estimate how many tokens Claude Code and Codex CLI actually used, and roughly
-what that cost, by reading the local logs those tools already write to your
+See Claude Code and Codex CLI token usage and estimated costs from the logs already on your
 machine.
 
-## Quick start
+Keep pricing gaps visible: unknown models are marked `unpriced`, and cache-write estimates
+with missing TTL information are marked `lower_bound`. JSON reports include the rate catalog
+version and SHA-256 used for the calculation; the catalog records its pricing sources.
+
+**These are estimates based on the selected rate catalog, not your actual bill.** The CLI makes
+no runtime network calls and has zero runtime dependencies. On macOS and Linux with IANA
+timezone data, no account, service, or project configuration is required. See the
+[platform note](#report-and-export) for minimal Windows Python environments.
+
+Need the result in another program? [`agent-cost measure`](#machine-consumption-agent-cost-measure)
+returns versioned `measure/v1` JSON for session IDs you supply. Task and PR attribution remain
+the caller's responsibility.
+
+[日本語サマリ](#日本語サマリ) · [Synthetic output example](#synthetic-output-example)
+
+## Try it on your machine
+
+With [`uvx`](https://docs.astral.sh/uv/guides/tools/), no persistent install is needed:
+
+```bash
+uvx --from coding-agent-cost agent-cost doctor
+uvx --from coding-agent-cost agent-cost report
+```
+
+The package runner may need network access to download the tool. The CLI itself does not upload
+your logs.
+
+The first command checks the expected local paths, the Codex database, and the bundled rate
+catalog. It does not open every Claude JSONL file. The second command performs the real scan,
+prints the result, and exposes unreadable inputs in `data_quality.skipped_files`. This path was
+exercised on macOS from a clean temporary directory against the published `0.1.0` package on
+2026-08-23.
+
+Prefer a persistent command? Install the PyPI distribution, then run the same two commands:
 
 ```bash
-pip install coding-agent-cost   # or, from a checkout: pip install -e .
-agent-cost doctor         # sanity-check log locations and the rate catalog
+pip install coding-agent-cost
+agent-cost doctor
 agent-cost report
 ```
 
-`agent-cost report` scans `~/.claude/projects/**/*.jsonl` and
-`~/.codex/state_5.sqlite` (+ its rollout files), turns every token-usage
-event into a canonical "fact" (one model, one token kind, one timestamp,
-one count), prices each fact against a bundled rate catalog, and prints an
-aggregated table:
+The PyPI distribution is named [`coding-agent-cost`](https://pypi.org/project/coding-agent-cost/),
+while the command remains `agent-cost` and the import remains `agent_cost`. The shorter PyPI
+name is unavailable because of PyPI's similarity rule; this project is not affiliated with the
+unrelated `agentcost` distribution.
+
+## Synthetic output example
+
+This is **synthetic Claude data**, not observed usage, a bill, or evidence of savings. The
+CLI output below was reproduced with PyPI package `coding-agent-cost==0.1.0` and source
+commit `d170ea301ed0c46351749214bd299e75ae8a7786` (only trailing space padding is omitted).
+The [two-event fixture](examples/synthetic-claude.jsonl) contains known-model input, cache
+writes without a TTL breakdown, and unknown-model input.
 
+```text
+Month    Agent   Model                 Token Kind           Tokens   Priced   Unpriced  Est. Cost (USD)  Credits  Status
+-------  ------  --------------------  -------------------  -------  -------  --------  ---------------  -------  -----------
+2026-06  claude  claude-opus-4-8       cache_write_unknown  1000000  1000000  0         6.2500           -        lower_bound
+2026-06  claude  claude-opus-4-8       input_nocache        1000000  1000000  0         5.0000           -        priced
+2026-06  claude  model-not-in-catalog  input_nocache        500      0        500       0.0000           -        unpriced
+
+Total tokens: 2,000,500   Total estimated cost: $11.2500
+Rates catalog: 2026-07-29  (sha256=5d86e11b3c95...)
+Data quality: malformed_events=0  skipped_files=0  negative_deltas=0  unpriced_tokens=500
 ```
+
+- `priced`: $5.0000 is a catalog-based estimate for 1,000,000 input tokens, not a confirmed charge.
+- `lower_bound`: $6.2500 uses the 5-minute cache-write rate because the TTL is missing. This is
+  a lower bound under the catalog and observed token data, not a lower bound on your actual bill.
+- `unpriced`: `0.0000` does not mean free. Read `pricing_status` (the `Status` column) together
+  with `unpriced_tokens` (`Unpriced`): 500 tokens have no price and are excluded from the dollar
+  total. The $11.2500 total also includes the cache-write lower-bound estimate.
+
+<details>
+<summary>Reproduce with synthetic data only (Python 3.9+ and uvx)</summary>
+
+Run from this repository's root. Both log locations and the config file are explicitly set
+inside a temporary directory. Inherited log-location overrides are removed for the child
+process, so it cannot fall back to your real Claude or Codex logs. The temporary data is
+removed when the command exits. The package runner may download the pinned package.
+
+```bash
+python3 - <<'PY'
+import json
+import os
+from pathlib import Path
+import shutil
+import subprocess
+import tempfile
+
+fixture = Path("examples/synthetic-claude.jsonl").resolve()
+with tempfile.TemporaryDirectory(prefix="agent-cost-example-") as directory:
+    root = Path(directory)
+    project = root / "claude" / "projects" / "synthetic"
+    project.mkdir(parents=True)
+    (root / "codex").mkdir()
+    shutil.copyfile(fixture, project / "session.jsonl")
+    config = root / "config.json"
+    config.write_text(json.dumps({
+        "claude_home": str(root / "claude"),
+        "codex_home": str(root / "codex"),
+    }))
+    env = dict(os.environ)
+    for key in ("CLAUDE_HOME", "CODEX_HOME", "PYTHONPATH"):
+        env.pop(key, None)
+    env["AGENT_COST_CONFIG"] = str(config)
+    subprocess.run([
+        "uvx", "--from", "coding-agent-cost==0.1.0", "agent-cost",
+        "report", "--agent", "claude", "--since", "2026-06-01",
+        "--until", "2026-06-02", "--format", "table",
+    ], cwd=root, env=env, check=True)
+PY
+```
+
+Use `--format json` in the same command to see the full catalog SHA-256 and field names.
+
+</details>
+
+## Why this exists
+
+Many usage trackers are designed as rich, convenient dashboards. That can be the right choice
+when broad agent coverage and interactive exploration matter most. `agent-cost` is for a
+different trust model:
+
+- **Local-first and zero-network:** source logs stay on disk; the package has zero runtime
+  dependencies.
+- **Auditable:** token facts, the versioned rate catalog, its digest, and data-quality counters
+  are available for inspection and machine consumption.
+- **Fail-closed pricing:** unknown models remain `unpriced`; ambiguous cache-write TTLs are
+  labeled `lower_bound` instead of silently receiving a best-guess price.
+- **Explicit attribution boundary:** `measure` accepts named session ids, but agent-cost does
+  not infer a task from a branch or PR. A workflow such as
+  [`spec-lane`](https://github.com/shiki-yusuke/spec-lane) can own the session-to-task binding
+  and consume the versioned JSON result.
+
+This is not a claim that local CLI accounting is universally better than a dashboard. It is a
+smaller primitive for environments where data egress, dependency surface, custom metrics, or
+task attribution need to remain under the operator's control.
+
+See [Choosing by use case and trust model](https://github.com/shiki-yusuke/agent-cost/blob/main/docs/comparison.md)
+for a dated, source-linked comparison with broader CLI trackers, local dashboards, and
+observability stacks.
+
+## Report and export
+
+`agent-cost report` turns each local usage event into a canonical fact (one model, one token
+kind, one timestamp, one count), prices each fact against the bundled rate catalog, and prints
+an aggregated table:
+
+```bash
 agent-cost report --since 2026-06-01 --until 2026-07-01 --format table
 agent-cost report --group-by month,agent --format csv
 agent-cost report --format json > usage.json
@@ -58,6 +190,11 @@ interpreted in `--timezone`, default UTC), `--agent claude,codex`,
 `--group-by month,agent,model,token-kind`, `--rates PATH` (use a different
 catalog entirely, see below), `--exclude-archived` (Codex threads).
 
+**Platform note:** `report`, `export`, and `measure` use Python's IANA timezone database even
+when the timezone is `UTC`. macOS and typical Linux installations provide it through the
+operating system. A minimal Windows Python environment may require `pip install tzdata` first;
+`tzdata` is not currently a declared package dependency.
+
 If another program wants to parse agent-cost's output for one or more
 specific session ids, see `agent-cost measure` below rather than
 scraping `report`.
@@ -75,10 +212,12 @@ network, never calls `gh`, and never resolves branches or PRs.
   at the 5-minute rate as an explicit **lower bound** and flagged
   `lower_bound` rather than guessed at the (more expensive) 1-hour rate.
 - **Codex CLI**: rollout files record a *cumulative* token count after each
-  turn; agent-cost turns that into per-turn deltas. Codex does not expose a
-  separate cache-write signal at all, so cache writes are never reported for
-  Codex (not zero -- simply not observable, and left out of the row rather
-  than implied). Codex's `output` fact is `output_tokens` alone: cross-checking
+  turn; agent-cost turns that into per-turn deltas. Under the Codex token-rate
+  tariff used here, cache writes are not added as a separate charge. The current
+  reader does not emit independent cache-write facts; newer logs can include
+  `cache_write_input_tokens`, but this field is not used to add a cache-write
+  charge. This differs from older logs where that signal was unavailable.
+  Codex's `output` fact is `output_tokens` alone: cross-checking
   real rollout files confirms `total_tokens == input_tokens + output_tokens`
   in every sample, which means `reasoning_output_tokens` is a breakdown of
   output tokens already counted, not an additional charge -- adding it in
@@ -115,6 +254,15 @@ network, never calls `gh`, and never resolves branches or PRs.
 
 ## Updating the rate catalog
 
+**GPT-6 Astra:** Codex Standard/Fast estimates use exact ID `gpt-6-astra` and
+owned settings matched to turn/model context in the public Codex 0.153.4 JSONL
+format. Explicit `default` uses Standard; `priority` uses 2.5x. Missing/ambiguous
+tier or model attribution stays `unpriced`, with tokens preserved. These are
+request-setting estimates, not confirmed processing tiers or actual bills.
+The catalog observation cutoff is not an official launch time; earlier usage
+is `unpriced`. See [evidence, synthetic tests and model-switch limits](docs/astra-pricing.md).
+API pricing and legacy message billing are outside this entry.
+
 Prices live in `agent_cost/rates.json`, not in code. It's a historical
 catalog: each model can have several time-bounded rate periods, so a price
 change is recorded as a new period rather than overwriting the old one (see
@@ -224,8 +372,10 @@ MIT. See [LICENSE](LICENSE).
 
 - 集計の最小単位は「1 イベント = 1 モデル × 1 token 種別」の fact であり、session 単位でモデルを
   丸めません。Claude の prompt cache は TTL 内訳（5分/1時間）が取れればそれを使い、取れない場合は
-  5分単価で **下限推計**（`lower_bound`）として明示します。Codex の cache write は観測不能なため
-  出力しません（0 とは区別）。
+  5分単価で **下限推計**（`lower_bound`）として明示します。対象の Codex トークン料金体系では
+  cache write を別料金として加算せず、現行 reader は独立した cache-write fact を出力しません。
+  新しいログに `cache_write_input_tokens` が含まれていても追加料金の計算には使用しません。
+  これは、旧ログでその情報を観測できないこととは区別します。
 - 出力フィールドは `estimated_cost_usd`（推計であることを明示）。Codex は `credits` も併記します。
   未知のモデル・単価表にない token 種別は `unpriced` として扱い、憶測の価格を出しません。
 - 単価表 (`agent_cost/rates.json`) は履歴型カタログで、値上げは新しい期間として追加します。

{coding_agent_cost-0.1.0 → coding_agent_cost-0.1.1}/README.md

@@ -1,29 +1,158 @@
 # agent-cost
 
-> PyPI distribution: [`coding-agent-cost`](https://pypi.org/project/coding-agent-cost/) — the
-> command (`agent-cost`) and import name (`agent_cost`) are unchanged. The name `agent-cost`
-> itself is blocked by PyPI's name-similarity rule against `agentcost`, an unrelated project
-> by a different author; this tool is not affiliated with it.
-
-Estimate how many tokens Claude Code and Codex CLI actually used, and roughly
-what that cost, by reading the local logs those tools already write to your
+See Claude Code and Codex CLI token usage and estimated costs from the logs already on your
 machine.
 
-## Quick start
+Keep pricing gaps visible: unknown models are marked `unpriced`, and cache-write estimates
+with missing TTL information are marked `lower_bound`. JSON reports include the rate catalog
+version and SHA-256 used for the calculation; the catalog records its pricing sources.
+
+**These are estimates based on the selected rate catalog, not your actual bill.** The CLI makes
+no runtime network calls and has zero runtime dependencies. On macOS and Linux with IANA
+timezone data, no account, service, or project configuration is required. See the
+[platform note](#report-and-export) for minimal Windows Python environments.
+
+Need the result in another program? [`agent-cost measure`](#machine-consumption-agent-cost-measure)
+returns versioned `measure/v1` JSON for session IDs you supply. Task and PR attribution remain
+the caller's responsibility.
+
+[日本語サマリ](#日本語サマリ) · [Synthetic output example](#synthetic-output-example)
+
+## Try it on your machine
+
+With [`uvx`](https://docs.astral.sh/uv/guides/tools/), no persistent install is needed:
+
+```bash
+uvx --from coding-agent-cost agent-cost doctor
+uvx --from coding-agent-cost agent-cost report
+```
+
+The package runner may need network access to download the tool. The CLI itself does not upload
+your logs.
+
+The first command checks the expected local paths, the Codex database, and the bundled rate
+catalog. It does not open every Claude JSONL file. The second command performs the real scan,
+prints the result, and exposes unreadable inputs in `data_quality.skipped_files`. This path was
+exercised on macOS from a clean temporary directory against the published `0.1.0` package on
+2026-08-23.
+
+Prefer a persistent command? Install the PyPI distribution, then run the same two commands:
 
 ```bash
-pip install coding-agent-cost   # or, from a checkout: pip install -e .
-agent-cost doctor         # sanity-check log locations and the rate catalog
+pip install coding-agent-cost
+agent-cost doctor
 agent-cost report
 ```
 
-`agent-cost report` scans `~/.claude/projects/**/*.jsonl` and
-`~/.codex/state_5.sqlite` (+ its rollout files), turns every token-usage
-event into a canonical "fact" (one model, one token kind, one timestamp,
-one count), prices each fact against a bundled rate catalog, and prints an
-aggregated table:
+The PyPI distribution is named [`coding-agent-cost`](https://pypi.org/project/coding-agent-cost/),
+while the command remains `agent-cost` and the import remains `agent_cost`. The shorter PyPI
+name is unavailable because of PyPI's similarity rule; this project is not affiliated with the
+unrelated `agentcost` distribution.
+
+## Synthetic output example
+
+This is **synthetic Claude data**, not observed usage, a bill, or evidence of savings. The
+CLI output below was reproduced with PyPI package `coding-agent-cost==0.1.0` and source
+commit `d170ea301ed0c46351749214bd299e75ae8a7786` (only trailing space padding is omitted).
+The [two-event fixture](examples/synthetic-claude.jsonl) contains known-model input, cache
+writes without a TTL breakdown, and unknown-model input.
 
+```text
+Month    Agent   Model                 Token Kind           Tokens   Priced   Unpriced  Est. Cost (USD)  Credits  Status
+-------  ------  --------------------  -------------------  -------  -------  --------  ---------------  -------  -----------
+2026-06  claude  claude-opus-4-8       cache_write_unknown  1000000  1000000  0         6.2500           -        lower_bound
+2026-06  claude  claude-opus-4-8       input_nocache        1000000  1000000  0         5.0000           -        priced
+2026-06  claude  model-not-in-catalog  input_nocache        500      0        500       0.0000           -        unpriced
+
+Total tokens: 2,000,500   Total estimated cost: $11.2500
+Rates catalog: 2026-07-29  (sha256=5d86e11b3c95...)
+Data quality: malformed_events=0  skipped_files=0  negative_deltas=0  unpriced_tokens=500
 ```
+
+- `priced`: $5.0000 is a catalog-based estimate for 1,000,000 input tokens, not a confirmed charge.
+- `lower_bound`: $6.2500 uses the 5-minute cache-write rate because the TTL is missing. This is
+  a lower bound under the catalog and observed token data, not a lower bound on your actual bill.
+- `unpriced`: `0.0000` does not mean free. Read `pricing_status` (the `Status` column) together
+  with `unpriced_tokens` (`Unpriced`): 500 tokens have no price and are excluded from the dollar
+  total. The $11.2500 total also includes the cache-write lower-bound estimate.
+
+<details>
+<summary>Reproduce with synthetic data only (Python 3.9+ and uvx)</summary>
+
+Run from this repository's root. Both log locations and the config file are explicitly set
+inside a temporary directory. Inherited log-location overrides are removed for the child
+process, so it cannot fall back to your real Claude or Codex logs. The temporary data is
+removed when the command exits. The package runner may download the pinned package.
+
+```bash
+python3 - <<'PY'
+import json
+import os
+from pathlib import Path
+import shutil
+import subprocess
+import tempfile
+
+fixture = Path("examples/synthetic-claude.jsonl").resolve()
+with tempfile.TemporaryDirectory(prefix="agent-cost-example-") as directory:
+    root = Path(directory)
+    project = root / "claude" / "projects" / "synthetic"
+    project.mkdir(parents=True)
+    (root / "codex").mkdir()
+    shutil.copyfile(fixture, project / "session.jsonl")
+    config = root / "config.json"
+    config.write_text(json.dumps({
+        "claude_home": str(root / "claude"),
+        "codex_home": str(root / "codex"),
+    }))
+    env = dict(os.environ)
+    for key in ("CLAUDE_HOME", "CODEX_HOME", "PYTHONPATH"):
+        env.pop(key, None)
+    env["AGENT_COST_CONFIG"] = str(config)
+    subprocess.run([
+        "uvx", "--from", "coding-agent-cost==0.1.0", "agent-cost",
+        "report", "--agent", "claude", "--since", "2026-06-01",
+        "--until", "2026-06-02", "--format", "table",
+    ], cwd=root, env=env, check=True)
+PY
+```
+
+Use `--format json` in the same command to see the full catalog SHA-256 and field names.
+
+</details>
+
+## Why this exists
+
+Many usage trackers are designed as rich, convenient dashboards. That can be the right choice
+when broad agent coverage and interactive exploration matter most. `agent-cost` is for a
+different trust model:
+
+- **Local-first and zero-network:** source logs stay on disk; the package has zero runtime
+  dependencies.
+- **Auditable:** token facts, the versioned rate catalog, its digest, and data-quality counters
+  are available for inspection and machine consumption.
+- **Fail-closed pricing:** unknown models remain `unpriced`; ambiguous cache-write TTLs are
+  labeled `lower_bound` instead of silently receiving a best-guess price.
+- **Explicit attribution boundary:** `measure` accepts named session ids, but agent-cost does
+  not infer a task from a branch or PR. A workflow such as
+  [`spec-lane`](https://github.com/shiki-yusuke/spec-lane) can own the session-to-task binding
+  and consume the versioned JSON result.
+
+This is not a claim that local CLI accounting is universally better than a dashboard. It is a
+smaller primitive for environments where data egress, dependency surface, custom metrics, or
+task attribution need to remain under the operator's control.
+
+See [Choosing by use case and trust model](https://github.com/shiki-yusuke/agent-cost/blob/main/docs/comparison.md)
+for a dated, source-linked comparison with broader CLI trackers, local dashboards, and
+observability stacks.
+
+## Report and export
+
+`agent-cost report` turns each local usage event into a canonical fact (one model, one token
+kind, one timestamp, one count), prices each fact against the bundled rate catalog, and prints
+an aggregated table:
+
+```bash
 agent-cost report --since 2026-06-01 --until 2026-07-01 --format table
 agent-cost report --group-by month,agent --format csv
 agent-cost report --format json > usage.json
@@ -35,6 +164,11 @@ interpreted in `--timezone`, default UTC), `--agent claude,codex`,
 `--group-by month,agent,model,token-kind`, `--rates PATH` (use a different
 catalog entirely, see below), `--exclude-archived` (Codex threads).
 
+**Platform note:** `report`, `export`, and `measure` use Python's IANA timezone database even
+when the timezone is `UTC`. macOS and typical Linux installations provide it through the
+operating system. A minimal Windows Python environment may require `pip install tzdata` first;
+`tzdata` is not currently a declared package dependency.
+
 If another program wants to parse agent-cost's output for one or more
 specific session ids, see `agent-cost measure` below rather than
 scraping `report`.
@@ -52,10 +186,12 @@ network, never calls `gh`, and never resolves branches or PRs.
   at the 5-minute rate as an explicit **lower bound** and flagged
   `lower_bound` rather than guessed at the (more expensive) 1-hour rate.
 - **Codex CLI**: rollout files record a *cumulative* token count after each
-  turn; agent-cost turns that into per-turn deltas. Codex does not expose a
-  separate cache-write signal at all, so cache writes are never reported for
-  Codex (not zero -- simply not observable, and left out of the row rather
-  than implied). Codex's `output` fact is `output_tokens` alone: cross-checking
+  turn; agent-cost turns that into per-turn deltas. Under the Codex token-rate
+  tariff used here, cache writes are not added as a separate charge. The current
+  reader does not emit independent cache-write facts; newer logs can include
+  `cache_write_input_tokens`, but this field is not used to add a cache-write
+  charge. This differs from older logs where that signal was unavailable.
+  Codex's `output` fact is `output_tokens` alone: cross-checking
   real rollout files confirms `total_tokens == input_tokens + output_tokens`
   in every sample, which means `reasoning_output_tokens` is a breakdown of
   output tokens already counted, not an additional charge -- adding it in
@@ -92,6 +228,15 @@ network, never calls `gh`, and never resolves branches or PRs.
 
 ## Updating the rate catalog
 
+**GPT-6 Astra:** Codex Standard/Fast estimates use exact ID `gpt-6-astra` and
+owned settings matched to turn/model context in the public Codex 0.153.4 JSONL
+format. Explicit `default` uses Standard; `priority` uses 2.5x. Missing/ambiguous
+tier or model attribution stays `unpriced`, with tokens preserved. These are
+request-setting estimates, not confirmed processing tiers or actual bills.
+The catalog observation cutoff is not an official launch time; earlier usage
+is `unpriced`. See [evidence, synthetic tests and model-switch limits](docs/astra-pricing.md).
+API pricing and legacy message billing are outside this entry.
+
 Prices live in `agent_cost/rates.json`, not in code. It's a historical
 catalog: each model can have several time-bounded rate periods, so a price
 change is recorded as a new period rather than overwriting the old one (see
@@ -201,8 +346,10 @@ MIT. See [LICENSE](LICENSE).
 
 - 集計の最小単位は「1 イベント = 1 モデル × 1 token 種別」の fact であり、session 単位でモデルを
   丸めません。Claude の prompt cache は TTL 内訳（5分/1時間）が取れればそれを使い、取れない場合は
-  5分単価で **下限推計**（`lower_bound`）として明示します。Codex の cache write は観測不能なため
-  出力しません（0 とは区別）。
+  5分単価で **下限推計**（`lower_bound`）として明示します。対象の Codex トークン料金体系では
+  cache write を別料金として加算せず、現行 reader は独立した cache-write fact を出力しません。
+  新しいログに `cache_write_input_tokens` が含まれていても追加料金の計算には使用しません。
+  これは、旧ログでその情報を観測できないこととは区別します。
 - 出力フィールドは `estimated_cost_usd`（推計であることを明示）。Codex は `credits` も併記します。
   未知のモデル・単価表にない token 種別は `unpriced` として扱い、憶測の価格を出しません。
 - 単価表 (`agent_cost/rates.json`) は履歴型カタログで、値上げは新しい期間として追加します。

{coding_agent_cost-0.1.0 → coding_agent_cost-0.1.1}/agent_cost/__init__.py

@@ -6,4 +6,4 @@ against a versioned rate catalog. Everything happens locally; there are no
 network calls.
 """
 
-__version__ = "0.1.0"
+__version__ = "0.1.1"

{coding_agent_cost-0.1.0 → coding_agent_cost-0.1.1}/agent_cost/aggregate.py

@@ -33,6 +33,14 @@ def price_fact(catalog: RateCatalog, fact: Fact) -> Tuple[Optional[Decimal], str
     if resolved_key is None or period is None:
         return None, "unpriced", None
 
+    # Astra's verified rates cover only the exact Codex model ID. Generic
+    # suffix normalization must not implicitly price unverified snapshots,
+    # speed variants or usage from a different agent with this new entry.
+    if resolved_key == "gpt-6-astra" and (
+        fact.agent != "codex" or fact.model_raw != "gpt-6-astra" or fact.mode == "unknown"
+    ):
+        return None, "unpriced", None
+
     rate_field = fact.token_kind
     lower_bound = False
     if fact.token_kind == "cache_write_unknown":

{coding_agent_cost-0.1.0 → coding_agent_cost-0.1.1}/agent_cost/rates.json

@@ -1,13 +1,14 @@
 {
   "schema_version": "1",
-  "catalog_version": "2026-07-29",
+  "catalog_version": "2026-09-06",
   "currency": "USD",
   "unit": "per_mtok",
   "usd_per_credit": "0.04",
   "notes": [
     "All Claude rates are standard (non-batch, non-priority) API prices. Batch API is roughly 50% off, but agent-cost's logs have no signal for whether a given request went through Batch, so batch usage is priced at standard rates and will overstate cost for anyone using Batch.",
     "claude-opus-5's launch date could not be confirmed from an authoritative source as of this catalog_version; its rate period's effective_from (2026-07-01) is a placeholder -- update it once confirmed.",
-    "gpt-5.6 (sol/terra/luna) credits could not be confirmed from the primary source (help.openai.com's Codex rate card returned HTTP 403 to automated fetches). The values here were cross-checked across several independent secondary sources that agreed with each other and are internally consistent (Sol's $5/$30 per-MTok API price x 25 = 125/750 credits, matching usd_per_credit=0.04); see the gpt-5.6-* sources entries below. Treat effective_from (2026-07-01) as a placeholder and re-verify against the primary rate card when it becomes reachable."
+    "gpt-5.6 (sol/terra/luna) credits could not be confirmed from the primary source (help.openai.com's Codex rate card returned HTTP 403 to automated fetches). The values here were cross-checked across several independent secondary sources that agreed with each other and are internally consistent (Sol's $5/$30 per-MTok API price x 25 = 125/750 credits, matching usd_per_credit=0.04); see the gpt-5.6-* sources entries below. Treat effective_from (2026-07-01) as a placeholder and re-verify against the primary rate card when it becomes reachable.",
+    "gpt-6-astra: Codex standard token-based rates only; not API pricing, legacy message metering, or actual contractual charges. Confirmed 2026-09-06 JST (2026-09-05T17:18:23Z). effective_from is this observation cutoff, NOT an official launch or price-start time; earlier events remain unpriced. No Codex long-context surcharge or cache-write charge is applied. Fast multiplier 2.5 applies to explicit priority request settings with matching owned turn/model context; explicit default uses Standard. Missing or ambiguous Astra modes remain unpriced. These settings do not confirm the backend processing tier. See docs/astra-pricing.md for compatibility limits and the separate GPT-5.6 discrepancy report."
   ],
   "sources": [
     {
@@ -32,6 +33,21 @@
       "url": "https://rohitai.com/blog/openai-gpt-5-6-codex-rate-card-agentic-credit-pool",
       "retrieved_at": "2026-07-29",
       "note": "secondary source, used only for gpt-5.6 (sol/terra/luna) credits pending primary confirmation"
+    },
+    {
+      "url": "https://help.openai.com/en/articles/11481834-chatgpt-rate-card",
+      "retrieved_at": "2026-09-05T17:18:23Z",
+      "note": "Primary source: ChatGPT Work / Codex token credits and Astra Fast 2.5x. Existing GPT-5.6 values are preserved pending historical-period review; see docs/astra-pricing.md."
+    },
+    {
+      "url": "https://help.openai.com/en/articles/20001415-chatgpt-rate-card-enterprise-token-based-pricing",
+      "retrieved_at": "2026-09-05T17:18:23Z",
+      "note": "Primary source: Work / Codex standard USD token rates, Astra Fast 2.5x, no Codex Astra long-context surcharge and no Codex cache-write charge. Individual agreements may differ."
+    },
+    {
+      "url": "https://developers.openai.com/api/docs/models/gpt-6-astra",
+      "retrieved_at": "2026-09-05T17:18:23Z",
+      "note": "Boundary check only: API cache writes, long context and API Fast rates are NOT applied to this Codex entry."
     }
   ],
   "models": [
@@ -485,6 +501,30 @@
           "output": "10.0"
         }
       ]
+    },
+    {
+      "model_key": "gpt-6-astra",
+      "aliases": [],
+      "fast_multiplier": "2.5",
+      "credits_per_mtok": {
+        "input_nocache": "250.0",
+        "cache_read": "25.0",
+        "cache_write_5m": null,
+        "cache_write_1h": null,
+        "output": "1250.0"
+      },
+      "rates": [
+        {
+          "rate_id": "gpt-6-astra-codex-observed-2026-09-05",
+          "effective_from": "2026-09-05T17:18:23+00:00",
+          "effective_until": null,
+          "input_nocache": "10.0",
+          "cache_read": "1.0",
+          "cache_write_5m": null,
+          "cache_write_1h": null,
+          "output": "50.0"
+        }
+      ]
     }
   ]
 }