shell-prompter 0.2.0__tar.gz → 0.6.0__tar.gz

{shell_prompter-0.2.0/shell_prompter.egg-info → shell_prompter-0.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: shell-prompter
-Version: 0.2.0
+Version: 0.6.0
 Summary: A natural-language shell agent powered by Claude, OpenAI, or Gemini.
 Author-email: radinkv <26397342+Radinkv@users.noreply.github.com>
 License: MIT
@@ -38,6 +38,8 @@ endpoint such as Groq or OpenRouter), or Gemini.
 ```
 $ prompter "make a folder called scratch, cd into it, then run claude"
 $ prompter "download uv if it isn't installed, then print its version"
+$ prompter "write a C++ linked list, compile it with clang++, and run it"
+$ prompter "make a venv in ~/Documents/ml, install scikit-learn, open it in VS Code"
 $ prompter
 ```
 
@@ -110,6 +112,10 @@ On first run prompter writes `~/.prompter/config.json`:
 | `api_key_env` | The environment variable that holds the API key. Defaults to `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, or `GEMINI_API_KEY`. |
 | `max_fix_attempts` | How many commands may fail in a row before prompter stops. See [Self-repair](#self-repair). |
 | `auto_approve_safe` | Set to `false` to confirm every command, including safe ones. |
+| `concise` | Set to `true` to make the model write code and scripts with no comments and skip the closing explanation. Saves tokens. Off by default; override per run with `--concise` / `--verbose`. |
+| `show_usage` | Set to `true` to print token usage after each run. Off by default; enable per run with `--usage`. |
+| `render_markdown` | Render the model's markdown (bold, headings, lists, code) as styled terminal text. On by default; disable per run with `--raw`. Always off when output isn't a TTY, so pipes stay plain. |
+| `pricing` | Optional per-model price map for a cost estimate, e.g. `{"claude-sonnet-4-6": {"input": 3.0, "output": 15.0}}` (USD per million tokens). Without it the usage line shows tokens only. |
 | `preferences` | Free-form lines passed straight to the model, such as `"Use pnpm, not npm."` |
 
 Run `prompter config` to print the path, or `prompter status` to see the
@@ -174,9 +180,9 @@ decides whether it runs on its own or asks first.
 
 | Tier | Examples | Default |
 |------|----------|---------|
-| 🟢 SAFE | `ls`, `cd`, `mkdir`, `git status`, `cat` | runs automatically |
-| 🟡 CONFIRM | `brew install`, `pip install`, `git clone`, `mv`, `rm`, `curl` | asks first |
-| 🔴 DANGER | `rm -rf`, `sudo`, `curl ... \| sh`, force-push, `dd` | asks first, shown in red |
+| 🟢 SAFE | `ls`, `cd`, `mkdir`, `git status`, `cat`, `... 2>/dev/null` | runs automatically |
+| 🟡 CONFIRM | `brew install`, `pip install`, `git clone`, `mv`, `rm`, `curl`, `git clean -f` | asks first |
+| 🔴 DANGER | `rm -rf`, `sudo`/`su`/`pkexec`, `curl ... \| sh`, force-push, `dd`, `shutdown`, `shred` | asks first, shown in red |
 
 When prompter asks, you answer:
 
@@ -185,8 +191,10 @@ When prompter asks, you answer:
 - `a`: run it and auto-approve the rest of this run
 - `q`: quit
 
-`curl ... | sh` is always DANGER, because it runs code you have not seen.
-`--yolo` removes the gate entirely. Use it only when you trust the task.
+Piping a download into an interpreter (`curl ... | sh`, and likewise `bash`,
+`python`, `perl`, `ruby`, `node`) is always DANGER, because it runs code you
+have not seen. `--yolo` removes the gate entirely. Use it only when you trust
+the task.
 
 ## Self-repair
 
@@ -199,6 +207,12 @@ commands that fail in a row. At the limit, prompter tells the model to stop and
 explain what went wrong. A command you decline does not count. Only commands
 that ran and failed do.
 
+This is separate from network resilience. When a *model request* fails
+transiently (a rate limit, timeout, or 5xx), prompter retries it automatically
+with exponential backoff before giving up; a bad request or auth failure is not
+retried. A retry only happens before any of the turn's response has streamed, so
+output is never printed twice.
+
 ## How it works
 
 1. Your request, plus your OS, shell, current directory, default workspace, and
@@ -234,6 +248,7 @@ opens a prompt.
 | `prompter status` | Show the current provider, model, workspace, and keys. |
 | `prompter config` | Print the config file path. |
 | `prompter completions <shell>` | Print a tab-completion script (zsh, bash). See [Tab completion](#tab-completion). |
+| `prompter version` | Print the installed version (also `--version`, `-V`). |
 | `prompter help` | Show usage. |
 
 ## Tab completion
@@ -281,6 +296,10 @@ Flags modify a single run and sit alongside the goal (`prompter --yolo "..."`).
 | `--max-fix N` | Override `max_fix_attempts`. |
 | `--ask-all` | Confirm every command, including safe ones. |
 | `--yolo` | Run everything with no confirmation. Dangerous. |
+| `--concise` | No code comments, minimal explanation, for this run. |
+| `--verbose` | Full comments and explanation, even if `concise` is set in config. |
+| `--usage` | Print token usage (and cost, if `pricing` is set) after the run. |
+| `--raw` | Print the model's text as-is, without rendering markdown. |
 
 ## Project layout
 

{shell_prompter-0.2.0 → shell_prompter-0.6.0}/README.md

@@ -11,6 +11,8 @@ endpoint such as Groq or OpenRouter), or Gemini.
 ```
 $ prompter "make a folder called scratch, cd into it, then run claude"
 $ prompter "download uv if it isn't installed, then print its version"
+$ prompter "write a C++ linked list, compile it with clang++, and run it"
+$ prompter "make a venv in ~/Documents/ml, install scikit-learn, open it in VS Code"
 $ prompter
 ```
 
@@ -83,6 +85,10 @@ On first run prompter writes `~/.prompter/config.json`:
 | `api_key_env` | The environment variable that holds the API key. Defaults to `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, or `GEMINI_API_KEY`. |
 | `max_fix_attempts` | How many commands may fail in a row before prompter stops. See [Self-repair](#self-repair). |
 | `auto_approve_safe` | Set to `false` to confirm every command, including safe ones. |
+| `concise` | Set to `true` to make the model write code and scripts with no comments and skip the closing explanation. Saves tokens. Off by default; override per run with `--concise` / `--verbose`. |
+| `show_usage` | Set to `true` to print token usage after each run. Off by default; enable per run with `--usage`. |
+| `render_markdown` | Render the model's markdown (bold, headings, lists, code) as styled terminal text. On by default; disable per run with `--raw`. Always off when output isn't a TTY, so pipes stay plain. |
+| `pricing` | Optional per-model price map for a cost estimate, e.g. `{"claude-sonnet-4-6": {"input": 3.0, "output": 15.0}}` (USD per million tokens). Without it the usage line shows tokens only. |
 | `preferences` | Free-form lines passed straight to the model, such as `"Use pnpm, not npm."` |
 
 Run `prompter config` to print the path, or `prompter status` to see the
@@ -147,9 +153,9 @@ decides whether it runs on its own or asks first.
 
 | Tier | Examples | Default |
 |------|----------|---------|
-| 🟢 SAFE | `ls`, `cd`, `mkdir`, `git status`, `cat` | runs automatically |
-| 🟡 CONFIRM | `brew install`, `pip install`, `git clone`, `mv`, `rm`, `curl` | asks first |
-| 🔴 DANGER | `rm -rf`, `sudo`, `curl ... \| sh`, force-push, `dd` | asks first, shown in red |
+| 🟢 SAFE | `ls`, `cd`, `mkdir`, `git status`, `cat`, `... 2>/dev/null` | runs automatically |
+| 🟡 CONFIRM | `brew install`, `pip install`, `git clone`, `mv`, `rm`, `curl`, `git clean -f` | asks first |
+| 🔴 DANGER | `rm -rf`, `sudo`/`su`/`pkexec`, `curl ... \| sh`, force-push, `dd`, `shutdown`, `shred` | asks first, shown in red |
 
 When prompter asks, you answer:
 
@@ -158,8 +164,10 @@ When prompter asks, you answer:
 - `a`: run it and auto-approve the rest of this run
 - `q`: quit
 
-`curl ... | sh` is always DANGER, because it runs code you have not seen.
-`--yolo` removes the gate entirely. Use it only when you trust the task.
+Piping a download into an interpreter (`curl ... | sh`, and likewise `bash`,
+`python`, `perl`, `ruby`, `node`) is always DANGER, because it runs code you
+have not seen. `--yolo` removes the gate entirely. Use it only when you trust
+the task.
 
 ## Self-repair
 
@@ -172,6 +180,12 @@ commands that fail in a row. At the limit, prompter tells the model to stop and
 explain what went wrong. A command you decline does not count. Only commands
 that ran and failed do.
 
+This is separate from network resilience. When a *model request* fails
+transiently (a rate limit, timeout, or 5xx), prompter retries it automatically
+with exponential backoff before giving up; a bad request or auth failure is not
+retried. A retry only happens before any of the turn's response has streamed, so
+output is never printed twice.
+
 ## How it works
 
 1. Your request, plus your OS, shell, current directory, default workspace, and
@@ -207,6 +221,7 @@ opens a prompt.
 | `prompter status` | Show the current provider, model, workspace, and keys. |
 | `prompter config` | Print the config file path. |
 | `prompter completions <shell>` | Print a tab-completion script (zsh, bash). See [Tab completion](#tab-completion). |
+| `prompter version` | Print the installed version (also `--version`, `-V`). |
 | `prompter help` | Show usage. |
 
 ## Tab completion
@@ -254,6 +269,10 @@ Flags modify a single run and sit alongside the goal (`prompter --yolo "..."`).
 | `--max-fix N` | Override `max_fix_attempts`. |
 | `--ask-all` | Confirm every command, including safe ones. |
 | `--yolo` | Run everything with no confirmation. Dangerous. |
+| `--concise` | No code comments, minimal explanation, for this run. |
+| `--verbose` | Full comments and explanation, even if `concise` is set in config. |
+| `--usage` | Print token usage (and cost, if `pricing` is set) after the run. |
+| `--raw` | Print the model's text as-is, without rendering markdown. |
 
 ## Project layout
 

{shell_prompter-0.2.0 → shell_prompter-0.6.0}/prompter/__init__.py

@@ -16,6 +16,7 @@ the tools it knows how to launch.
 from __future__ import annotations
 
 from .cli import main
+from .config import program_version
 
 __all__ = ["main"]
-__version__ = "0.1.0"
+__version__ = program_version()

{shell_prompter-0.2.0 → shell_prompter-0.6.0}/prompter/agent.py

@@ -8,17 +8,22 @@ changes nothing here.
 
 from __future__ import annotations
 
+import time
 from dataclasses import dataclass
 
 from .config import ApprovalMode, Config
+from .cost import estimate as estimate_cost
+from .history import compact
 from .prompts import build_system_texts
 from .providers.base import (
     AssistantMessage,
     AssistantTurn,
     ModelProvider,
+    ProviderError,
     ToolInvocation,
     ToolResult,
     ToolResultsMessage,
+    Usage,
     UserMessage,
 )
 from .risk import RiskAssessment, RiskTier, classify
@@ -41,6 +46,9 @@ _PAYLOAD_TEMPLATE = (
     "stderr:\n{stderr}"
 )
 
+MAX_RETRIES = 3
+_RETRY_BASE_DELAY_SECONDS = 2
+
 
 class QuitRequested(Exception):
     """Raised when the user chooses 'quit' at a confirmation prompt."""
@@ -102,14 +110,25 @@ class Agent:
         self._approve_all = False
         self.consecutive_failures = 0
         self._force_stop = False
+        self._run_usage = Usage()
 
     def run_turn(self, user_text: str) -> None:
         self.consecutive_failures = 0
         self._force_stop = False
+        self._run_usage = Usage()
         self.conversation.add_user(user_text)
         another_round = True
         while another_round:
             another_round = self._run_round()
+        self._report_usage()
+
+    def _report_usage(self) -> None:
+        if not self.config.show_usage:
+            return
+        usage = self._run_usage
+        cost = estimate_cost(self.config.resolved_model, usage, self.config.pricing)
+        self.console.usage(
+            usage.input_tokens, usage.output_tokens, usage.cached_tokens, cost)
 
     def _run_round(self) -> bool:
         """Run one model turn and return whether another round is needed."""
@@ -124,16 +143,46 @@ class Agent:
 
     def _complete(self) -> AssistantTurn:
         system_texts = build_system_texts(self.config, self.shell.cwd)
-        self.console.begin_stream()
-        turn = self.provider.complete(
-            self.conversation.history,
-            system_texts,
-            self._force_stop,
-            self.console.stream_text,
-        )
-        self.console.end_stream()
+        history = compact(self.conversation.history)
+        turn = self._complete_with_retry(history, system_texts)
+        if turn.usage is not None:
+            self._run_usage = self._run_usage + turn.usage
         return turn
 
+    def _complete_with_retry(self, history, system_texts) -> AssistantTurn:
+        """Call the provider, retrying transient failures with backoff.
+
+        A retry replays the whole turn, so it is only safe before any text has
+        streamed to the console; once output has started, a transient failure
+        propagates rather than risk printing the turn twice.
+        """
+        streamed = [False]
+
+        def on_text(text: str) -> None:
+            streamed[0] = True
+            self.console.stream_text(text)
+
+        attempt = 0
+        while True:
+            streamed[0] = False
+            self.console.begin_stream()
+            try:
+                turn = self.provider.complete(
+                    history, system_texts, self._force_stop, on_text)
+                self.console.end_stream()
+                return turn
+            except ProviderError as e:
+                self.console.end_stream()
+                if not self._may_retry(e, attempt, streamed[0]):
+                    raise
+                delay = _RETRY_BASE_DELAY_SECONDS * (2 ** attempt)
+                self.console.retry_notice(str(e), attempt + 1, MAX_RETRIES, delay)
+                time.sleep(delay)
+                attempt += 1
+
+    def _may_retry(self, error: ProviderError, attempt: int, streamed: bool) -> bool:
+        return error.retryable and attempt < MAX_RETRIES and not streamed
+
     def _process_tool_calls(self, tool_calls: list[ToolInvocation]) -> list[ToolResult]:
         results = []
         for invocation in tool_calls:

{shell_prompter-0.2.0 → shell_prompter-0.6.0}/prompter/cli.py

@@ -24,7 +24,9 @@ from .config import (
     default_model_for,
     load_config,
     normalize_provider,
+    program_version,
     save_config,
+    source_version,
 )
 from .completion import (
     CommandSurface,
@@ -54,7 +56,14 @@ _CMD_CONFIG = "config"
 _CMD_RUN = "run"
 _CMD_HELP = "help"
 _CMD_COMPLETIONS = "completions"
+_CMD_VERSION = "version"
 _HELP_FLAGS = ("-h", "--help")
+_VERSION_FLAGS = ("-V", "--version")
+_VERSION_LINE = "{program} {version}"
+_VERSION_DRIFT = (
+    "note: pyproject.toml declares {source}, but the installed package is "
+    "{installed}. Run `pip install -e .` to refresh."
+)
 
 _KEYS_LIST = "list"
 _KEYS_ADD = "add"
@@ -71,6 +80,10 @@ _FLAG_WORKSPACE = "--workspace"
 _FLAG_MAX_FIX = "--max-fix"
 _FLAG_ASK_ALL = "--ask-all"
 _FLAG_YOLO = "--yolo"
+_FLAG_CONCISE = "--concise"
+_FLAG_VERBOSE = "--verbose"
+_FLAG_USAGE = "--usage"
+_FLAG_RAW = "--raw"
 _ACTION_STORE_TRUE = "store_true"
 
 _LABEL_ADD_IT = "add it"
@@ -156,6 +169,10 @@ Run flags (modify one run):
   --max-fix N                       failures in a row before stopping
   --ask-all                         confirm every command
   --yolo                            run everything with no confirmation
+  --concise                         no code comments, minimal explanation
+  --verbose                         full comments and explanation (overrides --concise/config)
+  --usage                           print token usage (and cost if priced) per run
+  --raw                             don't render markdown; print the model's text as-is
 
 Manage:
   prompter keys add <provider> <key>    store an API key
@@ -165,6 +182,7 @@ Manage:
   prompter status                       show the current setup
   prompter config                       print the config file path
   prompter completions <shell>          print a tab-completion script (zsh, bash)
+  prompter version                      print the installed version
   prompter help                         show this help
 
 Providers: anthropic, openai, gemini
@@ -251,6 +269,10 @@ def _run_parser() -> argparse.ArgumentParser:
     parser.add_argument(_FLAG_MAX_FIX, type=int, default=None)
     parser.add_argument(_FLAG_ASK_ALL, action=_ACTION_STORE_TRUE)
     parser.add_argument(_FLAG_YOLO, action=_ACTION_STORE_TRUE)
+    parser.add_argument(_FLAG_CONCISE, action=_ACTION_STORE_TRUE)
+    parser.add_argument(_FLAG_VERBOSE, action=_ACTION_STORE_TRUE)
+    parser.add_argument(_FLAG_USAGE, action=_ACTION_STORE_TRUE)
+    parser.add_argument(_FLAG_RAW, action=_ACTION_STORE_TRUE)
     return parser
 
 
@@ -278,6 +300,14 @@ def _apply_overrides(args, config: Config) -> None:
         config.default_workspace = args.workspace
     if args.max_fix is not None:
         config.max_fix_attempts = args.max_fix
+    if args.concise:
+        config.concise = True
+    if args.verbose:        # an explicit --verbose wins over --concise and config
+        config.concise = False
+    if args.usage:
+        config.show_usage = True
+    if args.raw:
+        config.render_markdown = False
     pinned = config.model if config_provider == config.provider else EMPTY
     config.model = args.model or pinned or default_model_for(config.provider)
 
@@ -357,6 +387,7 @@ def cmd_run(argv: list[str], console: Console) -> int:
         agent, config, mode = _prepare(args, console)
     except _CommandError as e:
         return _fail(console, e.title, e.hints, e.detail)
+    console.set_markdown(config.render_markdown)
     console.banner(config, mode)
     return _run_agent(agent, config, console, SPACE.join(args.prompt).strip())
 
@@ -458,6 +489,15 @@ def cmd_help(console: Console) -> int:
     return OK_EXIT_CODE
 
 
+def cmd_version(console: Console) -> int:
+    installed = program_version()
+    console.info(_VERSION_LINE.format(program=PROGRAM_NAME, version=installed))
+    source = source_version()
+    if source is not None and source != installed:
+        console.note(_VERSION_DRIFT.format(source=source, installed=installed))
+    return OK_EXIT_CODE
+
+
 _COMMANDS = {
     _CMD_KEYS: cmd_keys,
     _CMD_USE: cmd_use,
@@ -499,6 +539,8 @@ def main(argv: list[str] | None = None) -> int:
     head = argv[0]
     if head in _HELP_FLAGS or head == _CMD_HELP:
         return cmd_help(console)
+    if head in _VERSION_FLAGS or head == _CMD_VERSION:
+        return cmd_version(console)
     if head == _CMD_RUN:
         return cmd_run(argv[1:], console)
     handler = _COMMANDS.get(head)

{shell_prompter-0.2.0 → shell_prompter-0.6.0}/prompter/colors.py

@@ -12,6 +12,7 @@ ENV_NO_COLOR = "NO_COLOR"
 _RESET = "\033[0m"
 _BOLD = "\033[1m"
 _DIM = "\033[2m"
+_ITALIC = "\033[3m"
 _RED = "\033[31m"
 _GREEN = "\033[32m"
 _YELLOW = "\033[33m"
@@ -36,6 +37,7 @@ class Palette:
         self.RESET = pick(_RESET)
         self.BOLD = pick(_BOLD)
         self.DIM = pick(_DIM)
+        self.ITALIC = pick(_ITALIC)
         self.RED = pick(_RED)
         self.GREEN = pick(_GREEN)
         self.YELLOW = pick(_YELLOW)

{shell_prompter-0.2.0 → shell_prompter-0.6.0}/prompter/completion.py

@@ -182,7 +182,13 @@ def _value_flag_cases(shell: str, surface: CommandSurface) -> str:
 
 
 def render(shell: str, surface: CommandSurface) -> str:
-    """Return the completion script for ``shell`` (one of SUPPORTED_SHELLS)."""
+    """Return the completion script for ``shell`` (one of SUPPORTED_SHELLS).
+
+    The surface's commands, flags, and provider names are interpolated into the
+    script unquoted. They are code-defined identifiers (dispatch keys, argparse
+    flag strings, registry names), never external input, so they are assumed
+    shell-safe -- keep them that way (plain ``[\\w-]`` tokens) if you add any.
+    """
     template = _ZSH_TEMPLATE if shell == ZSH else _BASH_TEMPLATE
     return template.format(
         program=PROGRAM_NAME,

{shell_prompter-0.2.0 → shell_prompter-0.6.0}/prompter/config.py

@@ -13,14 +13,63 @@ from __future__ import annotations
 
 import json
 import os
+import re
 import sys
 from dataclasses import asdict, dataclass, field, fields
 from enum import Enum
+from importlib.metadata import PackageNotFoundError, version as _distribution_version
+from pathlib import Path
 
 from .colors import palette
 from .constants import EMPTY, FILE_WRITE_MODE, JSON_INDENT
 
 PROGRAM_NAME = "prompter"
+DISTRIBUTION_NAME = "shell-prompter"
+_UNKNOWN_VERSION = "0.0.0+unknown"
+
+
+def program_version() -> str:
+    """The installed package version, read from distribution metadata.
+
+    Single source of truth is pyproject.toml; the metadata carries it into the
+    built package. Falls back when running from a tree that was never installed.
+    """
+    try:
+        return _distribution_version(DISTRIBUTION_NAME)
+    except PackageNotFoundError:
+        return _UNKNOWN_VERSION
+
+
+_PYPROJECT_NAME = "pyproject.toml"
+_PROJECT_HEADER_RE = re.compile(r"^\[project\]\s*$", re.MULTILINE)
+_TABLE_HEADER_RE = re.compile(r"^\[", re.MULTILINE)
+_PYPROJECT_VERSION_RE = re.compile(r'^version\s*=\s*["\']([^"\']+)["\']', re.MULTILINE)
+
+
+def _version_from_pyproject(text: str) -> str | None:
+    section = _PROJECT_HEADER_RE.search(text)
+    if section is None:
+        return None
+    next_header = _TABLE_HEADER_RE.search(text, section.end())
+    project_body = text[section.end():next_header.start()] if next_header else text[section.end():]
+    match = _PYPROJECT_VERSION_RE.search(project_body)
+    return match.group(1) if match else None
+
+
+def source_version() -> str | None:
+    """The version declared in a sibling pyproject.toml's [project] table, or None.
+
+    Only a source checkout has pyproject.toml next to the package; an installed
+    wheel does not ship it. So a non-None result that differs from
+    program_version() means the editable install's metadata is stale -- the dev
+    bumped the version but has not reinstalled. End users never hit this path.
+    """
+    pyproject = Path(__file__).resolve().parent.parent / _PYPROJECT_NAME
+    try:
+        text = pyproject.read_text()
+    except OSError:
+        return None
+    return _version_from_pyproject(text)
 
 PROVIDER_ANTHROPIC = "anthropic"
 PROVIDER_OPENAI = "openai"
@@ -88,6 +137,10 @@ class Config:
     api_key_env: str | None = None
     max_fix_attempts: int = DEFAULT_MAX_FIX_ATTEMPTS
     auto_approve_safe: bool = True
+    concise: bool = False
+    show_usage: bool = False
+    render_markdown: bool = True
+    pricing: dict = field(default_factory=dict)
     preferences: list[str] = field(
         default_factory=lambda: list(DEFAULT_PREFERENCES)
     )

shell_prompter-0.6.0/prompter/cost.py

@@ -0,0 +1,31 @@
+"""Estimate the dollar cost of a run from token usage and configured prices.
+
+Prices are not hard-coded: they move, and they differ per model and provider.
+The user supplies them in config under ``pricing``, keyed by model id, as USD
+per million input and output tokens::
+
+    "pricing": {"claude-sonnet-4-6": {"input": 3.0, "output": 15.0}}
+
+With no price for the active model, estimate() returns None and the usage line
+simply shows token counts. The estimate charges the whole prompt at the input
+rate (it does not discount the cached portion, which providers bill cheaper), so
+it is a slight upper bound -- hence "est".
+"""
+
+from __future__ import annotations
+
+from .providers.base import Usage
+
+PRICE_INPUT = "input"
+PRICE_OUTPUT = "output"
+_PER_MILLION = 1_000_000
+
+
+def estimate(model: str, usage: Usage, pricing: dict) -> float | None:
+    """USD estimate for one run, or None when the model has no configured price."""
+    rates = pricing.get(model)
+    if not rates:
+        return None
+    in_rate = rates.get(PRICE_INPUT, 0) or 0
+    out_rate = rates.get(PRICE_OUTPUT, 0) or 0
+    return (usage.input_tokens * in_rate + usage.output_tokens * out_rate) / _PER_MILLION

shell_prompter-0.6.0/prompter/history.py

@@ -0,0 +1,65 @@
+"""Context compaction: shrink the history re-sent to the model each turn.
+
+A multi-step task resends the whole conversation on every turn, so old command
+outputs pile up as input tokens. compact() keeps the goal, the most recent tool
+results, and any failures verbatim, and collapses older *successful* outputs to
+their first line (the exit-code header) with the body elided. The model
+already reacted to the full output when the command ran; a later turn only needs
+to know the step succeeded.
+
+It is a pure transform applied just before a provider call. The stored
+conversation is never mutated, so nothing is lost from the program's own record;
+only the copy handed to the model is slimmed. Failures survive untouched because
+self-repair reads their error text, and short outputs survive because collapsing
+them would save nothing.
+"""
+
+from __future__ import annotations
+
+from .constants import NEWLINE
+from .providers.base import HistoryItem, ToolResult, ToolResultsMessage
+
+# Tool-result messages kept verbatim, counting back from the latest turn.
+DEFAULT_KEEP_RECENT = 2
+# Successful outputs at or under this length are left alone; collapsing a short
+# output would replace it with a marker of about the same size.
+COLLAPSE_THRESHOLD = 240
+_ELISION = "... [output elided to save context] ..."
+
+
+def compact(history: list[HistoryItem],
+            keep_recent: int = DEFAULT_KEEP_RECENT) -> list[HistoryItem]:
+    """Return a copy of history with old, large, successful outputs collapsed."""
+    total = sum(isinstance(item, ToolResultsMessage) for item in history)
+    seen = 0
+    out: list[HistoryItem] = []
+    for item in history:
+        if isinstance(item, ToolResultsMessage):
+            seen += 1
+            recent = seen > total - keep_recent
+            out.append(item if recent else _collapse_message(item))
+        else:
+            out.append(item)
+    return out
+
+
+def _collapse_message(message: ToolResultsMessage) -> ToolResultsMessage:
+    return ToolResultsMessage([_collapse_result(r) for r in message.results])
+
+
+def _collapse_result(result: ToolResult) -> ToolResult:
+    if result.is_error or len(result.content) <= COLLAPSE_THRESHOLD:
+        return result
+    return ToolResult(result.call_id, _shorten(result.content), result.is_error)
+
+
+def _shorten(content: str) -> str:
+    """Keep the first line (the exit-code header) and elide the rest.
+
+    A single-line payload (no header/body split) has no body to elide, so cap it
+    by character budget instead of growing it with a marker.
+    """
+    head, separator, _body = content.partition(NEWLINE)
+    if not separator:
+        return content[:COLLAPSE_THRESHOLD]
+    return head + NEWLINE + _ELISION