Add commands

Document claude-thermos and hanky from Hacker News discovery.

Author: SimonSchubert

assets/commands/claude-thermos.md

@@ -0,0 +1,93 @@
+# TAGLINE
+
+Keep Claude Code prompt cache warm during long subagent waits
+
+# TLDR
+
+**Run Claude Code** through the warmer (interactive)
+
+```uvx claude-thermos```
+
+**Pass a one-off prompt** through to Claude
+
+```uvx claude-thermos -p "[fix the bug]"```
+
+**Tune idle threshold** before warming starts (seconds)
+
+```uvx claude-thermos --idle [270]```
+
+**Change interval** between warm cycles (seconds)
+
+```uvx claude-thermos --interval [270]```
+
+**Cap warm cycles** per idle episode (or use auto for unlimited)
+
+```uvx claude-thermos --max-cycles [4]```
+
+**Disable warming** for one run via environment
+
+```CLAUDE_WARMER_DISABLE=1 uvx claude-thermos```
+
+# SYNOPSIS
+
+**claude-thermos** [**--idle** _SECONDS_] [**--interval** _SECONDS_] [**-n** | **--max-cycles** _N_|**auto**] [**--subagent-window** _SECONDS_] [_claude_args_ ...]
+
+# PARAMETERS
+
+**--idle** _SECONDS_
+> Seconds the main agent must be idle before a warming cycle runs. Default: **270**. Environment: **CLAUDE_WARMER_IDLE_THRESHOLD_SEC**.
+
+**--interval** _SECONDS_
+> Seconds between warming cycles. Default: **270**. Environment: **CLAUDE_WARMER_WARM_INTERVAL_SEC**.
+
+**-n** _N_|**auto**, **--max-cycles** _N_|**auto**
+> Maximum number of warm requests per idle episode, or **auto** for unlimited. Default: **4**. Environment: **CLAUDE_WARMER_WARM_MAX_CYCLES**.
+
+**--subagent-window** _SECONDS_
+> How long a subagent still counts as active after its last traffic. Default: **540**. Environment: **CLAUDE_WARMER_SUBAGENT_ACTIVE_WINDOW_SEC**.
+
+**-V**, **--version**
+> Print version and exit.
+
+**-h**, **--help**
+> Show help.
+
+_claude_args_ ...
+> Remaining arguments are passed through unchanged to the **claude** CLI.
+
+# DESCRIPTION
+
+**claude-thermos** launches Anthropic's **Claude Code** CLI behind a small local reverse proxy that keeps the main agent's prompt cache alive while subagents run. Claude Code's prompt cache has a roughly **5-minute TTL**. When the main agent waits on a subagent longer than that, the main conversation prefix expires and the next turn re-encodes the full history at the more expensive cache-write rate. Warming sends cheap refresh requests (identical cacheable prefix, **max_tokens: 1**) so resume pays a cache read instead of a full rewrite.
+
+The wrapper points **ANTHROPIC_BASE_URL** at a loopback proxy that observes **/v1/messages** traffic. Lineages (cache prefixes) are keyed by model, tool set, and system text: the first tool-bearing lineage is treated as the main agent; others as subagents. When the main lineage is idle and a subagent is still active, the warmer fires on an interval under the TTL. Warm requests go **directly** to the Anthropic API (not through the proxy) so they do not interfere with real session traffic.
+
+Session telemetry is written under **~/.claude-thermos/logs/**_session_id_**/`events.jsonl`** and **summary.json**, including warm counts and estimated rewrite tokens avoided. Requires **Python 3.11+** and the **claude** binary on **PATH**. Typical invocation is via **uvx** so the package need not be installed permanently.
+
+# CONFIGURATION
+
+**CLAUDE_WARMER_DISABLE**
+> Set to **1** to run Claude without warming for that process.
+
+**CLAUDE_WARMER_IDLE_THRESHOLD_SEC**, **CLAUDE_WARMER_WARM_INTERVAL_SEC**, **CLAUDE_WARMER_WARM_MAX_CYCLES**, **CLAUDE_WARMER_SUBAGENT_ACTIVE_WINDOW_SEC**
+> Environment overrides for the matching CLI flags.
+
+**~/.claude-thermos/logs/**_session_id_**/
+> Per-session **events.jsonl** (request usage and warm decisions) and **summary.json** (rollup savings estimates).
+
+# CAVEATS
+
+Does not replace **claude**; it wraps it. Warming still incurs small cache-read cost and network traffic to Anthropic. Savings estimates in **summary.json** are in base-input-token units, not dollars—multiply by your model's input price. Behavior depends on Claude Code's request shape and Anthropic's cache TTL remaining about five minutes. Requires a working Claude Code install and valid Anthropic authentication.
+
+# HISTORY
+
+Created by **Iaroslav Zeigerman** as a Python package (**claude-thermos** on PyPI) that uses a local reverse proxy (via **mitmproxy**) to refresh Claude Code prompt-cache prefixes during long subagent waits. Released under the MIT license.
+
+# SEE ALSO
+
+[claude](/man/claude)(1), [uvx](/man/uvx)(1), [uv](/man/uv)(1)
+
+# RESOURCES
+
+```[Source code](https://github.com/izeigerman/claude-thermos)```
+
+<!-- verified: 2026-07-23 -->

assets/commands/hanky.md

@@ -0,0 +1,109 @@
+# TAGLINE
+
+CLI and pipeline framework for loading flash cards into Anki
+
+# TLDR
+
+**Import a CSV** of cards into a deck (Basic note type by default)
+
+```hanky pipe [words.csv] --into [english::vocab]```
+
+**Import every matching file** under a directory into nested decks
+
+```hanky pipe-dir [french/] ["*.csv"] -r```
+
+**Override the Anki note type** / model name
+
+```hanky pipe [cards.csv] -m [Basic] --into [deck]```
+
+**Dry-run** processors without writing the collection
+
+```hanky pipe [words.csv] --dry-run```
+
+**Verbose report** of each processed card
+
+```hanky pipe [words.csv] -v --into [deck]```
+
+**Run a custom pipeline script** the same way as the hanky binary
+
+```python3 [my_script.py] pipe [words.csv] --into [english::vocab]```
+
+# SYNOPSIS
+
+**hanky** **pipe** [**-m** _MODEL_] [**--into** _DECK_] [**--fail-fast**] [**--dry-run**] [**-v**] [**--args** _K=V_ ...] _file_
+
+**hanky** **pipe-dir** [**-m** _MODEL_] [**-r**] [**--fail-fast**] [**--dry-run**] [**-v**] [**--args** _K=V_ ...] _dir_ _pattern_
+
+# PARAMETERS
+
+**pipe** _file_
+> Load cards from a single file (**.csv**, **.json**, or a registered extension) and add them to Anki.
+
+**pipe-dir** _dir_ _pattern_
+> Load many files under _dir_ matching the glob _pattern_. Deck names are derived from relative paths (e.g. **french/animals.csv** → **french::animals**).
+
+**-m** _MODEL_, **--model** _MODEL_
+> Override the Anki model (note type) used for new cards. Standalone **hanky** defaults to **Basic**; pipeline scripts set the model in **HankyPipeline(...)**.
+
+**--into** _DECK_
+> Destination deck for **pipe**. Defaults to the input filename without extension.
+
+**-r**, **--recursive**
+> For **pipe-dir**, also descend into subdirectories.
+
+**--fail-fast**
+> Stop and raise on the first card that cannot be added instead of skipping and reporting at the end.
+
+**--dry-run**
+> Run card processors and report results without modifying the Anki collection (no decks, media, cards, or backup).
+
+**-v**, **--verbose**
+> Print each card dict after processors run, plus errors in the final report.
+
+**--args** _K=V_
+> Key=value pairs forwarded to custom card processors (pipeline scripts only). May be repeated.
+
+# DESCRIPTION
+
+**hanky** is a Python CLI and library that reads flash cards from files (or in-script sources), optionally transforms them through user-defined **card processors**, and writes notes into a local **Anki** collection. It is aimed at ETL-style workflows: scrape or OCR material, enrich fields (translations, examples, TTS audio), then bulk-load into decks.
+
+Built-in **pipe** / **pipe-dir** commands work with CSV and JSON out of the box when column or key names match the target note type fields. For custom logic, install the package, write a small script that registers **@hanky.card_processor** functions on a **HankyPipeline**, and call **hanky.run()** so the same CLI surface is available as **python3 my_script.py pipe ...**. Processors may attach audio media via **CardMedia** (**.mp3**, **.oga**, **.opus**, **.wav**, **.weba**, **.aac**). Custom file loaders can be registered for formats such as Excel.
+
+hanky only **adds** cards, media, and decks. Note types (models) must already exist in the collection; create them in the Anki UI first. Requires Python **3.11+** and a local Anki collection database. Not affiliated with AnkiWeb or the Anki project.
+
+# CONFIGURATION
+
+**~/.config/hanky/hanky.toml**
+> Optional TOML config. Needed mainly if you do not use the default Anki profile (**User 1**) or you are on Windows.
+
+**ANKI_DB_PATH**
+> Path to the Anki collection SQLite file (**collection.anki2**). Defaults to the **User 1** profile under the platform Anki2 data directory.
+
+**DO_SAFETY_CHECK**
+> When true (default), refuse to open the collection if another process (e.g. Anki) appears to hold it.
+
+**ALLOW_DUPLICATES**
+> When false (default), skip cards that duplicate existing notes.
+
+**BACKUP_FOLDER**
+> Directory for collection backups before writes. Default: **~/.local/share/hanky/backups**.
+
+A **Config** object passed to **HankyPipeline** in a script overrides file settings.
+
+# CAVEATS
+
+Close Anki (or disable the safety check only if you accept corruption risk) before writing the collection. Note types must pre-exist; hanky does not create models. Duplicate detection and field names depend on your note type. Windows and non-default profiles almost always need **ANKI_DB_PATH**. Custom processors that call external APIs (translation, TTS, vision) need their own credentials and dependencies.
+
+# HISTORY
+
+Written by **HBA** as an open-source Python package on PyPI under the MIT license. Provides both a **hanky** console script and a library API for programmable Anki import pipelines.
+
+# SEE ALSO
+
+[anki](/man/anki)(1), [python](/man/python)(1), [pip](/man/pip)(1)
+
+# RESOURCES
+
+```[Source code](https://github.com/Haeata-Ash/hanky)```
+
+<!-- verified: 2026-07-23 -->

assets/commands/index.txt

@@ -1208,6 +1208,7 @@ clang.md
 clangd.md
 claude-pee.md
 claude-pulse.md
+claude-thermos.md
 claude.md
 claumon.md
 clawk.md
@@ -3206,6 +3207,7 @@ handler.md
 handlr.md
 hangman.md
 hangups.md
+hanky.md
 hapless.md
 happy.md
 haproxy.md