refresh-model-context-windows

name	refresh-model-context-windows
description	Refresh sessions/models.json (the Claude model context-window table that bin/sessions reads for its CONTEXT column) from Anthropic's public models documentation. Fetches the live docs with WebFetch — no API key required — regenerates the table deterministically, and commits only if the model data actually changed.

Refresh model context windows

bin/sessions renders the CONTEXT column as N/M, where M is the model's context window. It reads those windows from sessions/models.json (symlinked to $XDG_CONFIG_HOME/sessions/models.json by install.sh). This skill refreshes that table from Anthropic's public models page — it needs no API key, only a WebFetch, so it runs fine inside any Claude Code session.

When a model isn't in the table, bin/sessions shows the raw token count in lavender instead of inventing a denominator, so a stale table degrades gracefully — but keeping it current is the point of running this weekly.

Steps

1. Confirm a clean tree for sessions/models.json: git status --porcelain sessions/models.json. If it has uncommitted changes, stop and tell the user — don't fold their edits into an automated refresh.

2. Fetch the live model data with WebFetch:

   • URL: https://platform.claude.com/docs/en/about-claude/models/overview.md
   • Prompt: ask for every Claude model with its Claude API alias, its context window / max input tokens, and its max output tokens, across both the current and legacy tables.

3. Build the new table in the exact shape of the existing file:

   • Top-level keys: _comment (keep the current one verbatim), updated, models.
   • Under models, one entry per model, keyed by its bare Claude API alias (e.g. claude-opus-4-8, not claude-opus-4-8-20251101). For a model with no alias, use its API ID. bin/sessions strips [1m] / -fast / -YYYYMMDD suffixes before lookup, so keys must be the bare alias.
   • Each entry: { "context": <max input tokens>, "max_output": <max output tokens> } as integers (1M → 1000000, 200k → 200000, 128k → 128000).
   • Sort the model keys alphabetically and use 2-space indentation, so diffs stay minimal and reviewable.
   • Include every model the docs list, current and legacy. Don't drop a model just because it's deprecated — old sessions may still reference it.

4. Diff against the current file, ignoring updated. Compare only the models object (and _comment). If the model data is unchanged, make no changes and do not commit — report "model table already current" and stop. Do not bump updated when nothing else changed (it would churn a no-op commit).

5. If the data changed, set updated to today's date (YYYY-MM-DD), write the file, then sanity-check it:

   • python3 -c "import json; json.load(open('sessions/models.json'))" parses.
   • Spot-check that claude-opus-4-8, claude-sonnet-4-6, and claude-haiku-4-5 are present with plausible windows (1M, 1M, 200k respectively as of this writing).

6. Commit just this file on the current branch: git add sessions/models.json && git commit. Use a message like chore: refresh model context-window table and summarize what changed (models added/removed, windows that moved). Don't push unless the user asks.

7. Report the net change: which models were added, removed, or had their window updated — or that the table was already current.

Notes

• The only external dependency is the docs page's table format. If WebFetch comes back without a recognizable model table (page restructured, fetch failed), stop and report it — never write a partial or empty table over a good one.
• This is a read-of-public-docs operation. It must not require, look for, or use an ANTHROPIC_API_KEY.