context-weight

name	context-weight
description	Reach for this when your context is huge, cache reads are high, you're asking should I compact, or you want to know what filled the window — it explains what this skill can and cannot do, ranks prevent/prune/measure in priority order, and gives the checkpoint move before you compact.

Context weight

What this skill cannot do

This skill cannot remove anything from your context window. It is text loaded INTO the window — read-only instructions the model follows — with no mechanism to mutate the message array a call submits. Only the harness can do that (compaction, or clearing). This skill's job is to tell you when to act and what to act on; it never acts for you.

Run the engine

python3 {{POLYTROPOS_ROOT}}/bin/context_weight.py session --harness copilot    # this session's per-call weight
python3 {{POLYTROPOS_ROOT}}/bin/context_weight.py overview --harness copilot   # cross-session working-set table
python3 {{POLYTROPOS_ROOT}}/bin/context_weight.py audit                        # resident config surfaces (incl. copilot-instructions.md) vs a token budget
python3 {{POLYTROPOS_ROOT}}/bin/context_weight.py demo                         # synthetic cross-harness smoke, no real data touched

All four take --json. This is the real argparse surface — confirm against python3 bin/context_weight.py --help before assuming any other flag exists; do not invent one.

Copilot's honest fidelity

Copilot's session logs record no per-turn input/cache split, so there is no growth curve on this harness — session --harness copilot reports a session-average weight ((input + cache_read + cache_write) / assistant turns) instead, and that is the honest substitute, not a curve in disguise. watch is Claude-only: it has no --harness flag, and passing copilot to it prints an honest refusal line rather than a fabricated live number. That means there is no live threshold to watch here — apply the three levers below on a schedule (e.g. once per session, or before a long task), not on a threshold crossing.

The three levers, in priority order

1. PREVENT — free and lossless. Delegate bulk reads to subagents that return conclusions, cap tool output before it enters the window, defer loading a file until it's needed. Nothing here ever costs anything to reverse, because the mass never enters the window.
2. PRUNE — cheap but lossy. Compaction. The only lever that can cost accuracy — a compacted summary is not the original. Before compacting, write decisions, constraints, and open questions to a notes file — that anchor survives even after the transcript behind it is folded away.
3. MEASURE — free. session, overview, and audit tell you which of the first two is the move right now; measurement changes nothing by itself.

Honesty rules

Every estimated figure is labeled est. — a rank or magnitude, never an exact token count and never priced. That holds for the attribution and audit figures. The session/overview cards are different: they also print a context carry cost line in USD and AIC, carrying the mandatory label API-equivalent dollars — an estimate, not a bill. Always relay that figure WITH its label — never strip it, and never add it to real spend figures (that's what the usage skill reports). Each harness is reported at its own honest fidelity: Copilot gets session-average, never a fabricated Claude-style curve.

Same-named agent

For persona-isolated runs — a separate dispatch that should carry its own model pin instead of this session's model — use the context-weight custom agent: pick it in the /agent picker, or run copilot --agent context-weight -p "<task>". This skill and that agent are the same capability on two surfaces; the agent's frontmatter carries the model pin, this skill runs on whatever model the session already uses.

Installed?

If the literal {{POLYTROPOS_ROOT}} text is still visible above, the bundle is not installed — tell the user to run python3 bin/harness_select.py install --harness copilot (then /skills reload picks the skills up in-session).