pi-cost-switch

Pi extension that estimates next-turn cost before you change model or thinking level.

Built-in /model and thinking controls only notify after the change. This extension owns a pre-change switch UX with:

• hit — warm continuation for the current model at the observed cache hit rate (~85% (assumed) when unknown); n/a for other models (prompt cache does not transfer)
• cold — base uncached total through cache-write-premium upper bound
• tax — extra cost from re-billing only the previous cacheable prefix

Install

From a local clone (development / sharing a path)

pi install /absolute/path/to/pi-cost-switch
# or from the package directory:
pi install .

From git (once published)

pi install git:github.com/BlockedPath/pi-cost-switch
# optional pin:
pi install git:github.com/BlockedPath/[email protected]

From npm (once published)

pi install npm:pi-cost-switch

Temporary try without installing

pi -e /absolute/path/to/pi-cost-switch

Reload after install:

/reload

Commands

| Command | What it does | |---|---| | /cost-switch [filter] | Pick model + thinking with $ estimates, then apply | | /cost-estimate [filter] | Comparison table only (no switch) | | /cost-switch status | Toggle status-bar next≈… · miss … (in-memory only; resets on /reload) | | /cost-switch help | Quick help |

Status bar

Shows something like:

next≈$0.162 · miss $1.35 · hit ~85% (assumed)

• next≈ — estimated warm (cache-hit) next-turn total
• miss — estimated base-cold total if the cache is lost
• hit — rate used for the warm estimate (72% observed, or ~85% (assumed) when unknown)

Toggle with /cost-switch status. The toggle is in-memory only — it resets to on after /reload.

How estimates work

| Field | Meaning | |---|---| | hit | Current model only: continue with observed session cache hit rate (~85% (assumed) when unknown). Other models show n/a — cache does not transfer across models/providers | | cold | Range: normal uncached total → cache-write premium upper bound | | tax | Extra vs cache-read for the previous cacheable prefix only |

Model switches and reasoning changes are treated as cache-miss risks. List estimates clamp thinking to each candidate’s supported levels. Output/reasoning is a heuristic from session averages × thinking effort multipliers — not provider-exact.

Subscription / zero-priced models show sub instead of $0.00.

Dev layout

pi-cost-switch/
├── package.json          # pi-package manifest + test script
├── README.md
├── LICENSE
├── test/
│   ├── estimate.test.ts  # pure cost helper unit tests (node:test)
│   └── hit-rate.test.ts  # observed vs assumed hit-rate policy
└── extensions/
    └── cost-switch/
        ├── index.ts      # extension entrypoint
        ├── estimate.ts   # pure cost math
        ├── format.ts     # pure $ / token formatters
        ├── hit-rate.ts   # observed vs assumed hit-rate policy
        └── rank.ts       # pure model filter/rank

Tests

npm test
# or:
node --test --experimental-strip-types test/**/*.test.ts

Local auto-discovery (optional): symlink into the global extensions tree:

ln -sfn "$(pwd)/extensions/cost-switch" ~/.pi/agent/extensions/cost-switch

Or use pi install . so settings point at this package path.

Security

Like all pi extensions, this runs with full system permissions. Review source before installing third-party copies.

License

MIT