fix: compute costUsd from model rates when the provider reports none (#67)

[stealthwhizz]

Problem

costs() reports costUsd: 0 even when tokens were consumed and billed. sdk.ts mapped cost straight from the provider:

costUsd: msg.usage.cost?.total ?? 0,

But msg.usage.cost.total is only populated when pi-ai's calculateCost runs, which:

• doesn't run on the openai-completions path (Bug 3 in bug: costs() always returns 0 — token usage tracked but costUsd never calculated correctly #67), and
• can't produce a real number for custom/unregistered models that carry no price table (Bug 2).

So the 0 was read verbatim, and callers couldn't tell "this was free" from "we don't know the cost."

Root-cause split

Bug	Where	Fixable in gitagent?
calculateCost emits 0 for unknown model IDs	pi-ai	❌ upstream
openai-completions never calls calculateCost	pi-ai	❌ upstream
EMPTY_USAGE zeros usage on the error path	pi-agent-core	❌ upstream
sdk.ts reports that 0 with no fallback / no signal	gitagent	✅ this PR

Fix (gitagent side)

• computeCostUsd(usage, rates) in cost-tracker.ts — the same per-million-token formula pi-ai uses ((rate / 1e6) * tokens). Returns null when tokens were used but the model has no pricing, so "unknown" is distinguishable from 0.
• sdk.ts — prefer the provider's cost.total when it's present and non-zero; otherwise recompute from loaded.model.cost (known models carry real rates, e.g. gpt-4o = $2.5/$10 per 1M). Pass a costResolved flag to the tracker.
• SessionCosts.costDataAvailable — flips to false when any request consumed tokens that couldn't be priced, so costs() consumers can surface "cost unavailable" instead of a misleading $0.
• Exported computeCostUsd + the new types.

Result

• Known models that hit the openai-completions path (or otherwise arrive with no provider cost) now get a correctly computed costUsd from their own rate table.
• Custom/zero-rate models report costDataAvailable: false instead of silently claiming $0.
• Token counts are untouched; this only changes how cost is derived and reported.

Tests

test/cost-compute.test.ts — pricing math (input/output/cache), null for unpriceable-with-tokens, 0 for no-tokens-no-rates, partial-rate models; plus costDataAvailable default/flip/reset. Full suite green.

Upstream follow-ups (out of scope here)

Worth filing against pi-ai / pi-agent-core: have calculateCost return null (not 0) for missing pricing, call it on the openai-completions path, and document/​accumulate EMPTY_USAGE on the error path.

Closes #67

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[shreyas-lyzr]

Clean fix for the misleading $0 cost reporting. The three-way outcome — provider cost wins when non-zero, fallback computation from model rates, and null/costDataAvailable=false for genuinely unresolvable cases — is the right design. The logic in computeCostUsd correctly mirrors pi-ai's formula, and the test coverage hits all the meaningful branches including the partial-rate and zero-token edges.

One suggestion below: null-guard loaded.model.cost in sdk.ts. It's safe today because both createCustomModel and getModel always produce a cost object, but a future model source that omits it would throw at runtime rather than fail gracefully.

[shreyas-lyzr]

loaded.model.cost is always set today (createCustomModel hardcodes it, getModel from pi-ai includes it), so this won't throw in practice. Worth adding a null-fallback anyway so a future model source that omits the field degrades gracefully rather than crashing:

Suggested change

	const computed = computeCostUsd(tokens, loaded.model.cost);
	const computed = computeCostUsd(tokens, loaded.model.cost ?? { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 });