fix(vscode): show upstream inference cost for OpenRouter BYOK sessions

[iscekic]

kwf trial run. This PR came from a workflow trial request, not from a tracked work item. Review it as you would any other PR; the label kwf-trial marks where it came from.

Request

Surface: the VS Code extension (Kilo-Org/kilocode issue 13829).

OpenRouter BYOK sessions always show $0, because providerCost reads usage.cost and ignores usage.cost_details.upstream_inference_cost.

OpenRouter returns, for every response it routes through a BYOK provider key:

"usage": { "cost": 0, "is_byok": true,
           "cost_details": { "upstream_inference_cost": 0.00000445 } }

cost is what OpenRouter charged the account, which is 0 by definition for BYOK routing. The real spend is cost_details.upstream_inference_cost. Reported on 7.5.14 and 7.5.15 with openrouter/z-ai/glm-5.3-flash.

Expected: a BYOK session shows the upstream inference cost, and a non-BYOK session keeps showing usage.cost exactly as it does today.

Rule: reproduce first. If it does not reproduce on main, answer no_change with the evidence.

Changelog for users

• OpenRouter BYOK sessions in the VS Code extension no longer show $0: the session cost now reports usage.cost_details.upstream_inference_cost — the real spend on the user's own provider key — whenever OpenRouter bills the account $0 or only its routing fee for BYOK routing (OpenRouter BYOK sessions always show $0 — providerCost ignores cost_details.upstream_inference_cost #13829).
• Non-BYOK OpenRouter sessions are unchanged: the displayed cost is still usage.cost, exactly as today.

Changelog for maintainers

• The OpenRouter chat-completions branch of the session providerCost helper now prefers costDetails.upstreamInferenceCost over the billed cost whenever the upstream figure is strictly greater; previously upstream was preferred only for the Kilo provider. That one expression is the whole fix — start the review there.
• Non-BYOK safety: a non-BYOK response always bills the account at least the upstream cost, so upstream > regular is false and usage.cost is returned untouched; unit tests pin both the BYOK and non-BYOK payload shapes.
• Edge case to review: when a response carries costDetails but no top-level cost, the new expression (regular ?? -Infinity) now returns the upstream figure, where the old code fell through to the token-based estimate. Only reachable if OpenRouter ever omits cost while still reporting cost details; not covered by a test.
• New unit suite for the OpenRouter chat-completions path covers five cases: BYOK with cost: 0, BYOK where cost is only the routing fee, non-BYOK unchanged, no costDetails unchanged, and the Kilo-provider preference.
• The transport doc comment in the session cost path now documents the BYOK semantics (account billed $0 or a routing fee; true spend on the user's own key).
• Verification: the reproduce-first rule was proved live; the run's decisive log excerpts are appended below under their own heading.

E2E proof — log excerpts

[e1] Rule: reproduce first. If it does not reproduce on main, answer no_change w -> pass :: Baseline (e1-baseline.log) fails with Expected: 0.00000445 Received: 0 and Expected: 0.06418625 Received: 0.0032093125; packed (e1-packed.log) reports 10 pass a

/Users/igor/.local/share/kwf/sections/surface-the-vs-code-extensio-5184/e2e-cli/e1-baseline.log

155 |     expect(result.cost).toBe(0.00000445)
                              ^
error: expect(received).toBe(expected)
Expected: 0.00000445
Received: 0
      at <anonymous> (/Users/igor/.local/share/kwf/wt/surface-the-vs-code-extensio-5184/packages/opencode/test/kilocode/provider-cost.test.ts:155:25)
(fail) KiloSession.providerCost — OpenRouter chat completions > uses upstream_inference_cost when OpenRouter reports $0 for BYOK routing [1.51ms]
169 |           },
170 |         },
171 |       },
172 |     })
173 |
174 |     expect(result.cost).toBe(0.06418625)
                              ^
error: expect(received).toBe(expected)
Expected: 0.06418625
Received: 0.0032093125
      at <anonymous> (/Users/igor/.local/share/kwf/wt/surface-the-vs-code-extensio-5184/packages/opencode/test/kilocode/provider-cost.test.ts:174:25)
(fail) KiloSession.providerCost — OpenRouter chat completions > uses upstream_inference_cost when OpenRouter bills only its BYOK routing fee [3.19ms]
 3 pass
 5 filtered out
 2 fail
 5 expect() calls
Ran 5 tests across 1 file. [2.15s]

/Users/igor/.local/share/kwf/sections/surface-the-vs-code-extensio-5184/e2e-cli/e1-packed.log

bun test v1.4.0 (34cbb9a40)
 10 pass
 0 fail
 10 expect() calls
Ran 10 tests across 1 file. [2.10s]

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• packages/opencode/src/kilocode/session/index.ts
• packages/opencode/test/kilocode/provider-cost.test.ts

Previous Review Summaries (4 snapshots, latest commit 6e32cd0)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit 6e32cd0)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• packages/opencode/src/kilocode/session/index.ts
• packages/opencode/test/kilocode/provider-cost.test.ts

Previous review (commit b57e37c)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• packages/opencode/src/kilocode/session/index.ts
• packages/opencode/test/kilocode/provider-cost.test.ts

Previous review (commit cb2326d)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• packages/opencode/src/kilocode/session/index.ts
• packages/opencode/test/kilocode/provider-cost.test.ts

Previous review (commit 04ae572)

Status: No Issues Found | Recommendation: Merge

Files Reviewed (2 files)

• packages/opencode/src/kilocode/session/index.ts
• packages/opencode/test/kilocode/provider-cost.test.ts

---

_{Reviewed by grok-4.6 · Input: 50.5K · Output: 9.6K · Cached: 468.5K}

_{Review guidance: REVIEW.md from base branch main}

[chrarnoldus]

I think technically it would have to be upstream+regular for OpenRouter BYOK