pnpm pack produces a 12 KB tarball with only the two halves and the bundle
patch; it was installed into a throwaway profile to prove it works and
that both halves parse. The notes make the composition step explicit,
because a package can sit in a profile without ever loading -- which is
the state dsh-bottom-info-bar is in on this machine.
Two real defects, both found by testing rather than reading:
1. The client sent sessionId in the POST body while the route read it from
the query string, so the host always saw an empty id and answered
total:null -- the pill was correct to stay silent. The client now sends
the id in both places and the route reads both.
2. The ledger stored only a session total, so a restart kept the sum but
dropped the per-purpose split and the call count (9.99 total against
6.24 in parts, seen on the live route). Sessions are now stored as
objects; a bare number from an older ledger is still accepted.
Also made the client report why it cannot read a total: swallowing the
error made 'nothing spent yet' indistinguishable from 'request failed'.
The host half is confirmed working end to end: the ledger holds 24 calls
and the authenticated route returns a real total (6.42 RUB, 14 calls, all
attributed, zero unpriced or unattributed). The plugin is present in the
browser boot manifest, so its bundle is served.
Remaining issue is browser delivery only: the pill stays silent, and the
client swallows the transport error. Recon scripts added to pin down
whether the route answers 200, 401 or 404 from the page origin.
Host half wraps globalThis.fetch to read usage.cost out of provider
responses (tee, never clone) and listens on the llm/stream seam for the
session; the two halves join on the generation id. Client half adds one
pill to the composer dock under its own id, so the harness's native stats
stay where they are.
Verified against the live API: three calls in one session (agent reply,
compaction, session title) attributed with per-purpose totals, an unknown
session answered with null, the ledger written and reloaded across
instances, and the pill rendering only when a cost exists.
The plugin manager accepts an absolute local path, so the package can be
installed from this workspace. Two requirements make or break it: a DSH
peer range (runtime is 0.2.0-rc.2) and a declared bundle patch. Notes that
HMR is absent from the desktop profile, so a restart is required and the
restart is therefore the last step.
Spike proves response.body.tee() splits a real routerai SSE stream: the
consumer sees chunks progressively (first at 1ms of 1643ms) and our branch
still recovers the generation id and usage.cost. It also found a real
defect -- a rebuilt Response loses url, which the OpenAI SDK reads -- and
that Object.defineProperty is the only way to restore it.
Records what was verified against the live routerai API and the harness
code: the provider reports usage.cost on every response, the harness
discards it in mapUsage, and the cost can still be recovered by wrapping
globalThis.fetch while session attribution rides the llm/stream seam.
antonov