Three additions to tighten the eval loop:
- offload-eval-styles.sh: new 4-arm eval (raw/refs/map/src) isolating the Worker's
output shape's effect on main-session tokens, latency, and accuracy. Delegation
blocked by default (DISALLOW=Agent) so variance from Haiku subagent spawning doesn't
contaminate the measurement.
- offload-eval-cost.mjs: cost/token analyzer that reads Claude Code's own per-model
accounting (modelUsage.costUSD) rather than re-deriving from raw token counts,
giving a correct main(Sonnet)/sub(Haiku) split with proper per-tier pricing.
- offload-eval-3arm.sh: adds DISALLOW env to block sub-agent delegation across all
arms, and REP_START to append reps to an existing run without clobbering earlier
jsonls (e.g. REP_START=4 REPS=3 → reps 4,5,6).
Also adds CODEGRAPH_OFFLOAD_STYLE forwarding to the managed gateway so the styles
eval can drive output shape end-to-end; the field is stripped before the upstream
model call and never sent to BYO endpoints.
`CODEGRAPH_OFFLOAD_DISABLE=1` immediately disables the offload for the current
process without touching the persisted config or stored login — useful for A/B
arms or sessions where raw source is preferred.
`CODEGRAPH_OFFLOAD_USAGE_LOG=` appends one JSONL entry per call with token
counts, charged credits, and derived cost (`creditsCharged / 100_000`) so a
harness can attribute CodeGraph AI spend to a single run independently of the
server's cumulative totals. Both features are best-effort and never disrupt the
degradable offload path.
Also fixes the `login` credit display to check `unlimited` before the numeric
balance, so comped/internal accounts don't incorrectly show "0 remaining".
Adds a `usage` subcommand that pings `/v1/usage` with the stored token and
displays balance, plan, 30-day explore/token counts, and allowance reset date.
Degrades quietly in all non-happy-path states — signed out, BYO endpoint, or
unreachable server — so managed reasoning remaining optional doesn't change.
Also extends `OffloadUsage` with the fields the endpoint already returns
(`unlimited`, `banned`, `tokensLast30`, `callsLast30`, `creditsLast30`) that
were previously untyped.
The old `offload` command family required users to paste a token manually (`offload login --token `) and exposed bring-your-own-endpoint plumbing (`set-endpoint`, `status`, `disable`) as top-level CLI surface. This replaces it with a standard OAuth device flow (RFC 8628 shape) against the CodeGraph dashboard.
`codegraph login` calls `/api/cli/device/start`, opens the browser to the returned URL, polls `/api/cli/device/token` until the user approves, then stores the minted token and enables managed reasoning. `codegraph logout` clears it. BYO-endpoint configuration moves entirely to env vars (`CODEGRAPH_OFFLOAD_URL` / `CODEGRAPH_OFFLOAD_KEY` / `CODEGRAPH_OFFLOAD_MODEL`), keeping the CLI surface minimal.
Adds the managed offload mode: point codegraph_explore at the CodeGraph AI metered
gateway (https://ai.getcodegraph.com) with an org token instead of a BYO provider key.
Same synthesis client, pointed at codegraph-ai-proxy (a metered OpenAI-compatible gateway).
- credentials.ts — org token in ~/.codegraph/credentials.json (0600); unlike a BYO
provider key it's a revocable org-scoped auth token (gh/npm-login style), kept out
of config.json
- config.ts — managed branch in resolveOffload: default gateway URL + public model id
(openai/gpt-oss-120b) + login token as bearer; managed requires a token to be enabled
- reasoner.ts — fetchUsage() reads the credit balance from /v1/usage
- bin/codegraph.ts — `codegraph offload login --token <t>` / `logout`; status shows the
managed tier + live balance
Proven GREEN end-to-end against a local wrangler-dev of the proxy: org token validated,
credits prechecked, real Cerebras synthesis returned, and credits metered + charged
(250,000 → 248,473). Graceful degrade on upstream failure; balance via /v1/usage.
Phase 3 (codegraph login device flow) replaces the manual --token.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
codegraph_explore can now hand the source it retrieved to a reasoning model you
point at — any OpenAI-compatible endpoint (Cerebras, OpenAI, a local vLLM/Ollama)
with your own key — and return that model's tight, cited answer instead of the
raw source dump. The agent's main context gets the answer in far fewer tokens, at
the cost of one network round-trip.
Off by default. Configure with `codegraph offload set-endpoint <url> --model <m>
--key-env <ENV>` (or the CODEGRAPH_OFFLOAD_* env vars); status/disable manage it.
The API key is never written to disk — the config stores the NAME of an env var
and the key is read from it at call time. Strictly degradable: any failure
(no endpoint, network, timeout, empty answer) returns null and the call falls
back to the local source, so the offload can never surface an error to the agent.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Colby McHenry
