ContextQMD
Back to Ruflo

ADR-314: Power saver mode (auto-route lesser requests) + sponsored-capacity abuse prevention

v3/docs/adr/ADR-314-power-saver-mode-and-sponsored-abuse-prevention.md

3.26.112.3 KB
Original Source

ADR-314: Power saver mode (auto-route lesser requests) + sponsored-capacity abuse prevention

Context

Two requests, and they turn out to share a root cause:

  1. "Include a power saver mode: if your Claude Code Max/Pro account is below 20%, auto-route lesser requests that don't need the extra intelligence to the meta-llm."
  2. "Include anti-cheating and abuse capabilities" for the sponsored-capacity system ADR-313 built.

The connection: widening the trigger for routing through Cognitum (from "only when actually rate-limited" to "proactively, whenever quota looks low") widens the abuse surface at the same time. This ADR designs both together rather than bolting anti-abuse on afterward.

A constraint carries over directly from ADR-312 and has to be stated up front: ruflo cannot read "your account is at 20%." ADR-312 already established, by direct inspection of the installed Claude Code CLI, that rate-limit/quota state is populated from response headers (anthropic-ratelimit-unified-status, -utilization, etc.) that are not exposed to hooks, the statusline JSON payload, or anywhere else a third-party CLI extension can read. That finding applies unchanged here — "20%" can't be measured any more precisely than "rate-limited or not" can be. Building this as if the number were readable would ship something that silently never activates, or activates on a fabricated value. The design below is honest about that gap.

Decision

A. Power saver mode is self-reported, not measured (Phase 0, same pattern as ADR-312)

ruflo settings notices quota-low [--clear] sets a manual flag — structurally identical to rate-limited in rate-limit-notifier.ts, same 6h TTL, same file-based mechanism (~/.ruflo/quota-status.json). This is not automatic detection; it's the same honest workaround ADR-312 already shipped for the binary case, generalized to a second, lower-severity signal: "not blocked yet, but I want to start conserving." A future Phase 1/2 (per ADR-312) that gets real quota data from Anthropic would replace the self-report with a real threshold check — no design here depends on staying manual forever.

Separate flag from rate-limited deliberately: "I'm running low" (proactive, user still has capacity) and "I'm blocked" (reactive, zero capacity left) are different urgencies with different correct responses — conflating them would mean either power-saver mode never fires until you're already blocked (too late to have saved anything) or a false-rate-limited claim silently piggybacks onto power-saver's less-scrutinized activation path.

B. Power saver mode routes through the Cloud plane with the model rewritten to cognitum-auto

This is the part that makes "auto-route lesser requests" actually work without ruflo building its own prompt-difficulty classifier — Cognitum's gateway already has one, and it exists for exactly this purpose (ADR-201's cost-Pareto thesis, docs/ARCHITECTURE.md in the meta-llm repo: "cheap traffic stays cheap and only the hard tail escalates").

The mechanism that makes this necessary, not just convenient: Cognitum's /v1/messages translation adapter maps a named Anthropic model to a tier by name — opus→high, sonnet→mid, haiku→low (confirmed reading src/routes/messages.ts in the meta-llm repo). The existing Cloud/Sponsored planes in the proxy forward whatever model Claude Code requested, unchanged. Since a Claude Code session runs on one model for its whole lifetime (typically opus or sonnet, chosen once, not varied per-message by task complexity), forwarding the name verbatim would map to the same tier — usually high — on every single request. That is not cost savings; it just moves the same spend to a different biller. Power saver mode instead rewrites the outgoing model field to the literal string cognitum-auto before forwarding, so Cognitum's real per-prompt difficulty scorer (not a name lookup) decides low/mid/high per request. Everyday messages get cheap; a genuinely hard reasoning turn still escalates to a comparable frontier model — that's the "lesser requests... don't need the extra intelligence" distinction, made by the router that already exists for it, not reimplemented client-side.

Power saver mode uses the Cloud plane (the user's own cognitum_api_key, their own Cognitum account/billing) — not Sponsored. This is the user choosing to spend their own Cognitum credit to conserve their own Anthropic quota; it is not free capacity, and shouldn't share Sponsored's "billed to Cognitum, not you" framing or its abuse surface. A user can independently also be rate-limited and sponsored-consented; if both apply, Sponsored wins (ADR-313's existing precedence is unchanged — being actually blocked outranks proactively conserving).

C. Sponsored capacity is now capped at low/mid tier, never high — a direct abuse-surface fix

Independent of power saver mode, this closes a real gap in the shipped ADR-313 sponsored plane: today it forwards the client's requested model name unchanged, which (per the mapping above) means a session running on opus forwards as high tier every time — Cognitum's most expensive tier, on Cognitum's own dime, for every sponsored request, with no client-side ceiling. Sponsored requests now always rewrite model to a fixed low/mid-capable alias, never a tier that can resolve to a frontier/high-cost model, regardless of what Claude Code originally requested. This bounds Cognitum's worst-case per-request exposure on the free plane independent of anything else in this ADR, and independent of whether the difficulty scorer itself could ever be gamed into over-escalating (see the vector list below) — the cap holds even if the scorer is fooled, because the tier ceiling is enforced before the scorer's decision is even relevant.

D. Anti-abuse: the vectors, ranked, and what's buildable now vs. blocked on Cognitum-side work

Ranked by severity — most severe first:

  1. Fabricated rate-limited/quota-low claims for free Sponsored capacity — the most severe, and currently the ONLY thing standing between "opt-in for genuine downtime" and "unlimited free inference on request." Phase 0 detection (ADR-312) is entirely self-reported with zero server-side verification; nothing today distinguishes an honest user who actually hit their limit from someone who runs ruflo proxy sponsor-enable --yes once and never looks back.
    • Buildable now (this ADR ships it): a cooldown on the rate-limited/quota-low toggle commands — no more than one flip per 10 minutes each. Cheap, purely client-side friction; it doesn't stop a determined abuser but it stops casual/accidental always-on gaming and raises the bar on scripted abuse.
    • Blocked on Cognitum-side work: per-identity sponsored quotas (see #2) are the real fix — a cooldown only slows down abuse from a single identity; it can't cap total abuse without a stable per-user identity to cap it against.
  2. The shared/unattributable API key. ADR-313's addendum already flagged this as open: there is no automated way to provision a real, per-user, scoped cognitum_api_key — every sponsored user today would share one key (in this session's testing, COGNITUM_TEST_API_KEY). A shared key means zero per-user accountability, no way to revoke one abuser without cutting off everyone, and unlimited blast radius if it ever leaks. This is the prerequisite for meaningfully addressing #1, #4, and #5 below — it is Cognitum-backend work (a key-minting endpoint sponsor-enable calls on first activation), not something this proxy or ruflo's CLI can build unilaterally. Flagging it here as the single highest-priority follow-up, not re-flagging it as "still fine to defer indefinitely."
  3. Power-saver's own difficulty-scorer gaming. A user could pad trivial prompts to trick Cognitum's scorer into escalating to a costlier tier than the content warrants. On the Cloud plane this is the user spending their own credit inefficiently — not a security abuse vector, just a self-defeated feature, out of scope to defend against. The reason this is listed at all: the same gaming, if it ever applied to Sponsored traffic, WOULD be a real abuse vector (tricking free capacity into the expensive tier) — which is exactly why mitigation C above (hard tier ceiling, enforced before the scorer runs) matters independent of whether the scorer itself can be foxed.
  4. Proxy config/token copied across machines to multiply effective throughput. Buildable server-side once #2 lands: per-key concurrent-request and requests/minute caps (not just a daily $ cap) bound this regardless of how many machines share one key.
  5. Scripted volume abuse — pure spam against the shared sponsored pool, no benefit to the attacker beyond degrading the pool for everyone. Same fix as #4: per-key rate limiting, Cognitum-side, blocked on #2.
  6. Sponsored anonymity as a lower-accountability vector for policy-violating content. Recommend Cognitum's existing opt-in safety:scan guardrail layer (documented in the meta-llm repo) become mandatory, not opt-in, specifically on Sponsored-plane traffic — pseudonymous free capacity is exactly the surface this class of abuse would target. Cognitum-backend policy change, not a proxy change.

E. What ships in this ADR's implementation vs. what's tracked as follow-up

Ships now (ruflo + meta-proxy, no Cognitum backend changes required):

  • ruflo settings notices quota-low [--clear] — the power-saver self-report flag
  • power-saver consent domain (mirrors sponsored-downtime's pattern exactly)
  • Proxy: Cloud-plane model rewrite to cognitum-auto when power-saver is active and Sponsored isn't (precedence: Sponsored > Power-saver > default Passthrough)
  • Proxy: Sponsored-plane model rewrite to a fixed low/mid-only alias, always, independent of power-saver — this is a standalone hardening fix to the already-shipped ADR-313 plane
  • Cooldown (10 min) on rate-limited/quota-low flag toggles

Tracked as follow-up, explicitly not solvable from ruflo's side alone:

  • Per-user Cognitum key minting (the real fix for abuse vectors #1, #2, #4, #5)
  • Per-key concurrent-request / requests-per-minute caps (Cognitum backend)
  • Mandatory safety-scan on Sponsored traffic (Cognitum backend policy)

Consequences

  • Power saver mode's actual value depends entirely on Cognitum's difficulty scorer being good — this ADR does not re-litigate or re-verify that scorer's accuracy, it just routes to it.
  • The Sponsored tier cap (C) is a behavior change to the already-shipped ADR-313 plane: sponsored users lose access to high-tier models entirely, unconditionally. This is intentional — sponsored capacity was never framed as "free frontier access," and ADR-313's own text already describes it as "best-effort... Cognitum may throttle or decline requests under load."
  • Every mitigation ranked "blocked on Cognitov-side work" above means the abuse-prevention story for Sponsored mode remains genuinely incomplete until that work lands — this ADR does not claim Sponsored mode is abuse-proof, only that it is more bounded than before (D3) and that the concrete missing piece is named (D2), not hand-waved.
  • No change to the Passthrough plane (ADR-313 addendum) — power saver mode and its abuse surface are additive, not a replacement for "normal usage still means your own subscription by default."

References