cost-aware-pipeline

Cost-Aware Pipeline

A harness that cannot finish under its own cost gates — or that burns a build-sized budget answering a question — is anti-enhancement. The economics rest on four levers: tokens as the primary ledger, the scale gate, model routing by job, and killing branches before the expensive passes see them.

Tokens first, USD second

PQA runs on a Claude Code subscription: USD is synthetic; tokens, rate limits, and context are real. Budget(max_usd, max_tokens, warn_at=0.8) trips on EITHER axis; run_budget_tokens (default 800k) is the number that matters and run_budget_usd (default 5.0) is the display/API-mode cap. Model identity flows through ONE translation point — pqa.cost.resolve_model (aliases fable|opus|sonnet|haiku) — so projection, dispatch, and recording can never price different models for the same call.

The scale gate (before anything is spent)

Ask	Mode	Spend shape
compare / choose / review / explain	decide	frames + ONE judge pass over idea digests; < 30k tokens; output flagged judgment — not verifier-backed
single-file fix, rename, config tweak	patch	n_branches=2, no scout, abbreviated frame
feature / refactor with real unknowns	build	full loop, N from config

Never silently upgrade a decide-ask into a build run — a question answered with a full superposition is the harness's worst failure mode: enormous spend for an answer one judgment pass could give. When unsure, ask the operator which mode they want.

Model routing (the biggest lever)

Tier	Work	Agents
fable	coding + quality-critical judgment	generator (every branch), unknown-scout, adversary, collapse-judge, baseline-runner (fair control)
sonnet	mechanical execution + orchestration	orchestrator, frame-loader, verifier, reconciler, self-reflector, instinct-synthesizer
haiku	bookkeeping	memory-curator, failure-taxonomist, eval-runner

Pass model explicitly on every Task dispatch. Never burn fable tokens on arithmetic, table rendering, or registry writes. The table is enforced by scripts/validate_components.py — adding an agent means making a routing decision.

Two gates around every dispatch

1. Pre-flight: governor.would_abort(model, projected_in, projected_out) — if True, do NOT dispatch: write the aborted RunReport (partial state, memories_injected included) and stop. Projections are pessimistic by design (max(len(prompt)//3, floor) input, fixed output floor): refuse a dispatch that might fit rather than admit one that won't.
2. Post-call: governor.record(branch_id, model, actual_in, actual_out) from the subagent's reported usage, then should_abort() as belt-and-braces. should_abort() alone fires after the money is spent — that asymmetry is why both gates exist.

Spend on survivors, not corpses

Order verification by cost: lint/type-check (seconds, zero tokens) kills compile-broken branches BEFORE the adversary, so the most expensive judgment pass only attacks branches that could win. The same shape applies everywhere: cheap gates first, fable last.

Worked example

N=3 build run, 800k-token budget: frame (sonnet, ~25k) → 3 generators (fable, ~3×90k) → static early-kill (0 tokens — b1 dies on pyright; taxonomy row verifier: failed static checks before collision — pyright: 4 errors) → 2 adversaries (fable, ~2×60k) → 2 verifiers (sonnet, ~2×30k) → judge (fable, ~20k) → bookkeeping (haiku, ~10k) ≈ 410k total: completes with headroom.

The same run WITHOUT early-kill pays ~60k fable tokens attacking a branch that could never merge. The same ask phrased as "which approach is better?" should have cost < 30k in decide mode — two orders of magnitude apart, governed by the gate, not luck.

A run that does abort mid-loop records every live branch as budget: aborted at collide, 612k/800k tokens spent — and retrieval never reads those rows as refuted (see failure-taxonomy).

Anti-patterns

• USD-primary thinking. Optimizing cents while blowing the context window or rate limit optimizes the wrong scarcity.
• Post-hoc-only gating. should_abort() without would_abort() bills you for the dispatch that broke the budget.
• Routing everything to fable. Bookkeeping on the top model is pure waste — and the validator rejects the frontmatter anyway.
• Adversary before static checks. Paying judgment prices to discover a syntax error.
• Optimistic projections. Projections exist to refuse; an optimistic floor turns the pre-flight gate into decoration.
• /cost or /budget dispatching a model. They are Bash wrappers over the engine's ledger — a model call to read a number is the disease in miniature.