Token Economy (FR-26)

Guardrail first (the rule's first clause)

Quality outranks tokens. These rules eliminate REDUNDANT consumption only — never skip a needed first-read, never thin verification, never truncate reasoning. On any conflict, the correctness/safety floor wins (same floor language as FR-21: ceremony drops, safety never). Citing FR-26 to avoid a read or a verification step inverts the rule.

Rules (each with its quality guard)

Context compression discipline

Extends FR-26 to large context and large output — when the input or a tool result is big enough that loading it whole is itself the waste. This is read-time and reasoning-time routing, not a budget: it slices and points to large artifacts and keeps compressed context honest. It never authorizes skipping a needed read or thinning verification — the Guardrail still governs.

Generated/vendored restraint and reference-once live in the Rules table as TE-03 / TE-07 — one canonical row each.

Residency and cache lifetime

Every request re-reads the agent's whole context, so cost grows with context × requests; a context idle past its cache lifetime is re-written in full on the next request. Like § Context compression discipline this is routing, not a budget — the Guardrail governs.

Fusebase-CLI grounding — known large-output surfaces (TE-17 instances)

Pre-identified so the FIRST invocation is scoped — not the rerun. /token-waste-audit's large-output class flags results ≥20k chars from these surfaces; cite the TE ID in fixes.

Guard: scoping serves the question — never exclude the evidence the diagnosis needs (TE-17's guard governs). Surface names/flags track the vendored CLI provider skills; on a FuseBase CLI refresh, re-check this table against remote-logs / dev-debug-logs / fusebase-dashboards / fusebase-gate skills.

Measure it

• Auto-classification is labeled, never silent (--probe-command CMD adds this ticket's documented gate probes, repeatable). Two conjunctive rules only: growing-source-tail (same full read key and monotonic growth or all-differing digests and no contradictory event — an intervening write to that path, a compaction between the reads, or an error-shaped result keeps it live; a size-difference-only re-read stays live) and possible-FR-10-triple (exactly-3 runs are labeled; dismissed only when the command is probe-shaped — a test runner, --dry-run, or a health/status/version verb — because three failed retries and three polls are count-identical to a reproduction triple; the count is runs since the last write, not truly consecutive). Every dismissal prints the rule that fired and its evidence, and is counted in a section separate from live findings. The run names one of four terminal states in words — candidates found · all auto-classified · no candidates above thresholds · no transcripts/parse failure. Silence is never proof of cleanliness.
• Other surfaces (Codex / Cursor / Copilot / Gemini): transcript metrics are unavailable — say so explicitly ("transcript metrics unavailable on this surface") and degrade to the repo-side summary the parser also produces: largest tracked source files, docs/tmp/handoff.md size, optional bash hooks/local/check-module-size.sh --all. Never fabricate transcript numbers.

Growth rule

A waste pattern that recurs across audits and matches no row above → add one rule row (with its quality guard and the next free TE-xx ID — IDs are append-only, never renumbered: audit reports and problem-catalog entries cite them) via skill-authoring. This includes recurring large-output / large-context patterns surfaced by the large-output audit class — add the row clean-room and dependency-free (never reach for a third-party compression tool). Project-specific waste patterns stay in project docs/skills, not here unless they generalize across Flow use cases.

Anti-patterns

• Citing FR-26 to skip a needed first-read, a reproduction run, or a verification step — that inverts the guardrail; the correctness/safety floor wins.
• Treating audit findings as verdicts. They are candidates: a warranted FR-18 full-Write supersede (structure/mode/ticket change, or most sections changed), mirror regeneration, and deliberate FR-10 reproduction look identical to waste in the metrics.
• Hard token budgets, caps, or gates on token counts — a budget gate trains truncation (intelligence damage); FR-26 is deliberately write-time discipline + retrospective audit only.
• Grep-and-edit blind: editing a file whose invariants you never read because "scoped reads are cheaper".
• Treating a summary as the source of truth for a final edit or an approval — the original is reopened first (§ Context compression discipline).
• Compressing away the evidence verification needs — reproduction runs, smoke output, acceptance evidence, or a stated uncertainty.
• Using token economy as an excuse to avoid opening the authoritative file before deciding.
• Adding a third-party compression dependency to Flow core — FR-26 economy is behavioral discipline + a deterministic stdlib audit, not a tool to install.
• Copying a third-party compression implementation, prompt, or doc into Flow — canonical Flow content is clean-room original.
• Rewriting always-on / governance files for cosmetic phrasing under the banner of "optimization" — governance edits are explicit, reviewed, diff-based.

Clean-room note

Original Fusebase Flow content. Designed after reviewing public AI coding workflow patterns; no third-party code, prompts, skill files, or hook scripts are copied. See docs/source-map.md.