cost-routing-discipline

name	cost-routing-discipline
description	Use when starting any new task to select the right GSD mode (fast/quick/plan-phase) and the right model tier (Haiku/Sonnet/Opus) for subagents — prevents 5-10× cost overpay on routine work

Cost Routing Discipline

Why this skill exists

Default Anthropic UI sends everything to Opus 4.7 + max effort. This is correct for architecture but 5-10× overpay for grep, file-search, and routine coding. Without discipline, monthly cost is $300-500 for a solo dev. With discipline, same work costs $50-100.

When this skill activates

• Before invoking /gsd-* commands — match mode to task risk
• Before dispatching subagents — match model to task type
• When user describes a task — auto-route based on keywords

Mode selection

Task                                          Mode
─────────────────────────────────────────────────────────────
Trivial (typo, 1 line, doc update)         →  /gsd-fast
Small feature (<100 LOC, single concern)   →  /gsd-quick
Big feature (auth, payments, public API)   →  /gsd-plan-phase
Breaking change / migration                →  /gsd-plan-phase + /council

Trigger keywords (force-route)

Keyword	Auto-route to
"опечатка", "typo", "одна строчка"	/gsd-fast
"обновить доку", "doc update"	/gsd-fast
"добавить тест", "add test"	/gsd-quick
"новый endpoint", "new endpoint"	/gsd-quick (small) or /gsd-plan-phase (with auth)
"auth", "payments", "schema migration"	/gsd-plan-phase + /council (always)

Subagent model routing

When dispatching subagents (via better-model installed):

Task type                              Model              Effort
──────────────────────────────────────────────────────────────────
Search / grep / pattern match        →  Haiku 4.5         (no effort)
Implementation / coding (1-2 files)  →  Sonnet 4.6        medium
Test writing                         →  Sonnet 4.6        medium
Single-file debug                    →  Sonnet 4.6        medium
Multi-file refactor (3+)             →  Opus 4.7          xhigh
Cross-file debug                     →  Opus 4.7          xhigh
Architecture design                  →  Opus 4.7          max
Security audit                       →  Opus 4.7          max
Code review                          →  Opus 4.7          xhigh (max overthinks)
Novel algorithm                      →  Opus 4.7          max

Haiku does not accept the effort field (better-model v0.7.0+ warns on haiku+effort combos and treats haiku-no-effort as correct). Do not pass effort to Haiku — it is rejected upstream.

Observability — better-model stats

better-model v0.8.0+ ships a read-only better-model stats CLI subcommand for cost / routing analysis. Run after a heavy session to see per-tier token spend, route hit rates, and which Sonnet/Haiku calls escalated to Opus. Use to validate that your routing rules above are actually saving the predicted 5-10× cost.

better-model stats             # summary for the last 24h
better-model stats --since 7d  # weekly cost rollup

Toolkit-pinned better-model version is now v0.8.1 (manifest.vendor_pins.better-model).

Edit operations

If Serena MCP installed (look for mcp__serena__* tools):

• Use Serena's symbol-level operations (replace_symbol_body, insert_before_symbol, insert_after_symbol, rename_symbol) for refactoring — local LSP, no network round-trips, dramatically fewer tokens than full-file rewrites
• Use native Edit for plain-text single-line changes

Morph Fast Apply was removed in v6.1 (no public source repo for the SDK, paid SaaS with no privacy guarantee). There is no plug-and-play equivalent — native Edit covers ~95% of cases honestly.

Discovery operations

"find function/symbol/all callers of foo" →  Serena (mcp__serena__*)
"find code that does X" semantically       →  claude-context MCP (>100k LOC)
exact-string search                        →  ripgrep
small project (<10k LOC)                   →  ripgrep + Read
big project (>100k LOC) + frequent queries →  claude-context MCP

Budget reality check

Track via Anthropic Console → Usage:

≤ $100/mo   →  use any mode freely
$100-300/mo →  default to /gsd-quick for ~70% of work
$300-500/mo →  force /gsd-fast for trivial; /gsd-quick mid-tier
> $500/mo   →  STOP. Audit last week's transcripts. Likely retrying same task multiple times.

Anti-patterns

• ❌ Using /gsd-plan-phase for typo fixes (waste $4 per typo)
• ❌ Sending grep tasks to Opus (waste 5× on each search)
• ❌ Using native Edit when Morph available (waste 10× on each edit)
• ❌ "Just one more iteration" — if cost >$10 on single task, root-cause why
• ❌ Avoiding /gsd-plan-phase on auth/payments to save money (wrong place to cheap out)

When NOT to optimize

Pay full price for:

• Security-relevant changes (auth, crypto, file uploads)
• Payment processing
• Public API breaking changes
• Database migrations on prod data
• Multi-week features touching 10+ files

Cost discipline ≠ cheaping out on high-stakes work.

Cross-references

• components/cost-discipline.md — full cost layer breakdown
• components/external-tools-recommended.md — Morph + better-model install
• rules/cost-discipline.md — auto-load trigger keywords
• scripts/setup-cost-routing.sh (PR 5) — better-model installer