# Cost Aware Coding > Propose a cost-aware coordination plan or execute an accepted one. Use only when requested for coding work; exclude ordinary implementation and standalone model comparisons. - Skill: `stevennitesh/cost-aware-coding` (Agent Skill, multi-file: 9 files) - Install (CLI): `npx skillmds@latest add stevennitesh/cost-aware-coding` - Raw SKILL.md: https://api.skillmd.com/api/skills/stevennitesh/cost-aware-coding/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: stevennitesh (https://skillmd.com/u/stevennitesh) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/stevennitesh/cost-aware-coding --- # Cost-aware coding Complete authorized work with justified model allocation and evidence that the result is acceptable. Model choices here are experimental starting policies, not demonstrated savings or capability guarantees. The engineering contract owns coding quality; this skill owns routing, execution custody, and escalation. ## 1. Establish the outcome and cost constraint Use two modes: - **Coordination plan:** analyze the work, propose ownership and routing, and obtain acceptance before product implementation. This is the default for a new task. - **Execution:** carry out the accepted coordination plan. An explicit execution request with a settled route counts as acceptance; do not ask again. If the route is unsettled, present it for acceptance first. A route is settled by a previously accepted proposal or explicit user-selected ownership and model routing, with no material routing decision outstanding. An accepted feature spec, a model name alone, or the agent's inference does not establish route acceptance. A bare invocation without an identifiable task needs the task first. Resume is execution after state reconciliation, not a third mode. Establish the task from current repository guidance, requirements, code and checks. Resolve consequential missing product decisions with the user. Reuse existing discovery, decisions, and any plan or ticket. Use [shape-work](../shape-work/SKILL.md) for unresolved substantial decisions about feature behavior or approach, without requiring a new spec artifact. An Astra Medium root shapes directly; other roots delegate that pass to Astra Medium with the shape-work reference and relevant discovery context. Routine sequencing and implementation choices within settled requirements stay with the implementer. Include the [feature review gate](#4-accept-the-actual-result) when selecting the route. Discussion-only requests stop at their requested outcome. Identify any user-selected model, spending limit, latency requirement, or minimum assurance. Distinguish API spend, subscription allowance, and elapsed time; infer the priority from the request and clarify only when a material tradeoff remains unresolved. Choose a sufficient route considering total work through acceptance: context transfer, execution, repairs, verification, and integration. Without reliable accounting, use qualitative judgment; do not invent a budget or promise savings. Read [Model policy](references/model-policy.md) when selecting a route. Enforce a hard budget only with reliable accounting and bounded dispatch; otherwise present a coordination proposal without executing, or obtain agreement to an observable proxy. These limits also apply to planning work. ## 2. Propose and accept the coordination plan The root owns coordination planning: who works, with which model, and in what order. Choose implementation ownership: - **Root implements:** coherent work fits the root's model/effort and benefits from shared context. Specialist assistance and independent review do not change this ownership. The accepted plan may assign a bounded sequential implementation unit to a specialist while the root remains primary implementer. Transfer exclusive writer custody and return unit-local corrections to its current implementer under [Repair allowances](references/repairs.md). Schema, API, UI, and tests can be one assignment. - **Root coordinates:** managing dependencies, integration, or competing work units warrants dedicated coordination. Size, step count, or specialist assistance alone does not establish this need. The root plans, coordinates, integrates, and accepts; delegate product implementation and repairs. One reusable worker may execute several sequential units. Keep root integration to landing and verification; return product-code corrections to their implementer. Coordination protects the root's attention; it is not automatically cheaper. For root implementation, retain sufficient permitted work at the root unless a cheaper actor is likely to repay transfer and coordination costs over the remaining work. Reuse a suitable actor before spawning. Record each delegation's benefit in the proposal, or its assignment if not previously planned; reassess only when circumstances change. Capability, independence, useful concurrency, sufficient savings, or dedicated coordination can justify delegation when direct work or an existing actor cannot supply the benefit. Phases, role labels, file count, and repository boundaries alone cannot. No separate ledger or table is required. Planning permits justified read-only specialists and isolated scratch checks, but no product implementation or execution-worker pre-spawning before acceptance. Only the root delegates; workers must not delegate. During execution, default to one write-capable actor at a time, including the parent. Up to two independent read-only children may run when useful. They may use isolated scratch checks, but must not mutate the reviewed candidate or shared product state. Read [Runtime selection](references/runtime.md) once when establishing the route; reuse it until the runtime, route, or relevant configuration changes. Retain a sufficient root; include any justified model/effort switch in the proposal. For explicitly requested concurrent implementation, use [parallel implement](../parallel-implement/SKILL.md). Supply model choices to its scheduler. It owns admission, custody, integration, and safe recovery mechanics; this skill retains accepted implementation ownership, model and budget restrictions, repair allowances, and the feature-delivery review gate. Present a compact proposal: reference the accepted feature approach, name the root model/effort and implementation ownership, assign coherent work and reusable actors with dependencies, and include acceptance checks, review, and permitted recovery from [Repair allowances](references/repairs.md). Identify unresolved constraints. A small task may need only two sentences; no separate artifact or roster of pre-spawned workers is required. End with one acceptance request and, if needed, ask the user to switch this conversation's root model/effort before execution. Reuse existing acceptance; it covers the proposed route and recovery allowances, not unrelated effects. ## 3. Execute the accepted plan Verify root settings against the accepted route using runtime guidance; resolve a missing required switch rather than silently substituting a route. - **Changes requiring acceptance:** changes to scope, approach, root settings, root implementation versus coordination-only ownership, accepted implementation responsibility boundaries, model/effort allocations, concurrency, budget, review requirements, repair allowances, or explicitly required actors outside accepted alternatives. Obtain acceptance before dependent work. Available policy models are not blanket role permissions: an Astra reviewer allocation does not authorize Astra implementation. - **Root discretion:** adapt scheduling, coherent subdivisions, actor reuse, equivalent replacements, cross-unit repair ownership, and approved recovery within those boundaries without renewed acceptance. Subdivisions and actor IDs need not be prelisted. Apply the benefit, reuse, and custody rules to new or replacement actors; preserve unaffected work and counters. Record material adaptations briefly in existing run context, without a new plan artifact. Use [Telemetry](references/telemetry.md) for cheap best-effort start/end capture when available. Missing telemetry does not delay work unless required accounting makes it necessary. Use the runtime guidance for worker packets, context selection, actor reuse, and waits. Workers implement directly under the engineering contract. After dispatch, work only on useful independent responsibilities or wait for the return. Do not duplicate its assigned investigation while it runs. If a concrete gap requires intervention, coordinate with the worker rather than silently doing the same work again. For ownership transfer, interruption, replacement, or a run that must be resumed after the current context ends, read [Continuation and evidence](references/continuation.md). Preserve partial work and confirm prior writers and subprocesses have stopped before reassignment. Unknown writer state blocks reassignment, not evidence preservation. If execution or candidate acceptance fails, or evidence invalidates the approach, read [Repair allowances](references/repairs.md) before retrying or escalating. It owns separate implementation and review limits and their persistence. If a supplied budget is exhausted or cannot be enforced as required, stop before further discretionary work, preserve state and report the remaining outcome. Safe cancellation and custody reconciliation still take precedence over saving an extra turn. Do not silently exceed a model restriction or premium-use cap. ## 4. Accept the actual result Cost optimization operates within the required independent Astra review gate; Medium is the default. A feature-delivery run means this workflow owns a feature from its idea or accepted requirements through completion. Whether the root implements or only coordinates, once the requirements and any existing plan are implemented and required checks pass, delegate [change review](../change-review/SKILL.md) before declaring completion. Use an independent Astra Medium child even when the root is Astra Medium and did not implement. Astra XHigh requires explicit selection; it does not activate high-assurance review, which remains a separate user choice. Supply the accepted requirements/plan, candidate and proof; request a gate decision that challenges plan assumptions as well as code. Change-review owns inspection, finding standards, and rules for residual-risk acceptance. Hold the candidate stable under its custody rules until review returns. Resolve required corrections and any residual-risk acceptance with its authorized owner before completion; a review verdict does not grant that authority. If review requires corrections, read [Repair allowances](references/repairs.md) before dispatching repairs. If independent review cannot run, report the unmet gate rather than substitute self-review or claim completion. Bounded fixes and edits do not acquire this gate merely by being routed; their review remains conditional on the user's request or a concrete correctness concern. The root verifies decisive evidence for the final candidate, including additions, deletions, untracked changes, material interactions, and requirements omitted from worker packets. A worker's PASS is a claim to verify. Reuse valid evidence under the engineering contract and repository checks; investigate concrete gaps, contradictions, drift, or required assurance without repeating the worker's investigation or the delegated diff review. Independent review supplements executable proof. ## 5. Finish with evidence and limits Report the outcome, decisive checks, material limits, and a brief actor/repair summary. Include measured usage with its scope and cutoff; mention missing usage only when measurement was requested or its absence affected a decision. Follow the telemetry reference for interpretation. Omit empty fields and routine narration. Authorization carries across phases, actors, and continuation of the same scoped objective; pass its boundaries to the receiver. It does not cover unrelated future work. Commit, publish, install, or create a new app task only when authorized.