Claude Code provides no in-session determinism mechanism, forcing automation users onto the metered Agent SDK path

Problem Statement

Anthropic's own legal page describes subscriber-included use as:

Advertised usage limits for Pro and Max plans assume ordinary, individual usage of Claude Code and the Agent SDK. — https://code.claude.com/docs/en/legal-and-compliance

"Ordinary individual usage" includes automation of one's own workflow on one's own machine. Automating a personal quality process is a normal individual activity, not third-party product development.

The problem is that Claude Code as currently designed offers no first-class mechanism for deterministic execution inside the interactive session. Same prompt produces different actions across runs. Slash commands and skills are prose loaded into context — they shape the model's behavior but do not constrain it to a fixed step sequence. There is no protocol/playbook mode that commits to executing N steps in order, no schema-validated tool outputs that fail hard on deviation, no idempotent step machinery, no built-in transactional/checkpoint semantics, no "no-LLM-judgment" mode that dispatches deterministic operations from a registered registry.

The architectural gap is well understood: a multi-step quality process at 99% per-step accuracy is not a quality process — it is a partially-supervised assistant. Determinism requires code, not prose.

The only available path to the determinism that "automation" implies is therefore to leave the interactive session and wrap claude -p in a deterministic Python or TypeScript harness — exactly what the Agent SDK is designed for. This is not an alternative architecture I chose for convenience; it is the only mechanism the product permits to deliver on the "individual automation" promise its own legal page makes.

The June 15, 2026 pricing change then meters that forced path separately:

Starting June 15, 2026, Claude Agent SDK and claude -p usage no longer counts toward your Claude plan's usage limits. [Affected:] Claude Agent SDK usage in your own projects (Python or TypeScript); the claude -p command in Claude Code (non-interactive mode); the Claude Code GitHub Actions integration; third-party apps that authenticate with your Claude subscription. Eligible Pro, Max, Team, and Enterprise plan users can claim a separate monthly credit. Pro: $20/month. Max 5x: $100/month. Max 20x: $200/month. When your monthly credit runs out, additional Agent SDK usage flows to extra usage at standard API rates — but only if you've enabled extra usage. If extra usage isn't enabled, Agent SDK requests stop until your credit refreshes. — https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan

The net effect is that Anthropic's pricing model penalizes the only mechanism it provides for the "ordinary individual usage" capability its legal page implicitly promises. The third-party-product use case the change is appropriately targeted at (and which already requires API key auth per the legal page) is not the use case I am describing.

Existing related signals from this repo

• #37686 — claude -p suggested to a Max subscriber, caused $1,800+ unintended API billing
• #45572 — Claude Code CLI usage incorrectly classified as API billing on Max subscription
• #56250 — Subscription quota eligibility for child_process.spawn("claude", ...) from owner-only local backend is undocumented

These are surface symptoms of the same underlying issue: there is no clean place in the product for a solo subscriber who needs deterministic workflow automation against their own projects.

Proposed Solution

Either of the following — ideally both:

(1) Provide first-class determinism mechanisms inside Claude Code interactive, so users do not have to leave the session to get repeatability:

• A protocol/playbook mode that commits the session to executing a named sequence of steps in order, with no skip, no reorder, and no paraphrase. Each step has structured inputs and outputs.
• Schema-validated tool outputs: a step declares the JSON shape it must produce, and the session fails hard (rather than continuing or improvising) if the model produces something else.
• Idempotent step machinery: re-running step N from a checkpoint produces the same effect, allowing the session to be safely resumed after a crash, interrupt, or human review pause.
• A "deterministic dispatch" mode in which the LLM's role is to choose from a registered, fixed set of operations and the session executes them as code — no free-form generation in the action path.

This would let "individual automation" happen inside the subscription-pool interactive session, where the legal page already says it belongs.

(2) Or, recognize that as long as (1) is missing, subscription-pool claude -p use is the natural consequence, and price it accordingly:

• Keep solo, single-user, owner-only claude -p invocations on the regular subscription limit when the auth context is the subscriber's own OAuth token, the cwd is the subscriber's own checkout, and the calls are not being routed on behalf of other users.
• The third-party-credential-routing prohibition already in the legal page is the correct enforcement boundary; the -p flag itself is not a reliable proxy for "this is a third-party product."

Use Case Example

A solo developer maintains a private codebase with a multi-step quality process: pre-commit hooks, internal review, automated review-thread classification and resolution, premerge checks, post-merge retrospective. Each step has well-defined inputs, outputs, and failure modes; missing or misordering a step is a defect the process is specifically designed to catch.

The LLM's role is bounded subtask execution: generate a diff for a specific review thread, classify a thread by intent, draft a retrospective section. The orchestrator decides what to call, validates each result against a schema, and fails hard if the contract isn't met. The same single subscriber, on the same machine, is at the keyboard for both the interactive Claude Code session and the orchestrator that spawns claude -p. Nothing is being routed on behalf of any other user. No third-party product is being built.

This pattern was built specifically because interactive Claude Code is not deterministic enough to drive a quality process directly — there is no in-session machinery to commit to a step sequence, validate outputs against a schema, or fail hard on deviation. The Python+claude -p layer is doing the work the product itself does not yet do.

Alternative Solutions

• Treat claude -p invocations from a verified OAuth-authenticated context, on the subscriber's own machine, against the subscriber's own checkout, as ordinary subscription use. The auth context already distinguishes this case from third-party services routing through subscriber credentials.
• Provide first-class cost telemetry surfaced in the subscriber's account dashboard. Today, total_cost_usd is in every claude -p --output-format json envelope but is not aggregated or reported back to the user. Subscribers have to instrument their own pipelines to see what they're spending — which means the first time most users learn they have crossed a billing threshold is when the threshold has already been crossed (#37686 is the canonical case).
• Document explicitly which claude -p invocation contexts count as subscription use vs. Agent SDK metered use. The current categorization is keyed off the -p flag, which is not informative about whether the invocation is third-party productization or owner-attributed automation. Issue #56250 is asking for exactly this clarity.

Additional Context

The pattern this issue describes — Python orchestration around claude -p with schema validation and fail-hard gates — is not unusual or marginal. It is the standard pattern for obtaining reliable, repeatable execution from an LLM in any context, and Anthropic ships and supports the Agent SDK precisely because of that. The disagreement is not whether the pattern is legitimate, but whether solo subscribers exercising this pattern on their own projects should be priced into the same metered pool as commercial product builders.

The cleanest resolution is to close the gap that forces them out of the interactive session in the first place. Until then, recognizing solo claude -p use as part of the "ordinary individual usage" the legal page already names would honor what the policy already says.