Coding-runtime research

This research was checked 15 October 2026 against the primary sources below. It informs Development bootstrap, while account entitlement and LychD implementation still need their own evidence. Recheck provider behavior before purchase, authentication integration, or runtime qualification. The CovenantsDesign Rule (ADR)sDesign Rule (ADR)s retain architecture authority and State of Work retains delivery authority.

Recommended road and alternatives

Use the existing CodexSettings FolderSettings Folder application and requested Sol setting for the first bounded campaign. Qualify the existing serial CLI protocol behind a trusted LychD host before adding a product-facing app-server integration. Compare another reviewer on frozen cases before buying ongoing capacity. This recommendation fits the present bootstrap constraints; no benchmark establishes a universally best agent stack.

Road	What the primary source establishes	Bootstrap decision
CodexSettings Folder app with a goal	A durable objective continues across turns toward a verifiable stopping condition. Goal guide.	Start here.
CodexSettings Folder exec	Noninteractive jobs, JSONL events, output schemas, explicit sandbox settings and saved CLI authentication. Guide.	Smallest first foreign-runtime job.
CodexSettings Folder SDK	Programmatic CodexSettings Folder threads and streaming through official TypeScript and Python surfaces. SDK guide.	Consider when typed integration reduces actual work; avoid redundant layers.
CodexSettings Folder app-server	Bidirectional thread/turn protocol, steering, interruption and approval handling. Protocol.	Add for a demonstrated persistent-session or interactive-product requirement.
OpenAI Agents SDK	Application-owned agent loop, handoffs and integration. Runtime comparison.	A distinct orchestration choice that cannot absorb LychD's owners.
OpenAI Agents API	Managed CodexSettings Folder harness, durable sessions and compaction/recovery; model usage is API-billed, with applicable tool/container costs. Overview.	A later managed-server option, rather than the default way to spend app allowance.
Claude Code / Agent SDK	Programmatic coding and another provider's harness; supported authentication varies by mode. Programmatic guide, SDK.	Optional comparison and separately qualified adapter.

The distinction also appears in OpenAI's platform engineering account: exec serves bounded jobs, SDKs serve workflows, and app-server serves product clients. Reusing a harness saves implementation work, but its private loop acquires no LychD promotion, credential, or lifecycle authority.

Sol, Ultra and allowance

The model guide names gpt-6.1-sol and distinguishes Max's additional single-model reasoning from Ultra's use of subagents. Record the requested Sol identifier and effective client settings; never silently substitute another model. Availability still depends on client rollout and workspace/account controls.

The Sol API page lists reasoning efforts through max, without an ultra effort. Ultrafast is a separate speed/service choice. An app's Ultra selection therefore cannot be copied blindly into a raw Responses reasoning field or CLI override. The adapter needs a reviewed mapping and must refuse unsupported settings.

The current pricing guide says Work and CodexSettings FolderSettings Folder share subscription usage, while API prices are separate. Model, speed, task size and retained context affect consumption. Allowance is not a fixed number of coding hours. Read actual usage and reset windows; keep account balances outside public docs. Unused allowance authorizes no paid overflow.

The noninteractive guide documents saved CLI authentication and API-key alternatives. Its advanced ChatGPT-account CI credential-seeding workflow excludes public/open-source repositories. That specific restriction does not prohibit local coding against an open-source checkout. Do not copy personal tokens into shared/fork runners.

Subscription-backed integration has a specific admission path

Avoid the outdated blanket claim that a ChatGPT subscription can never back an application call. OpenAI documents a Sign in with ChatGPT plan-usage preview for open-source and locally hosted apps, with client registration and host identity. Paid or remotely hosted apps have a separate interest path. This is a provider integration to qualify, not permission to extract an existing CodexSettings FolderSettings Folder session token.

The preview limitations require streaming with storage disabled and client-owned HTTP history. Several fields and hosted capabilities are excluded, including HTTP multi_agent, hosted MCP and native computer use. Local harness orchestration is a different layer. Match required capabilities to the exact road. The same preview rejects max_output_tokens and max_tool_calls. A wall deadline, usage snapshot or captured token count is not a strict per-job token/credit cap. Required native ceilings need separately qualified reservation/enforcement; if unavailable, that launch must refuse.

The plan-usage app-server guide warns that a bundled model catalogue is not an entitlement test: successful inference is needed to prove request access. Metadata-only preparation cannot establish the whole model/billing path.

For operator/platform-controlled personal bootstrap, use the existing supported client authentication. Native CoffinSandboxSandbox launch remains a separate qualification: the current protocol supplies fresh homes with neither provider credentials nor an auth file, and needs a reviewed Gate bridge for every child provider request. Mounting a personal auth store into the guest does not meet that requirement. Until a supported credential-separated, bounded provider road is qualified, keep lifecycle control in the external operator campaign. A distributed LychD offering must independently qualify registration, per-user credentials, host identity, consent, revocation, and provider restrictions. Shared subscription pools and copied account sessions are not proposed.

Continuous work needs a host and a stopping rule

Long-running work documents goal mode with the same sandbox/approval boundaries and separate worktrees for parallel coding. Local scheduled tasks need the computer on, app running and project available. These are wakeups, not an always-on server guarantee.

OpenAI's long-horizon example uses durable specifications, milestones, validation commands and recorded decisions/status. Its reported long run is an example, not a general completion guarantee.

Anthropic's effective-harness experiment reports that compaction alone did not solve continuation; incremental features, explicit checks and recoverable handoffs mattered. A later harness-design experiment finds evaluator separation useful near the generator's capability boundary, while some scaffolding became overhead as models improved. The inference for LychD is to measure each extra role and remove ceremony that does not improve accepted outcomes.

Anthropic's multi-agent research system reports substantial token overhead and notes that tightly dependent coding parallelizes less readily than research. This supports bounded independent panels, not a universal cost multiplier or a claim that more agents always improve coding.

Containment version selection

The 15 October review retires Security's former 0.66-series baseline without admitting a replacement. Upstream's DNS-before-admission advisory and same-port destination bypass list affected versions through 0.73.0 and fixes at 0.74.0. Those published ranges include 0.66; Landlock ABI 6+ alone does not dispose of the destination-address defect. Conversely, the older D-Bus advisory lists versions before 0.55.0, so it does not establish affectedness of 0.66. These are upstream reports, not reproduced exploits against an admitted LychD profile.

Later metadata needs source resolution. The endpoint advisory lists the ordinary dot-segment fix at 0.73.0 while retaining an unresolved semicolon variant. The tool-shim advisory uses a commit cutoff and an unreleased-fix field. The 0.78 release associates these advisories with that release, but association cannot silently repair the patch fields. The 0.80 release is a candidate for review, not an approved version.

A bounded read-only study pinned 0.80 source commit 8bb041356d48f771a221e58c98d9b1ff442e76fe. Its endpoint policy contains dot-segment handling, ambiguous-path rejection and separator-aware globs; the inspected reverse-proxy path calls the policy before forwarding. The Linux entrypoint and shim contain a refusal path for recognized shim invocations whose broker context is missing. These observations narrow the findings for the inspected paths. They establish neither the first fixed release nor complete caller, launcher, policy, dependency or forwarding closure. The internal child-launch path and its authority still need review. Embedded tests were not run; no downloaded code was executed or installed.

Host qualification begins with one exact source and policy. Resolve every applicable advisory through inspected fixes and regressions or an enforced exclusion, bind that review to the built executable and dependencies, then prove actual containment and ceilings separately. Requested systemd properties cannot replace missing effective CPU/memory files. Execution stays refused while any required qualification is unknown. The Security owner retains this law; the study delivers no CoffinSandboxSandbox, Provider Gate, native launcher, or platform qualification.

Whether to add Claude

Claude's pricing page lists Pro at US$20 monthly and Max from US$100 before applicable checkout differences. These are dated observations. Purchase only for independently observed value within the actual total cost envelope.

Pro/Max Code guidance describes shared Claude/Claude Code usage windows and API keys selecting separate API billing. Current Agent SDK plan guidance says SDK, claude -p and third-party app use may still draw plan limits; earlier announced changes were paused. Do not assume the SDK is universally API-only.

The monthly API credit policy, changed on 7 October, describes eligible Max/Team credits and claim requirements. Those credits are distinct from interactive subscription allowance and apply to the policy's eligible API use. Buying a plan does not prove a particular adapter will draw those credits.

The programmatic guide states that bare mode does not use subscription login and requires supported API/provider credentials. The SDK overview prohibits third-party developers from offering claude.ai login/rate limits in products without prior approval. Personal operator use and a distributed LychD authentication offering are separate decisions; isolation flags that change billing must be qualified explicitly.

First trial a fresh Sol verifier; then, optionally, compare Claude on the same frozen diffs, acceptance criteria, and total-budget envelope. Keep the answer key protected and use separate confirmation cases. Another provider needs measured incremental benefit before adoption. This research bought or configured neither.

Verify the road before first use

1. Read local CLI --version/--help; compare flags and wire format with the reviewed adapter. Here the available CodexSettings FolderSettings Folder binary reported 0.159.2, matching the repository's reviewed CodexSettings Folder protocol. Help launched no agent and proves no provider access.
2. Inspect supported client model choices, settings, auth mode and usage without exposing secrets. App-server documents account/read, account/rateLimits/read and model/list; review the versioned protocol before calling them.
3. Select the exact personal-local, registered plan-usage or API road. Record missing entitlement and accounting evidence instead of inferring it from catalogue presence.
4. Qualify containment with an inert test executable, then admit one harmless bounded inference to prove model access, effective settings, usage attribution, output and cancellation.
5. Freeze runtime/policy identity and compatibility fixtures. Version, authentication, plan or protocol changes reopen affected qualification.

The repository's CLI protocol constructs command intent without a physical launcher. Limits cover wall time/capture; it emits no reasoning-effort setting. Official provider features do not close these implementation gaps.

Reopen the recommendation when interactive approvals require app-server, explicitly chosen server uptime justifies managed API costs, or matched review evidence earns another provider. Revisit model/effort mapping and plan eligibility after provider changes, and agent count after measured integration or review bottlenecks. None permits silent runtime fallback.