Mo-first coding harness: bounded first deliverable

Orientation

Robert approved the lead/oracle recommendation on 18 Sep 2026: build a small Mo-first coding harness in Mo, initially maintaining existing Mo applications. Robert’s 19 Sep overnight instruction now authorizes the Astra lead to decide and drive bounded setup, implementation and verification while he is AFK. The lead establishes technical readiness and records each slice. This does not seal an experiment or accept an outstanding audit obligation.

Current worker instruction (19 Sep): keep Astra as lead in this Mac session and spawn fresh Astra workers at low reasoning in Herdr panes. The former oracle requirement and Amp worker configuration are superseded. See mo-lead for ownership and verification. The lead/worker model does not select the harness’s reference model.

Historical relocation instruction, 18 Sep 2026, 11:44 PM ET: move the Astra lead to Robert’s Mac, where he confirms OrbStack works. Keep orb worker configuration unchanged. HANDOFF.md owns arrival checks. Nine final credential-free executor probes passed in this orb; the two earlier failed probe runs are also retained at audit/evidence/2026-09-18/executor-feasibility/. This is evidence of a tested Linux container configuration, not a built Mo executor, protected verdict, Mac validation or general security certification. Temporary privileged setup was dismantled and controller changes restored. A dedicated delegated subtree and bounded output-drain behavior are requirements learned from the probe.

Readiness is scoped by work unit: full model-comparison budgets need not block a provider-independent fixture runner, but each brief must explicitly name its prerequisites, bounds and remaining unmet obligations. No implementation or login is authorized by the completed probe approvals or this relocation.

Mac inspection, 18 Sep 2026, 11:58 PM ET: the clean handoff checkout, native arm64 Zig 0.16.0 and OrbStack Docker endpoint are verified. No named Linux machine exists; a dedicated executor, Linux checkout/toolchain and protected verdict remain unverified. The Codex session lacks Amp oracle/thread tools; Robert’s workflow choice is pending. Evidence is at audit/evidence/2026-09-18/mac-arrival/. No setup or implementation had started at that checkpoint. The 19 Sep Herdr/overnight instructions above resolve the workflow and start authority.

The useful result is: task + isolated checkout → candidate patch + independent behavioral verdict, with command outcomes, time, usage and failures recorded. Two questions stay separate: is Mo useful for implementing this developer tool, and does Mo-specific guidance help the agent change Mo applications reliably? Neither result by itself proves a new language/runtime is necessary.

Pi is a design reference and the practical baseline, not a feature checklist. The librarian inspected it; the lead also cloned and read revision 36b60d2. Borrow its separation of provider integration, agent loop and coding session, and its read/write/edit/command tool set. Pi explicitly delegates real isolation to the OS or an external executor; project trust is not containment.

Existing examples/programs/agent supplies a custom HTTP model loop, file tools, budgets and transcripts. It lacks command execution in its tool dispatch; restart fails interrupted runs rather than resuming them. Reuse selectively, without assuming the recorded guarantees have been independently accepted.

Write scope and ownership

The lead chooses each bounded implementation slice under Robert’s overnight authority. Before launch, append an exact path allowlist for each part and its independent acceptance owner. The designated worker owns future code under toolchain/ and examples/; the lead owns requirements and acceptance readings. Do not choose a new worker environment implicitly or assume private restoration.

During a coding trial, the target agent can change only a disposable application checkout. Freeze and protect the running harness and its source, compiler, guide, task requirements, executor policy, and final verification code/results. The target Agent application may be a separate copy of the harness’s ancestor; editing that copy must not change the harness actually executing the trial.

Parts, in order after implementation approval

A. Truthful evidence and bounded execution

First bounded implementation brief: mo-executor-foundation, with exact worker write scope, isolated-machine setup and lead-owned acceptance controls.

Define a protected external runner that binds a verdict to the exact candidate, toolchain, configuration and expected check inventory. Development tests in the writable checkout are feedback, not final authority. Acceptance checks assert observable behavior independently of the implementation; verified: and a successful compiler exit alone do not establish correctness.

Before scoring agent work, demonstrate rejection of failed builds/processes, empty or incomplete check selections, timeouts, stale candidate outputs and known-wrong candidates. Include a control where the requested fault never occurs: absence of the stimulus is incomplete evidence, not a pass.

A writable checkout alone is not isolation. Select an established OS-isolated executor with bounded CPU, memory, wall time and output, descendant cleanup, restricted filesystem/network access and credentials outside the task boundary. Protect verifier outputs from the target agent and target programs. Do not build a new sandbox or rely on Mo source capabilities to constrain child processes.

Treat candidate code and commands as potentially malicious, including during final verification. Execute them inside isolation; keep the authoritative runner and verdict writer outside that boundary. A negative control must show that forged success output or an attempted result overwrite cannot produce acceptance.

B. One complete coding path

mo-provider-foundation is the bounded offline provider slice: pin/artifact identity, actual SSE parser, usage presence and sanitized outcomes. The Mo bridge and interactive authentication follow independent acceptance.

One session, one provider and one reference model; thin headless CLI plus machine-readable events. Read/search, targeted edits, writes and bounded command execution must support check → inspect failure → repair. Keep the model/tool loop separate from provider and execution adapters without a plugin framework.

Supply version-matched Mo documentation, available command help, diagnostics and existing runtime inspection where useful. Preserve diagnostic detail; do not invent commands or require an interface redesign before use. Return distinct success, refusal, failure, timeout and cancellation outcomes, exit status when available, and explicit output truncation. Do not equate transport success with tool success. Stop explicitly at step, context, time or usage limits.

Use a narrow trusted adapter for provider HTTPS/auth and external execution where needed, with maintained TLS and verified certificates. Pin and disclose that adapter and its dependencies: this is not an end-to-end zero-dependency claim. Its exact transport, provider and implementation require a readiness decision.

Provider candidate identified; readiness incomplete (18 Sep 2026): Robert requires independent OpenAI subscription OAuth for the harness. At the Pi pin above, packages/ai/src/auth/oauth/openai-codex.ts implements device and browser login, and providers/openai-codex.ts selects the dedicated subscription Responses transport. The lead/oracle recommend a pinned pi-ai adapter for authentication and inference only; Mo retains tool dispatch and the agent loop. Do not invoke the full Pi/Codex agent as a substitute or reuse Amp credentials.

Prefer device login: forward the real verification link and short-lived code through an operator-only channel when an authorized flow starts. It avoids the orb/laptop localhost callback mismatch. If needed, browser fallback must accept the complete redirect through private input, check the expected callback and nonempty matching state, and never ask for tokens or redirect URLs in chat.

Keep a harness-specific credential store inaccessible to candidate processes, not merely outside the checkout; preserve serialized refresh/deletion and atomic replacement. Pi’s default AI credential store is in-memory; its coding-agent file store is permission-restricted plaintext, not a keychain. Pi login returns credentials and some upstream errors include raw response bodies: expose only allowlisted auth status fields, never raw results/errors in transcripts. Local logout deletes the credential; server-side revocation is not established.

The code demonstrates Pi’s use of the Codex OAuth registration and subscription backend, not blanket authorization for independent client reuse. Resolve client identity/support expectations, account eligibility, reference-model access and network/storage policy before live integration. OpenAI’s authentication guidance recommends API keys for programmatic Codex workflows; that does not establish either a ban or permission for this personal third-party subscription integration. No login or entitlement test has occurred. Subscription allowance is not API dollar cost.

Record completed model/tool results before subsequent work. Interrupted runs fail explicitly; no automatic replay of commands with uncertain outcomes. A timeout does not prove a command did nothing. Report usage from trusted provider/executor observations; missing usage remains unknown, not zero. Do not claim exact billing caps or crash durability without evidence for those claims.

C. Calibration and useful application tasks

mo-agent-terminal-auth records the lead-selected v1 Agent specialization, separate recipe binding and exact fixture controls under overnight authority.

Start with the Agent’s terminal-401 policy change as workflow calibration, not a language-value test. The lead recommends an Agent-specific requirement for the test copy: after receiving 401, make no further request in that call and return the status error; preserve other retries and deadlines. The shared recipe currently promises retries on model errors. Before implementation, version the new requirement and explicitly resolve its recipe binding; no silent exemption or shared-recipe rewrite. Approval of this project is not a completed policy spec.

Independently observe request counts and outcomes for immediate 401, transient failure followed by 401, transient recovery, exhaustion, deadline and normal success. Reject solutions that skip all requests or disable all retries. Then choose two or three bounded fixes/features in existing Mo applications, with frozen requirements and behavioral checks before model runs. Include a task exercising failure diagnosis or required progress if appropriate; do not inflate the first task into a durability or distributed-systems project.

D. Practical comparison, then affordable-model evaluation

Compare with actual Pi under matched model, task, guidance, available tool capabilities, budgets and isolation. Pin Pi and document unavoidable differences in schemas, editing semantics, prompts, context policy and tool integration. The operator’s tools are not automatically tools available to either harness. Match conditions before interpreting outcomes; no comparison of unlike scaffolds as proof about language syntax. Do not build a second Elixir harness first.

Calibrate reliability with the reference model, then add one affordable model. Any stronger-model assistance is a separately recorded treatment. The future onboarding comparison remains a separate question: use the same guidance in both harnesses rather than confounding it with this product comparison.

Numbers and interpretation

Report three separate ledgers: harness correctness, independently accepted task outcomes, and effort/cost to develop and maintain the harness in Mo. Record all planned attempts, failures, timeouts, incomplete runs, tokens, available billed cost, human intervention and stronger-model assistance. Record end-to-end time and tool/model/verification latency separately, with cache and environment state.

Before scored runs, freeze task identities, provider/model versions, budgets, repetitions, ordering and verdict rules. Those numerical limits are unresolved, not invented in this brief. No 2× threshold, production assurance, or superiority claim is adopted. A negative result can motivate a repair or narrower scope; inconclusive evidence stays inconclusive. Neither substitutes for retained audit obligations. Shared compiler bugs require independent behavioral checks even when both Pi and the Mo harness agree.

Done when and readiness

The first implementation is complete only when its controlled negative cases are rejected, one full coding path produces a correctly bound external verdict, and cancellation, interruption and limit behavior are exercised. Report failures honestly; all tasks need not succeed to finish an evaluation. The broader pilot finishes with its full attempt ledger, not when a preferred outcome appears.

Before a worker brief is ready, the lead must resolve and record:

  • Pi provider candidate: resolve registration/support, account/model access and maintained TLS/auth adapter contract; permitted network boundary.
  • Astra/low Herdr worker configuration is selected; verify executor isolation, resource limits, credential separation and protected storage before readiness.
  • Exact code write scopes, command/API feasibility and acceptance ownership.
  • Versioned first-task policy and recipe relationship, budgets and trial plan.
  • Required environment restoration and instrument repairs; any omitted legacy obligation explicitly remains unmet rather than being silently waived.

No TUI, dynamic plugins, multi-agent coordination, automatic resume, automatic compaction, self-modification or compiler changes in the first deliverable. Unexpected missing runtime support returns to the lead for a bounded decision. Program 7 stays suspended; this project does not inherit its number, seal or retirement gates. Step 39 remains unaccepted. Containing the pilot is not acceptance of Mo TLS, Darwin full-sync or any outstanding durability claim.