Codex interview framework

Run every normal-mode start, triage, research, requirements, design, and tasks phase through references/algorithm.md and references/domain-modeling.md.

Completion criterion

Delegate only after all of these conditions hold for one identity tuple:

• phase, interviewId, discoveryRevision, and contextDigest match in state and in the delegation packet.
• The selected skill manifest has status complete or partial_warned.
• The phase interview has status complete or skipped after explicit final approval.
• phase_gate.py check-delegation succeeds.

Hard-transition invariant

Run the current check-delegation immediately before each affected phase transition or child dispatch. A failed check-delegation in either mode stops this invocation before state transition, child dispatch, or target-artifact write. After a normal-mode failure, the next explicit invocation uses a fresh manifest/interview identity. After an exact --quick delegation failure, the next explicit invocation reruns discovery, records a fresh phaseSkillLoad and interview identity, and does not reuse a terminal bypassed_quick interview or its discovery revision. Only a matching in-progress collecting or awaiting_confirmation interview is resumable. A matching in-progress interview that has not reached a failed delegation boundary remains valid for resume.

Exact --quick preserves discovery, manifest, parent-delegation provenance, check-delegation, receipt recording, and check-agent-write. Quick mode bypasses interview questions only. Exact --quick separately bypasses final approval. It still runs check-delegation immediately before dispatch, uses the same identity tuple, requires a current complete or partial_warned manifest plus an interview receipt with status bypassed_quick and quickAuthorization.source: "--quick", and has no other bypass.

Hard boundaries

• Ask only critical user decisions that can change phase scope, externally visible behavior, acceptance, architecture, sequencing, or material risk.
• Inspect code, configuration, state, prior artifacts, and selected skill contracts for facts. Do not ask the user for discoverable facts or setup and administration choices.
• Ask the whole currently unblocked critical frontier. Split a frontier only when Codex's native user-input tool limit requires another batch. The current limit is three questions per call.
• Put the recommended option first and state its tradeoff. Offer only viable alternatives.
• Persist each partial answer before asking the next frontier.
• Treat control-only replies such as apply the changes, continue, proceed, and go ahead as no answer to an active interview question.
• Treat bare skip during an active question as an instruction to finish the remaining interview with stated defaults and assumptions. Present the final approval gate before delegation.
• Delegate in the same turn after the user explicitly chooses approve and delegate.
• During artifact review, treat apply the changes as a revision request. Delegate the revision and remain at the artifact approval gate.
• Apply domain-language modeling during the grill. Challenge conflicting or vague terms, use concrete boundary scenarios, and record resolved domain terms without creating an ADR.

Preload boundary

Load selected skill bodies and required current-work references as contracts. During preload, collect constraints, vocabulary, checks, and conflicts. Start no task action prescribed by a loaded domain skill until the phase interview is approved and the delegation gate passes.

If this core skill, its algorithm reference, or its domain-modeling reference cannot be read and hashed, record phaseSkillLoad.status: "core_failed", report the failure, and stop. Record domain-skill load failures in warnings and failures, use partial_warned, and continue. Turn unresolved material conflicts between loaded contracts into first-layer interview decisions.