Dual Pass Workflow

Dual Pass Workflow

Overview

Use this skill as an outer shell. Let the domain skill handle task-specific method, canonical outputs, and detailed verification. Use this skill to define a first-pass mode and, when needed, a second-pass mode.

Step 1: Select The Domain Skill

• Identify the task-specific skill first. Examples: cs-paper-review, cs-meta-review, a bug-fix workflow, a frontend-edit skill, or a writing skill.
• Keep the domain skill as the source of truth for the main artifact and its verification.
• If the domain skill already has an outputs/ folder or other task-local artifact layout, reuse it. Do not create a second top-level workflow folder unless the task has no obvious artifact directory.

Step 1.5: Run Environment And Input Preflight

• If the domain skill defines a workspace contract (a set of paths or files that must exist before work starts), verify that contract now. If any required path is missing, stop and ask the user instead of improvising. See workspace_contract in references/contracts.md.
• If the domain skill defines an input integrity check (such as PDF injection scanning, adversarial input detection, or untrusted-input handling), run it before any drafting. Record the result as a preflight artifact under the artifact directory.
• If the domain skill defines an ordered preflight pipeline (multiple sequential checks that each produce an artifact), run them in the declared order. Do not collapse them into a single flat verification step. See the ordered verification form in references/contracts.md.
• This step is a dispatch point, not a duplication point. The dual-pass workflow triggers the domain skill's preflight; it does not redefine what the preflight checks.
• If the domain skill has no workspace contract, no integrity check, and no ordered preflight, skip this step.

Step 2: Write The Task Packet

• Prefer workflow.yaml in the task root when the repo already has a task folder. If it is missing, infer the packet from the domain skill's standard layout before asking for more setup.
• For review folders in this repo, default to:
  • task_root: reviews/<paper-id>/
  • primary_result: outputs/review.md
  • artifact_dir: outputs/
  • sources_of_truth: review.yaml, intake/requirements.md, intake/extra-text.md, intake/extracted-paper.md
• If the task has no dedicated folder and no obvious standard layout, create .agent/<task-id>/workflow.yaml.
• Use the required fields in references/contracts.md.
• Fill only the fields needed for a clean handoff: task_type, goal, selected_skills, task_root, primary_result, artifact_dir, sources_of_truth, constraints, acceptance_criteria, and verification.
• If the task is already packetized by a domain skill, keep the packet thin and point to the domain files instead of duplicating them.

Step 2.5: Build Intermediate Representation

• If the domain skill defines a fact extraction or evidence mapping step before drafting (such as an evidence map, normalized fact table, or consensus map), run it now and record the output as a preflight artifact.
• The intermediate representation belongs to the domain skill. Dual-pass only ensures it exists before first-pass drafting starts.
• If the domain skill has no such step, skip this.

Step 3: Run first-pass

• The current agent is the default first-pass agent unless the user says otherwise.
• first-pass uses the domain skill and updates the primary result.
• first-pass may also write short workflow artifacts under the artifact directory:
  • build.<agent>.md
  • handoff.<agent>.md
• build.<agent>.md should record what changed, which files moved, and what verification actually ran.
• handoff.<agent>.md should contain: what_i_did, assumptions, verification, open_issues, and suggested_audit_focus.
• When the domain skill already produces evidence or reasoning artifacts, reference them instead of repeating their full contents.
• If the task is low risk or the first pass is already sufficient, stop here.

Step 4: Run second-pass When Needed

• Use second-pass only when you want a real check, refinement, or adversarial review.
• If there is no first-pass record (no handoff.<agent>.md, no build.<agent>.md, and the primary result does not exist or is empty), stop and ask the user before proceeding. The user may override to run an independent audit on a human-written artifact.
• The second-pass agent reads workflow.yaml when present and otherwise infers the packet from the task layout.
• If independence matters, let the second-pass agent write a short precheck.<agent>.md with expected strengths, risks, or failure modes before opening the draft.
• The second-pass agent then reads the primary result, the first-pass handoff when available, and the domain verification artifacts.
• The second-pass agent writes audit.<agent>.md with the categories from references/task-mappings.md:
  • factual_or_behavior_issue
  • unsupported_claim_or_assumption
  • missed_requirement
  • clarity_or_structure_issue
• If the primary result or any output file contains protected region markers defined by the domain skill (such as % refiner:body-* in nsf-proposal-composer), the second-pass agent must not rewrite content inside those regions unless the user explicitly requests a full overwrite. The audit may comment on protected regions, but reconciliation must respect their boundaries. See protected_regions in references/contracts.md.
• Default rule: audit first, rewrite second. Do not let the second pass collapse into unstructured polishing.
• If the issues are clear and safe to fix immediately, the second-pass agent may also reconcile the draft in the same turn.

Step 5: Reconcile Only If second-pass Ran

• Reconcile is optional because second-pass is optional.
• If second-pass ran, update the primary result, keep or reject audit points explicitly, and rerun the domain verification steps.
• Write reconcile.md with three short sections: adopted changes, rejected points, and verification after reconcile.
• The primary result remains the canonical artifact. Workflow notes are supporting context only.

Output Contract

• Use references/contracts.md to map the workflow onto an existing task layout.
• Use references/task-mappings.md to tune the audit emphasis by task type.
• Copy the templates under assets/ only when the repo or skill does not already provide an equivalent file.
• Treat workflow.yaml as an override file, not a hard prerequisite, when the domain skill already defines a clear task folder layout.

When Not To Use Dual-Pass

• Do not use dual-pass when the task is a single atomic operation with no meaningful check (e.g., renaming a file, running a script with no judgment).
• Do not use dual-pass when the domain skill already has a complete internal verification pipeline and the user has not asked for an external audit.
• Do not use dual-pass when the second pass would lack the context needed to audit meaningfully (e.g., a code review where the checker cannot run the tests).
• Do not use dual-pass as a substitute for domain-skill-internal quality steps. If the domain skill has built-in cross-checks, dual-pass adds value only when a separate agent or separate context provides genuine independence.

Stop And Ask

• Ask when there are multiple plausible primary results.
• Ask when the domain skill has no clear verification step and the task is high risk.
• Ask when the checker would need context that the packet does not record and cannot infer safely.