Ghidra Headless Progressive Decompilation

Progressive Decompilation

Use this phase skill when the earlier evidence, target-selection, and compare boundary surfaces are already reviewable and you need the standalone Stage 6 contract for Selected Decompilation And Incremental Compare.

The canonical contract is ./planning-brief.md. That brief is the portable handoff surface for speckit, then becomes the audit checklist for generated spec.md, plan.md, and tasks.md.

When To Use This Phase Skill

Use this skill when the request has already reached Stage 6 and the current outside-in step is concrete enough to classify as ready, qualified, or blocked.

Canonical Stage 6 Name

• Phase skill name: Progressive Decompilation
• Canonical workflow stage name: Selected Decompilation And Incremental Compare

The skill name is the independently invocable entrypoint. The longer Stage 6 name stays unchanged across the repository's existing reverse-engineering artifacts.

Full Stage 6 Scope

This phase covers:

• one selected outside-in decompilation step at the current frontier
• the incremental compare required for the replaced boundary
• explicit carry-forward of any caveat that keeps the step qualified
• audit of generated planning artifacts against the same Stage 6 contract

Non-Negotiable Stage 6 Constraints

• Stage 6 covers selected decompilation plus incremental compare as one contract.
• Decompilation remains selected-only and late-stage.
• Outside-in ordering remains the default progression rule.
• A current target, selection_reason, and question_to_answer must already be visible before direct invocation is ready.
• The current step must keep replacement_boundary, fallback_strategy, and a reviewable compare record explicit.
• Direct invocation does not waive headless-only, evidence-backed, reproducible, or Markdown-reviewable workflow expectations.
• No downstream speckit extension or constitution change is required.

Required Prerequisite Artifacts

• Stage guidance and canonical Stage 6 rules: ../ghidra-headless/examples/analysis-selection-playbook.md
• Current Stage 6 selection and compare-input surface: ../ghidra-headless/examples/artifacts/sample-target/input-inventory.md
• Replayable Stage 6 command surface: ../ghidra-headless/examples/artifacts/sample-target/command-manifest.md
• Reviewable incremental compare record: ../ghidra-headless/examples/artifacts/sample-target/comparison-command-log.md

When a source-derived claim is part of the current step, the upstream reference posture must also already be reviewable. Route back to source-comparison artifacts before proceeding when that posture is still deferred or stale.

Direct Invocation Readiness Model

How To Use This Phase Skill

1. Re-open the Stage 6 source surfaces and confirm the current step is ready, qualified, or blocked.
2. Fill in ./planning-brief.md with the current selected target, compare boundary, and output expectations.
3. Hand that brief into speckit as a file or inline paste.
4. Re-open this skill and audit generated spec.md, plan.md, and tasks.md with the same Stage 6 checklist.
5. If any blocking mismatch appears, refine or regenerate the planning artifacts rather than weakening the contract.

Stage 6 Output Expectations

Every Stage 6 step that proceeds must leave a reviewable output that keeps the selected target, selection reason, current interpretation, incremental compare posture, and remaining uncertainty explicit.

Worked Examples

• Happy-path handoff: ./examples/progressive-decomp-speckit-handoff.md
• Blocked direct invocation and audit findings: ./examples/blocked-direct-invocation.md

Audit Checklist For Generated Artifacts

• spec.md still describes the full Stage 6 scope rather than decompilation alone.
• plan.md keeps direct invocation states, prerequisite artifacts, and the compare boundary visible.
• tasks.md includes work for the happy path, blocked path, audit flow, and reviewer evidence.
• Generated outputs keep the selected target, selection reason, current interpretation, incremental compare posture, and remaining uncertainty explicit.
• Outside-in ordering and route-back behavior remain visible instead of being flattened into generic late-stage analysis.
• If any generated artifact drops the compare obligation or hides a caveat, the result is a blocking finding that must be corrected by refinement or regeneration.

Audit Finding Format

• artifact_name
• severity
• violated_rule
• evidence
• required_correction when the finding is blocking

Post-Speckit Audit Routing

• Re-open the generated artifact that violated the contract.
• Refine or regenerate the planning output so the Stage 6 rule becomes visible again.
• Do not accept a local shortcut that weakens Stage 6 in order to make the artifact pass.

Next Step Routing

• Use the umbrella skill when the real question is still which phase applies.
• Return to intake when target identity, archive normalization, or planning scope is still unstable.
• Return to evidence when the replay or artifact surface is not reviewable yet.
• Return to source-comparison artifacts when a source-derived claim still lacks a reviewable upstream posture.
• Return to semantic reconstruction when role, name, or prototype evidence is not strong enough for Stage 6.

Local Rule Policy

• Local repository rules may require extra review evidence, stricter naming, or additional compare notes.
• Local repository rules may not remove selected-only gating, compare obligations, or Markdown-visible audit expectations.
• Treat stricter local rules as additive overlays.
• Treat any local rule that weakens this Stage 6 contract as invalid.