Pre-existing Skills Scaffold Grill Log

Branch: Agent-Owned Skill Reconciliation

Q1 Question: How should HI handle pre-existing repo skills that overlap scaffolded packs?

Accepted answer:

• HI baseline has precedence when a pre-existing repo skill overlaps a scaffolded HI baseline skill.
• Exact name/id overlap is the first deterministic overlap signal.
• Dynamic or semantic overlap is model-guided work at dp report time, not command-time behavior.
• Non-overlapping local skills should stay available and be exposed through harness mirrors/symlinks.

Evidence:

• dp scaffold setup currently migrates legacy .claude/skills, .codex/skills, .cursor/skills, and .opencode/skills into .agents/skills before writing selected baseline skills.
• Current setup behavior renames collisions to *-local; this preserves an intersecting local override instead of letting the agent decide.
• Current setup writes selected baseline skills and runs .agents/scripts/sync-subagents.mjs, which regenerates harness-native mirrors.
• dp scaffold init currently copies staged skills directly into .agents/skills/<skillId>.

Q2 Question: Who owns overlap detection and reconciliation: the command or the agent following the handoff?

Accepted answer:

• The agent owns overlap detection and reconciliation.
• dp scaffold init, dp scaffold setup, and dp update should not detect overlapping skills in command runtime.
• Command output, .devpunks/AGENT-HANDOFF.md, and $dp-cli guidance must clearly instruct the agent to inspect pre-existing skills.
• The agent checks .agents/skills, .claude/skills, .codex/skills, .cursor/skills, and .opencode/skills against scaffolded HI baseline skill names.
• If exact-name overlap exists, the agent treats HI baseline as precedence and does not keep the local skill active.
• If no exact-name overlap exists, the agent keeps the local skill and ensures harness mirrors/symlinks expose it correctly.
• If archived or removed local skill knowledge should be integrated into HI baseline, the agent suggests dp report.
• Reason: keep the CLI simple and deterministic; make the agent perform repository-specific reconciliation.

Q3 Question: Should dp scaffold init, dp scaffold setup, and dp update share this behavior?

Accepted answer:

• Yes. All three commands share one agent-owned post-command skill reconciliation policy.
• The command does not check overlap.
• The handoff and $dp-cli guidance tell the agent to check the same skill homes and apply the same HI precedence rule.
• Reason: operators should not need separate mental models for first scaffold, repo-aware setup, and baseline refresh.

Superseded Q4 Question: Should the command itself detect exact-name overlap?

Superseded answer:

• Earlier grill answers said command runtime should detect exact-name overlap, archive intersecting local skills under .devpunks/replaced-skills, and report replacements in dp update --check.

Superseding answer:

• No. Commands should not detect overlapping skills.
• Commands should preserve enough post-command context and provide clear guidance.
• Agents following .devpunks and $dp-cli guidance own the check, decision, and dp report suggestion.

Q5 Question: If commands should not detect overlap, should they still snapshot pre-existing skill homes for agent inspection?

Accepted answer:

• Yes. Commands may take a blind snapshot of existing skill homes before writing scaffold/update assets.
• Snapshotting must not classify overlap, decide precedence, or report intersections.
• Suggested snapshot target: .devpunks/pre-existing-skills/.
• Agent uses snapshot plus final scaffolded skills to decide overlap after command finishes.
• Reason: preserve evidence for agent-owned reconciliation without moving judgment back into command runtime.

Q6 Question: When agent finds exact-name overlap, should it both archive local skill evidence and remove active local copies/mirrors?

Accepted answer:

• Yes. Agent archives overlapping local skill evidence under .devpunks/replaced-skills/<skill-id>/....
• Agent removes overlapping local active copies from .agents/skills and harness mirrors.
• Agent keeps HI baseline skill active.
• Agent suggests dp report when archived local skill knowledge should be considered for HI baseline integration.
• Reason: this makes "we don't keep them" true operationally without losing evidence.

Glossary

Glossary Q1 Question: What terms describe this policy?

Accepted answer:

• Canonical term: HI baseline skill. A skill selected from the resolved Harness Intelligence scaffold baseline for the current command.
• Canonical term: Local skill. A pre-existing repo-authored skill found in repo skill homes before or after an HI scaffold/update command.
• Canonical term: Overlapping local skill. A local skill whose canonical directory id/name matches a scaffolded HI baseline skill.
• Canonical term: Non-overlapping local skill. A local skill whose canonical directory id/name does not match scaffolded HI baseline skills.
• Canonical term: Harness mirror. Tool-specific compatibility surface generated from neutral .agents sources, such as .claude, .codex, .cursor, or .opencode.
• Avoid: command reconciliation, local baseline, local override.
• Axiom: HI baseline skills have precedence over overlapping local skills.
• Axiom: Commands do not detect overlapping skills.
• Axiom: Commands may snapshot pre-existing skill homes without classifying them.
• Axiom: Agents following .devpunks and $dp-cli guidance own overlap checks and reconciliation.
• Axiom: Agents inspect .agents/skills, .claude/skills, .codex/skills, .cursor/skills, and .opencode/skills.
• Axiom: Agents archive overlapping local skill evidence and remove active local copies/mirrors.
• Axiom: Dynamic or semantic skill overlap belongs to model-guided dp report work, not command runtime.