Upgrade

When to invoke

Invoke when the user says "upgrade goc", "pull the latest goc", "sync goc templates", "run goc upgrade", or asks how to upgrade the GoC version in this repo. The engine's safety guarantee — never overwrite authored project state — holds with or without this skill; the skill exists to reconcile evolving upstream content into the local copy where the engine's "preserve" verdict alone leaves real upstream changes on the floor.

Upgrade GoC

goc upgrade is deterministic and safe by construction — it never overwrites authored content under .game-of-cards/. That guarantee holds for every consumer regardless of whether an agent is present (headless CI runs, scheduled cron, scripted upgrades).

But "never overwrite" is not the same as "stay current". Two files under .game-of-cards/ ship real, evolving content that the upstream template may have changed across versions:

• README.md — the hook-point catalogue documenting which content stubs and workflow hooks the skills consume, and where.
• config.yaml — the runtime-neutral closure attestation config (layer_2_project_dod, layer_3_goc_dod) plus workflow options.

If the engine has preserved a diverged copy of either, the engine itself has nothing to say about whether real upstream changes need to land in the consumer's local copy. That is this skill's job.

For the 12 user-owned content stubs and workflow hooks, the engine's "preserved" verdict is the whole story — the templates ship blank by design, so there is nothing to reconcile. This skill confirms that to the user so they know nothing was lost.

For the AGENTS.md / CLAUDE.md GoC marker block, the engine regenerates the goc-owned region wholesale via the marker-bounded merge (<!-- BEGIN GOC vX.Y.Z --> … <!-- END GOC -->). That region is goc-owned, not user-owned, so nothing to reconcile — this skill summarizes what methodology guidance changed across the upgrade so the user knows what new behavior the agents will now follow.

Workflow

1. Run goc upgrade and capture stdout. The engine emits a sentinel-marked JSON divergence report after its normal upgrade output:

   GoC project-state divergence report (JSON):
   {"version": 1, "templates_root": "...", "files": [...]}
   

   Parse the JSON on the line immediately following the marker. Each entry has path (relative to .game-of-cards/), status (create / unchanged / preserved), and ownership (user-owned / evolving).

2. For each evolving file with status: preserved — read the local file at .game-of-cards/<path> AND the shipped template at <templates_root>/<path>. Drive a 2-way reconcile:

   • Diff the two files mentally. Identify what changed upstream (lines present in the template that are absent or different in the local copy).
   • For each upstream change, decide whether it should land:
     • Mechanical updates (new key with default value, new comment, new doc paragraph that does not contradict local customization) → propose the merged content and AskUserQuestion for confirmation.
     • Conflicting with the user's customization → describe both sides and ask the user how to resolve. Default to keeping the user's value; surface the upstream change as informational.
   • When the user approves, write the merged content back to .game-of-cards/<path> directly. Do not invoke goc upgrade again — the engine has already done its work; this is a one-shot reconciliation.
   • If the local file and template are obviously compatible (e.g., local added a new top-level key the template doesn't know about, and the template added an unrelated comment), apply both sides and report the merged result.

3. For each user-owned file with status: preserved — confirm to the user, in one line per file: "kept your <path> (nothing upstream)". The shipped template is permanently blank by design, so there is no upstream content to bring across. The confirmation exists so the user does not have to wonder whether the upgrade silently overwrote anything.

4. For files with status: create — these are absent locally and were scaffolded by the engine. Mention them so the user knows new files appeared (e.g., a new workflow-hook stub for a skill that landed in this version).

5. For files with status: unchanged — silent. The local copy already matches the template; nothing to report.

6. Summarize AGENTS.md / CLAUDE.md GoC block changes. Diff the <!-- BEGIN GOC vX.Y.Z --> … <!-- END GOC --> region before and after the upgrade (the engine already rewrote it; git diff on the working tree surfaces the delta), and present a short list of substantive methodology changes — new verbs documented, new conventions enforced, behavioral defaults flipped. This is informational, not a merge — the marker-bounded region is goc-owned and the upgrade contract is that it gets regenerated.

7. Commit follows repo policy. If the reconciliation pass modified any files, the user typically wants those staged alongside the engine's own changes (the bumped .goc-version, the rewritten methodology block, any newly-scaffolded files). The engine's goc upgrade does NOT auto-commit; commit the union of engine writes + reconciliation edits with a subject like chore(goc): upgrade to vX.Y.Z.

Why the engine is safe without this skill

The engine's _sync_game_of_cards_config classifies every shipped file against the on-disk copy and only writes when the destination is absent. Diverged files are preserved unconditionally — no prompt, no confirmation, no agent in the loop. This means:

• A scripted upgrade in CI does not lose authored content.
• A scheduled cron goc upgrade does not lose authored content.
• A user running goc upgrade in a terminal without invoking this skill does not lose authored content.

What the user loses without this skill is informational reconciliation of upstream changes to the two evolving files. The local copies stay exactly as the user left them; the upstream changes sit in the template payload until the next time someone runs this skill (or manually diffs .game-of-cards/README.md against <templates_root>/README.md).

What this skill does NOT do

• Does NOT re-run goc upgrade. The engine has already done its work by the time this skill is reading the report. Re-running would be a no-op (version is now current).
• Does NOT overwrite user-owned content stubs or workflow hooks. The templates are blank by design — there is nothing upstream to bring across, so the reconcile pass for those files is purely a confirmation.
• Does NOT touch the AGENTS.md / CLAUDE.md GoC marker block. The engine regenerates that region wholesale on every upgrade; the contract is "do not edit between the markers."
• Does NOT bisect or git revert the engine's writes. If the reconciliation pass produces a result the user dislikes, the user edits the file by hand or invokes the create-card skill to file the disagreement as a card for follow-up.

Cross-references

• the kickoff skill — first-time setup in a fresh repo. The kickoff skill writes the initial .game-of-cards/ scaffold; this skill is the follow-up that runs every time the user wants to pull a new goc version into an existing repo.
• the card-schema skill — frontmatter and DoD conventions the upgraded engine may have evolved.
• the deck skill — the methodology surface the regenerated AGENTS.md / CLAUDE.md block documents.