Context Sync

Context Sync (kenn-forge)

Keep Kenn Forge's context aligned with durable maintainer decisions and cross-cutting invariants. Commit capture is narrow; audits require explicit scope.

Modes

• --commit: inspect only the intended commit, current conversation, and mapped context areas. Use before every agent-created commit.
• <area>: audit one Area Map entry.
• --changed [--base <ref>]: audit only areas mapped from changes since the supplied base, or the merge base with the repository default branch when omitted.
• --all: run the repository-wide audit, including orphan discovery, recent decision research, and cross-area invariant-guard review.
• --check: run scripts/context-sync --check and stop after structural validation.
• --audit-claims: verify anchored claims only within the selected area or changed areas. Combine with --all only when a repository-wide claim audit was requested.

With no mode or area, show this list and ask for explicit scope. Never infer --all from a general request to improve or review context.

Commit Mode

Use commit mode immediately before the repository's normal commit skill. Do not stage files, construct a commit message, or create the commit.

1. Run scripts/context-sync --check before semantic inspection. Repair clear structural drift and rerun the check; ask the maintainer only when the correct routing cannot be derived from repository evidence.
2. Inspect git status --short, the intended diff against HEAD, relevant untracked files, and the current conversation. Ignore unrelated user changes.
3. Look only for durable signals introduced or clarified by this work:
   • a maintainer correction or explicit design decision;
   • a changed cross-cutting invariant, workflow, or policy;
   • a gotcha that existing context should have prevented.
4. Map each real signal to the smallest Area Map scope. Read only that area's topic docs, the governing CLAUDE.md section, and context-guide.md.
5. Apply the guide's grep test and per-addition budget.
6. Choose one outcome:
   • Continue: no durable signal exists, or existing context already captures it. Proceed silently without a marker or no-update report.
   • Updated: the required addition, correction, or removal is clear. Apply the smallest context edit, then return control to the commit workflow.
   • Needs maintainer input: materially different durable rules remain plausible after inspecting the evidence. Stop and ask one focused question.

Do not dispatch subagents, scan history, validate unrelated anchors, or audit invariant guards in commit mode. Wording, routing, document placement, and clear factual deletions are agent-resolvable work, not reasons to block. Block only for missing maintainer knowledge that changes what future agents should do.

Area Map

Mode, notification, and inline-review scopes intentionally overlap generic areas. For --changed, select every matching focused and generic area rather than choosing only the broadest or most specific row.

Audit Workflow

Follow every step in order for <area>, --changed, and --all. Do not collapse the workflow into a structural check or a general code review.

Step 0: Select the Scope

• <area> selects exactly one Area Map row.
• --changed [--base <ref>] resolves the merge base, maps changed files to candidate areas, then drops candidates with no durable context signal.
• --all selects every area and enables the explicitly broad work called out below.

When --base is omitted, use the merge base with the repository default branch. If --changed leaves no selected area, report that no applicable audit scope remains and stop. Never add areas merely because their code is adjacent.

Step 1: Load the Guide

Read context-guide.md completely. Use its grep test, anchored-claim format, sorting test, size limits, invariant-guard rules, and context-poisoning safeguards throughout the audit. Do not propose or apply context changes before loading it.

Step 2: Validate Routing

For --all, validate the full context manifest:

• every context/*.md and routed docs/**/*.md reference in CLAUDE.md exists;
• every context document an agent needs is reachable from CLAUDE.md or another reachable document;
• AGENTS.md resolves to CLAUDE.md.

For <area> and --changed, validate only the selected topic docs and their routing references. Report broken or unreachable routing before continuing; unreachable context is a high-priority finding.

For --check, run scripts/context-sync --check, report its result, and stop. The script validates structure only; it is not a semantic audit or commit decision.

Step 3: Scan the Selected Areas

Dispatch one read-only subagent per selected area when subagents are available. Give each agent:

• the selected topic docs and governing CLAUDE.md sections;
• the mapped code paths from the Area Map;
• the anchored-claim and four-tag rules from context-guide.md.

Collect the same artifacts from every area:

• a proposed diff for the topic doc or governing CLAUDE.md section;
• one paragraph describing what drifted and why;
• knowledge-gap questions the code cannot answer;
• per-anchor results: resolves, moved, or gone;
• documented invariants that lack an appropriate Go guard or analyzer.

Do not dispatch area subagents in --commit mode. For focused audits, do not give agents unselected docs or code paths.

Step 4: Check Design Decisions

Compare the selected code and current conversation with docs/adr/ and the selected topic docs. Distill feature specs and implementation plans into current contracts in a topic doc or CLAUDE.md; never preserve those transient artifacts by converting them into ADRs. Reserve an ADR for independently required dated rationale that still binds the architecture.

Only --all may search broader recent conversation history and recent commits. Focused audits use the current conversation and selected change range only.

Step 5: Research, Suggest, Then Ask

For every knowledge gap:

• research current best practice first when the question is general and technical;
• ask the maintainer directly when the answer is Kenn Forge-specific domain knowledge;
• present a confidence-tagged recommendation before asking for confirmation.

Do not ask a bare question that safe local inspection or focused research can answer.

Step 6: Check Invariant Guards

For each invariant asserted by a selected context doc, confirm an appropriate Go test or custom analyzer protects it. Examples include provider identity, capability gating, UTC datetimes, stable error codes, Huma-only routing, migration history, and testify-only assertions.

Flag documented-but-unguarded invariants and propose the smallest useful guard. When a guard and a doc disagree, treat it as a high-priority finding. Only --all may review guards outside the selected areas.

Step 7: Present or Apply Diffs

Compress every addition to the guide's per-addition budget before presenting or applying it. State the constraint future agents must respect, not the implementation walkthrough.

For <area> and --changed, apply clear scoped additions, factual corrections, moved anchors, and removal of claims whose subjects are gone. Ask the maintainer only when the correct durable rule cannot be derived from evidence.

For --all, present all proposed changes together. Separate deletions and modifications from additions, justify each removal, and wait for approval before broad changes that reinterpret policy.

Step 8: Route Changes and Return to the Commit Workflow

When a new topic doc is created, route it from CLAUDE.md or another reachable context doc. Keep AGENTS.md resolving to CLAUDE.md, keep the canonical provider list in its single existing home, and add skill symlinks under both .agents/skills/ and .claude/skills/ only when a new repository skill is created.

After context edits are ready, return to the repository's normal verification and commit workflow. Do not stage, commit, amend, or push from this skill.

Claim Audit

When --audit-claims is requested, read claim-verifier.md completely and add the four-tag verification to Steps 3 and 7. Tag each anchored claim in the selected scope as VERIFIED, OUTDATED, GONE, or UNVERIFIABLE. Refresh moved anchors, describe behavioral deltas, and identify important unguarded invariants. Never widen claim auditing beyond the selected scope unless --all was requested.