Verify

/verify — orchestrated IDE-local code-review + build gate

You are the main agent. You orchestrate all sub-agent dispatches. Sub-agents under .claude/agents/ cannot dispatch further sub-agents (Claude Code architectural constraint). All parallel fan-out happens from your context.

This command is the single verification gate for any finished piece of work in this repo — a module addition, a bug fix, a scratch experiment. It diffs the working branch against a base ref and runs a tight, high-signal review grounded in this repo's AGENTS.md files. Only issues that any contributor who knows AGENTS.md would say "yes, that has to be fixed before this is pushed" reach the final report.

This is heavier than the "don't run these after every edit" guidance in root AGENTS.md — that guidance is about not gating local iteration. /verify is what you run once, deliberately, before opening a PR, in place of running the gates by hand.

Arguments

The user invoked: /verify $ARGUMENTS

$ARGUMENTS is optional. Recognized flags (any order, both optional):

• --base=<ref> — git ref to diff against. Default: develop (this repo's main branch).
• --full — also run the whole-app check (spotlessCheck detektDebug :build-logic:convention:detektMain lintDebug :app:assembleDebug) in Stage 5, matching what CI gates on. Without it, Stage 5 only runs spotlessApply plus a compile of the touched modules — fast, matching the "iterate on a single-module compile" guidance in root AGENTS.md.

Any other token, any malformed flag, or any unrecognized form → emit INVALID_INVOCATION (Stage 8 format) with the offending token in Notes and stop.

Sub-agent registry

Dispatch them with subagent_type: "<name>" on the Agent tool.

---

Stage 0 — Parse arguments

1. Tokenize $ARGUMENTS. Empty → defaults (base=develop, full=false).
2. For each token, match --base=<ref> or --full. Anything else → INVALID_INVOCATION (Stage 8 format) and stop.

Stage 1 — Skip check

Compute changed_files = git diff <base>...HEAD --name-only. If empty, emit SKIPPED (Stage 8 format) with reason "no diff vs <base>" and stop.

Otherwise dispatch one skip-checker Agent call. Pass diff_command (e.g. git diff <base>...HEAD) and changed_files (newline-separated).

If the agent returns {"skip": true, ...}, jump straight to Stage 8 with verdict SKIPPED and the agent's reason in Notes. Do not proceed past Stage 1.

Stage 2 — Context discovery (orchestrator, sequential reads)

Read into your context (these are ground truth for downstream leaves). Collect their absolute paths in a discovered_files list:

• Root AGENTS.md (always).
• docs/engineering/coding-standards.md and docs/engineering/architecture.md (always — these expand the root AGENTS.md rules with the worked examples reviewers can quote).
• For every file in changed_files, the nearest module-scoped AGENTS.md: walk up from the changed file's directory until you find an AGENTS.md other than the root one (same algorithm validateAgentContext uses to bind a Gradle module to its AGENTS.md). Deduplicate.

Also record, from the same scan:

• changed_modules — the set of Gradle module directories touched (walk up from each changed .kt/.kts file to the nearest directory containing a build.gradle.kts).
• context_files_touched — true if any changed file is an AGENTS.md, a CLAUDE.md, anything under .claude/skills/, .github/copilot-instructions.md, anything under build-logic/, settings.gradle.kts, or a library/plugin/bundle entry (not just a version bump) in gradle/libs.versions.toml.

Stage 3 — Diff summary (orchestrator, no sub-agent)

Run git diff <base>...HEAD <files> and produce a 5–10 line factual summary of what changed. Keep it short — it's an aid for the reviewers, not user output.

Stage 4 — Parallel review fan-out (SINGLE MESSAGE, MANY AGENT CALLS)

In one assistant turn, dispatch the following in parallel — separate Agent tool calls within the same message:

1. bug-reviewer × 1. Payload: diff, diff_summary, discovered_files.
2. convention-reviewer × 1. Payload: same.
3. security-reviewer × 1. Payload: same.

If any *Test.kt file is in the diff (an explicitly-requested exception to the "no test infrastructure" rule), it's part of diff and is reviewed by all three under the same rules as production code.

Hard rule: parallel Agent calls in one assistant turn. Sequential dispatch is a bug.

Collect every JSON response. Build a single findings list of issues (each tagged with its origin category).

Stage 5 — Local build gates

1. Always: ./gradlew spotlessApply — the one gate root AGENTS.md requires before every commit. If it rewrites files, note which ones in Notes; those rewrites are not separately re-reviewed by Stage 4.
2. Always, per module in changed_modules: ./gradlew <module>:compileDebugKotlin (use <module>:compileKotlin for a module with no Android debug variant).
3. Conditional — context_files_touched is true OR --full was passed: ./gradlew validateAgentContext.
4. Conditional — --full was passed: the whole-app check — ./gradlew spotlessCheck detektDebug :build-logic:convention:detektMain lintDebug :app:assembleDebug.

Capture PASS / FAIL and the full failing-task output verbatim for every task actually run.