A1 — Parallel Repository Analysis

Principal-architect, multi-agent analysis of a large Paytm Money Android monorepo (android-monorepo: a Kotlin equity-trading SDK + a centralized Room database module + a Flutter add-to-app client). Six independent specialist agents fan out across the codebase, a coordinator cross-verifies their claims against source, and the findings are consolidated into a single master report.

Posture: every claim cites a real file:line or command output and is labeled VERIFIED (read/grepped) · INFERRED (naming/convention) · UNVERIFIED (not confirmed). Independence between lanes is what makes the Phase-3 cross-verification meaningful.

Quickstart

cd Advanced/parallel-repo-analysis

# 1) Validate every committed deliverable (no target repo needed):
bash scripts/validate_a1_reports.sh            # exits 0 when all checks pass

# 2) Same, via the orchestrator entrypoint:
bash run_a1_analysis.sh --validate-only

# 3) (Optional) capture live grep evidence from the analyzed repo:
cp .env.example .env && $EDITOR .env           # set TARGET_REPO=/path/to/android-monorepo
bash run_a1_analysis.sh                         # capture evidence -> artifacts/repro, then validate

# From the monorepo root, the same gate is wired into make:
make a1-validate

Results at a glance

Metric	Value	Status
Files scanned	~6,629 (4,279 Kotlin equity src + 2,350 Dart)	VERIFIED (find)
Gradle modules	36 unique :module refs (44 include lines)	VERIFIED
Outbound endpoints	78 Retrofit interfaces (equity_sdk) + 1 (base_app); ~363 methods	VERIFIED (78)
Room entities	27 tables / 2 DBs (24 EquityDatabase v19 + 3 LoggingDataBase v7); 0 FKs	VERIFIED
Tests discovered	520 *Test.kt (equity) + 212 *_test.dart; 1,684 repo-wide	VERIFIED (counts)
Contradictions resolved	2 (Retrofit 71→78; module-count reconciled) + 1 risk reworded	Coordinator + independent agent
Lanes	6 core + 2 optional (security, performance)	—

Top finding: a CI test-coverage gap — Bitbucket runs unit tests for :base_app only (bitbucket-pipelines.yml:713), so equity_sdk's ~303 and Flutter's ~212 tests never execute on that pipeline (GitLab scope flagged UNVERIFIED). See the master report's Risks section.

Deliverable tree

parallel-repo-analysis/
├── README.md                          ← this file
├── run_a1_analysis.sh                 ← orchestrator: --validate-only / evidence capture / --help
├── .env.example                       ← TARGET_REPO template
├── scripts/
│   ├── validate_a1_reports.sh         ← structural + content gate (CI entrypoint)
│   ├── capture_evidence.sh            ← grep-count evidence from TARGET_REPO (Phase 3)
│   └── generate_api_inventory.sh      ← Retrofit/route inventory -> A1_api_inventory.yaml
├── prompts/                           ← the agent + coordinator prompts (reproducibility)
│   └── remediation/                   ← PROMPT-023..032: target-repo fix specs
├── artifacts/repro/                   ← captured evidence + run logs (reproduce headline counts)
└── docs/
    ├── A1_engineering_evaluation_audit.md   ← category scores + 32-issue register
    ├── A1_remediation_tracker.md            ← AUDIT-001..032 with Status
    └── agent-analysis/
        ├── A1_plan.md                       (Phase 1 — decomposition)
        ├── A1_inventory.md                  (Agent 1 — modules/artifacts)
        ├── A1_api_map.md                    (Agent 2 — outbound API surface)
        ├── A1_entities.md                   (Agent 3 — Room data model)
        ├── A1_tests.md                      (Agent 4 — test discovery)
        ├── A1_architecture.md               (Agent 5 — architecture/deps)
        ├── A1_flow_trace.md                 (Agent 6 — end-to-end flow)
        ├── A1_security.md                   (Lane 7 — optional)
        ├── A1_performance.md                (Lane 8 — optional)
        ├── A1_dependencies.md               (dependency/version table)
        ├── A1_verification_report.md        (Phase 3 — cross-verification)
        ├── A1_repository_master_report.md   (Phase 4 — consolidated)
        ├── A1_manifest.json                 (machine-readable metrics)
        └── A1_provenance.yaml               (commit/branch/date)

How the analysis is structured

Phase	What happens	Output
1 — Plan	Decompose into 6 independent lanes; define scope, evidence rules	A1_plan.md
2 — Fan-out	6 specialists work alone (no cross-reading), each writes one report	6 lane reports
3 — Cross-verify	Coordinator diffs overlapping claims, resolves contradictions, re-greps headline claims; a blind adversarial agent confirms/refutes	A1_verification_report.md
4 — Consolidate	Merge in order (Inventory→Architecture→Entities→API→Flow→Tests), attach metrics, risks, recommendations	A1_repository_master_report.md

The lane decomposition, merge order, and conflict policy are defined in skills/tasks-parallel-repo-analysis/SKILL.md.

Acceptance criteria

• scripts/validate_a1_reports.sh exits 0 (9 files present; master has Metrics + all completion criteria checked; verification has contradictions + spot-checks; no absolute paths; no stale "25 entities"; every lane has a VERIFIED label + Agent header).
• run_a1_analysis.sh --validate-only exits 0.
• make a1-validate exits 0 from the monorepo root.
• Every headline metric is reproducible from artifacts/repro/ (or documented as blocked).
• CI (.github/workflows/a1-parallel-repo-analysis.yml) runs the gate on path-filtered pushes/PRs.

Notes & honest limitations

• This is a client app: there are no server endpoints — the entire API surface is outbound.
• Test execution was not run (Android SDK + Flutter build blocker); counts come from file reads/find, never fabricated pass/fail.
• TARGET_REPO is intentionally not committed. Report headers reference $TARGET_REPO (android-monorepo) instead of an absolute path. Validation works without it; live evidence capture requires it.