Files
orca/.ciagent/RESEARCH_v0.7.md
Jon Chery 7c4b603811 docs(P00): v0.7 research findings + persona assessment
Codebase audit: cert command unreachable, no config parser, low coverage
(engine 8.3%, transport 26.3%, proxmox 5.1%, audit 0%), pprof deferred.
5 architectural decisions (AD-022..AD-026). Zero new deps.

---ci---
project: orca
phase: 0
milestone: v0.7
status: research
---/ci---
2026-08-03 20:29:07 +00:00

161 lines
8.3 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Research: Orca v0.7 — Hardening & Completion
## 1. Codebase audit findings (RESEARCH stage)
A full codebase audit surfaced the gaps that define the v0.7 scope.
Each finding is grounded in a specific file/coverage measurement.
### 1.1 `orca cert` command tree is unreachable (critical)
- `internal/cli/cert.go:44` exports `NewCommand(log *slog.Logger)
*cobra.Command` which builds the full `cert ca-init | gen | show |
renew | fingerprint` tree (5 subcommands, all implemented, all
spec-compliant per REQ-033/035/036).
- **No file in the repo calls `NewCommand` or registers it on
`rootCmd`.** `grep -rn "rootCmd.AddCommand" internal/cli/` lists
daemon, init, audit, version, job, node, doctor, status — `cert` is
absent. `./bin/orca cert` returns `error: unknown command "cert"`.
- The function is named `NewCommand` (not `newCertCmd`), so it is not
picked up by any init-based registration convention.
- **Impact**: every cert operation the spec promises (REQ-023, REQ-025,
REQ-033, REQ-035, REQ-036) is unreachable from the CLI. Operators
cannot bootstrap a CA, issue a server cert, or rotate one without
hand-crafting calls into the `security` package. This is the single
highest-impact bug in the v0.1v0.6 line.
- **Fix**: one-line `rootCmd.AddCommand(NewCommand(log))` in
`internal/cli/cert.go` (or a new `init()`), plus a regression test
that asserts `rootCmd.Commands()` contains a child whose `Use ==
"cert"`.
### 1.2 `internal/store/cert_repo.go` has no test file
- `internal/store/cert_repo.go` exists (the `certs` table from
migration 0004) but `internal/store/cert_repo_test.go` does not.
- Every other repo in `internal/store/` has a `_test.go`:
`node_repo_test.go`, `job_task_repo_test.go`, `capacity_repo_test.go`,
`audit_repo_test.go`, `migrate_test.go`.
- **Fix**: add `cert_repo_test.go` covering Insert/Get/List/rotation
history (N=3 per REQ-025) + serial_hex uniqueness.
### 1.3 Low test coverage in core packages
| Package | Coverage | Missing tests for |
|---------|----------|-------------------|
| `internal/engine` | 8.3% | `executor.go`, `dispatcher.go`, `peer.go` (only `scheduler_test.go` exists) |
| `internal/transport` | 26.3% | `mtls.go`, `dispatch.go`, `handshake_log.go` (only `idempotency_test.go` exists) |
| `internal/proxmox` | 5.1% | `bootstrap.go` SSH path (only `bootstrap_test.go` exists, exercises the no-op dry-run) |
| `internal/audit` | no test files | `audit.go` (Emit, EmitWithErr, LogHandshake*) |
- Target per D-042: 50% floor per package, 70% for new code in P02/P04.
- Strategy: table-driven tests + `httptest.NewTLSServer` for transport;
interface-based mocks for the SSH dialer (already an interface in
`proxmox/bootstrap.go:211` `defaultSSHDialer` with `DialContext`).
### 1.4 No HCL config file parser
- D-009 specified `~/.orca/config.hcl` and `/etc/orca/orca.hcl` as
config locations. `find . -name "*.hcl"` returns only testdata
(`testdata/hello.hcl`, `testdata/fail.hcl`) used by jobspec tests.
- The CLI relies entirely on flags + env vars (`ORCA_HOME`,
`ORCA_DB`, `ORCA_PROXMOX_PASSWORD`). There is no `internal/config`
package.
- `internal/jobspec/spec.go:40` already uses
`hclsimple.Decode(filename, data, nil, &spec)` — the exact same
pattern works for a `Config` struct. No new dep required (hashicorp/hcl/v2
is already a direct dep).
- **Fix**: new `internal/config` package with a `Config` struct (HCL
tags: `db_path`, `listen_addr`, `ca_path`, `server_cert_path`,
`server_key_path`, `node_capacity`), a `Load(paths ...string)`
function, and a `--config` flag on the root command. Precedence per
D-039: flag > env > file > default.
### 1.5 pprof endpoint (I-308, deferred since v0.2)
- I-308 was deferred in v0.2 IDEATE ("keep v0.2 lean") and never
revisited. The daemon (`internal/daemon/server.go`) has no pprof
surface today.
- `net/http/pprof` is stdlib — zero new deps. Mount on a separate
`*http.ServeMux` so it never touches the mTLS daemon listener.
- **Fix**: `--pprof <addr>` flag on `orca daemon` (default disabled).
If set, start a second `http.Server` on `<addr>` with
`pprof.Index`/`pprof.Cmdline`/etc. registered. Log a WARN that the
endpoint is unauthenticated + operator-only.
## 2. Prior art & patterns
### 2.1 HCL config in HashiCorp tools
Nomad, Consul, and Terraform all use HCL for config with the same
`hclsimple.Decode` + struct-tag pattern. The precedence model (flag >
env > file > default) is the de-facto standard; Viper implements it but
adds a large dep. Orca's `internal/config` will implement the 4-layer
merge by hand (~80 LOC) to stay minimal-deps.
### 2.2 pprof in Go daemons
Standard pattern: `import _ "net/http/pprof"` registers handlers on
`http.DefaultServeMux`. Best practice for production daemons is a
**separate listener** (not DefaultServeMux) so pprof is never exposed
on the public port. Orca will use a dedicated `*http.ServeMux` +
`http.Server` on the `--pprof` addr, default disabled.
### 2.3 Test coverage for concurrent Go
`internal/engine` (executor, dispatcher) and `internal/transport`
(mtls, dispatch) are concurrent. Coverage strategy:
- `httptest.NewTLSServer` for transport — exercise real TLS handshakes
against an in-process server.
- Interface-based mocks for the SSH dialer (proxmox) and the peer
client (transport) — both already have interface seams.
- `sync.WaitGroup` + channel assertions for executor/dispatcher
lifecycle.
- `-race` is already on in CI (REQ-031) — new tests inherit it.
## 3. v0.7 Architectural Decisions (AD-022..AD-026)
| ID | Decision | Rationale |
|----|----------|-----------|
| AD-022 | `orca cert` registered via `init()` in `cert.go` calling `rootCmd.AddCommand(NewCommand(slog.Default()))` | Keeps registration co-located with the command definition; matches the pattern in `daemon.go`/`audit.go` where each command file self-registers. Avoids a central registration function that would drift. |
| AD-023 | `internal/config` package: `Config` struct + `Load(paths ...string) (*Config, error)`; no global singleton | Config is passed explicitly to `daemon.NewServer`, `cli` commands, etc. No package-level state — testable, no init-order surprises. |
| AD-024 | pprof on a separate `*http.Server` + `*http.ServeMux`, default disabled | Never co-mingles with the mTLS daemon listener. Operator opts in via `--pprof :6060`. Matches Go daemon best practice. |
| AD-025 | Coverage floor measured per-package via `go test -cover ./<pkg>` | No aggregate threshold (aggregates hide low-coverage packages). CI gate added in P03: `go test -cover ./internal/engine ./internal/transport ./internal/proxmox ./internal/audit` and assert each ≥ 50%. |
| AD-026 | No new direct dependencies in v0.7 | `net/http/pprof` (stdlib), `hashicorp/hcl/v2` (already direct). v0.7 preserves the minimal-deps ethos. |
## 4. PERSONAS assessment
v0.7 is an NFR milestone touching CLI, config, tests, and daemon. The
default 3-persona roster (lead-developer, backend-engineer,
data-engineer) is sufficient:
- **lead-developer**: owns P01 (cert registration) + P04 (pprof) — CLI/
daemon territory.
- **backend-engineer**: owns P02 (config package) — internal/config +
CLI integration.
- **data-engineer**: owns P01 cert_repo tests + P03 store coverage —
`internal/store` territory.
- **lead-developer** also owns P03 engine/transport/proxmox/audit
coverage (test-only phase, no schema changes).
No new personas needed. No phase-specific personas. Territory
enforcement stays `warn`. See `.ciagent/PERSONAS.md` (updated).
## 5. Dependencies
v0.7 adds **zero** new direct dependencies:
- HCL parsing: `hashicorp/hcl/v2` (already direct, used by jobspec).
- pprof: `net/http/pprof` (stdlib).
- Tests: `net/http/httptest` (stdlib), existing interfaces.
`go.mod` is unchanged by v0.7.
## 6. Risks
- **P01 cert registration** may surface latent bugs in the cert
subcommands (they've never been exercised end-to-end). Mitigation:
P01 includes a smoke test that runs `cert ca-init` + `cert gen` +
`cert show` + `cert fingerprint` against a temp `ORCA_HOME`.
- **P02 config precedence** is easy to get wrong (flag/env/file/default
merge order). Mitigation: table-driven test covering all 4 layers.
- **P03 coverage** on concurrent packages may reveal race conditions
(already hidden by the 8.3% coverage). Mitigation: `-race` is on; P03
fixes any races it uncovers as part of the same phase.