emcli/specifications/PHASE1-STATUS.md

emcli — Phase 1 Status Report

Date: 2026-06-22 Branch: main (pushed to gitea.dcglab.co.uk/steve/emcli, HEAD 6061bd2) Phase 1 scope: Foundation + read path (encrypted config, IMAP read, agent commands, password auth)

TL;DR

Phase 1 is complete and merged to main. All 14 planned tasks were implemented test-first, each reviewed (spec compliance + code quality) by a separate agent, with fixes applied and re-verified. The binary builds as a single static CGO-free executable, go vet is clean, and the full unit-test suite passes. An end-to-end smoke test of the real binary confirms the agent (JSON) and admin (human-readable) surfaces behave per spec.

go build (CGO_ENABLED=0)  → OK, single static binary
go vet ./...              → clean
go test ./...             → all packages pass

What was built

Package	Responsibility	Status
internal/version	Build version string + version command	✅
internal/crypto	AES-256-GCM field encryption keyed from EMCLI_KEY (fail-closed)	✅
internal/store	Pure-Go SQLite: schema v1, accounts CRUD (encrypted secrets), whitelists, seen-set read state (floor + acked + compaction + UIDVALIDITY reset), audit log + retention purge	✅
internal/policy	Pure enforcement: case-insensitive address/domain matching, inbound whitelist + subject-regex filter	✅
internal/mail	RFC822 parsing (headers/body/attachments) + IMAP client (select/fetch/search, read-only, Peek)	✅
internal/cli	JSON envelope, agent commands list/get/search/ack (policy-gated), flag-based admin (account, whitelist), command router	✅
cmd/emcli	Entry point routing to cli.Run	✅

Working commands (verified via smoke test)

• Agent (JSON-only output, exit code mirrors error):
  • emcli list --account <a> [--new] [--before/--since <uid>] [--limit N]
  • emcli get --account <a> --uid <uid>
  • emcli search --account <a> [--from --subject-contains --text --since-date --before-date]
  • emcli ack --account <a> --uid-list 1,2,3
• Admin (human-readable):
  • emcli account add|list, emcli whitelist in|out add|remove|list
  • emcli version

Security invariant — independently verified in the final review

Mail that fails the inbound policy (whitelist_in / subject_regex) is invisible on every path: excluded from list/search, returns generic not_found on get (body never fetched), and cannot be acked (no state manipulation for unseen mail). Filtered vs. genuinely-absent UIDs are indistinguishable to the agent. Credentials/secrets never appear in stdout, the JSON envelope, or the audit log; passwords are sealed at rest (ciphertext ≠ plaintext confirmed by test).

Process

Executed via subagent-driven development: a fresh implementer per task (TDD, commit per task), a separate reviewer per task (spec + quality gates), fix sub-agents for Critical/Important findings, and a final whole-branch review on the most capable model. Durable progress tracked in .superpowers/sdd/progress.md. 20 commits on main (04d3b61..6061bd2).

Fixes applied during review (all re-verified):

• store: pinned SQLite pool to one connection so PRAGMA foreign_keys (cascade deletes) stays effective; added a cascade test.
• mail: propagate io.ReadAll errors when parsing message parts and reading bodies off the network (prevents silent truncation); apply the search limit.
• cli: AckCmd reuses the folder UID-validity from setup (was swallowing a second SelectFolder error → epoch-0 acks); agent commands now exit non-zero when the JSON envelope reports an error (spec §8); search limit now counts visible results (filter before cap), consistent with list.

Known limitations / deferred (not defects)

Deliberately out of scope for Phase 1 (have their own future plans)

• Phase 2 — Send path: SMTP, MIME building, reply threading, outbound policy (RO/RW + whitelist-out), send.
• Phase 3 — OAuth2: Gmail loopback consent + token refresh (schema columns already present).
• Phase 4 — Admin TUI + doctor: interactive init/reconfigure, connectivity diagnostics.

Carry-over items to address (logged from reviews)

1. [Important, Phase 2 priority] Header listing fetches full message bodies. mail.fetchByUIDSet downloads each whole RFC822 message even for list/search, then discards the body. Functionally correct but slow/heavy on real mailboxes. Fix = request a header-only body section / ENVELOPE + BODYSTRUCTURE when full=false. Deliberately not fixed yet because the change to the IMAP fetch path could not be validated against a live server in this environment (see below); it should be done with a real/TLS-configured IMAP server in front of it. Phase 2's --reply-to should build that header-only fetch path too.
2. [Minor] Audit completeness: list/search don't write a per-filtered-message audit row; some IMAP-error envelope paths don't write an audit row; imap_error is an audit reason not enumerated in the spec.
3. [Minor] CLI polish: account add opens the store before parsing its own flags (--help touches the DB); invalid --since-date/--before-date are silently ignored rather than erroring; whitelist add/remove don't reject an empty --address at the CLI layer.
4. [Minor] Performance: ack fetches each UID's header in its own round-trip; get fetches the header then re-fetches the full message. Fine for a per-invocation CLI; batch ops in later phases may want to optimize.

Validation gap

• Live IMAP integration was not exercised. The internal/mail integration test (build tag integration) compiles and self-skips without a server. An attempt to validate against a GreenMail container failed on TLS/port friction: GreenMail 2.1.0's IMAP STARTTLS didn't negotiate here and the IMAPS port wasn't exposed, while our client (correctly) requires verified TLS and the test config has no skip-verify option. The IMAP code was reviewed by inspection and compiles under the integration tag, but has not been run end-to-end against a real server. Recommended follow-up: add a dev-only --insecure/skip-verify option (or point the integration test at a properly-certificated server) so the read path can be validated against live IMAP before relying on it operationally.

Suggested next steps

1. Validate the read path against a real IMAP account (e.g. your own mailbox over TLS) — this is the highest-value confidence step and would also surface the full-body-fetch performance issue (#1) in practice.
2. Decide whether to fix the header-only fetch (#1) before or as part of Phase 2.
3. Proceed to Phase 2 (send path) — the structure is ready: policy is set up to host outbound checks, Account.Mode is loaded for RO rejection, and MatchAddress already implements the domain matching whitelist-out needs.

Repository layout

emcli/
├── Makefile                 # build/test/vet (CGO_ENABLED=0)
├── go.mod / go.sum
├── cmd/emcli/main.go
├── internal/
│   ├── version/
│   ├── crypto/
│   ├── store/               # store, schema, settings, account, whitelist, seenset, audit
│   ├── policy/              # policy (address), inbound
│   ├── mail/                # message, imap, testdata/*.eml
│   └── cli/                 # envelope, agent, dispatch, run, admin
└── specifications/
    ├── PRD.md
    ├── SPEC.md
    ├── PHASE1-STATUS.md     # this file
    └── plans/2026-06-21-phase1-foundation-and-read-path.md