Run the Connector Plans with a VP Eng and Subagents
One VP Eng holds the spec and the ledger; engineer subagents each take a plan from failing tests to green, changelog, and push.
- Path
- loops/Run-the-Connector-Plans-With-a-VP-Eng-and-Subagents.md
- Authors
- Michael Staton
- Augmented with
- Claude Code on Claude Opus 5.5 (1M context)
- Tags
- Loop · MemoPop · Claude-Connector · TDD · Subagents
Run the Connector Plans with a VP Eng and Subagents
context-v/loops/is an experimental folder. This loop adopts corpora-builder’s [[Spec-to-Shipped-With-TDD]] almost whole: test IDs in the spec, failing tests first, status derived by the ledger, never written. What it adds is a VP Eng who farms plans to engineer subagents and verifies their work independently before anything counts as done.
Roles
┌─ VP ENG (the main session) ──────────────────────────────────────────┐
│ OWNS the spec, the plans, the gh issues, merges into development, │
│ this loop's log, and the operator's attention │
│ DOES reviews each plan before handing it out; re-runs the ladder │
│ and the ledger itself after every engineer reports; reads │
│ every changelog; asks the operator at RED gates │
│ NEVER takes an engineer's "it's green" on trust │
└──────────────────────────────────────────────────────────────────────┘
⇅ one plan per engineer, one branch per plan
┌─ LEAD ENGINEER (a subagent) ─────────────────────────────────────────┐
│ WRITES src/connector/, tests/connector/, scripts/, CI, its plan's │
│ status, its changelog entry │
│ NEVER edits a test to make it pass · changes the spec's intent · │
│ deploys or spends money · pushes to main or master · │
│ touches io/ │
└──────────────────────────────────────────────────────────────────────┘
Per plan
- VP Eng reviews the plan against the live code and the spec. Anything the plan got wrong is fixed in the plan before it’s handed out.
- VP Eng opens the gh issue for the plan, linking the plan file.
- Engineer writes the failing tests first, one per owned ID, each marked
@pytest.mark.spec("CONN-…"). The ledger must show every owned ID RED or MISSING-free before implementation starts (the TDD floor). - Engineer implements in small steps, climbing the ladder
(
scripts/check.sh) after each. - Iterate until green. If a test can’t go green against honest code, the spec is wrong: the engineer stops and reports it, and the VP Eng amends the spec with the operator. The test is never edited to fit.
- Engineer writes the changelog entry, commits, and pushes its branch. The entry says plainly what was not tested and why.
- VP Eng verifies: merges into
development, re-runs the full ladder and the ledger, reads the changelog against what the code does, and closes the gh issue withFixes #N.
Branches
Each plan runs on connector/plan-N-<name>, cut from development after the
plans it depends on have merged. Plans 4, 5, and 6 run in parallel in separate
worktrees. Only the VP Eng merges into development. main is promoted once,
when the spec is done, with the operator’s yes.
Gates the VP Eng takes to the operator
- A test that can’t go green honestly (the spec is wrong).
- Anything outward-facing or that costs money: the Railway deploy, DNS for
memopop.didi.sh, deploying id.didi.sh. - Promoting to
main.
Everything else runs without asking.
Done
The ledger reports every CONN- ID GREEN except those that need a live
service or credentials (GATED), each of which the VP Eng has run by hand and
recorded in the log below; every plan’s gh issue is closed; the operator has
walked the connector in Claude Desktop.
Log
One entry per plan as it lands: the ledger result, what the VP Eng’s own run found, and anything decided that the plan didn’t cover.
Phase 3, didi.sh OAuth — merged to id-didi-sh main 2026-10-10, not deployed
Engineer reported: OAUTH-01..10 green, mix test 106 → 125.
VP Eng re-ran: 125 passed, 18 tagged OAUTH cases passed, format and
--warnings-as-errors clean. Reviewed the one test edit after the red commit
(c6b3b2b): it made two hash-storage checks able to fail and fixed a header
read; nothing was loosened. Accepted.
Found: Claude requires the protected-resource resource to equal the
connector URL including /mcp. Spec amended (99ce648, CONN-AUTH-02), Phase 1
told. id.didi.sh resolves /mcp to the registered https://memopop.didi.sh
audience.
Open: Fly deploy (operator), a real Claude sign-in, id_token for OIDC
clients, and a pre-existing open redirect in AccessController.safe_next/1
the engineer found and left out of scope. Session rolling on refresh was the
engineer’s call; flagged to the operator.
id.didi.sh redirect fix and deploy — 2026-10-10
An open redirect the OAuth engineer found (//evil.example after sign-in, and
an unauthenticated POST /api/magic-links that stored any next) was fixed on
its own branch with 12 tests (125 → 137), verified, merged, and deployed to Fly
with the OAuth server. Live discovery checked: every URL https://.
Phase 1, foundation — merged 840e5b9
Engineer: 33/33 GREEN, CI green. VP Eng re-ran: ladder green, 33 green, 597
passed with io/. Post-red test edits added tests and one stricter ledger
expectation; nothing loosened. Four spec questions settled in ba46275.
Phase 4, materials — merged 7a11801
40 green, 627 passed. Post-red change shrank a synthetic PDF so the asserted
last page survives the 400k text cap; assertion unchanged. Spec records the
skipped material status (§The tools, “Materials never block”). Follow-up handed to
Phase 2: sweep materials left queued.
Phase 6, history and snapshots — merged 8f3c023
API-changelog conflict with Phase 4 resolved keeping both. 45 green, 638 passed. jj adds ~120 ms per save; the connector suite now takes minutes.
Phase 5, enhancements, compile, flow — merged e45ab21
Conflicts in app.py, config.py, and the API changelog, all resolved keeping
both sides. 54 green, 665 passed. Post-red commit 2778b4c fixed canned data
and added an assertion. Spec gained drafts_incomplete, link_expired,
step_skipped, and the ~200 s compile budget. Open: the Claude-side skip and
earlier-version reads have tests but no spec IDs yet; deck-image placement
contract unreconciled with materials extraction.