design: parallelism map + /loop driver prompt + followups triage

- design/dependencies.md: per-milestone parallelism map for M2–M6 with
  per-phase agent budgets (peak 4 in parallel for M5 framing wave;
  7-agent maximum if M2 wave 1 + M5 framing run concurrently).
- design/prompt.md: self-contained /loop driver. Step 0 triages
  design/followups.md (auto-resolves items whose preconditions are met,
  shelves the rest). Step 3 spawns parallel general-purpose agents per
  design/dependencies.md when the active wave has multiple lanes.
  Sequential lanes (M4 Session core, M5 client integration) run directly.
  Local-commit-only by default; explicit stop conditions; Q7 hasDetailStatus
  audit reminder for any new conditional-read codec port.
- design/README.md: index updated to reference prompt.md, followups.md,
  dependencies.md, and review.md.

design/followups.md is intentionally not pre-created — prompt.md Step 0
bootstraps it on first /loop run.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Joseph Doherty
2026-05-05 06:34:30 -04:00
parent fe2a6db786
commit 16f2c148e5
3 changed files with 469 additions and 0 deletions
+294
View File
@@ -0,0 +1,294 @@
# `/loop` driver — autonomous M2M6 implementation
You are inside a `/loop` iteration for the **mxaccess Rust port** at
`c:\Users\dohertj2\Desktop\mxaccess`. Your goal each iteration: advance the
project by one cohesive unit of work, verify it doesn't regress, commit it
locally, and either schedule the next iteration or stop and surface.
This file is re-fed to you on every iteration with no carry-over state. Read
it top-to-bottom each time. Discover everything else from the project itself.
---
## Iteration protocol
### Step 0 — Triage `design/followups.md`
Read `design/followups.md`. If it does not exist, create it with this
skeleton:
```markdown
# Followups
Open work items deferred during /loop iterations. Triaged at the top of
every iteration. New items are appended under `## Open`; resolved items
move to `## Resolved` with a date + commit hash.
## Open
(none yet)
## Resolved
(none yet)
```
Then for each item under `## Open`:
1. Read the **`Resolves when:`** clause.
2. If its preconditions are met **now** (the gating commit landed, the
ambiguity it flagged is now resolved, etc.) → solve it as part of this
iteration's work, then move it to `## Resolved` with today's date and
the resolving commit hash (the commit you make in Step 5 below).
3. If preconditions are **not** met → leave it under `## Open` and move on.
If `## Open` exceeds **10 items**, STOP and surface to the user — that's a
drift signal that needs human triage, not more iterations.
### Step 1 — Discover state
Run in parallel (single assistant message, multiple tool calls):
- `git log --oneline -20`
- `cd rust && cargo test --workspace 2>&1 | grep "test result:" | grep -v "0 passed" | tail -5` — must show all-pass results.
- Read `design/60-roadmap.md`, `design/dependencies.md`, `design/review.md`.
If `cargo test` is **not green** at the start of the iteration:
1. Diagnose with `cargo build` and the failing test name.
2. Apply **one** targeted fix.
3. If the fix doesn't recover green: `git reset --hard HEAD`, log a
followup describing the failure mode, and STOP.
Never proceed past Step 1 with a red baseline.
### Step 2 — Identify the current phase and unblocked lanes
From the git log + recent commits, determine which milestone is in flight:
- Most recent `[M0]` / `[M1]` / `[M2]` / ... commits → that's the active phase.
- If the active phase's DoD (per `design/60-roadmap.md`) is fully met, the
next iteration's work is the **next** milestone. Advance the phase
marker mentally.
Then consult `design/dependencies.md` for the active phase's parallelism map.
Identify which lanes are unblocked **right now** (their dependencies have
landed in earlier commits).
The summary table for fast lookup:
```
M2 wave 1: 3 agents — NTLMv2 client / DCE/RPC PDU codec / OBJREF parser
M2 wave 2: 2 agents — OXID resolution / IRemUnknown::RemQueryInterface
M2 wave 3: 1 agent — mxaccess-callback (the INmxSvcCallback exporter)
M3: 2 agents — mxaccess-galaxy / mxaccess-nmx
M4 wave 1: 1 agent — Session core + RecoveryPolicy types (sequential)
M4 wave 2: 2 agents — write family / subscribe family
M4 wave 3: 7 agents — examples (one each)
M5: 4 agents — MS-NMF framing / MC-NBFX codec / MC-NBFS dictionary / DH+HMAC+AES
M5 client: 1 agent — mxaccess-asb operations (sequential after framing)
M6: 4 agents — mxaccess-compat / perf / metrics / docs
```
Pick the smallest unit that:
- Is currently unblocked (dependencies landed).
- Is not already covered by an open followup that's deferred.
- Can be completed in one iteration's work (one commit's worth).
### Step 3 — Execute
**If the active wave has multiple parallel streams** (any row above with
`N agents` where `N > 1`), spawn that many `general-purpose` agents in **one
single message** containing N parallel `Agent` tool calls. Each agent owns
one `.cs` source file (or one logical unit) and emits one Rust module.
Each agent's prompt MUST include:
- Project context (one paragraph: this is the mxaccess Rust port,
`src/` is the executable spec, CLAUDE.md forbids fabrication).
- The exact `.cs` source path to port.
- The exact Rust output module path.
- Reference to existing M1 modules as the pattern to follow
(`reference_handle.rs`, `envelope.rs`, `status.rs`).
- Test requirements: round-trip, boundary checks, citation-bearing parity
vectors against `tools/Compute-Crc.ps1` style helpers where applicable.
- Hard rule: do NOT edit `lib.rs`. The driver wires up modules after agents finish.
- Audit reminder: any conditional read pattern (`hasDetailStatus`-style)
must mirror the .NET reference's unconditional/conditional split exactly
(`design/70-risks-and-open-questions.md` Q7 — the M1 wave-1 audit defect).
**If the work is sequential** (M4 Session core, M5 client integration after
framing lands, any wave with `1 agent`), do it directly. Read the .cs source,
port it inline, write tests inline. Do not spawn an agent for sequential
single-stream work.
After parallel agents return: wire up `lib.rs` (mod declarations + re-exports)
and remove any stub types they replaced.
### Step 4 — Verify
Run all four DoD gates:
- `cargo build --workspace --all-targets`
- `cargo test --workspace --no-fail-fast`
- `cargo clippy --workspace --all-targets -- -D warnings`
- `cargo fmt --all -- --check`
For codec changes also verify the .NET parity test still passes:
- `cargo test -p mxaccess-codec --test dotnet_codec_parity`
If any gate fails:
1. Try **one** targeted fix.
2. If still failing → `git reset --hard HEAD`. Append a followup. STOP.
### Step 5 — Commit (local only)
- `git add -A`.
- Commit message format:
```
[M<n>] <crate>: <one-line summary, ≤ 70 chars>
<body bullets, what changed and why, citing src/...:LINE for protocol claims>
- Test count delta: NNN → MMM (+K)
- Open followups touched: F<N>, F<N> (or "none")
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
```
- **Do not push.** The user pushes manually unless they've explicitly
enabled auto-push for the loop. If auto-push has been enabled (you'll
know because they invoked `/loop` with that flag or said so in this
iteration's pre-amble), then `git push` after the commit.
### Step 6 — Log new followups
For each item this iteration **discovered but did not solve**, append to
`design/followups.md` under `## Open` using this schema:
```markdown
### F<N> — <one-line title>
**Severity:** P0 | P1 | P2 | P3
**Source:** commit <hash>
**Why deferred:** <reason — e.g. "needs M3 NMX client to verify wire round-trip", "ambiguous protocol question, no evidence in src/ or captures/">
**Resolves when:** <preconditions — what must land or change for this to be actionable>
**Notes:** <optional cross-links to design/*.md sections>
```
`<N>` is the next free integer (highest existing F-number + 1, or 1 on
first followup).
Common follow-up sources:
- An agent reported a deviation from the .NET reference that requires a
separate design-decision turn.
- A risk register item (`R1``R16` / `Q1``Q7`) was hit but did not block
this iteration's work.
- A test fixture is missing and would require a new live capture.
- A dep-version bump or feature-gate decision arose that needs a workspace-level
agreement.
### Step 7 — Decide next
| Condition | Action |
|---|---|
| Milestone DoD fully satisfied (per `60-roadmap.md`) | Make a final `[Mn-done]` commit summarising the milestone, then `ScheduleWakeup` for next milestone start. |
| Hit ambiguous protocol question (no evidence in `src/` / `docs/` / `captures/`) | Log followup, STOP, surface to user. |
| Hit P0/P1 blocker tracked in `70-risks-and-open-questions.md` | Log followup, STOP, surface. |
| 3 consecutive iterations with zero net progress (no new commit, no resolved followup) | STOP, surface to user. |
| `## Open` followups list > 10 items | STOP, ask user to triage. |
| M6 DoD fully satisfied | Project complete. STOP. Do not schedule. |
| Otherwise | `ScheduleWakeup` with `delaySeconds` in 60270 (cache stays warm). |
When you call `ScheduleWakeup`, pass the literal sentinel
`<<autonomous-loop-dynamic>>` as `prompt` so the runtime re-resolves these
instructions. Use a one-sentence `reason` describing what the next iteration
will pick up.
---
## Hard rules (do not negotiate)
- **No fabricated protocol behaviour.** Every wire-byte, IID, opnum, HRESULT,
byte-offset, or layout claim must cite `src/MxNativeCodec/*.cs:LINE`,
`src/MxNativeClient/*.cs:LINE`, `src/MxAsbClient/*.cs:LINE`, `docs/*.md:LINE`,
`analysis/frida/*.tsv`, or `captures/0NN-frida-*`. If you can't cite, you
can't claim. Log a followup instead.
- **No `--force`, `--no-verify`, `--no-gpg-sign`** on git commands.
- **No amending pushed commits.** Always create new commits.
- **No deleting or rewriting** files in `captures/`, `analysis/frida/`,
`analysis/proxy/`, `analysis/decompiled-*/`, or `analysis/ghidra/exports/`.
These are evidence per CLAUDE.md.
- **No editing `lib.rs` from an agent.** The driver wires up modules.
- **No skipping verification.** All four cargo gates green or revert.
- **No pushing without authorization.** Commit locally only by default.
- **Preserve unknown bytes.** Match the .NET reference round-trip for any
field whose semantics are not yet decoded. Use `[u8; N]` preservation
fields and document with a `:LINE` citation.
- **Tests over assertions.** Do not add `assert!(true)` or `assert!(<const expr>)`
at runtime; use `const _: () = assert!(...)` for compile-time checks.
- **Conditional reads must match the .NET reference exactly** — see
`design/70-risks-and-open-questions.md` Q7 (`hasDetailStatus` audit). Any
field the .NET reads unconditionally must be read unconditionally in Rust.
---
## Self-check before scheduling next iteration
Before calling `ScheduleWakeup`, verify each:
- [ ] `cargo build --workspace --all-targets` exited 0.
- [ ] `cargo test --workspace` exited 0 with all-pass results.
- [ ] `cargo clippy --workspace --all-targets -- -D warnings` exited 0.
- [ ] `cargo fmt --all -- --check` exited 0.
- [ ] One commit landed this iteration (verify with `git log -1 --oneline`).
- [ ] If any followup was deferred this iteration, `design/followups.md`
has the new entry.
- [ ] `delaySeconds` is in [60, 270] for cache-warm continuation, or in
[1200, 3600] if waiting on a genuinely slow external process.
If any item fails, do **not** schedule next iteration. Surface to the user.
---
## Useful commands reference
```powershell
# State discovery
cd c:\Users\dohertj2\Desktop\mxaccess
git log --oneline -20
git status --short
# Rust gates (run from rust/)
cd rust
cargo build --workspace --all-targets
cargo test --workspace --no-fail-fast
cargo clippy --workspace --all-targets -- -D warnings
cargo fmt --all -- --check
cargo test -p mxaccess-codec --test dotnet_codec_parity
# Live-probe gating (M3+ only)
. ..\tools\Setup-LiveProbeEnv.ps1
cargo test -p mxaccess --features live -- --ignored
# .NET parity helpers
dotnet build src\MxNativeCodec\MxNativeCodec.csproj
pwsh -NoProfile -File tools\Compute-Crc.ps1
```
---
## Why this design
- **Step 0 first** keeps `followups.md` from rotting. Items that became
solvable get cleaned up immediately.
- **Step 1's red-baseline check** prevents iterations from compounding bugs.
- **Step 3's parallel-agent fan-out** is the throughput lever — `M2 wave 1`
and `M5 framing` both run 34 agents concurrently, cutting wall-clock.
- **Step 5's local-commit-only** default is reversibility. A bad iteration
can be `git reset --hard HEAD~1` without affecting any remote.
- **Step 7's stop conditions** are explicit and disjoint. There's no "if it
feels right, stop" phrasing — every stop condition has a measurable
trigger.
- **Hard rules** are lifted from `CLAUDE.md` so the loop cannot drift even
if a single iteration loses context.