docs: nested-sic client design + chain-resolution decision (audit trail)

Records the 2026-07-23 bullpen design session (room msgs 865-877): the v2 client model
(client builds the whole onion, servers peel one hop each), and two decisions with their
rejected alternatives so a future reader sees the reasoning, not just the outcome.

Decision 1 — chain resolution = static per-client hosts.toml, fail-loud, no mutable store in
the routing path. mneme-as-router was proposed and REJECTED: mutable shared store = a
supply-chain RCE vector (poison one chain entry, reroute a hop's exec); staleness fails
silently into the wrong container vs a loud connection-refused; dies when boltzmann sleeps;
and adds nothing over a flat file since the client needs a local mneme address regardless.
mneme stays the fleet MEMORY store — this is narrowly about routing.

Decision 2 — migration = daemon dual-read (sniff first byte: 0x00=v2, digit=v1-legacy),
daemon-first, no flag day.

Escape hatch (documented, not built): signed manifests out-of-band if redeploy friction ever
exceeds the trust cost.

Scribed by noether from the room debate; @architect/@skeptic (deepseek) argued it, Markus
initiated and made the calls. Not code — no review gate.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Claude (noether)
2026-07-23 08:30:58 +02:00
parent b28f9d3722
commit 77c58beb93
+81
View File
@@ -0,0 +1,81 @@
# sic — nested (multi-hop) design & chain-resolution decision
Status: **Accepted** — 2026-07-23, bullpen design session (room msgs 865877;
@architect / @skeptic on deepseek-v4-flash-dspark, initiated by Markus).
This records a decision and its rejected alternatives *for the audit trail* — a future
reader should see that mneme-as-router was considered and killed for cause, not overlooked.
## Goal
Make `sic host/hop1/hop2 … argv` real: run a command inside a container reached through a
chain of hops (`incus exec`, `docker exec`, `pct exec`, …) without a shell re-parsing argv at
any layer. The v2 wire-protocol **daemon** (`cmd/sicd`, `gateway/sicd`) already implements
one-layer-per-hop peeling; this covers the **client** and the rollout.
## Model (confirmed)
The **client builds the whole onion up front**: wrap the innermost command in a frame for hop
N, wrap that for hop N-1, … outermost for hop 1, and ship the nested blob to hop 1. **Each
`sicd` peels exactly one layer** (its own), runs its runtime verb with `sic-run` as the
command, and `sic-run` reads the next frame from stdin and repeats. Construction is recursive
and up-front; unrolling is iterative and distributed.
One frame = preamble (`MAGIC 0x00`, then 4-byte big-endian `len32`) + a djb netstring whose
content is `<command> 0x00 <payload>`; the payload of a non-terminal frame is itself a full
frame for the next hop.
## Decision 1 — chain resolution: **static per-client config**
`/etc/sic/hosts.toml` declares, per host, the ordered runtime verbs for its hops
(`nest = ["incus", "docker"]`, etc.). The client reads it to turn `host/hop1/hop2` into the
concrete verb + address at each layer. **Fail loud on any mismatch** (unknown host, more hops
than the config declares) — never guess.
### Rejected: mneme as the chain router
Resolving chains by querying the fleet-memory store (mneme) was proposed and **rejected**.
Even with the bootstrap circularity removed (mneme's own address is a flat local setting, so a
lookup is one direct call, not a nested one), it loses on four counts:
- **Trust (dispositive).** mneme is a *mutable shared* store. Any container with write access
poisons one chain entry and every subsequent `sic` call through that chain execs the
attacker's verb on the target host — a supply-chain RCE vector. Static config is
immutable-after-deploy and auditable once.
- **Staleness fails the wrong way.** A moved/renamed container makes the stored chain lie, and
the client *silently* execs into the wrong container. Static config fails loud (connection
refused); mneme turns a hard error into quiet wrong-target execution.
- **Availability.** mneme lives on boltzmann, which sleeps; resolution would die with it, and a
local cache to cover that is just a slower, staler config file.
- **No net benefit.** The client needs a local mneme *address* regardless, so mneme only adds a
fragile network hop + cache to do what a flat file read already does.
mneme remains the fleet **memory** store — this decision is narrowly about *routing*, nothing else.
## Decision 2 — migration: **daemon dual-read, daemon-first**
The v2 daemon hard-requires the `0x00` magic; the deployed v1 client sends a digit first, so a
naive v2 deploy breaks every existing call on 30+ live hosts. Instead the daemon **sniffs the
first byte** — `0x00` → v2 parser, digit → v1-legacy parser — a once-per-connection branch,
~20 lines, zero perf cost. Ship the v2 daemon everywhere first; roll clients independently, no
flag day.
## Backward compat & the two original bugs
- A **bare `--`** must survive into the innermost argv untouched (the payload is opaque bytes;
no hop interprets it) — closes marfrit/sic#1.
- **stdin must be forwarded** into/through the frame — closes the silent zero-byte-file bug.
- A **bare host** with no `/` stays a single v1-style hop (backward compatible).
## Escape hatch (documented, not built)
If redeploying clients on chain changes ever costs more than the trust of a live store,
distribute **signed manifests out-of-band** — deterministic like static config, updatable like
mneme, without the mutable-store poison vector. Future work; explicitly not today's problem.
## Consequences
- Chain changes require redeploying clients (accepted: moves are infrequent, rollouts controlled).
- No network dependency and no mutable store in the routing path — routing is deterministic and
auditable.
- Next: implement `cmd/sic` (config read + recursive frame build + `--`/stdin fixes) and the
daemon dual-read; both are Go, grindable via @godev.