Files
alkssh/AGENTS.md
T
glm-5.3-flash f4424b7e58 docs(research): vault's SSH role is at-rest secret storage, not key derivation
A conflation in the research set: alkvault's derive-and-custody surface
(derive_ed25519, SLIP-0010 paths) was being sketched as the SSH key
path. Corrected across the docs — held SSH keys (client outgoing keys
and the server's own host key) are bearer-token-class secrets: opaque
bytes the family holds, vault-encrypted at rest, decrypted at use (the
alkhttp ADR-014 posture). No derivation story, no vault-custody
design, and the Signer-injection idea is withdrawn and recorded with
its dismissal rationale.

- auth-bridge.md: client-side gap re-framed as at-rest storage — the
  at-rest shape (OpenSSH encoding inside alkvault EncryptedData), the
  at-use lifecycle (decrypt-at-dial, one-time decode_secret_key,
  zeroized for the dial), the KeySource-like source shape
- pocs.md: POC-2 rewritten to at-rest key handling (formats comparison,
  decrypt-at-use lifecycle, locked-vault fail-closed, Signer dismissal)
- vision.md: principle 5 scoped (secret store, not key generator/custodian)
- russh.md: Signer seam noted as not presumed to be the vault shape
- AGENTS.md/CHANGELOG: convention 6 + context updated to match

Verification: cargo doc --no-deps, cargo fmt --check, cargo test
(docs-only change)
2026-10-25 10:37:37 +00:00

251 lines
13 KiB
Markdown

# AGENTS.md
Operating instructions for opencode agents working in this repo. opencode
auto-loads this file as instructions, overriding the built-in defaults for
this project. Custom agents in `.opencode/agents/` inherit these rules
unless their own prompts say otherwise.
## Git Workflow
**Commit and push when reasonable.** When a change is complete and
verified (build + lint + tests pass), commit and push to `origin/main`
without asking. This overrides the built-in default of "only commit when
explicitly asked."
Commit in small, focused units — one commit per unit of work (a fix, a
feature, a doc change), not one large commit covering many topics. Push
regularly so work is never stranded locally.
The workflow:
1. Make the change
2. Verify: `cargo test`, `cargo clippy --all-targets -- -D warnings`,
`cargo fmt --check`, `cargo doc --no-deps` if docs changed
3. Inspect `git status` and `git diff` before staging — stage only the
intended files, never secrets
4. Write a concise commit message in **conventional commits** style:
`<type>(<scope>): <summary>` — types are `feat`, `fix`, `docs`,
`chore`, `refactor`, `test`, `perf`, `build`, `ci`; the scope is
optional. For multi-point changes, use a summary line plus a body with
bullet points and a verification block.
5. `git push origin main`
6. Report the commit hash and the verification summary
Exceptions — **do not** commit or push without asking:
- The change is exploratory / speculative (you're not sure the user wants
it kept)
- The user is actively reviewing the diff and may ask for changes
- The change touches wire-relevant or semver-relevant public API (the
`alk/ssh` ALPN pin, the capability feature graph, and the mapping
contracts become stable one-way doors once consumers exist; see the
ADRs in `docs/architecture/decisions/` once they exist)
- You'd be force-pushing, amending a published commit, creating an empty
commit, or skipping hooks
Never commit secrets, keys, or credentials. If a commit fails or hooks
reject it, fix the issue and create a new commit — do not amend the
failed one.
Git identity is preconfigured (`glm-5.3-flash <glm-5.3-flash@alk.dev>`).
Do not change `git config`, skip hooks, or use `git commit -i`.
## Project Conventions (Rust / SSH adapter crate)
This is alkssh — the family's SSH adapter (family ADR-003): a thin
wrapper around russh terminating the vanilla-ssh wire and bridging
to/from alkcall channels. The SSH counterpart of alkhttp's adapter
role — the vanilla-ssh front door for vanilla clients. No binary until
an assembly needs one; downstream apps embed the crate and bring their
own deployment. The conventions below apply to all work in `src/` and
`tests/`. They mirror `.opencode/agents/implementation-specialist.md`
§Project Conventions and are repeated here so they apply to every
session, not just spawned implementation agents.
1. **No comments in code** unless the user explicitly asks. This is a
project-wide convention. Doc comments (`///`, `//!`) are fine and
expected on public API. Inline `//` comments only when the user asks
or when a non-obvious safety/correctness constraint would otherwise
be missed (e.g., "the decline path must run before any per-channel
state is allocated — a leaked channel alloc is a capability leak").
2. **Error handling** — `thiserror` for library error types. No panics
in library code. No `unwrap()` or `expect()` outside tests. If you
reach for `unwrap`, the error path wasn't specified — stop and decide
what should actually happen. For poisoned `RwLock`/`Mutex`, use
`unwrap_or_else(|e| e.into_inner())` so a panic in one operation does
not cascade to other operations.
3. **`tokio` is the async runtime** — all I/O is async. SSH sessions are
long-lived streaming conversations; do not introduce blocking I/O on
the async path. Use `tokio::sync` primitives (`oneshot`, `mpsc`) for
correlation; `parking_lot` for short-held internal locks.
4. **russh owns the SSH wire** — no in-family SSH framing (family
ADR-003 decision 1). All SSH protocol framing, key exchange, and
channel multiplexing come from russh; the crate's substance is the
glue to/from alkcall and the channel-type mappings. Third-party-
upstream posture: file asks to russh upstream; never silently absorb
a workaround.
5. **Fail-closed capability features (the binding constraint)** — every
channel-type mapping is a cargo feature. A session opening an
unfeatured or unimplemented channel type is declined with a clean
error — never a half-capability, never a silent fallback (family
ADR-003 decision 3). A default build (no features) has no
capabilities; a git-only build cannot serve a shell, tunnels, or
socks. This is alkforge's ssh endpoint shape. Every disabled
feature's decline path is contract territory and gets tested too —
feature-gated code that is never exercised in a given build rots
silently otherwise.
6. **Key reuse, no env vars** — keys reuse the alknet `main` key model
(`russh::keys`, authorized-key/KeySource handling — see
`/workspace/@alkdev/alknet-main/crates/alknet-core/src/auth/`) via
alkvault-backed sources — keys held encrypted at rest, decrypted at
use (the vault is a secret store on this path, not a key generator)
(family ADR-003 decision 2). No secret
material on the wire, in plaintext at rest, or in env vars — the
standing family security posture (family ADR-002 decision 5).
7. **Visible-surface = authorized-surface** — every channel type and
request exposed on the wire is checked against alkcall's
`AccessControl::check(peer_identity)`; ops with
`Visibility::Internal` are never wire-callable. There is no
unauthenticated capability; the fail-closed feature declines are part
of this invariant. This is a spec-level invariant.
8. **Wire formats are stable one-way doors** — the `alk/ssh` ALPN pin is
its own ADR (ledger rule 1, family ADR-003 decision 4) and, once
published, does not change. The russh version pin is a one-way door
in practice (SSH key exchange/kex algorithm negotiation crosses
crate boundaries); bumps are deliberate, tested, and recorded.
Protocol capability advertisement is honest: never advertise what we
don't serve — an unfeatured channel type is declined, not advertised.
9. **Bounded resources** — every session carries wall-clock, size, and
channel limits (max concurrent channels, max channel window, session
timeout, auth attempt limits). SSH servers are internet-facing;
unbounded loops and unbounded buffers are bugs.
10. **License hygiene** — this project is MIT OR Apache-2.0. russh is
MIT/Apache-2.0-compatible; check before adopting any new dependency,
and never copy or derive code from incompatible-licensed reference
projects. Reference code is for reading and facts, never copying.
11. **Naming** — Rust standard: `snake_case` for functions/variables/
modules, `PascalCase` for types/traits, `SCREAMING_SNAKE_CASE` for
constants.
12. **Module structure** — one module per file under `src/`, re-exported
from the crate root (the alktty crate-root pattern). Public API
surface is `lib.rs` re-exports. The mapping table
(channel-type → family protocol) is the crate's spine; each mapping
lives behind its feature and speaks only its family protocol's
types.
13. **Composability boundary** — alkssh is a door, not a backend: it
terminates vanilla ssh and forwards mapped channel payloads to
family protocol halves (alktty/alktunnels/alksocks/alkgit). It
depends on alkcall types at the wire layer and the mapped crates'
session/client surfaces, and carries no deployment policy — a
downstream app (alkforge) embeds it, enables the features it wants,
and brings identity/ACL configuration. No feature implies another;
the feature graph is independent and flat until architecture says
otherwise.
## Verification Commands
Run these before committing. All must pass.
```bash
cargo test # full suite
cargo clippy --all-targets -- -D warnings
cargo fmt --check
cargo doc --no-deps # if docs changed
cargo publish --dry-run --allow-dirty # before a release
```
If feature flags are added, also run `cargo test --all-features` and
`cargo clippy --all-features --all-targets -- -D warnings`. Once the
feature graph lands, each disabled feature's decline path gets a test —
the fail-closed invariant is contract territory on both sides of every
feature.
## Task files (taskgraph)
Any file under `tasks/` is a taskgraph task. Frontmatter is required
(`id`, `name`) and its enum fields are **closed sets** (taskgraph
0.1.3): `status: pending|in-progress|completed|failed|blocked` (never
`done` or `finished`), `scope: single|narrow|moderate|broad|system`,
`risk: trivial|low|medium|high|critical`, `impact:
isolated|component|phase|project`, `level:
planning|decomposition|implementation|review|research`. A file without
frontmatter is silently dropped from the graph (validate does not flag
it) — prose-only trackers belong under `docs/`. Run `taskgraph
validate` after writing or editing anything under `tasks/`; full field
semantics live in `docs/sdd_process.md` §"Task File Format" /
§"Categorical Estimates".
## Lifecycle Status
The project is in **SDD phase 0** (exploration; docs only — no
architecture committed, no implementation). Family ADR-003 (2026-10-25,
Accepted) records the structural decision already made: russh-wrapped
adapter, no in-family SSH framing, fail-closed capability features,
alknet-main key model via alkvault-backed sources; OQ-08's ordering
question is resolved (alkssh leads alkforge's phase 0) and OQ-08's
remaining scope — the channel-type → family-protocol mapping table
details, the auth design tying russh's authorized-key handling to
alkcall auth, and the `alk/ssh` ALPN pin (ledger rule 1) — is this
repo's phase 0 work. Phase 0 output lands in `docs/research/` (the
phase 0 document is the current work item); `docs/architecture/` +
ADR-001 (the `alk/ssh` ALPN pin) come with phase 1. POC code never
merges to main — two modes: branch mode (POC builds on repo code; git
branch, findings merged into research docs, branch dropped) or
standalone mode (POC independent of repo code; scratch project at
`/workspace/<poc-name>`). POCs have relaxed constraints
(comments/unwrap acceptable) — they are exploration tools, not
production code.
## Architecture Context
- `docs/sdd_process.md` — the SDD process (phases, roles, task format)
- `docs/research/` — phase 0 research (current source of truth for
design direction):
- `vision.md` — vision, guiding principles, non-goals, the phase 0
scope questions
- `russh.md` — russh crate inventory (server/client Handler, Config
knobs, keys module) and pin considerations (RC-pinned dep tree,
MSRV, license) — no alknet-derived bias; the pin is verified
against current russh at architecture time
- `auth-bridge.md` — russh auth ↔ alkcall `IdentityProvider` mapping
(server: verified-key fingerprint → `resolve_from_fingerprint`;
client: alkvault-backed key sources — held keys encrypted at rest,
at-rest/bearer-token-class storage, not key derivation)
- `mapping-table.md` — the crate's spine: SSH channel type / request
→ family protocol seam + cargo feature; the alktunnels
absence note; the `alk/ssh` ALPN topology question
- `pocs.md` — POC-1..3 plan (auth bridge, client-side at-rest key
handling, mapping bake-off) gating phase 0 completion
- family docs: `/workspace/@alkdev/alkfamily/docs/` —
`architecture/decisions/003-alkssh-russh-adapter.md` is the structural
decision this repo builds from (and `002-alkforge-gitea-replacement.md`
explains why: alkforge's git door, fail-closed posture);
`architecture/open-questions.md` OQ-08 holds the remaining scope
- Sibling crates (published, MIT OR Apache-2.0, each with per-repo
AGENTS.md + `docs/architecture/decisions/`):
`alkcall 0.10` (call + channels RPC with ACL; the `alk/ssh` ALPN and
the channel-wire notes are alkcall ADR-004 and ADR-034),
`alktty` (session/pty mapping target), `alktunnels`
(direct/forwarded-tcpip target), `alksocks` (`-D` target),
`alkgit 0.1.0` (git exec/subsystem target), `alkvault 0.1`
(key/secret storage). The alkhttp adapters are the closest structural
analog for the adapter shape.
- alknet `main` key model reference:
`/workspace/@alkdev/alknet-main/crates/alknet-core/src/auth/`
(russh 0.49, `keys.rs` / `server_auth.rs` / `identity.rs`) — the
authorized-key/KeySource construction alkssh reuses; a reading aid
and correctness reference, not a path dependency.
- russh: pin per phase 0 research (the alknet pin — 0.49 — is a data
point, not a decision).