Peat ecosystem architecture, security invariants, dependency rules, cross-repository workflow, and verification requirements. Use when editing or reviewing the peat workspace, coordinating changes across peat-* repositories, designing shared types or FFI surfaces, or evaluating ecosystem-wide architectural decisions.
Peat ecosystem architecture, security invariants, dependency rules, cross-repository workflow, and verification requirements. Use when editing or reviewing the peat workspace, coordinating changes across peat-* repositories, designing shared types or FFI surfaces, or evaluating ecosystem-wide architectural decisions.
Peat Ecosystem SKILL
Peat is an interoperability-first mesh registry sync platform built for heterogeneous autonomous systems in defense and edge environments. Its core value proposition is interoperability that enables scale — Peat connects systems that don't speak the same language across transport and protocol boundaries. Peat is developed under the Defense Unicorns GitHub org: https://github.com/defenseunicorns
When this skill applies
Any session touching files in a peat-* repo or the top-level peat crate
Cross-repo changes affecting more than one peat repo
Reviewing a PR in any peat repo
Invoke the relevant repo-scoped skill from the router below. Per-repo skills
are authored against SKILL_TEMPLATE.md.
Skill router
Read only what's relevant to the current task. Do not preload every per-repo skill.
The ecosystem comprises one Rust workspace repo (peat) plus several sibling repos. Internal workspace subcrates do not have separate per-repo skills — they share this file.
Sibling repos (each its own git repo, each with its own per-repo skill):
Repo
Purpose
Skill
peat
Rust workspace. Hosts the ecosystem skill (this file) and the per-repo skill for the workspace. Active.
Lightweight implementation for constrained environments.
../peat-lite/.agents/skills/peat-lite/SKILL.md
peat-atak-plugin
One consumer-plugin integration (Android/Kotlin). Pure Kotlin/Gradle — no Rust here. Consumes pre-built JNI/UniFFI bindings from peat/peat-ffi. Repo name is historical; this repo treats it as a generic consumer surface.
ContainerLab-based network simulation harness. Not on the production path.
../peat-sim/.agents/skills/peat-sim/SKILL.md
Workspace subcrates (members of the peat repo, share this skill):
Subcrate
Role
peat
Top-level crate; eventual home for shared types and traits. Currently a placeholder (see "peat repo-specific skill" below).
peat-schema
Schema definitions.
peat-protocol
Protocol logic (Translator, ChangeEvent, etc.).
peat-transport
Transport abstractions used by the workspace.
peat-persistence
Persistence layer.
peat-ffi
FFI bindings for Kotlin/Swift (UniFFI 0.28, proc-macro mode) + direct jni 0.21 for Android. The only routinely-unsafe Rust in the ecosystem.
peat-mavlink, peat-tak, and other repositories without repo-scoped skills
still use their checked-in AGENTS.md guidance when present. Confirm status
before adding a new skill for a repository that lacks one.
Hard invariants (cross-cutting)
These rules apply in every repo. Violating one without explicit user approval is out of scope, full stop.
Language. Rust everywhere except Kotlin in consumer plugins (currently peat-atak-plugin). No new language dependencies. No Python. No shell scripts for anything that belongs in Rust.
FIPS-approved cryptographic primitives only. Every algorithm used anywhere in the peat ecosystem must be on the FIPS 140-3 approved list. AEAD: AES-GCM (not ChaCha20-Poly1305). Signatures: Ed25519 (FIPS 186-5) or ECDSA-P256/P384. Key agreement: ECDH-P256/P384 (X25519 only with explicit review). KDF: HKDF-SHA-2. MAC: HMAC-SHA-2. TLS/QUIC must run under a FIPS-mode crypto provider (e.g. aws-lc-rs for rustls; the default ring is not FIPS-validated). MLS suites must be FIPS-aligned (e.g. MLS_128_DHKEMP256_AES128GCM_SHA256_P256). Existing ChaCha20-Poly1305 references in ADR-006/044/048/049 + spec docs are tracked for amendment; do not propagate them. Canonical reference: ADR-060 §5 "Cryptographic primitives (FIPS posture)" + AGENTS.md § "Hard rule: FIPS-approved cryptographic primitives only".
No consumer-specific references in peat. peat is the generic mesh substrate; consumers (mobile-app plugins, wearable firmware, CLI tools, server bridges) live in sibling repos. Code, comments, examples, READMEs, operational docs, JNI symbol names, package paths, and test fixtures in this repo MUST NOT name a specific consumer (ATAK, WinTAK, iTAK, WearTAK, etc.) or vendor-derived identifiers. Use "consumer", "consumer plugin", "CoT consumer", "mobile-app plugin", "wearable", "CLI tool", or "server bridge". Protocol-name appearances (e.g. CoT, TAK Server wire protocol) are allowed where the protocol itself is structurally load-bearing; consumer-name appearances are not. The only exception is ADRs in docs/adr/ citing real-world use cases that motivated a design decision. See AGENTS.md § "Hard rule: no consumer-specific references in peat" for the full rule + rationale. Verification gate (below) includes a grep check for ATAK / vendor names in new diffs.
Dependency flow.peat is the dependency anchor. Common types, traits, error handling flow down from peat. Repos depend on peat, never on each other directly. Circular dependencies are rejected.
Transport agnosticism. Peat protocol logic must not assume a transport. BLE, mesh, IP, serial are all interchangeable. Transport-specific code stays in transport repos (peat-btle, peat-mesh) or the peat-transport workspace subcrate — never in core / protocol / persistence layers or non-transport sibling repos.
Interoperability first. Every feature decision answers: does this make Peat more or less interoperable with external systems? Peat must never require a counterpart to run Peat software to integrate.
Unsafe Rust. Requires explicit justification in a code comment. The FFI boundary in peat-ffi (workspace subcrate inside the peat repo) is the only routinely legitimate unsafe zone.
FFI boundary direction. All Peat protocol logic stays in Rust (peat-ffi and the rest of the workspace). Kotlin in peat-atak-plugin is UI and Android lifecycle only — never a destination for protocol, mesh, transport, persistence, or serialization logic. Native libraries and Kotlin bindings flow from peat-ffi into peat-atak-plugin, not the reverse.
Async runtime. Tokio is the ecosystem standard. The peat workspace pins tokio = { version = "1", features = ["full"] } (or feature subsets where appropriate); sibling repos do the same. Do not introduce alternative async runtimes (async-std, smol) without explicit user approval.
Error handling. Library crates use thiserror for crate-level error enums. Application/binary code uses anyhow for ergonomic error chaining. Both are pinned at 1 in workspace deps and in sibling repos — use those, do not pull alternative error crates without justification.
Commit signing. GPG-signed commits are required across the ecosystem. Do not configure git to bypass signing or use --no-verify to skip pre-commit hooks. Fix the signing setup; do not work around it.
Branch + merge convention. Trunk-based development on main with short-lived feature branches. Every PR squash-merges to main — write the PR title as if it will become the merge commit subject (because it does).
Workflow
For any task in a peat repo:
Orient. Read this file, invoke the relevant repo-scoped skill from the router, follow AGENTS.md, and inspect git status plus git log -10.
Locate the spec. Confirm the task has a GitHub issue with the required sections (see "Issue format" below). If not, ask the user before implementing.
Plan. Produce a short plan. Check it against the hard invariants and the per-repo skill's scope guards.
Implement following the per-repo workflow. Vertical slices, one concern per commit.
Verify per the per-repo skill's exit criteria.
Hand off. Open PR referencing the issue. Summary states what changed and why. Flag any cross-repo implications.
Changelog and GitHub Release notes contract
CHANGELOG.md and GitHub Release notes serve different scopes:
A changelog entry records the delta assigned to one immutable version. Keep it
accurate and version-scoped, including a failed release's intended source
change even when publication later fails.
GitHub Release notes tell operators what the complete published release
delivers. A tag-specific changelog extraction is an input, never sufficient
evidence that the release notes are complete.
Before publishing or editing a GitHub Release:
Choose the real baseline. Normally compare with the previous release. If
any intermediate tag was failed, partial, incompatible, withdrawn, or
superseded, compare from the last complete usable release and state the
disposition of every intermediate tag and any immutable artifact it
published.
Inventory the shipped work. Inspect the tag comparison, merged commits,
linked issues and PRs, cross-repository dependency releases, qualification
trackers, and retained nonzero outcomes. GitHub-generated notes may help
enumerate PRs; they do not replace this review.
Write operator-facing notes. Include, as applicable:
release status, upgrade guidance, and corrective-train history;
behavior grouped by user/operator impact rather than commit order;
issue/PR traceability, including cross-repository work and open trackers;
package checksums or image digests needed to identify qualified artifacts;
waivers, failed gates, known limitations, and claims the release does not
make; and
a full comparison link from the selected baseline.
Verify after publication. Read the release back from GitHub and confirm
the expected title, baseline, traceability, warnings, artifact identities,
and complete body. A successful workflow or gh release edit exit is not
proof that the audience-facing notes are complete.
Release work is not complete while the GitHub Release describes only the final
patch in a multi-tag corrective train or omits the issues and evidence that
justify recommending—or withholding recommendation of—the artifact.
Verification (ecosystem-level)
Beyond the per-repo verify checklist, an ecosystem-level change is not done until:
Each affected repo's CI is green
No new cross-repo cycle introduced (peat does not depend on its consumers; sibling repos do not depend on each other)
PR references a GitHub issue with Context / Scope / Acceptance Criteria / Constraints / Dependencies sections
If a hard invariant was waived, the PR description names which one and quotes the user approval
For changes inside the peat repo: the new diff is free of consumer-specific identifiers (ATAK, WinTAK, iTAK, WearTAK, etc.). Run git diff main -- ':!docs/adr' ':!docs/whitepaper' ':!AGENTS.md' ':!.agents/skills/peat-ecosystem/SKILL.md' ':!CHANGELOG.md' | grep -E '^\+' | grep -iE '\b(atak|wintak|itak|weartak)\b' | grep -vE 'peat-atak-plugin|com\.atakmap|atakmap\.app|ATAKActivity' — must be empty before merge. The pipeline checks ADDITIONS only (^\+); excludes the rule documents themselves (which legitimately enumerate the forbidden names as part of the rule definition), the release-notes archive (CHANGELOG.md, which by purpose records the history of vendor-name removals and may need to name what was removed to be useful), and ADR / whitepaper material; and tail-filters the sibling repo name peat-atak-plugin (its actual repo name, historical) plus the third-party host app's real Android identifiers (com.atakmap.*, ATAKActivity) that operational adb commands genuinely target. If a citation is genuinely load-bearing it belongs in an ADR (docs/adr/) or a CHANGELOG entry, not in code or operational docs.
"Seems right" or "the diff looks correct" is never sufficient.
Anti-rationalization
Excuse
Rebuttal
"This change spans repos but they're tightly coupled — one big PR is cleaner."
One PR per repo, linked through a tracking issue. Atomicity across repos isn't real; reviewability is.
"I'll just import directly from peat-mesh into peat-node for this."
Cross-repo direct deps are a circular-dependency factory. Add the trait to peat core.
"It's only a tiny shell script for a build helper."
No shell scripts for things that belong in Rust. Justify the language choice with the user first.
"I'll write a quick TS/Python utility for this."
No new language dependencies without explicit approval.
"peat core doesn't have the type I need; I'll just put it in this repo for now."
Surface the gap as an issue against peat. Don't fork the type system.
"I'll move this Peat protocol logic into Kotlin to make the FFI simpler."
All Peat protocol logic stays in Rust. Kotlin is UI and Android lifecycle only.
"This change makes Peat assume the counterpart is also running Peat — fine for now, we'll generalize later."
Interoperability-first is the product. Generalize before merging, or don't merge.
"Just one mention of ATAK in the comment is fine — that's the consumer everyone knows about."
No consumer-specific references in peat. Ever. Use "consumer plugin" / "CoT consumer" / "mobile-app plugin". One mention becomes 382 mentions in two years. Verification gate greps for it.
"The JNI symbol name has to encode the Java package path of the calling class."
True for the suffix; the Java package path itself is a choice. Pick a generic namespace (com.defenseunicorns.peat.PeatJni), not a consumer-specific one (com.defenseunicorns.<vendor>.peat.PeatJni).
"The workflow copied this version's changelog, so the Release notes are done."
A changelog stanza is version-scoped; Release notes are operator-scoped. Inventory the real baseline, corrective tags, issues/PRs, cross-repo work, evidence, artifacts, and open limitations, then read the published body back.
Scope guards
Kit is the general contractor across all repos. Codex sessions are sub-contractors, scoped to one repo at a time.
Cross-repo changes are coordinated through linked GitHub issues, not by reaching across repos.
Use the public API/traits of other peat repos. Never assume their internals.
Do not add a dependency on peat core that assumes types or traits not yet stabilized — flag assumptions in the PR.
Do not add new repos to the ecosystem without explicit user approval.
Issue format
Each issue used as a Codex spec must include:
## Context
Which repo(s) this touches and why.
## Scope
What is in scope. What is explicitly out of scope.
## Acceptance Criteria
Specific, testable conditions for done.
## Constraints
Architecture invariants, performance requirements, conventions.
## Dependencies
Links to related issues or PRs in other repos.
Gotchas
Populate as sessions run. One line per gotcha plus a Why: line.
Cross-repo changes: develop with [patch.crates-io] path overrides pointing at local sibling checkouts; run the release chain only after the change works end-to-end against the consumer.
Why: serializing through publish-per-layer (PR → merge → tag → crates.io → bump consumer) costs hours per round-trip and you can't iterate the consumer until the upstream publish lands; the override workflow caught a tokio::spawn concurrency race and a cross-version legacy-read BLOCKER before either reached a release (peat#864 chain).
A single surface symptom can have multiple independent root causes at different layers — keep bisecting after the first fix lands instead of declaring victory.
Why: peat#864 ("subscribe_progress emits no terminal frame") was three separate defects — a peat-mesh sync_cooldown silent-drop, a missing peat-protocol watcher, and a wholesale-scalar Automerge merge race — each fixed in a different repo; closing the issue after the first fix would have been wrong.
Don't close an issue when the wire-up lands but the end-to-end acceptance test still fails; reopen with a narrowed title rather than filing a vague follow-up.
Why: peat#864 was prematurely closed at the rc.7 wire-up while its named acceptance test still stalled; reopening with the narrowed scope kept the bisect trail intact.
A corrective release train cannot use only the final tag's changelog stanza as its GitHub Release notes.
Why: peat-node v0.4.18 initially documented only its protoc workflow fix, hiding the sync, packaging, dependency, simulation, qualification, and issue/PR work accumulated since the last complete usable release; the published notes required a full v0.4.15...v0.4.18 reconstruction.
peat repo-specific skill
This repo is a Rust workspace with seven internal subcrates plus example crates. It also hosts the ecosystem skill above.
Workspace members
Subcrates inherit the workspace Cargo.toml's pinned deps (Tokio, serde, thiserror, anyhow, etc.). Cross-subcrate changes coordinate within this repo; cross-repo changes follow the ecosystem skill's "one PR per repo" rule.
peat — top-level crate; eventual home for shared types and traits. Currently a placeholder (see below).
peat-transport — transport abstractions used by the workspace.
peat-persistence — persistence layer.
peat-ffi — FFI bindings (see "FFI conventions" below).
examples/peat-ble-test — workspace example. (examples/peat-tak-bridge moved to the standalone peat-tak repo.)
examples/m5stack-core2-peat is excluded from the workspace (separate toolchain — embedded ESP32 / xtensa-esp-none-elf).
peat top-level crate (placeholder)
The peat subcrate is currently a placeholder. Its intended role is to be the shared dependency that gives the ecosystem a single source of truth for types, traits, and errors.
Intended contents:
Core data types (messages, identities, capabilities)
Shared traits (e.g., Transport, Node)
Error types
Does NOT belong in peat core:
Transport implementations (live in peat-transport, peat-mesh, peat-btle)
Hardware-specific code
Platform-specific code (Android, ROS2)
Until peat core stabilizes, sessions should be conservative about adding dependencies on it and must flag any assumptions about peat-core types in PR descriptions.
peat-ffi subcrate — FFI conventions
The FFI boundary uses two complementary tools:
UniFFI 0.28 in proc-macro-only mode (no UDL file) — generates Kotlin and Swift bindings from Rust types annotated via #[uniffi::export] etc. Async-supported via UniFFI's tokio feature.
Direct jni 0.21 bindings for Android native methods that bypass JNA's symbol-lookup issues. These coexist with UniFFI; use UniFFI as the default and jni only where JNA fails.
JNA (net.java.dev.jna:jna:5.14.0@aar) is the runtime layer UniFFI depends on at the Kotlin side and is required in peat-atak-plugin/app/build.gradle.kts — do not remove it.
peat-atak-plugin consumes pre-built native libraries (copied into app/libs) and copied Kotlin bindings (into app/src/main/java/uniffi/peat_ffi) from this crate's build output. Do not duplicate FFI types or bindings inside peat-atak-plugin — peat-ffi is the single source of truth.
Workflow guards for changes to peat core or peat-ffi
Any new public type or trait in peat core requires a brief design note in the PR description.
Any breaking change to a public item in peat core or to peat-ffi's FFI surface requires a list of downstream consumers that need updating, in the PR description.
Removing a public item requires confirming via cargo check -p <consumer> (and, for FFI, rebuilding peat-atak-plugin) that no consumer breaks, OR coordinating an update PR in each consumer first.
One-sentence statement on Peat's relationship to UDS (peer / complement vs. subset).
Confirm status (active / planned / deprecated / renamed) for repos flagged status unknown in the router: peat-registry, peat-node, peat-gateway, peat-rmw, peat-mavlink. If they exist privately, the router should note that; if they were renamed, the router should reflect the current name; if deprecated, drop them.
WearTAK integration location — inside peat-atak-plugin or a separate repo?
FFI threading model — once stabilized, document the Kotlin-coroutine ↔ Rust-async convention via UniFFI's tokio feature in the peat-ffi section above.
Required reviewers per repo / CODEOWNERS — none currently in peat repo. Adopt CODEOWNERS, or document the review-routing convention in the ecosystem skill.
Last updated: 2026-07-29Maintained by: Kit Plummer, VP Data and Autonomy, Defense Unicorns
