bare_address sliced trimmed[..7] after a len() >= 7 check — byte
count against a byte slice is fine, but the guard and the slice
disagree once the text is lossy-decoded elsewhere; an ATTENDEE line
whose value opens with a multibyte run panicked the reminder source
on a feed the operator does not control. The slice is get(..7) now,
and a test feeds multibyte prefixes through both the ATTENDEE and
ORGANIZER paths.
The watch key encoded source, repo, and number but not the thread
kind, and on GitLab an issue and a merge request share one number
space: iid 7 twice produced one key, one snapshot, and comments on
the merge request landing in the issue (or a 404 fallback that
misrouted the other way). The key carries the kind now —
{source}:{kind}:{owner}/{name}#{number}, with GitLab spelling the
pull variant mr — comment and snapshot route outright by kind, and a
kind-less payload parses to the unparsable key instead of guessing
issue. The claw's callers thread the kind from the payload and strip
the segment where the bare repo is wanted; pre-kind stored keys parse
as None rather than misroute. Four route-pinning wiremock tests cover
the GitLab split, and the GitHub/Gitea clients take the kind without
using it (their issues and PRs share the same REST space).
There is no native tool for this, and images are shell data like any
other, so the surface is a bash function the model pipes into:
curl -sL https://example.com/photo.png | see, or see path.png. The
name reads as a sentence in the pipe position, keeps working as a
plain argument for local files, and is one const in vision.rs if it
ever needs to change.
Two halves, joined by a staging directory. Provision writes the
BASH_ENV init file one non-interactive bash sources; see() execs the
rhema binary's hidden __vision mode, which sniffs magic bytes, stages
them under a 0700 per-process dir, and says so on stdout. 5 MB per
image (the strictest API ceiling), png/jpeg/gif/webp only, and a
refusal exits 2 with the reason on stderr so the model reads the
limit in-band rather than losing the bytes to a silent drop. Attach
runs after every bash command that reaches an exit: ShellTools drains
the staging dir into ToolResult.images, a new core field. Bytes never
ride stdout, so output caps and truncation cannot eat them, and a
staged image attaches even when the command failed. The model staged
it, so it belongs in the context. The caps: 5 MB staged, 8 images per
result, the rest named in the result's text, because a for loop
staging 30 screenshots must not cost a context window.
serde keeps the field default, so old checkpoints load unchanged.
Anthropic carries native image blocks in the tool_result content
array, text first. OpenAI-dialect tool messages (llama.cpp, Copilot,
OpenRouter, DeepSeek) and Gemini carry text only, so each withheld
image is named in a trailing note the model can act on. The
transcript's byte accounting and summaries count the images; the
coding-agent config carries the prompt section, and the render adds
the section only when wired.
Frontends provision only where their bash runs locally: the TUI's
session_agent_config, the headless one-shot, the ACP adapter, the
claw's daemon, the same gate the MCP wire applies, because a session
whose bash runs elsewhere cannot write where the drain reads. The
vision wiring lands before the MCP wire so BASH_ENV names the MCP
file last and its baked predecessor line reaches the vision init;
both functions end up defined. Subagent inner shells inherit the
parent's wiring, so a spawned agent keeps the macro in its own
commands. SPEC's decision bullet, AGENTS.md's scope line, and the
README row land with the feature.
The claw's side of §10: one bridge per process, established at first
agent build and never re-read — the claw is single-instance and its
servers are stateful across calls, so a mid-process rewire would
orphan running sessions' socket paths. The global `mcp.yaml` is the
claw's config scope (no profiles here); the off switch is
`disabled_mcp_servers` beside the rest of claw.yaml, and both agent
construction sites — chat turns and `coding_agent` jobs — layer the
wiring on: the init file and socket and token into `shell_env`, the
post-filter lines into the prompt's `mcp` section, the persisted set
into the config. A frontend without wiring changes nothing, so the
daemon without `mcp.yaml` behaves exactly as before.
The test configs gain the new field; the fd-limit discoveries from
the workspace run are already handled by the harness runs above.
The spec's §10 consumer steps land. The rhema binary routes `mcp`
and everything unrecognized (the external subcommand, which is the
road the init file's stubs take: `rhema __server <name> ...`) to
rhema_mcp::cli::run, whose exit code owns the process; outside a
wired shell it exits 4 silently, exactly the spec's contract. The
claw gains its first subcommand the same way (`claw mcp ...`).
The frontends wire: session_agent_config (the TUI's three call
sites) and the headless one-shot merge the wiring's shell_env into
the agent config and the lines into the prompt's mcp section --
always behind the sandbox gate, so a session whose bash runs
elsewhere advertises nothing it cannot run. The registry keys by
profile name: a profile switch rewires, staying put reuses the
bridge and its already-spawned servers. The bridge itself moved to
a dedicated thread with its own runtime in this pass, which is what
makes wire() synchronous -- a dedicated runtime dropped at return
took the accept loop with it the first time, and a blocking CLI read
starving a current-thread runtime taught the same lesson from the
other side.
Muting persists: disabled_mcp_servers lands beside skills_disabled
in the config file, the settings toggle writes both the live set and
the config through save_config, and the running agent rebuilds so
the very next turn sees the change -- the MCP twin of the skill
toggle, with the round-trip test extended to assert the config.
The process-level pin the spec's §9 asked for: a bash child sources
the provisioned init file from BASH_ENV, calls the generated
function, the stub execs the CLI bin (the same dispatch the rhema
binary routes to), and the envelope prints -- plus the unknown-tool
path exiting 2 through clap's own text. provision and wire grow
provision_with_bin/wire_with_bin test seams so the exec chain can
run without building the frontend; the standalone rhema-mcp-cli
bin target is the seam's exec side.
The Vercel frontend existed to prove the shared harness extraction:
two frontends speaking the same wire protocol, one assembly, one
gate. The proof landed; the deployment didn't, and the crate cost a
full axum server, a Postgres store, and a Sandbox REST client to
keep carrying. It leaves the tree; git history keeps it for the day
a second frontend returns.
The Cloudflare DO is now the harness's one consumer, and nothing
downstream depended on the crate. README, AGENTS.md, and the two
dockerfiles lose their vercel entries; the cloud-agent and
cloudflare-code doc comments name the remaining frontend. SPEC.md
section 11 still tells the two-frontend story; it stays until the
direction settles.
The old ladder doubled from one second and capped at two: three
attempts, three seconds — enough for a hiccup, nothing for a
reassociation or a real outage. The policy now takes two explicit
ladders and picks by error class. A dead link (Network) is the local
machine's problem and waiting can fix it, so it walks 1s, 3s, 30s,
60s, 300s — six attempts, six and a half minutes all told. A
struggling server (Timeout, 408, 5xx, an exchange that closed before
any content) fixes itself in seconds or not at all inside a turn,
and every timeout attempt also burns the request's own timeout, so
its ladder stays at 1s and 2s. Each ladder's length sets its attempt
count, so the policy can no longer disagree with itself.
Each wait is announced as an Event::RetryWait (error, delay,
attempt, total) before it runs. rhema-code carries it as a deadline
and the status strip counts it down — the error in the alarm color,
the attempt, the seconds left — while any event from the next
attempt clears it. Esc still cancels mid-wait: the executor drops
the future at the sleep. ACP, claw, subagents, and the cloud wire
ignore the event; from_event returns Option now, so a frameless
event sends no frame. The hosted frontends keep the short ladder on
both classes: their invocation outlives no five-minute wait.
The workspace manifest globs crates/*, so cargo fetch parses every
member's Cargo.toml; rhema-sandbox-ecrou joined the tree without its
manifest joining the COPY list, and both image builds died at manifest
parse ("failed to read /build/crates/rhema-sandbox-ecrou/Cargo.toml").
Add the line to claw.Dockerfile and sandboxd.Dockerfile, in the
alphabetical position the list keeps.
The images resolve the whole workspace graph before cargo fetch, so
every member manifest has to be in the build context by then, and
the copied list had fallen six crates behind the workspace: the
backend split added sandbox-local, sandbox-lxd, and sandbox-server,
and the cloud frontends added cloud-agent, cloudflare-code, and
vercel-agent. rhema-code path-depends on the sandbox backends, so
the claw image was the first to break — the fetch died reading the
lxd manifest.
The pinned toolchain's clippy grew opinions the older lint let stand:
nested if/if-let pairs it wants as let-chains, and a modulo test it
wants as is_multiple_of. Machine-applied collapses, fmt-settled, no
behavior change — the gauntlet runs clean again.
The (+) door stages gallery picks and camera shots as removable
chips against the daemon's own budget (eight a message, 5 MiB
decoded each, refusals surfaced, never silent drops); send reads
the files back as bare base64, the old {"text"} body unchanged
when there are none. Sent rows preview their staged files at once
and a replay fetches by attachment id like any attachment card.
Model options carry the vision badge only when the listing knows.
The (+) door and the clipboard door both stage files as removable
chips, the claw's own budget enforced client-side (eight a message,
5 MiB decoded each, errors surfaced, never silent drops). The sent
row renders local previews at once; the stream's user_message echo
carries the stored ids and dedupes against it, so a replayed message
renders its thumbnails through the same bearer-guarded fetch as an
attachment card. Model options mark vision when the listing knows.
The models route's cards serialize as rhema-provider's ModelCard
spells them, so supports_vision rides exactly when the adapter knows
it and is absent when it does not — unknown stays absent, never
false. The route's doc now pins that contract, and a listing builder
on Backend lets the surface tests mount one offline.
The shim's last user message now cuts the image_url parts along with
the text: an inline data URI with a parseable mime rides to the turn
as an image block, a remote http(s) url and an unknown mime are
skipped, and the messages route's image budget applies. Text-only
requests are unchanged.
A chat message may carry inline images: the messages route checks the
budget (5 MiB decoded per image, eight a message) and the dispatch
copies the bytes into the attachment store, so the replay streams ids
and media types over SSE, never megabytes. The turn's input carries
the inline image blocks the providers take, base64 verbatim — the
session checkpoint holds what the model saw.
Providers map images to per-backend content blocks, so the block rides
the message seam beside the tool blocks and round-trips through session
checkpoints. Untagged serde keeps old payloads loading; the new variant
sits last so the field sets disambiguate. Exhaustive matches on Content
across the adapters and views gained a mechanical Image arm; sending
images is a later effort.
privacy.rs bound the opened handle in a name only the unix build
used, so `-D warnings` failed the wasm target; the bind is now
`_file` and the unix branch uses it.
The docker images resolve the whole workspace manifest graph before
`cargo fetch`, so they now copy the tools contract with its workspace
root plus the sandbox manifests rhema-code path-depends on, and the
stub loop covers tools/*/ too.
The diagnostics page showed the status job "waiting" while its runs
and the report-now kick died without a trace: any exception before the
first log line was invisible, indistinguishable from a job that never
starts. Every worker now wraps its run — a crash logs its type, message,
and first frame; a cancellation logs as such and rethrows. The screen
gains the report-now kick's own row (a one-time job's empty list is
"never run", not an error) and the tap lands in the log, so the next
stale-row mystery names its cause instead of hiding it.
A stale "last seen" had no client-side answer: a dead worker, a down
claw, and a claw that drops writes all looked identical from the
settings screen. The new page separates them: the WorkManager jobs'
live state (a cancelled schedule renders as "not scheduled"), the two
clocks — the phone's last landed report against the claw's last stamp
— whose disagreement names the broken end, auto-detected system checks
(battery optimization, background restriction, notifications) with
jumps into the exact settings page, a report-now kick, and a 200-line
in-memory flight recorder where every worker run logs its outcome.
One [workspace.lints.rust] table replaces the deny attribute
scattered across every lib.rs, and each crate opts in through
[lints]. The lint now also reaches the integration-test crates,
whose roots gain their one-line docs.
BackendError and WatchShapeError replace the String registry and
watch seams, the tool operations report ToolError, the accessors
drop the get_ prefix, the too-many-arguments allows become
expects, and the llama catalog passes through unchanged.
The last forge content in config.rs was vocabulary, not data:
`ForgeKind` (name/parse/token_env/default_base) and the pure
`resolve_token` helper. By the same rule that moved `Watch` to the
watch table and left `CalendarWatch` in the calendar integration,
the kind belongs to the crate that speaks the forges.
- `rhema-forge-integration` gains `kind::ForgeKind` and
`resolve_token`, with the parse and token-fallback tests moved
alongside — the claw's config tests were testing the integration's
type through a re-export.
- The claw's surfaces import the type from the integration directly;
the crate root re-exports it, so `rhema_claw::ForgeKind` keeps
resolving for everything downstream.
config.rs is now config: credentials, the surface, paths, cadence,
and nothing that knows what a forge or a calendar is.
The seam crate carried every backend behind features — local, client,
server — so a wasm worker linked the axum daemon and a host daemon
linked reqwest it never dialed. The family mirrors the provider family
now, as the design always said it would: rhema-sandbox keeps the
contract, the ShellSandbox adapter, the wire protocol, the seeding
helper, the scripted mock, and — behind the new `spawn` feature — the
process family's shared machinery (exec_util, fsroot); the backends
are sibling crates, rhema-sandbox-local, -docker, -ssh, each dependent
only on the seam; the HTTPS client is rhema-sandbox-http (reqwest
behind its `native` feature, off on wasm, whose worker brings its own
HttpTransport over host fetch); the daemon is rhema-sandbox-server,
and the rhema-sandboxd binary moves with it.
Consumers rewire, semantics don't: rhema-code names its backends
directly; rhema-cloud-agent keeps its transitional re-exports;
rhema-cloudflare-code and rhema-vercel-agent take the HTTP client
reqwest-free and native respectively; the sandboxd image builds
-p rhema-sandbox-server. Cargo paths only — no behavior, no protocol,
no spec change.
The streaming progress seam grows from a bare output callback to
ToolProgress: output chunks as before, plus an optional ToolStatus —
the phase a tool's inner work is in and, when the tool runs model
calls of its own, their cumulative spend and the model they ran on.
The engine stamps the call's id on it and the sink receives it as an
ephemeral ToolStatus event, replacing any earlier one for the call;
status is screen-facing, never history.
The subagent tool is the first emitter: its inner turn's lifecycle
becomes phase transitions (thinking, running, writing output) and a
running token total. Every sink decides what a status is worth: the
cloud wire forwards it, ACP and the claw have no slot and drop it.
The calendar table gets the same contract treatment as every other
data domain: a named route constant, three allowlist rows (GET, POST,
DELETE /v1/calendars/{name}), mock-router arms that enforce the
door's loudest rules — a named calendar, an ics with a URL — and
conformance cases (add, the invalid refused, listed, removed) that
run against the mock and the real daemon alike, so the two cannot
drift. The claw's auth middleware reads the same list, which is how
the rows land there with no claw-side code.
The spec names the doors and the pattern they establish: domains the
daemon reconfigures at runtime live in tables behind their routes,
validation runs at the door, and claw.yaml carries only what boot
needs.
Same move as the watches, and the config's last piece of integration
state goes with it:
- `rhema-calendar-integration` owns the calendar table now
(`store.rs`): SQLite in the shared dispatch database, the pattern
the claw's watch table established. `CalendarWatch` gains the
`validate` the config's boot check used to run — per-kind URL and
credential rules, plus the lead-under-horizon check — so an invalid
calendar is loud at the one door every writer shares.
- `CalendarSource::from_store` reads the table fresh on every detect
pass (the forge source's pattern): a calendar added over the route
is live next tick, a removed one stops polling, no restart in the
story. The static constructor stays for tests.
- `POST/DELETE/GET /v1/calendars` on the daemon: the POST body is the
integration's `CalendarWatch` shape, validated at the table; DELETE
takes `?name=`; the api listing names the door.
- `ClawConfig.calendars` is gone. A config still carrying it is
refused at parse time, like the retired `watches:` seed. The
config-side validation and its tests moved with the type.
The config file now carries only what boot needs: credentials, the
surface, paths, the poll cadence. Everything the daemon reconfigures
at runtime lives in its tables, behind its HTTP doors.
The YAML `watches:` seed was the last piece of forge state the config
file carried. Everything else about watches already lived in the
extension: the table (SQLite, runtime-managed), the HTTP routes
(`GET/POST/DELETE /v1/watches`), the chat tools, the poller that reads
the table fresh every pass. A static seed list on top meant two homes
for one fact — and the seed's own semantics (import-once, ignored
once the table has rows) had to be explained to every reader.
Now the config carries none of it:
- `ClawConfig.watches` is gone, with the `Watch` seed type and the
boot-time validation loop that guarded it. A config still carrying
`watches:` is refused at parse time — `deny_unknown_fields` turns
the retirement into a loud migration, not a silent no-op.
- `Watch` moved to `watch_store` (re-exported from the crate root),
where the row's vocabulary belongs; `validate_watch` keeps guarding
the one door every writer shares — the routes and the tools both
call it, so the moved validation tests test it there.
- `seed_from_config` and the daemon's seeding are gone; the operator
seeds once over HTTP or `watches add`, and the table stays true.
The one guarantee the seed carried that nothing else did — boot
refuses a tokenless watch — goes with it, deliberately: a tokenless
watch now fails at its first poll, loudly, like any other watch the
agent adds at runtime. Boot-time certainty was paying for config
state the daemon no longer owns.
The agent token is deliberately barred from /v1/ingest: the events
are the inbound channel, and the model must not feed it itself. But
the same bar closes the door on every honest integration too — a
sidecar poller with events to hand in had to hold the human token,
which is wider than its job. This adds the missing class: a third
bearer, the ingest token, answering the event door and nothing else.
`ingest_allowed` is the whole credential in one function — one
route row, `POST /v1/ingest` — symmetric with `agent_allowed` and
suite-covered the same way: the mock and the daemon's real router
both answer the new checks (a push succeeds with the contract header;
a watch read and a consent answer are 403). The middleware classifier
gains the third arm, boot refuses an ingest token equal to either
other credential — one value playing two roles silently widens the
scoped one — and the daemon provisions it to tool shells as
`RHEMA_INGEST_TOKEN`, so a tool that only pushes events never sees
the wider agent token at all.
The door is the first half of the extensions story (rhema-tools spec
§8): a self-timed sidecar is now a tier-1 tool with a credential
narrow enough to hold.
The claw is the server half of the client tools. The middleware now
knows two bearer classes: the human token answers every /v1 route as
before; the agent token answers only the routes the contract
allowlists, and only with the contract header at the matching version
— consents, sessions, ingest, and the phones answer it 403, so a
model holding the token cannot prompt-inject its way past the scoping
to answer its own secret asks. Boot refuses the configuration where
agent_token is the human token: the scoping is the safety, and one
value playing both roles silently unscopes everything the shell
holds.
Three new routes write the tables the resident already reads: push to
the notification queue, schedule and cancel tasks the resident's tick
fires — the tools get native-shaped writes without the daemon growing
a second brain. A JSON fallback gives the one error shape everywhere,
unknown routes included. The daemon composes the tools' environment
(RHEMA_URL, RHEMA_TOKEN, the tools' bin dir on PATH) and hands it to
the resident's shell through the agent-tools env layer; with no
agent_token it provisions nothing, exactly as before.
The proof is the contract's conformance suite run against this router
over real HTTP — same answers as the mock, checked at test time.
Desktop notifications are the wave-1 built-in the resident has no
native way to send: the daemon's pushes land in the poll queue, and
this is the program that turns one into an OS notification —
`notify push --title --body [--device]` on the shared plumbing, the
claw's `POST /v1/notifications` on the far end.
The binary is the template for every later client: connection from
the environment the daemon provisions, one contract-headered call,
JSON or error shape, exit code. The e2e test drives the real binary
against a mock claw — a wrong token is a 401 with the code to match,
a consent route is a 403, a good push prints the row — so a packaging
regression cannot pass silently. `skills/` ships the markdown that
teaches the agent the tool exists and what its answers mean.
A client tool is three moves: read the connection from the
environment, make one HTTP call carrying the bearer and the contract
header, and print either the answer as one JSON line or the one error
shape — then exit with the code that says which happened. Common is
those three moves, so no client reimplements them and all clients
fail identically: `Config::from_env` (`RHEMA_URL`, `RHEMA_TOKEN`),
the `Client` call with the contract envelope, `print`/`fail` and
the `ToolError` exit codes.
The HTTP client is `ureq`, blocking, no TLS: the contract speaks to
a loopback or otherwise trusted daemon, and a remote TLS endpoint is
a stated non-goal in v1 — a thin client stays thin.
The client tools need one shared definition of what a tool may do,
and the daemon needs to enforce exactly that definition, not a
hand-copied copy of it. This crate is that definition, pure: no HTTP
client, no process — the allowlist as data, the version header as
constants, the request/response and error types, and the conformance
suite that drives any transport through the checks the contract
promises (the bearer gate, the version handshake, the allowlist's
denials, round-trips through every allowed route, one error shape).
`agent_allowed` is the whole security surface in one function:
method and path segments with `*` and `**`, exact otherwise — a
daemon's middleware asks it and nothing else, so the claw's scoping
and the tools' expectations cannot drift apart. The suite runs
against a mock router in this crate's tests and against the daemon's
real axum router in the claw's; same results is the definition of
implementing the contract.
A multi-stage build shipping the daemon: bash, git, and a CA store in
the runtime so git and tarball seeds work; the base directory is a
volume, the token is a flag with no env default, the port must be
passed because the binary defaults to an OS pick. Smoke-tested against
a real daemon: healthz auth, exec, put/get, session dirs under the
mounted base.
Fold thinking deltas into the turn's stream-ordered entries as a
thinking entry, merged back-to-back the way the stream delivers
them, instead of one flat string rendered above every call. Each
thinking block now sits where it streamed — between the calls it
precedes. Live turns keep their thinking expanded; finished turns
render it collapsed.
phone loads its spec from an embedded manifest under
crates/rhema-claw/tools; the Rust side keeps only the invoke half.
It stays replay-unsafe: a resumed run must not re-queue a control
command or re-retire a device. With this every registry-served tool
spec in the daemon comes from a manifest; only the ToolResolver-style
tools (identity_exec) keep their specs inline.
coding_agent loads its spec from an embedded manifest under
crates/rhema-claw/tools; the Rust side keeps only the invoke half. It
stays replay-unsafe: a resumed run must not re-enqueue a coding job
against an instance that is already running it.
ask_secret loads its spec from an embedded manifest under
crates/rhema-claw/tools; the Rust side keeps only the invoke half. It
stays replay-unsafe: a resumed run must not re-prompt the human or
re-store the credential the first ask already landed.
attach_file loads its spec from an embedded manifest under
crates/rhema-claw/tools; the Rust side keeps only the invoke half. It
stays replay-unsafe: a resumed run must not re-attach a file the
transcript already carries a card for.
identity_set loads its spec from an embedded manifest under
crates/rhema-claw/tools; the Rust side keeps only the invoke half. It
stays replay-unsafe - a resumed run must not re-write the identity.
identity_save, identity_list, and identity_forget load their specs
from embedded manifests under crates/rhema-claw/tools; the Rust side
keeps only the invoke half. identity_list declares replay_safety:
safe - the store answers with its own redacted records - while save
and forget stay unsafe: a resumed run must not re-store an account or
drop one twice. identity_exec keeps its spec inline; it is a
ToolResolver, the exec seam itself, not a registry tool.
A failed post returns retry(), which moves the next run onto
WorkManager's exponential backoff — 30s doubling, capped at five hours.
A couple of failures while the claw is unreachable is enough to leave
the fleet without a report for most of a day, which reads exactly like
the phone being gone.
The snapshot is only saved when every post landed, so the next
scheduled run re-derives the whole transition set and posts it again.
The fifteen-minute schedule is the retry; return success and keep it.
The fleet card reads last-seen from the claw, and only a landed report
moves it. Periodic work is deferred while the device idles — Doze
stretches the fifteen-minute floor to hours — so a phone sitting in the
human's hand could show itself as "last seen 3h ago" with nothing wrong
on the claw.
Kick a status run on resume, the same way the commands worker already
is, and reload the fleet whenever a run lands. The landed-run signal is
a new clock in StatusRepository, written in the same edit that saves the
snapshot (which only happens once every post of the run succeeded), so
the card reloads on a report that actually reached the claw.
notify loads its spec from an embedded manifest under
crates/rhema-claw/tools; the Rust side keeps only the invoke half.
The drift test pins the shipped set: the manifest loads against its
builtin, the registry serves the manifest spec, and notify stays
replay-unsafe - a resumed run must not re-send a notification.
watch_repo, unwatch_repo, and list_watches load their specs from
embedded manifests under crates/rhema-claw/tools; the Rust side keeps
only the invoke half. The manifests stay name-sorted, which is the
order the registry serves and the roster shows. list_watches declares
replay_safety: safe - re-reading the table on a resumed run is free -
and the drift tests pin the shipped set: every manifest loads against
its builtin, every builtin is named by a manifest, and replay safety
reaches the resolver from the file.
bootstrap_save now loads its spec from an embedded manifest, one
folder under crates/rhema-claw/tools; the Rust side keeps only the
invoke half, behind the run.builtin id. Drift tests follow in the
registry's own suite shape once the family grows; for now the
registry asserts the manifest loads, and the render tests exercise
the manifest-served spec through the roster.
The section-header restructure renamed the onboarding offer from
"## Setup" to "## Onboarding" in ONBOARDING.md; three prompt
assertions still looked for the old heading and failed. They follow
the file now.
vhaudiquet
