# omp-spt capability parity record > This record compares `omp-spt` with its sister adapter, `claude-spt`. > Parity means the same user outcome, including agent-facing outcomes, through > the best OMP-native seam or a discoverable public `spt` command. It does not > require Claude command names, hooks, PTY translation, or source structure. > Worker endpoints remain intentionally excluded. ## Versioned comparison baseline Current baseline: `omp-spt v0.9.4` → `BigscreenVR/claude-spt-bs v0.38.6`. `omp-spt` v0.8.0 was published at 2026-09-06 10:56:17 UTC. The first repository commit after that release, `ebb3bf77d8f0e0f8556d7c5a4791a95a1540d075`, landed at 2026-09-06 10:57:00 UTC and fixes the start of the work interval that produced v0.9.0 (the context ledger, `cfee63f` at 2026-09-06 11:43:15 UTC — operator design, replacing every per-request and per-turn injection with one durable block). `claude-spt-bs` v0.38.6 was still the latest published sister release when that interval opened (published 2026-09-06 09:53:41 UTC); no sister change was triaged in it. The previous baseline (`omp-spt v0.3.32` → `claude-spt-bs v0.25.1`, sister published 2026-07-19 04:55:53 UTC) identified the sister behavior used to define v0.3.20 parity. This baseline records the sister behavior triaged by parity pass v2 (below). It is not a claim that every later `claude-spt` capability is covered — the matrix says which are. Any subsequent feature-parity work must first select and record the exact published `claude-spt` release being consulted. Every parity-informed `omp-spt` release must update this baseline to its own version; the release version-consistency gate rejects a stale `omp-spt` baseline. ## Parity pass v2 — sister v0.25.1 → v0.38.6 (closed 2026-09-06) Sixty `claude-spt-bs` releases (v0.25.2 … v0.38.6) were triaged against the native OMP host. Dispositions: **covered** (same outcome here, by the same or a stronger mechanism), **N/A** (Claude Code plumbing with no OMP analogue), **divergence** (deliberately different, recorded), **unverified** (plausibly covered, needs a test before it is claimed), **gap → phase N** (built by the numbered phase of `PARITY-V2-PLAN.md`; each phase advances this baseline when it ships). | Sister change | Releases | Disposition | |---|---|---| | Interrupt/stuck-session watcher reads only real agent work | 0.25.2, 0.25.5 | N/A — CC transcript watcher; OMP turns are observed natively | | Skill short forms, inline skill-size limits, picked-for-you skill instructions | 0.25.3, 0.25.4, 0.28.0 | N/A — CC skill-stub mechanics; OMP `omp-plugins` skills load whole | | Hook additional-context cap (~2 KB) spill files, per-delivery files, spill-note failure modes | 0.25.9, 0.25.11, 0.25.17 | N/A — no injection cap on the native boundary splice | | Borrowed/inner CC sessions must not act as the agent; stale host pid after daemon restart; digest follows config root | 0.25.10, 0.25.12, 0.25.13, 0.25.18, 0.25.20, 0.25.22 | N/A — one endpoint ↔ one OMP session, bound in-process at activation (REQ-OMP-SESSION-IMMUTABLE) | | Psyche/summarizer sessions run without user hooks or plugins | 0.25.7 | covered — `psyche-omp` / `echo-commune-omp` are bounded helper turns with isolated context roots (REQ-PSYCHE-EPHEMERAL-SHIM) | | Companion failure carries the inner session's stdout/stderr | 0.25.14 | covered — `psyche-omp` / `echo-commune-omp` failure lines carry a bounded tail of BOTH inner streams and name a silent channel explicitly (REQ-COMPANION-FAILURE-TAILS, 0.6.0) | | Account/usage-limit refusal reported as outage-not-fault; endpoint held unavailable until the stated reset, re-asserted against stray recoveries, respected across daemon restart | 0.25.15, 0.25.25, 0.25.26, 0.25.27, 0.25.28 | covered — a turn refused for account reasons (OMP `errorId` UsageLimit / 402 / refusal wording) is reported as an outage, not a fault; the endpoint is held `busy` until the provider's stated reset + 1 min, re-asserted every 5 s, with deliveries custodied unsubmitted and the liveness ladder paused; a human prompt releases it; the session record re-arms it after a restart (REQ-USAGE-LIMIT-HOLD, 0.6.0). The Psyche turn maps the same refusal to the published exit 96 (REQ-PSYCHE-ACCOUNT-REFUSAL) | | End-of-turn shortform send fires at turn end, on tool-using turns too; first-turn send after setup not dropped | 0.25.16, 0.25.21 | covered — dispatch runs on OMP's `agent_end` with the final assistant text in hand (REQ-PARITY-PEER-SHORTFORM) | | Spooled vs delivered vs local-failure reported distinctly; sends logged; end-of-turn confirmation shown next turn | 0.25.23, 0.25.32, 0.29.2, 0.29.4 | covered — since 0.4.0 the adapter carries no shortform parser; spt-core reads `@<…@>` from the turn feed and reports every dispatch outcome (delivered, spooled, no perch) in the now-signal's `DISPATCH_RESULTS` at the next boundary, and the startup brief says outcomes arrive only there (REQ-PARITY-PEER-SHORTFORM). Recorded as covered 2026-09-06; the row had kept its phase-2 wording after phase 2 shipped | | Inline sends carry the same sender proof as shell sends | 0.28.2, 0.28.3 | covered — core dispatches inline sends from the session's own turn feed (`[io] compliance = true`, 0.4.0), so they carry the session's proof exactly as a shell `spt send` does. Recorded as covered 2026-09-06 | | Message arriving just before `/compact` no longer stalls delivery | 0.25.6 | covered — deaf-idle self-heal ladder (REQ-HAZARD-DELIVERY-TURN-LIVENESS) restores receivability regardless of cause | | Messages surface only in the main conversation, never a subagent's | 0.25.8 | covered — OMP re-runs every extension factory for an in-process subagent session (no subagent marker exists); the nested copy used to run a second `api bind` (field: hertz 2026-09-06, refused by core's ownership guard — F-037). Since 0.7.0 exactly one binding per process activates and every later one is inert (REQ-HAZARD-NESTED-ACTIVATION, KNOWN-HAZARDS #15) | | Hook log names its endpoint | 0.25.19 | covered — every extension log line is prefixed `[omp-spt ]` through one wrapper (REQ-ENDPOINT-NAMED-LOGS, 0.6.0) | | `spt ring` guidance withdrawn, then restored with a version floor for perchless sessions only | 0.25.24, 0.25.30 | covered — CI.md acceptance floor (installed spt ≥ v0.46.0); startup brief offers ring only to a perchless session | | Summarizer never lists machines/hosts as peer agents | 0.25.29 | covered — the echo-commune policy prompt names another agent only on a direct exchange and never presents a machine, host, or node name as an agent or teammate (REQ-ECHO-NO-MACHINE-PEERS, 0.6.0) | | Linux members packaged executable; reproducible archives | 0.25.31 | covered — archive writer sets explicit `0755` modes and `SOURCE_DATE_EPOCH`; packer refuses a non-executable Linux member | | `/knock` skill; sender-declared send-only / send-receive wording; invite-code subnet sealing | 0.26.0, 0.28.1 | covered — packaged knock skill + `[[hints]]` row (REQ-KNOCK-SKILL, 0.5.0) | | Long foreground command → background-it reminder | 0.26.0 | divergence — omp-spt nudges AFTER the fact from OMP's measured wall time (foreground `bash` ≥ 30 s, one `[spt]` line appended to the result, ≤ 3 per turn, advice only) instead of a pre-run heuristic (REQ-LONG-FOREGROUND-NUDGE, 0.6.0). Field-verified on hertz 2026-09-06 (35 s `sleep`, line quoted verbatim). Window: OMP itself backgrounds a foreground `bash` at `bash.autoBackground.thresholdMs` (default 60 s) and backgrounds early when a steer arrives, so the nudge fires for completed foreground runs of 30–60 s; longer ones OMP already backgrounds | | Deliveries during a save-and-clear turn held for the next session; hold self-releases | 0.26.0 | covered — from arming `spt_checkpoint` until its wake (or failure notice) is queued, inbound deliveries stay custodied unsubmitted and the deaf-session ladder pauses; the release resubmits them and OMP orders the wake turn against the delivery turns (queued steers survive native compaction) (REQ-CHECKPOINT-DELIVERY-HOLD, 0.7.0). Field-run on hertz 2026-09-06: a peer message sent after arming surfaced only in the post-reset turn, right after the wake | | Keyword tips actually delivered, once per session | 0.26.1 | covered — targeted hints unit-tested (REQ-PARITY-TARGETED-HINTS); phase 2 migrates them to the core briefing | | Matched monic note arrives ahead of the message | 0.26.2 | covered — `[spt]` note per matched record ahead of the opening tag, attribute still rides (REQ-MONIC-NOTE-AHEAD, KNOWN-HAZARDS #13, 0.5.0) | | "You are reachable" only after registration succeeds; refusal quoted | 0.27.0 | covered — activation announces only after a successful bind; a failed bind reports the reason and leaves the session unactivated | | Trust warning rides the message envelope | 0.27.0, spt-core #170 | covered — REQ-HAZARD-ENVELOPE-ATTRIBUTE-PASSTHROUGH (0.3.33) | | Alt Anthropic accounts: profiles, shared work history, plugin store, update from alt session | 0.29.0, 0.29.1, 0.29.3, 0.29.5 | N/A — OMP's native provider/profile routing already covers alternate accounts (explicit non-goal below) | | Sealed message arrives with its seal shown and verify commands named | 0.30.0 | covered — `[spt] SEALED by … seal=` note names `spt api seal verify` / `describe` (REQ-SEAL-SURFACE, 0.5.0) | | Every envelope attribute survives delivery | 0.31.0 | covered — REQ-HAZARD-ENVELOPE-ATTRIBUTE-PASSTHROUGH | | Damaged envelope: body delivered with a warning | 0.31.0 | divergence — omp-spt fails closed on protocol corruption (listener-fatal, REQ-tested); a corrupt frame is never presented as a delivery | | Turn text (user prompt, every assistant message, closing text) reported to spacetime's feed, exactly once | 0.32.0, 0.33.0, 0.35.0 | covered — REQ-IO-TURN-FEED (0.4.0): `api state --payload-stdin`, mid spans `--mid`, stub turns carry no USER_INPUT | | Per-turn "what changed?" briefing at turn start and mid-turn; agent's own words feed keyword matching | 0.33.0, 0.33.1, 0.34.0 | covered — REQ-NOW-SIGNAL-INJECT (0.4.0): `api now-signal --spec-manifest` at turn start and every in-turn boundary; hints are manifest `[[hints]]` | | `@<…@>` and `;;` read by spacetime, not the adapter; results in `DISPATCH_RESULTS`; reset-turn send still goes out | 0.36.0 | covered — REQ-IO-COMPLIANCE + REQ-HAZARD-SHORTFORM-DOUBLE-FIRE (0.4.0): `[io] compliance = true` landed with the parser deletion; the message that invokes `spt_checkpoint` is a `toolUse` stop, reported at `message_end` before the tool compacts, so a tag in it goes out | | Commune-across wake marker detected whatever tool wrote the file; marker near the top; quoted markers inert | 0.37.0, 0.38.0, 0.38.1, 0.38.2 | N/A — omp-spt's checkpoint is the explicit native `spt_checkpoint` tool (REQ-PARITY-CHECKPOINT); no file-marker parsing exists to misfire | | spt-core floor raised (v0.64.0, v0.66.0) for the feed | 0.32.0, 0.35.0 | covered — omp-spt floor is v0.67.0 (0.3.33) | | Commune drop no longer mistaken for a Psyche summary (boundary armed only from a self-written commune); one more tool call after a commune protects the wake | 0.38.3 | N/A — omp-spt's checkpoint is the native `spt_checkpoint` tool; no drop-file marker or race with the Psyche summary | | Boundary armed only from a commune written during the current session; the newest commune in an io-events replay wins; the record names the commune and matched text | 0.38.4, 0.38.5 | N/A — omp-spt reads no `io-events` and arms no boundary from communes (its checkpoint is the native `spt_checkpoint` tool), so a replayed or re-numbered history cannot clear a live session | | "One more tool call after a commune" guidance narrowed to across-communes | 0.38.6 | N/A — omp-spt has no commune drop file to race and no wake marker to arm; its checkpoint is the explicit native `spt_checkpoint` tool, which since 0.7.0 arms in the tool and resets after the turn ends (KNOWN-HAZARDS #16) | Not sister-driven but absorbed by the same pass: spt-core v0.67.0 retired `spt endpoint run` (`ENDPOINT_RUN_RETIRED`); every bringup this adapter teaches now names `endpoint create` / `start` / `resume` / `spt go` (0.3.33). ## Proven shared endpoint surface The current adapters already share release install/update, broker-hosted fresh launch and attachment, native resume, ReadyAgent and LiveAgent roles, post-spawn identity, inbound delivery, explicit outbound messaging, busy/idle lifecycle, graceful shutdown, Psyche turns, commune/signoff storage, echo-commune, opaque history, digest extraction, and Windows x86-64 plus GNU Linux x86-64 release payloads. Hosted OMP matches the sibling session-name shape, ` @ (/)`, and improves fleet scanning with an idle glyph or animated busy spinner in the terminal window title. OMP keeps its stronger native invariants: immutable endpoint/session binding, one bounded custody queue, finite listener recovery, fail-closed exhaustion, real OMP executable validation, and strict separation between local assistant output and explicit outbound peer messaging. Capability parity does not weaken those guarantees. ## Agent-facing parity decisions Ordinary OMP sessions gain extension-native `/ready` and `/live` activation commands with argument completion or native selection dialogs. Explicit live auto-resume selects the most-recent compatible live identity; ordinary activation never guesses silently. The extension binds the current session and owns the authenticated listener, after which binding is immutable. Messages accepted during an active OMP turn enter at the first proven boundary before the next tool or model continuation. Custody and arrival order remain intact; a turn with no injectable boundary falls back to ordinary next-turn delivery rather than interruption. Cancelled, interrupted, and failed turns settle affected custody and restore receivability automatically. Only exhausted listener recovery closes the endpoint. An activated session pulls its durable mind — role, cross-project live context, and the current project's context, plus any commune or signoff drop not yet synthesized into them — and receives it as opening context ahead of the startup brief. The pull is authenticated by the bind token, so reading the mind never re-pins the perch; a fresh endpoint with nothing stored yet injects nothing, and a failed pull is reported without failing activation. Activated sessions receive a concise startup brief for identity, roster, messaging (shortform-by-core, `DISPATCH_RESULTS`, seals), continuity, lifecycle, subnet, version, and update operations. Its messaging line spells the shortform grammar out (`@`: the opener is `@<` glued to the id, `@`), after todlando opened with `@doyle … @>` on 2026-09-06 and had to resend. The brief is the head of the **context ledger** (below), so every provider request of the session carries it. Everything situational after that arrives through spt-core's one funnel, `api now-signal --session --spec-manifest`: the extension polls it at turn start with the user's words (none on a delivery-stub turn), and again at every in-turn context boundary with the assistant's words since the last poll, and logs every non-empty answer — every kind, verbatim, timestamped — in the context ledger, which the next provider request carries. It is delta-only, so a quiet boundary adds nothing; a failed poll is a soft comms fault and never a broken turn; inline words are clipped to 8 KB. Targeted hints for clear live-activation, identity, messaging, subnet, and checkpoint intent are declared as manifest `[[hints]]` (regex keywords, first match wins) and surface once per session through the `HINTS` category; the versions in play (spt-core, this adapter) surface through `UPDATES` whenever they change, so the adapter runs no version probe of its own. ### The context ledger (0.9.0) OMP applies an extension's `context` hook per provider request (`transformContext` inside `prepareProviderCall`) and never writes its output back to the session store, and it discards the `before_agent_start` system-prompt override when the turn ends (`agent-session.ts`: "the per-turn before_agent_start override lives only for this turn"). claude-spt's hook `additionalContext` lands in the Claude Code transcript and persists; on omp-spt, everything added those two ways — the brief, the durable mind on resume, every now-signal, every delivery polled while busy — was seen for one turn or one request and then gone. KNOWN-HAZARDS #10 was the first symptom (delivered bodies); the ledger closes the class (#18). The ledger is one hidden custom message (`spt-context-ledger`) re-spliced as the **last** message of every provider request of an activated session, in-turn and off-turn alike: a head — the durable mind (prepended when its pull resolves), then the startup brief — and an ordered log of everything spt added since the last reset, each entry stamped `[YYYY-MM-DDTHH:MM:SSZ] `: `now-signal` (every kind, verbatim, turn-start and boundary alike), `delivery` (arrival time and sender; the body stays durable through the stub, #10), `delivery (polled while busy)` (the envelopes verbatim — they had no other durable home), `reset`. Order and timing are part of the memory: nothing is filtered or summarised (operator ruling, 2026-09-06). Tail placement keeps the transcript prefix cache-stable; a changing system prompt would invalidate the provider cache for the whole session on every change. The log is bounded (32 KB by default, `contextLedgerBytes`), evicts oldest first with a visible count, clips a single entry at 16 KB, and clears on `session_compact` — the head stays and the reset is the new log's first entry. A new session gets a new ledger at activation. The old per-turn system-prompt override and the per-request `spt-now-signal` / `spt-event` messages are gone. Field check on hertz (0.9.0, stopped + resumed 11:49:50Z, 2026-09-06): asked to quote the block, it returned the header sentence verbatim and, on turn 1, two entries — `[2026-09-06T11:50:20Z] delivery`, `[2026-09-06T11:50:20Z] now-signal`; on turn 2, four — the same two followed by `[2026-09-06T11:51:10Z] delivery`, `[2026-09-06T11:51:10Z] now-signal` — the turn-1 entries persisted, in order, timestamped. A native checkpoint then compacted the session and the wake turn quoted a one-line log: `[2026-09-06T11:55:00Z] reset` — the head intact, the log restarted. Every turn's text is reported to spt-core's IO feed exactly once per span: the user's prompt opens the turn as `USER_INPUT` on `api state busy --payload-stdin`, each completed assistant message goes out as a mid-turn `AGENT_OUTPUT` span (`--mid` — at `message_end` for a `toolUse` stop, so a peer named in it never waits on a long tool, and at the next context boundary as the catch-all), and the closing message rides `api state idle --payload-stdin`. Payloads travel on stdin only. A payload-carrying call is an event and never takes the same-state short-circuit; a failed report is logged and the turn completes. A delivery-stub turn (a peer's `` woke the model) carries no `USER_INPUT`: that body is already core's `MSG_IN`. Partial output of an aborted or errored turn is not reported. Public OMP `omp-plugins` packaging supplies model-driven skills for commune, signoff, and durable-role inspection/editing. Commune checkpoint mode saves continuity, resets OMP context through native session APIs, and wakes the same endpoint without operator intervention. The packaged knock skill (0.5.0) teaches the published access-control grammar: bare `spt knock ` is `send` and asks for `MSG`; `send`/`redeem` require exactly one of `--send-only` | `--send-receive` while `approve`/`new-code` take none; `--send-only` never bars replies; approving opens nothing in reverse (knock back); invite codes are sealed to a subnet; `--monic` imparts a standing note; the inbox is quiet, so look with `spt knock list`. A manifest `[[hints]]` row surfaces the skill on knock intent. A provider usage-limit refusal (0.6.0) is an outage, not a fault. When a turn ends with an account-level refusal — OMP's `errorId` UsageLimit flag, HTTP 402, or refusal wording without per-interval wording — the extension says so once (nothing is broken, the session and its context are intact, only a human or the stated reset changes anything, the provider's line quoted) and then holds the endpoint `busy` until the provider's stated reset plus one minute, re-asserting `api state busy` every five seconds so a stray recovery is corrected within seconds. Peer deliveries stay custodied and unsubmitted (a submission would spend the message into a refused turn) and the deaf-session ladder is paused for the hold. A real human prompt releases the hold early; the deadline releases it and resubmits everything queued, in order. Across a restart the session's own record re-arms the hold when its last message is still the refusal and the anchored deadline lies ahead; a refusal that is merely old history never triggers. No parseable or plausible reset means no hold beyond the notice — the failure direction is always "reachable too early". The Psyche runner (`psyche-omp`) maps the same refusal on its inner OMP turn to the published reserved exit code `96` — spt-core then paces retries, keeps custody, never reseeds and spends no strike — with a `PSYCHE_ACCOUNT_REFUSAL` line that says outage-not-fault and quotes the provider. Every other inner failure (Psyche or echo-commune) exits 1 with a bounded tail of BOTH inner streams in the line, naming a silent channel explicitly. Both LLM-turn roles declare their invocation budget in the manifest (`psyche_resume` 240 s, `echo_commune` 180 s) instead of riding the 90-second default. Phase 5 polish rides 0.6.0: the echo-commune summarizer names another agent only where the transcript shows a direct exchange and never presents a machine, host, or node name as an agent or teammate; every extension log line is prefixed `[omp-spt ]`; and a foreground `bash` call that measured 30 s or more gets one `[spt]` nudge appended to its result (at most three per turn, advice only) to run such commands with `async: true` so peer deliveries keep landing at the boundaries. The 0.7.0 close-out of this pass turned the two `unverified` rows into covered ones and fixed what the audit found underneath them. The native checkpoint is now armed by `spt_checkpoint` and run by the extension after the arming turn ends — OMP's compaction aborts the run and waits for it, and the run waits for the tool, so an in-tool reset was a hang (KNOWN-HAZARDS #16) — and from arming until the wake is queued every inbound delivery stays custodied unsubmitted, the status rail reads ` · checkpoint pending`, and the release resubmits them into the post-reset turns. A subagent's copy of the extension is inert: exactly one binding per OMP process activates (KNOWN-HAZARDS #15), so a message can only ever surface in the main conversation. A failed `spt` command is reported by the line that explains it (0.7.1). spt-core prints diagnostics and its status line on the same streams and a diagnostic can come first — `api bind` prints the eager reserved-id probe `ER_HOSTED_PROBE:…` ahead of a `CONFLICT: is live under session …` refusal (spt-bs-releases#279) — so the summary skips such diagnostics and the log line carries the command's whole output. The 0.7.0 subagent field report cited the probe line as the refusal because the first-line summary had dropped the CONFLICT line (F-037). Setup is hybrid: an extension-native command owns deterministic probes and activation, while a packaged skill owns diagnosis, repair strategy, published installer use, subnet onboarding, and explicit operator handoffs for authentication or elevation. Assistant output may dispatch the explicit cross-harness peer-message shortform `@` and mint wax seals with a `;;…;;` pair — read by **spt-core**, not the adapter: the manifest declares `[io] compliance = true`, so core parses the `AGENT_OUTPUT` spans the feed reports (a tag inside backticks or a fence is a quotation; a tag that never closes sends nothing) and reports the outcome only through the now-signal's `DISPATCH_RESULTS`. The extension carries no shortform parser and issues no `send` at turn end (KNOWN-HAZARDS #12). Ordinary `@` mentions have no side effect. Normal outbound messaging, roster, stop/shutdown, version, and subnet commands are taught through setup or startup context rather than duplicated as skills. OMP's native provider/profile routing satisfies alternate provider, account, and model routing parity; no `omp-spt:` analogue of Claude's CCS overlay is required. ## Release-target parity The supported target set is Windows x86-64 MSVC, GNU Linux x86-64, and a static Linux x86-64 musl helper compatibility tier for OMP-capable hosts where the GNU helper is incompatible. The musl tier requires full native endpoint acceptance on a representative target host; it does not imply generic Alpine support when OMP itself is unavailable. Neither sister adapter packages macOS or Arm64, so those remain explicit shared limitations rather than current parity gaps. The manifest version is canonical. Release gates reject disagreement in the helper crate, changelog heading, and versioned public documentation. Deterministic local documentation generation and drift checks gate changes. GitHub Pages publication runs the same gate before publishing. Each supported target retains a machine-readable release record naming the target, source commit and tag, artifact digests, acquisition result, and native fresh/resume, message/explicit-reply, lifecycle, shutdown, and ReadyAgent/LiveAgent acceptance outcomes. ## Explicit non-goals - Worker endpoint advertisement or OMP-subagent endpoint topology. - macOS, Arm64, or 32-bit release payloads in this parity program. - A CCS adapter overlay where native OMP provider/profile routing supplies the same outcome. - Claude hooks, PTY translation, plugin reconciliation, or model-managed background polling. - Claude's retired `>>commune<<` output marker. - Public support promises for the helper binary's manifest-role subcommands. - Separate skills for send, roster, stop, version, or subnet. ## Delivery order Coverage lands in dependency order: 1. repair version and documentation truth, then prove packaged-skill discovery and injectable OMP event seams; 2. implement extension-owned ready/live activation, live auto-resume, safe-boundary delivery, and abnormal-turn receivability; 3. add startup briefs, targeted hints, update notices, packaged skills, checkpoint continuity, hybrid setup, and peer-message shortform; 4. add the musl target and durable machine-readable acceptance records for all supported targets.