Skip to content

Changelog

View as Markdown

Every change that reached Durable Runner's main branch and that you can see or use, newest first. Each entry is dated the day it landed, labelled with its type and the areas it touches, and links to the docs. Durable Runner hasn't made its first release yet, so entries are grouped by month rather than by version. Subscribe with RSS.

November 2026

The first stable release, and what a reader can now rely on.

FeatureAPICLIInspectorJournalServerSelf-hosting

The first stable release, and what a reader can now rely on. The core promise is unchanged: a submit is acknowledged only after a durable commit, acknowledged runs survive kill -9 and reboots, and the framed append-only journal — one writer, exclusive lock — is the only authority. Journal records are now versioned: new records carry v: 2, readers accept v1 forever and refuse unknown future versions rather than guess, and --journal-format plus deploy/format-guard.sh refuse an upgrade or rollback to a binary that can’t read the journal — failing closed, not eating state.

Per-client limits ship. One address gets 100 requests/s by default (429 rate_limited with Retry-After), connection caps for untrusted peers and separately for trusted proxies, and --trusted-proxy CIDRs that decide whose CF-Connecting-IP is believed — anyone can set that header, so it’s believed from no one else. IPv6 keys at /64 with a /48 aggregate.

Capacity is measured, not assumed. Two honest numbers: a 12-core laptop sustained 100 submits/s — 1,000,000/1,000,000 done, p99 41 ms — and a 4-core CI runner topped out near ~85 completions/s, so a 100/s offer had 835,397 accepted then refused the rest 503 at_capacity at the 100,000-live-run bound. The CI miss is kept as evidence, not rerun until green: over-offered, the server keeps completing at its own rate and refuses the overflow by name. Your own host gets the same probe before traffic — see Capacity.

The Inspector covers everything now — submit, list, detail, spec read-back, cancel, diagnostics, compact, backup — with two-step confirms on mutations, keys in sessionStorage only, light/dark tokens and a phone layout that stays unclipped at 390 px. And a deploy kit ships with the repo: hardened systemd unit, Caddyfile with lb_try_duration 15s, Cloudflare ranges feeding both trust boundaries, daily encrypted backups under a 7-day lock, and upgrade/rollback scripts behind the format guard.

Read more: Capacity · Configuration · The Inspector · Link to this change

Every call the runner makes to a step handler now carries x-durable-run-id, x-durable-step and x-durable-attempt alongside the input and the idempotency key, so a handler can log, trace or reconcile an individual attempt without parsing the body.

FeatureAPI

Every call the runner makes to a step handler now carries x-durable-run-id, x-durable-step and x-durable-attempt alongside the input and the idempotency key, so a handler can log, trace or reconcile an individual attempt without parsing the body. The reconciliation probe carries the same headers.

Link to this change

The load harness completed a sustained run of 1,000,000 acknowledged submits against a single runner — every acknowledgment committed to the journal first — with the group's batching and bounds holding throughout.

ImprovementPerformanceJournal

The load harness completed a sustained run of 1,000,000 acknowledged submits against a single runner — every acknowledgment committed to the journal first — with the group’s batching and bounds holding throughout. Result records no longer carry the machine’s hostname.

Link to this change

GET /jobs/{id}/spec returns the spec exactly as it was submitted — the version, the steps and the base64 input — so a controller can resubmit the same work under a new op_key, or audit what a run was asked to do.

FeatureAPI

GET /jobs/{id}/spec returns the spec exactly as it was submitted — the version, the steps and the base64 input — so a controller can resubmit the same work under a new op_key, or audit what a run was asked to do. Because the input may be job content, the read-only key cannot use this route: it answers 403 to the read key and to open reads.

Read more: API reference · Link to this change

Restore and backup verification now reject a manifest that names the same file twice or lacks a checksum for a file it lists — and the error names the file, so a damaged backup is diagnosable instead of mysteriously unrestorable.

FixServer

Restore and backup verification now reject a manifest that names the same file twice or lacks a checksum for a file it lists — and the error names the file, so a damaged backup is diagnosable instead of mysteriously unrestorable. The checks happen before a byte of the target is trusted.

Read more: Backups · Link to this change

inputb64 in a submit now has to be canonical, padded standard base64 (RFC 4648).

FixAPI

input_b64 in a submit now has to be canonical, padded standard base64 (RFC 4648). Anything else — unpadded, URL-safe alphabet, stray whitespace — comes back 400 bad_input with a single stable message instead of the decoder’s own wording, which varied between versions.

Link to this change

The server now bounds its connection count and enforces deadlines on request headers and bodies, so a client that drips bytes or never reads its response can no longer hold a connection forever.

SecurityServer

The server now bounds its connection count and enforces deadlines on request headers and bodies, so a client that drips bytes or never reads its response can no longer hold a connection forever. Bodies cut off at their deadline answer 408 body_timeout instead of a parse error.

Link to this change

A clean stop is supposed to mean clean.

FixServer

A clean stop is supposed to mean clean. The server now exits 0 on SIGTERM only if it finished within --drain-timeout-secs — requests answered, decisions made durable. If the deadline cut anything off, it exits non-zero, so systemd and operators can tell a drained stop from a forced one instead of assuming both look the same.

Read more: Deployment · Link to this change

The recovery drill now stops the server with SIGKILL in the middle of real submits and replays the journal afterwards: every acknowledged run came back, zero lost.

ImprovementJournal

The recovery drill now stops the server with SIGKILL in the middle of real submits and replays the journal afterwards: every acknowledged run came back, zero lost. Frame, snapshot and index formats are pinned by golden files too, so an accidental format change fails a test instead of reaching a journal.

Read more: Concepts · Link to this change

--open-reads was already refused at startup on a non-loopback listen.

SecurityServer

--open-reads was already refused at startup on a non-loopback listen. Now it also checks the request’s Host header against the listener, so a web page that rebinds an attacker-controlled name to 127.0.0.1 (DNS rebinding) can’t read the run list through the victim’s browser. It changes nothing about the rule that matters most: never put --open-reads behind a proxy.

Read more: Authentication · Link to this change

Listing and inspecting runs now requires a key.

SecurityAPIServer

Listing and inspecting runs now requires a key. Servers can set a second, read-only key that may list and inspect but is refused (403) on submits, cancels, admin routes and spec read-back. For trusted local tooling, --open-reads still opens the two read routes — but only for requests that arrive on loopback addressed to a loopback Host, so a browser on the same machine cannot be talked into reading through it. Never place --open-reads behind a proxy.

Read more: API reference · Link to this change

durable restore now re-checks each file after opening it — so it restores the bytes it verified, not whatever landed in the meantime — and it fsyncs the target directory before publishing the restored manifest.

FixJournal

durable restore now re-checks each file after opening it — so it restores the bytes it verified, not whatever landed in the meantime — and it fsyncs the target directory before publishing the restored manifest. A backup list with a duplicate entry or a missing checksum is refused by name instead of producing a partial restore.

Link to this change

Everything before this window, in one card: the framed append-only journal with one writer and an exclusive lock; the engine that dispatches steps to allowlisted HTTP handlers at-least-once under a single idempotency key and reconciles lost replies by probe; the nine-route HTTP API with its error envelope; the operator CLI; consistent backups and verified restores; snapshots and compaction; and the read-mostly Inspector served same-origin.

FeatureJournalServerAPICLIInspector

Everything before this window, in one card: the framed append-only journal with one writer and an exclusive lock; the engine that dispatches steps to allowlisted HTTP handlers at-least-once under a single idempotency key and reconciles lost replies by probe; the nine-route HTTP API with its error envelope; the operator CLI; consistent backups and verified restores; snapshots and compaction; and the read-mostly Inspector served same-origin. The pages on this site cover each in depth — start with Concepts.

Link to this change

Was this page helpful?

Either answer opens a short issue on GitHub, so you can say what worked or what was missing.