Repository navigation
This release ships the code of v0.2.0-rc.1 unchanged.
Added
- A sandboxed job can reach a list of declared destinations and nothing else:
sandbox.network: proxywithsandbox.destinations(#25) - Add signed unit checkpoints for fast replay, preemption at slot safe points, and refusal of diverged log copies (#11)
- BREAKING: Interval and event job triggers in supervisor, refusing unexecutable schedules or restart policies on triggered jobs at config load (#126)
- Take up interrupted slot runs with epitropctl slot resume, adopting the running or ended job without a second start and recording the result as an uninterrupted run (#127)
- Traces and metrics for slot runs, supervised jobs, side effects and task-graph queue depth, exported over OTLP when an endpoint is set (#34)
- Start a slot run as a systemd user service of its own with epitropctl slot run --detach, carrying the template's memory_max and memory_high, so the run outlives the session that started it
- units frontier, assign and release: list the graph nodes a slot may take, claim one and open its unit in one record, and give it back with a reason
- units handover: the open units of one repository with their pending decisions and unread notes, written as a slot's resume note on request
epitropctl watchaccepts--github-base,--huggingface-base, and--ntfy-baseflags to point API requests at local stand-ins (#145)- A sandboxed job can declare paths inside its workspace read-only, and cannot rename them away (#128)
- End a hung job whose heartbeat, the last write to its log, is older than heartbeat.max_age_seconds, recording job.stale and leaving the restart to its policy (#21)
- Sync reads the calling LiteLLM key's budget window, and epitropctl headroom reports a stale or missing usage reading as unknown, never as zero (#24)
- Cutover 1 plan: the fleet's background jobs declared as Epitrop jobs in
docs/cutover-1.example.yaml, with the parallel run and switch-over criteria indocs/cutover-1.md(#36) docs/config.example.yaml: running the release watch's supervisor as a systemd user unit that loads its token withLoadCredentialEncrypted=, with the token's minimal rights (#23)- Release watch: every pass ends with one summary line (sources polled and failed; releases new, filed and deferred; whether the rate-limit throttle stopped it), so a pass without new releases still shows in the job log (#271)
- Catalog: the usage window carries
burn_percent_per_hourandactive_agents(optional, at least zero; schema and validation refuse anything else); codex-local measures both over the last 15 minutes of session logs, and litellm-key-budget names both as absent. The schema stays at version 1, so a reader that validates against the v0.1.0 schema refuses snapshots carrying the new fields until it takes the new schema (#274) - BREAKING: Sync: the claude-local source reads Claude Code session logs and writes the five_hour and seven_day usage windows; used_percent and the burn rate are derived from a ceiling observed at a rejected request and stay absent with a reason until one is seen, and epitropctl claude-limits lists the limit types the logs name; a snapshot may now hold 25,000 records, which a reader from an earlier release refuses (#276)
Changed
- BREAKING: Loading a configuration refuses a job or watch whose
router_aliasis not listed inrouter.aliases - BREAKING: Allow a sandboxed job's workspace and inputs only inside the directories sandbox_input_roots declares, and refuse login locations below HOME unless a root names exactly one
- Cutover 1 parallel run: shadow jobs for the landing daemon, auto-rebase, the update intake and the monitor run each script's dry mode with a heartbeat bound (#36)
docs/cutover-1.md: the shadow supervisor's host unit tags its journal lines with its own name, so its exit reason can be found byjournalctl -t(#263)- BREAKING: The /events HTTP stream serves signed event-log records in sequence order, replacing unit and job diff summaries; each event carries the record sequence number as its id and the event type as its event name, and clients can resume from any record with Last-Event-ID (#278)
Fixed
- A sandboxed job gets its arguments as its config wrote them:
systemd-runno longer replaces${NAME}and$$in them, and the value of an allowed variable no longer lands on the argv (#25) - A run-record, task-graph or effect-resolution write that loses a race to another writer is now refused instead of reported as written (#108)
- A sandboxed job with egress network no longer logs pasta's "Couldn't get any nameserver address" line (#137)
- A release watch run without GITHUB_TOKEN now names the missing token when a sink search is refused, and logs once when rate_limit_floor is at or above the GitHub rate limit (#138)
epitropctl watchrefuses a release watch seen-set file whose version is not 1, instead of reading it and rewriting it as version 1 (#148)- The signed event log and the epitropctl commands that write it work on Windows, and Windows test suites run every package (#152)
- Stopping a job whose shim is exiting no longer fails with "job lock is not held"; the job counts as stopped (#227)
- A sandboxed job start is no longer refused when the kernel still shows the shim's program for the job right after the start. (#228)
epitropctl jobs superviseno longer ends after 30 seconds with a deadline error; it runs until SIGTERM or an interrupt and then exits 0 (#36)
Migration
- Snapshots (#276, #277): upgrade every program that reads snapshots before a writer selects sources whose caps sum past 20,000 records. Code that only compares against
catalog.MaxSnapshotRecordsneeds no change. - Event stream (#278, #279):
/eventsno longer sends the unit and job summary events. Read event names as record types (unit.recorded,job.started, ...), parsedataas the record, and read the current state back with the tools after a 410.