Skip to content

v0.3.0-alpha.20

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 24 Oct 09:31
· 70 commits to main since this release
7ccce6b

Changed

  • Breaking: melra workflow cancel exits 0 when the run it returns is cancelled and 1 when it is not, as melra approve --deny does (#403). It used to exit 2 for a cancel that worked and 0 for a run already verified_complete, which it leaves as it is. The other workflow commands keep their exit codes, now listed in the README.

  • Breaking: melra run exits 5, instead of 2, for a task that ran and now waits in waiting_user for a semantic judgement, and its output's next holds the melra judge <task-id> --judgement <judgement-id> --verdict <yes|no|cannot_tell> command for each of the task's judgements, one per line, with both ids whole and with --mode enforced when the task ran enforced (#404). The verdict is left for the operator to fill in: as printed, a shell refuses the line rather than recording a verdict. It also exits 4, instead of 2, when policy blocks a task after its prompt, as when a grant expired or was spent while the prompt was open. melra help and docs/INSTALLATION.md list its exit codes: 0, 1, 2, 3 for an approval it cannot ask for, 4 for policy_blocked, and 5.

  • Breaking: --mode with nothing after it, or with a blank value (what --mode "$MODE" sends with MODE unset), is refused by every command with --mode requires developer or enforced, instead of running in developer mode. A value starting with --, already refused, now gets the same message. A second --mode is refused too (--mode given more than once): only the first was read, so --mode developer --mode enforced ran in developer mode. An unknown value is still refused, now naming where it came from and what it may be: deployment_mode_unknown: --mode must be developer or enforced, or MELRA_MODE in place of --mode for the variable.

  • Breaking: melra run, melra workflow and melra demo exit 1 if the effect or advance they wait on never returns, instead of 0, the code melra run uses for verified_success.

Fixed

  • melra inspect, melra export, melra demo, melra policy test and the melra workflow subcommands find their action and id wherever they sit among the flags, as melra approve and melra judge do (#402). A flag written before them, such as --mode enforced, was read as the id and refused with task_not_found:--mode or workflow_not_found, or as the action and refused as one the command does not support.

  • Memory search counts each record's tokens once instead of once per query term, and looks up only the query terms a record could hold, from whichever of the two is smaller, so a long query no longer holds the process for seconds, over a few large records or many small ones (#405). Scores are unchanged.

  • melra help lists MELRA_PAYLOAD_KEY and MELRA_HARNESS_TOOLS in its Environment block. pnpm docs:check now fails when a variable documented in docs/INSTALLATION.md is missing from it, and when melra help names a variable that has no row there (#388).

  • A workflow planned and never advanced stays planned when another process runs startup recovery, as melra workflow inspect does, instead of turning running with a workflow.recovered event (#406). Recovery now records nothing for a run whose nodes, status and error it finds as they were left, so a run between advances or waiting on an approval no longer gains an event on every process start. A recovery_required run an advance parked without naming why still gains that reason, and one event, at the next start. The Durable Core evaluation now reads evals/manifests/durable-core-alpha-v2.json: the same eight scenarios, except that its planned-task, node-boundary and post-approval restarts no longer expect that event. durable-core-alpha-v1.json stays byte-identical as the record of the alpha.1 run.

Security

  • http_resource_matches refuses a {{token}} whose value is empty, as it already refused . and .. (verification_token_empty:<path>), so the independent read always names the resource the URL means. The item is unanswerable rather than passed, and planning a reconciliation reports it.
  • An HTTP effect, and the independent read http_resource_matches makes, now settle however the exchange ends. A 101 Switching Protocols left the call unsettled whatever its timeoutMs, with the task running and no receipt; it is now effect_outcome_unknown:http_upgrade_refused, so a mutation is left recovery_required, and melra run no longer exits 0 for a task whose effect never returned. The effect is bounded by its timeoutMs, budget.maxDurationMs and a cancel. Each independent read is bounded only by its predicate's own timeoutMs (at most 60 s), as before: the budget and a cancel do not reach it, and a task's reads run one after another, so verification can outlast the budget, and hold a cancel, by up to that much per read (#407).
  • A --mode with a missing or blank value, or a repeated --mode, no longer leaves a process in developer mode, so a launcher or script that loses the value or overrides it cannot start a process its operator believes is enforced. Leaving --mode out still means developer, unless MELRA_MODE or the policy says enforced.
  • Overrides raise two transitive floors that pnpm audit --prod reported through @modelcontextprotocol/sdk: body-parser to 2.4.0 (a denial of service through bytes after a compressed request body) and fast-uri to 3.1.9 (scheme spoofing and host confusion). The SDK's ranges already allow both; only the lockfile pinned the vulnerable releases.