Skip to content

v0.3.0-alpha.19

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 23 Oct 17:17
· 78 commits to main since this release
810ea08

Added

  • The Python SDK gains judge(task_id, judgement_id, verdict, *, identity, rationale=None), which answers a waiting task's semantic question through melra_task_judge as the TypeScript SDK's judge() does (#378).
  • A settled semantic certificate's answered item carries executedBy, the action receipt's declared principal, beside judgedBy, so the two declared chains can be compared without the receipt (#378). Both are claims, and neither marks the harness: link an OAuth client's token adds. It is redacted with the certificate; certificates written before it simply lack it, and nothing is migrated.

Changed

  • Breaking: melra_task_judge also refuses a judge whose declared delegation chain names the task's principal, such as a grader the acting agent started (semantic_judge_is_actor, #378). Another session under the same harness, or another agent of the person the work was for, may still answer, unless the task's principal is that harness or person itself. The harness: link an OAuth client's token puts above a judge's chain is recorded in judgedBy but not compared. The check still rests on declared identities; melra judge is unchanged.

  • Breaking: melra approve with an empty task id, as an unset "$ID" sends, is refused with task_not_found: instead of taking the one task awaiting approval, if only one was; a prefix no waiting task has is refused with task_not_found:<prefix> (#386).

  • A task's goal is stored as declared, with only recognisable secret shapes taken out, instead of being replaced whole (#385). The stored request, the certificate's goal, and a workflow definition's step and delegation goals used to read [REDACTED_SENSITIVE_FIELD] whatever the goal said; they now hold the goal itself, so a certificate states what its task was for. The trade-off: a secret with no recognisable shape written into a goal is now kept in those records, so keep secrets out of goals; the goal's schema description now says so to whoever writes one. The browse harness tool's goal no longer includes the URL's userinfo, query or fragment. Tasks planned before this keep the placeholder, as do certificates later written for them. A definition stored before this keeps it in its stored copy (bump its version to store the goal there), but a new run of it reads the sealed definition, so that run's tasks, certificates and delegation prompts hold the goal. Nothing is migrated.

  • Breaking: an approval challenge's actionDigest (and so its phrase, still APPROVE plus its first twelve hex digits), the renderDigest an approval is bound to, and a receipt's inputDigest, which now also binds the task id, can no longer be recomputed without the home's payload key. A file read's or write's sha256 and a failed file_hash item's digest are stored as [REDACTED_SENSITIVE_FIELD]. The caller of melra_execute still receives the sha256 in output, but a failed file_hash item's digest is masked in the receipt and certificate and is no longer returned anywhere; run a file hash operation to read the actual hash. In a home shared with 0.3.0-alpha.17 or 0.3.0-alpha.18, this version still runs challenges and accepts decisions those versions recorded, and a decision it records for a challenge one of them issued keeps the plain renderDigest their executor checks; an older process refuses a challenge this version issued with approval_action_digest_mismatch before anything runs, and the task waits for a newer process. Upgrade every process sharing a MELRA_HOME together and re-run melra setup (compatibility).

Fixed

  • melra inspect and melra export take the eight-character id prefix the plain melra tasks table prints, in any status, as melra approve already did. Before, they passed it on as a whole id and reported task_not_found. A prefix two tasks share is refused with task_id_ambiguous:<prefix>, and one no task has with task_not_found:<prefix>; an empty argument names no task. melra judge still takes ids whole. (#386)
  • examples/README.md said the API example succeeds unauthenticated without GITHUB_TOKEN; it fails with credential_source_missing:github before any request is sent, as it always has. The example now says it needs the token, and docs/INSTALLATION.md#credentials documents credential_source_missing beside credential_source_invalid: a credential whose source is missing fails every call to its hosts and never falls back to an unauthenticated one. (#387)
  • The recall harness tool no longer fails a query it accepts that is longer than 9,993 characters: the goal built from it passed the 10,000 characters a goal may hold. A harness tool's goal that would pass that keeps only its verb, as in Recall (input too long to restate), rather than being cut through its input; the operation still carries the input (#385).
  • A verdict sent in band where only an operator's counts, from a developer-mode process on a task that ran enforced (melra judge without --mode enforced, or melra_task_judge) or from an agent on an enforced server (melra_task_judge), is checked for the task's status, the judgement id, an earlier answer and expiry first, so it gets task_not_awaiting_judgement:<status>, judgement_not_found, judgement_already_answered or judgement_expired where one applies; only a verdict on a question still waiting gets judgement_in_band_refused. Before, every one got judgement_in_band_refused, with melra judge advice to rerun with --mode enforced, and only the rerun showed the refusal it would meet. Every one of these refusals still records nothing. (#384)
  • melra judge on a question past its expiresAt reports judgement_expired (exit 1), as documented. Its own startup settles the expired wait first, so it reported task_not_awaiting_judgement:partial (or :recovery_required). melra tasks and melra status count a waiting_user task as waiting only while one of its open questions has not expired, and count the expired ones apart, with a hint that they settle closed when a runtime next opens the data directory; status --json adds awaitingJudgement and expiredJudgements. Before, only the stored waiting_user count was reported, so an expired wait counted as waiting until something opened the data directory. (#383)
  • MelraClient.connect() in the TypeScript SDK gives the stdio server it spawns this process's whole environment except MELRA_UNHINGED and MELRA_SOCKET, and except MELRA_WORKSPACE when cwd is given, with env on top: the Python SDK's rule. It passed only MELRA_HOME, MELRA_POLICY, MELRA_MODE, MELRA_PAYLOAD_KEY and MELRA_WORKSPACE, plus the MCP SDK's HOME, PATH and a few more, so a credential a policy reads from source.env was missing and every credentialed call failed (credential_source_missing), MELRA_BROWSER* were dropped without a word, and terminal commands lost LANG, LC_ALL, CI and NO_COLOR. Both SDK READMEs now say what a spawned server inherits. (#382)

Security

  • Digests stored beside redacted records are now keyed per home, with a key derived from the payload key, and a content digest stored beside redacted content is masked like it. A workflow step's idempotency key stays plain, so every version sharing a home derives it alike; the threat model lists what else stays plain. Records written before this are not rewritten.
  • A receipt's principal, the declared delegation chain, is redacted like the rest of the record, as the task row, judgedBy and a certificate's executedBy already were, so it reads the same as executedBy, and the kinds taken out are listed in the receipt's redactions. Receipts written before this are left as they were.
  • The prompt a human_input or delegation node records in a run's state, and so in its events and snapshots, is redacted like the rest of the record, as the stored definition's copy already was. Only secret shapes are taken out, so a delegation's prompt still says what to do, and it stays within the 2,000 characters node state allows. A run's state, events and snapshots written before this keep the prompt they had; a node still waiting takes the redacted prompt at the run's next advance, which records one more workflow.node_changed.