Skip to content

v1.5.12

Choose a tag to compare

@github-actions github-actions released this 11 Nov 03:01
· 117 commits to main since this release
Immutable release. Only release title and notes can be modified.
cfafde5

OMK v1.5.12

OMK 1.5.12 is a patch release from main. It bounds subagent work by effort,
keeps ownership of workers that have not confirmed termination, restores skill
instructions lost to compaction, makes MCP recovery reach the next tool call and
fixes two provider-failover defects. Rust-core and v2 development are not
included.

Contributors

Thank you to @Abhinav3499 for the test work
merged in this cycle: coverage for the durable file-mode helpers (#382) and a
stabilized timing assertion in the RPC client lifecycle test (#383).

The README now opens with everyone who has sent a pull request to OMK (#405).

Subagent execution

  • Every effort, including Ultra, has finite limits per subagent call. Ultra no
    longer removes the task, concurrency, internal deadline or outer tool-timeout
    caps (#403).

    Parent effort Concurrent workers Logical tasks Default budget
    off / minimal 1 4 2 minutes
    low 1 4 3 minutes
    medium 2 6 4 minutes
    high 2 8 6 minutes
    xhigh 3 8 8 minutes
    max 4 12 10 minutes
    ultra 4 16 14 minutes
  • Behavior change outside Ultra. Earlier releases ran up to 4 workers at once,
    accepted 8 parallel or graph tasks and used a 14 minute default budget at every
    effort. Concurrency is now 1 at low and below and 2 at medium and high, the task
    limit also applies to chain calls, and the default budget follows the effort.
    A call with many independent tasks can take longer, and a short budget can
    return a long task as incomplete. Set executionBudgetMs (60,000 to 840,000)
    to choose the budget yourself (#403).

  • Workers receive the parent's effort as --thinking; the worker clamps it to
    its resolved model (#403).

  • When the host run has a deadline, worker budgets leave 30 seconds for the lead
    to integrate results, run required checks and report. The remaining time is
    checked again immediately before each spawn; with no capacity left the call
    returns an incomplete result instead of starting a worker (#403).

  • Single and chain launches use the same session lane authority as graph and
    parallel work. A worker whose process termination is unconfirmed keeps its
    ownership and resource reservation; a timeout alone does not mark it complete
    (#403).

These limits are engineering bounds. No latency or memory measurement backs
them, and they do not guarantee that every delegated task finishes before the
lead does. See
subagent execution budgets.

Skills

  • Automatic skill loading checks the current retained context before skipping a
    body admitted earlier. Instructions removed by compaction or branch navigation
    are loaded again in their existing admission slot. A body still present is not
    duplicated, a changed body needs a new slot, and the lifetime admission cap is
    never reset. Restoring a body can raise prompt tokens for that turn (#402).
  • skill-autoload.jsonl records contextChecked, contextAbsent and
    restoredFromHistory. They describe delivery, not whether the skill was applied
    correctly (#402).
  • Skill discovery no longer probes for .gitignore, .ignore and .fdignore in
    every directory; it uses the directory listing it already read. The discovered
    skill set is unchanged (#394).

MCP

  • After mcpCheckHealth({ reconnectFailed: true }) reconnects a server, the
    session uses the replacement server's tool catalog, including changed names and
    schemas. Disabled tool selections are kept. A call made against the old
    connection fails and is not replayed against the new one (#401).

Provider failover

  • An automatic failover or route rotation is recorded in the session only. It no
    longer becomes the saved default model, so the next session starts on the model
    you chose (#395).
  • When the model a failover switched to rejects the request outright, for example
    403 Access to model denied, OMK marks it failed for the turn and tries the next
    authenticated candidate. With a --model/--provider pin, a denied candidate
    still ends the turn (#395).

Not in this release

  • The OMK_SUBAGENT_ADAPTORCH routing-result consumer (#404) is deferred to a
    later version.

Verification scope

The subagent, skill and MCP changes are covered by offline tests with local
fixtures. They were not exercised against a live model end to end, on macOS or
Windows process termination, or with before/after memory and latency
measurements.

All seven public workspace packages are released at 1.5.12. No dependency
changed.