Skip to content

v1.5.9

Choose a tag to compare

@github-actions github-actions released this 26 Oct 04:29
· 96 commits to main since this release
Immutable release. Only release title and notes can be modified.
c5e0b01

OMK v1.5.9

OMK 1.5.9 is a patch release. It turns on seventeen harness flags by default, stores
evidence commands verbatim, adds project context documents, refreshes the model catalog,
clears the production dependency audit and fixes lock, goal and clock edge cases.
Every new default can be turned off again; see "Behavior changes" before upgrading.

Behavior changes

  • Seventeen OMK_* harness flags are on by default (#279, #281):
    OMK_EDIT_DIAGNOSTICS, OMK_EXEC_TRUNCATE_MIDDLE, OMK_JSON_SESSION_INFO,
    OMK_GOAL_AUTONOMY, OMK_SKILL_ADMISSION, OMK_SKILL_AUTO_LOAD,
    OMK_EVIDENCE_LIFECYCLE, OMK_ADAPTIVE_EFFORT, OMK_DOMAIN_ROUTING,
    OMK_GOAL_OUTPUT_RECALL, OMK_LANE_READY_QUEUE, OMK_LANE_FAILURE_SCOPE,
    OMK_RTK_OUTPUT, OMK_SESSION_CONTROL, OMK_VERIFIED_MEMORY,
    OMK_RESPONSE_REASONING_CAP and OMK_PREFIX_BOUNDARY. Set any of them to 0,
    false or off to restore the 1.5.8 behavior. OMK_EVIDENCE_LIFECYCLE deletes only
    verified goal-recall files; observe deletes nothing.
  • Evidence receipts store commands verbatim (#223). Command credential redaction is
    removed and older v3 receipts stay readable. Keep secrets out of command lines.
  • GitHub Copilot default model is gpt-6.1-sol, because Copilot retired gpt-5.4.

Added

  • Project context documents: omk init --project, and a bare omk launch in an empty Git
    root, create PRD.md, ARCHITECTURE.md, DESIGN.md, RULES.md, TASKS.md and
    PROGRESS.md without overwriting files (OMK_PROJECT_CONTEXT_BOOTSTRAP=0 disables the
    launch bootstrap).
  • DeepSeek harness, default on for DeepSeek models (OMK_DEEPSEEK_HARNESS=0 disables it).
  • Acceptance diagnostics behind OMK_GOAL_DIAGNOSTICS (flag, default off) (#170).
  • Kimi K3.1 Pro in the kimi-coding and moonshotai catalogs (#280).

Fixed

  • A passed acceptance check completes its goal on a loaded host (#291).
  • Repeated skill searches reuse unchanged document terms (#290).
  • Monotonic clock for durations, caches and deadlines (WSL2 clock steps).
  • Lock ownership for the goal journal, auth, trust store and session doctor.
  • Per-call bash timeouts, local context overflow recovery, status sidebar usage after
    /new, omk run cancel during live appends, and stale autocomplete acceptance.

Dependencies and catalog

  • npm audit --omit=dev is clean: undici 8.12.3, zod 4.6.7, and the optional Gondolin
    example on 0.14.0 without node-forge (#289).
  • Model catalog regenerated; upstream-deprecated models are removed (#288).

Not yet measured

  • There is no on/off benchmark for the new defaults, so this release makes no
    performance claims about them.