真相之源：GitHub 上的 CHANGELOG.md

Changelog

All notable changes to OwlCoda public releases are documented here.

[0.15.12] — 2026-06-23

Runtime truth + supervised job replay reliability release.

Added

  • Added supervised runtime job tools: JobList, JobGet, JobCancel,

    • BrowserJob, ApiJob, and ServiceJob.

  • Added browser replay evidence capture across fetch/headless/CDP-oriented job

    • providers, with saved HTML/text/screenshot/console/network artifacts where

    the provider supports them.

  • Added service lifecycle supervision for local dev services, including PID,

    • port, health, log artifact, restart, and graceful stop metadata.

Changed

  • Command-backed TaskCreate work now mirrors into the job supervisor, so long

    • tasks have a queryable job snapshot in addition to task and runtime lifecycle

    state.

  • TaskVerify now records verification strength and refuses to complete

    • parser, snapshot, browser, and replay-sensitive steps from weak evidence such

    as a plain pattern match, typecheck, or build-only signal.

  • Runtime-sensitive final reports now have to mark unresolved browser/business

    • replay as pending_user_replay or runtime_replay_pending instead of

    claiming full completion while asking the user to refresh or manually retest.

Fixed

  • Fixed a release packaging regression in the transient 0.15.11 npm package:

    • 0.15.12 is rebuilt from the public 0.15.10 source line and preserves the

    runtime lifecycle, runtime supervisor, AgentControl, and mailbox control-plane

    files while adding this release's supervised job tooling.

Notes

  • Published as npm [email protected] with GitHub source tag v0.15.12.
  • This release improves runtime evidence, replay supervision, and report

    • honesty. It is not a claim that long-running degradation is permanently or

    completely solved.

[0.15.10] — 2026-06-22

Runtime truth + multimodal routing reliability release.

Added

  • Added image input handling for local paths, @image.png, Markdown image

    • references, and file:// image links, converting them into multimodal

    content blocks when the selected model is image-capable.

  • Added provider capability routing and probe support so Kimi K2.7 / Moonshot

    • and other OpenAI-compatible providers can enable image recognition through

    model capabilities instead of a hard-coded provider branch.

  • Added runtime lifecycle tools for task commands, Agent runs, supervisor

    • processes, and mailbox messages: RuntimeLifecycleList/Get,

    RuntimeSupervisorList/Get, AgentControlList/Get, and

    AgentMailboxSend/List/Get/Resolve.

Changed

  • Command-style long tasks now record supervisor-process snapshots with process

    • identity, parent run linkage, and inspect/recovery policy.

  • Agent run history now exposes parent/child run views and inspect-before-retry

    • recovery state outside of the transcript.

  • Parent/child agent messages now flow through a mailbox queue and runtime

    • lifecycle snapshots instead of living only in conversation text.

  • TaskVerify safe-readonly classification now allows read-only process and

    • hashing checks such as ps, pgrep, shasum, and sha256sum.

Fixed

  • Fixed runtime recovery / task-state divergence after a

    • verification-repair checkpoint resolves: the post-recovery-overrun guard now

    checks the real task-store step state before skipping TaskUpdate(completed),

    so a step that is still in_progress can converge to completed.

  • Kept the verification-repair hard-stop guard strict; the release does not

    • weaken checkpoint or overrun protections.

  • Fixed the runtime event contract so legal long_task_wait_policy

    • interventions do not require violation_kind when the event is a wait-policy

    record rather than a violation.

Notes

  • Published as npm [email protected] with GitHub source tag v0.15.10.
  • This release improves runtime evidence, recovery convergence, and multimodal

    • routing. It is not a claim that long-running degradation is permanently or

    completely solved.

[0.15.8] — 2026-06-18

Long-run reliability release.

Added

  • Added runtime recovery ledger surfaces and inspection tools so long-running

    • sessions can expose unresolved checkpoints instead of losing them behind

    generic completion or retry output.

  • Added long-task lifecycle tracking and tools for inspecting, waiting on, and

    • replacing managed long tasks.

  • Added runtime event persistence into saved sessions so recovery context can

    • survive resume paths.

Changed

  • Hardened TaskVerify, TaskUpdate, completion, finalization, and headless/TUI

    • gates around long-run degradation and recovery states.

  • Updated root and Admin dependencies to keep release-candidate audits clean.

Notes

  • The npm package remains the public install surface; the demo/ World Cup

    • predictor and private execution prompts are excluded from the npm tarball.

  • This release is intended to be paired with public source tag v0.15.8.

[0.15.7] — 2026-06-14

Packaging-hygiene release. No runtime change from 0.15.6.

Fixed

  • The published tarball is now built from a clean dist/. A prebuild step

    • removes dist/ before every build, so compiled output whose source has been

    deleted can no longer linger. This removes six stale dead-code files from the

    removed stub tools (repl, schedule-cron, send-message) that had been

    shipping — unregistered and unreachable — since 0.15.5. 0.15.6 remains

    published but carries those phantom files; 0.15.7 supersedes it with a clean

    tree.

[0.15.6] — 2026-06-14

Command-surface refinement release: fewer, clearer slash commands and keyboard

mode switching.

Added

  • **Shift+Tab cycles the operating mode** (normal → auto → plan; yolo stays

    • explicit), with the mode rail updating live as you cycle.

Changed

  • Consolidated the slash-command surface. /config now absorbs everything

    • /settings showed (approval mode, theme, persistent always-allow), and five

    pure-duplicate commands were removed: /settings, /color, /tokens,

    /reset-circuits, /reset-budgets. The old names still work — they print a

    friendly "use X instead" redirect rather than erroring.

  • /reset now combines both former reset commands, and observability output

    • folds into /status. /mode with no arguments explains every mode.

Fixed

  • Slash-command fuzzy search matches the command **name** first, so typing a few

    • letters of a command finds it instead of being buried under description

    matches.

Notes

  • Ships FIFA Phase 2 for the worldcup-predictor demo to the public source

    • mirror: deterministic post-match data backfill that feeds an honest tactical

    prior into the pre-match brief (combines with, never overrides, pre-match

    evidence). The demo lives under demo/ and is excluded from the npm package.

[0.15.5] — 2026-06-14

Safety release: closes the gates through which a code-executing command could run

without the approval the active operating mode promised.

Security

  • **HIGH — unattended headless --mode yolo ran dangerous bash without the

    • deny-gate.** In headless runs, mode auto-approve was overriding the headless

    safety deny-gate, so a destructive command could execute with no human present.

    The deny-gate now survives mode auto-approve: dangerous bash is blocked even

    under --mode yolo, while read-only and workspace-test commands still pass.

  • **Task sub-agents no longer bypass the approval gate by delegation.** A parent

    • could hand a dangerous command to a spawned sub-agent, which ran with no

    approval callback at all. Dangerous bash in a sub-agent is now gated like the

    parent's, closing the delegation bypass.

  • **Wrong-case tool names no longer slip past the risk and mode gates.** A model

    • emitting Bash (instead of canonical bash) sidestepped the destructive-

    command gate. Tool names are now canonicalized in the risk, mode, headless,

    write-scope, intent, and TUI gates, and in the persistent "Always allow" store

    (a wrong-case grant previously never stuck).

Fixed

  • Startup --mode yolo no longer desyncs the auto-approve mirror, and /mode

    • and /plan clear that mirror so you can switch back out of yolo.

  • /mode copy advertises yolo, and the /plan hint is accurate and gated on

    • whether modes are enabled.

Notes

  • Ships the worldcup-predictor daily auto-review (self-grading loop) v1 demo to

    • the public source mirror. The demo lives under demo/ and is excluded from the

    npm package; the runtime change in this npm release is the safety hardening

    above.

[0.15.4] — 2026-06-13

Dogfood-driven harness fixes, operating-mode consolidation, and a sub-agent model override.

Added

  • Sub-agents can run on a different model than their parent: the Agent tool

    • takes an optional model, resolved as input.model > OWLCODA_SUBAGENT_MODEL

    > parent — so an orchestration sub-agent need not share the parent's backend.

  • Admin "test connection" now shows the endpoint's real reported model version.

Changed

  • Unified the operating-mode surface: /mode, /yolo, /approve, and /plan

    • all write one shared mode state that the permission gate reads.

  • Removed three non-functional stub tools (repl, schedule-cron,

    • send-message).

Fixed

  • Tool dispatch is case-insensitive, so a model emitting Bash instead of the

    • canonical bash no longer fails with unknown tool.

  • A bare cd is classified as a safe read-only command instead of being gated.
  • Tightened bash risk classification so commands that execute code are not

    • mislabeled read-only: python -m pytest … -v (pytest's *verbose* flag, not a

    version check) and env VAR=val <cmd> now classify by what they actually run.

  • ReadMcpResource recovers from common mistakes: it coerces parameter aliases,

    • lists the keys it received, and redirects a file:// URI to Read.

  • The loop guard now catches cross-turn accumulation of the same failure class

    • (keyed by tool + failure category), and TaskVerify flags unsatisfiable

    checks so the guard stops futile re-verification.

  • "Step not found" errors from the task tools now list the available step ids.

[0.15.3] — 2026-06-12

Interrupt-recovery and rendering hotfix.

Fixed

  • Interrupting a tool loop and submitting a new message no longer poisons the

    • session with deterministic 400s: merged consecutive user turns now keep

    tool_result blocks first, and an outbound wire guard re-orders any

    non-conforming body as a last line of defense.

  • IME pre-edit no longer renders one row above the composer: the declared

    • cursor origin is clamped to the terminal's physical bottom row.

  • Ordered-list numbering resets at headings, code fences, and tables, and

    • honors an explicit start number — long CJK documents no longer continue

    a previous list's counter.

Added

  • Render-incident capture: on a render-path throw, the raw-chunk ring buffer

    • is persisted to a dump for diagnosis.

[0.15.2] — 2026-06-11

Transcript chrome, compaction-resilience, and protocol-hygiene release.

Added

  • Transcript chrome S1–S3: collapsed tool results with a unified ok/err shape

    • and an /expand toggle; narration ● gutter with merged action+result

    groups and hanging-indent wrapping; one-line notices, a merged turn footer,

    and shared key-value slash panels (including /cost).

Fixed

  • Emergency heap-pressure compaction no longer erases task context: task

    • anchors stay pinned, an ineffective-cut breaker stops repeated zero-value

    cuts, and a heap-significance gate skips conversations too small to matter,

    with pressure diagnostics for each decision.

  • Orphaned tool_use/tool_result pairs are stripped at the send chokepoint,

    • preventing deterministic 400 loops after interruptions; the daemon now dumps

    4xx request shapes for diagnosis.

  • Long CJK ordered-list items are no longer split mid-item and renumbered by

    • the fallback sentence splitter.

  • Headless runs fail loudly when a model emits tool-call markers that never

    • executed, instead of reporting silent success.

  • Unknown models now default to a 200k context window instead of 32768, and

    • mimo-v2.5 models are recognized at 1M.

[0.15.1] — 2026-06-11

First npm package release on the GPL source line.

Changed

  • Moved the public npm install line from the historical 0.14.x stream to

    • 0.15.x.

  • Added the post-source-open runtime fixes already shipped through 0.14.64

    • to the public source line, including mode visibility, submission recovery,

    terminal width hardening, headless exports, third-party skill hardening, and

    streaming usage accounting.

  • Synced the bilingual README and Admin model screenshot into the npm package

    • surface.

  • Updated package metadata, lockfile metadata, Admin display version, and

    • corresponding-source wording for 0.15.1.

Notes

  • Paired with public source tag v0.15.1.

[0.15.0] — 2026-06-04

License boundary and public source availability.

Changed

  • Relicensed the OwlCoda core package from Apache-2.0 to

    • GPL-3.0-or-later starting with the 0.15.0 boundary.

  • Added SOURCE.md to make the corresponding-source requirement explicit for

    • npm packages that ship compiled dist/.

  • Updated package metadata, lockfile metadata, OpenAPI license metadata,

    • README distribution posture, product truth, NOTICE, CONTRIBUTING, and

    SECURITY docs for the GPL source line.

Notes

  • Commercial, OEM, or embedded distribution is handled through a separate

    • maintainer license path.

  • Historical published versions remain under the license terms that accompanied

    • those versions when they were published.

Older Releases

Historical package versions remain under the license terms that accompanied those versions when they were published.