v0.15.1

What's Changed

• Hotfix: drop agent tool allowlists that block session spawn and notify by @dragoshont in #14

Full Changelog: v0.15.0...v0.15.1

v0.15.0

What's Changed

• Add OpenCode session bridge and upfront challenge rule by @dragoshont in #13

Full Changelog: v0.14.4...v0.15.0

v0.14.4

What's Changed

• docs: clarify setup and keep delivery product-first by @dragoshont in #11
• Reduce orchestration ceremony and release 0.14.4 by @dragoshont in #12

Full Changelog: v0.14.3...v0.14.4

v0.14.3

What's Changed

• Add authoring skills and benchmark scenarios by @dragoshont in #10

Full Changelog: v0.14.2...v0.14.3

v0.14.2

Architrave v0.14.2

A portable Copilot / Claude / Codex plugin update. Includes the previously
unpublished local 0.14.1 fixes and Session instrument work.

What changed

• The lead inherits host-available tools instead of maintaining a restrictive
  allowlist. Specialist scopes and host permissions remain intact.
• Ordinary plugin use no longer requires repository initialization/adoption.
  Adopt only for configured gates, repo-local agents, durable Runs or cloud setup.
• Independent implementation tracks prefer supported sidebar child sessions;
  CLI clients can use native subagents. Single-trace work remains direct.
• Continue within the user mandate, research unknowns narrowly, and avoid
  routine intermediate-status stops or unnecessary supervision.
• Explicit child skill handoffs and one lightweight architrave-work skill
  cover bounded development, evidence-backed research and author-voice writing.
• Preserve source-bound native review, managed installer safety, and retirement
  of automatic quality/tool-blocking hooks. Executable focused checks remain.

Install on another PC

For GitHub Copilot CLI:

copilot plugin marketplace add dragoshont/architrave
copilot plugin install architrave@architrave
copilot plugin list
copilot skill list

For Claude Code:

claude plugin marketplace add dragoshont/architrave
claude plugin install architrave@architrave

For an exact-version checkout, offline/source installation, or Codex:

git clone --branch v0.14.2 --depth 1 https://github.com/dragoshont/architrave.git architrave-0.14.2
codex plugin marketplace add ./architrave-0.14.2
codex plugin add architrave@architrave

GitHub's source ZIP and tar.gz downloads contain the same versioned kit.
The Python harness and optional repository adoption require Python 3.11+.
Shell and PowerShell entrypoints forward to the same Python implementation.

Start a fresh host session after installation. Five skills should be available:
architrave, architrave-cto, architrave-review,
architrave-tournament, and architrave-work.

Boundaries

Fresh Mac CLI skill loading, unadopted source editing and one native child
writing handoff were exercised. Existing desktop chats may retain older skill
paths or instructions; installation does not prove those chats refreshed.
Host-specific capabilities, permissions, model selection and child lifecycle
remain host-owned. Fixture/browser checks are not device deployment acceptance.

v0.14.0

Architrave v0.14.0

One coherent upgrade for truthful adaptive supervision, grounded in actual host
capabilities rather than an invented second runtime.

• On-demand feasibility reset estimates a finite time/turn/output window from
  evidence, risk, dependencies and the next discriminating test. Owner/global
  ceilings win; original clocks remain; unchanged evidence cannot renew;
  expired/unknown-budget lanes return partial/PARK rather than silently retry.
• Executing runtime, adopted kit, installed plugin and session-loaded
  instructions are separate identities. Missing host/loaded-context evidence
  stays UNKNOWN. Empty canonical workers or a quiet parent never prove idle.
• Product milestones require current source/task/criterion/objective-bound
  product observations. Control tests, chat claims, stale receipts and duplicate
  substantive observations cannot advance them. Milestones do not complete
  acceptance criteria.
• One-use owner primary-path correction preserves baseline, policy, human holds,
  candidates, acceptance, side effects and adaptive clocks. It rejects source/
  revision drift and symlink/junction escape; it cannot certify another repo.
• Product gate registration and completion reject source/objective drift inside
  their transactions. CTO advice distinguishes hypothesis/reproduction/verified
  evidence and uses a cheapest parity probe before risky advice.
• A generic process-growth manager was deliberately deferred. Existing exact
  process-group/PID timeout cleanup remains; risky experiments need supported
  owner control, not broad process-name kills or mandatory safe-command ceremony.

Verification and honest limits

Final independent OpenAI and Anthropic source reviews passed. Full existing
POSIX/manifest/fixture suite passed on the candidate; final published main and
tag workflows run POSIX and Windows release gates on the exact commit.
Mac targeted native suite: 53 tests, one version skip; product legibility:
19 tests passed. Windows junction escape regression passed.

One real joined task reported version 0.14.0 and the tiny fixture answer, returned
a compact read-only candidate, and passed an independent configured gate. The
first attempt used an incorrect path; it was not promoted. One evidenced
same-task recovery succeeded. This is native transport evidence, not product
UI/library certification. The six-file smoke Run occupied 30,912 bytes, including
the failed attempt; it is not a comparable 0.13 performance A/B.

The same frozen deterministic artifact fixture measured on the same Mac/Python
reports v0.13.0: ten files / 34,585 bytes; v0.14.0: ten files / 34,656 bytes
(+71 bytes, no new files). Both configured fixture criteria completed. This is
scripted artifact accounting, not native-agent or token/cost benchmarking.

LF-normalized entry text: 7,799 -> 7,978 bytes (+179).
Canonical CTO body: 1,825 -> 5,847 bytes (+4,022), loaded only through that skill;
with CTO loaded, total is 9,624 -> 13,825 bytes. These are honest added source
instructions, not a token/cost/context saving claim. Parent-context growth and
credit comparisons are unavailable.

Published-source installs are verified by actual runtime asset hashes and native
pins. Per-owner adoption receipts distinguish fresh instruction reads/host events
from files merely installed. Busy product work, authentication prompts and
user-paused/inactive chats are not restarted for adoption. A claim that every
chat loaded the latest context requires actual supported evidence; unavailable
observations remain UNCONFIRMED with a safe fresh-session handoff.

Final scoped adoption receipts: the builder, CTO, app owner, Xbox owner and
Exodus UI owner freshly read the published 0.14 conductor/CTO instructions.
The CTO additionally observed a supported host CTO-skill load in the current
turn. Exact adopted runtime/conductor/gate hashes and 0.14 kit stamps were
independently checked for the active Xbox and Exodus UI worktrees; app/CTO
copied-kit adoption is not applicable. Inactive backend, transport/backport/
inventory children and user-paused Portal are deferred until actual resume,
not awakened or labelled PASS.

Historical/system-prompt replacement is still UNKNOWN in existing chats.
Fresh-read adoption is not a claim that every old token or inactive session was
replaced. No product/config/custom-agent/Run mutation or game/authentication
restart was required by these checks.

The native status bridge's executing-root version/stamp can be null: that pinned
module-only root has no plugin manifest or copied-kit stamp. Its separately
verified nativeInstallation version is 0.14.0. The target repository's public
adoption-status receipt is the copied-kit identity source; a null bridge-root
stamp is not evidence of a missing/stale consumer kit.

Mac desktop native execution remains UNOBSERVED under the original permitted
install/discovery-plus-owner-prompt fallback. Existing sessions may retain prior
context. No private app hooks, agent CLI workers, credential/provider changes,
large model downloads, unrelated product edits or game interruption.

Exact optional desktop owner smoke prompt

In a new desktop thread in a fresh disposable test repository, with the installed
Architrave v0.14.0 enabled, use this one bounded prompt (not an instruction to
interrupt an active product/game/authentication session):

Use installed Architrave v0.14.0. Verify plugin discovery and executing/copied
runtime hashes separately from loaded instructions; unavailable proof is
UNKNOWN. Work only in this disposable repository. A: directly implement and
test whitespace-to-single-hyphen lowercase slug normalization, including empty
input. B: pre-decompose record validation and deterministic status-summary
functions into exactly two independent, non-overlapping native children.
Use the host's supported native tools (Copilot joined architrave_native_batch
with canonical WorkPackets, or Codex's own native subagents). Each receives one
objective, exact nearby tests, depth one, 240 seconds, 12 turns, 2000 output
bytes; no child spawning, agent CLIs, reports, credentials or provider changes.
If that host transport is not available, report NOT DISPATCHED, never fake it.
Integrate only scoped candidates and independently execute the tests. Report
actual native owners/concurrency, acceptance, effective model only if reported,
and unknown context/cost. Control-test success is not product UI/library proof.
Stop after this smoke and preserve any failure/uncertainty.

v0.13.0

Architrave v0.13.0

Architrave is now a thin mission supervisor: direct work first, host-native
children only when useful, progressive disclosure and compact candidate returns.
The host owns sessions, context, worktrees, lifecycle, permissions and model
selection. Python remains canonical for substantive runtime and installer logic.

Verified changes

• Startup text including CTO: 30,622 -> 9,770 UTF-8 bytes, 68.09% smaller.
  Forced durable-task text: 71,434 -> 37,291 bytes, 47.80% smaller.
  These are source-text measurements, not tokenizer, context-growth or cost claims.
• Maximum three active children, depth one, exclusive mutable ownership,
  bounded time/turn/output budgets and compact returns.
• Two identical failures without substantive new evidence stop the lane.
  Receipt timestamp/provenance churn cannot reset it; failure causes remain
  specific in primary-criterion diagnostics.
• Event-driven native transport subscribes before admission, preserving early
  turns and safely accepting unrelated authenticated sibling lifecycle changes.
  Policy, human holds, task ownership, source and evidence remain fail-closed.
• Optional user/host model requests inherit or report actual effective selection
  and fallback honestly. No provider SDK workers or nested agent CLI adapters.

Qualification

Final-source Windows Task A passed directly, with zero children. Task B passed
with two actual concurrent native children and independent integration/gates:
78 seconds, 2,204 bytes of compact worker records, ten artifacts / 52,026 bytes.
The host reported 295,861 combined child tokens; baseline usage, split usage,
parent-context growth and comparative cost are unavailable.

The final source has independent OpenAI and Anthropic PASS verdicts.
Required full gates passed on the reviewed runtime. Benchmark and review evidence
is in benchmarks/results/thin-supervisor.json; architecture and owner prompts
are in docs/orchestration-audit.md.

Honest host limitations

• Mac official CLI Task A was observed only on the earlier precommit hybrid
  installation. That is not final-source Mac-app Task A evidence.
• Mac CLI Task B was NOT DISPATCHED / UNOBSERVED, not a product failure.
• Final Mac Copilot-app and Codex-desktop Task A/B remain UNOBSERVED.
  Supported plugin installation/configuration/discovery and exact owner prompts
  are provided under the mandate's documented fallback. Publication does not
  turn those unobserved tasks into PASS.
• Existing sessions may retain old prompt context; start a new plugin session
  to load the published build. No private app hooks, credentials/provider changes,
  model downloads or unsafe app restarts are used.

Release assets include the source package and its SHA-256 checksum.

v0.12.4

Full Changelog: v0.12.3...v0.12.4

v0.12.3

Full Changelog: v0.12.2...v0.12.3

v0.12.2

Full Changelog: v0.12.1...v0.12.2