LoopX 1.4.1 — Clearer workspaces and reliable semantic control

LoopX 1.4.1

At a Glance

中文说明

Illustrated release guide: Feishu walkthrough. Personal-user ownership, two released UI images and native-rendered guide verified. Reading follows current document sharing permissions.

A clearer Goal workspace, explicit research decisions, and safer execution/recovery across local hosts.

Upgrade

loopx update check
loopx update plan
loopx update apply
loopx --version
loopx doctor
loopx extension doctor --all-enabled --execute

Highlights

• See current tasks, owners and adopted results together in the Goal workspace. #6191, #6469
• Make optional research stride and source-basis decisions explicit. #6298, #6338
• Recover original effects and validate local hosts/extensions without widening authority. #6243, #6418

Replan Fixes

• Receipt-owned retries retain their original effect while monitor/dependency waits keep independent clocks. #6243, #6265, #6299, #6331
• Stride compares goal headroom and full experiment cost; configuration preserves effective inheritance and source-verified Explore references. #6278, #6377, #6379, #6440

Release Decision

Who should upgrade: 1.4.0 users of Goal workspaces, delegated results, settlement/retry recovery, extension lifecycle, or Windows/local hosts. Operators on a separately frozen experiment should preserve its pinned source until deliberately opening a new run.

What this release solves: Missing or stale task context, original-operation recovery gaps, hidden transport failures and absent current-basis checks; optional research decisions become explicit and versioned.

Breaking changes: No. Existing new-Goal SQLite/hard-lease defaults remain. Invalid or unreadable policy/source/API state now refuses where previous paths could appear clean; optional stride, material/governed-work, checkpoint and native-host activation remain explicit. Existing Goals are not automatically migrated. Separate provider distributions and experimental boundaries remain separate.

How to verify: Expect loopx1.4.1; read installation/runtime diagnostics, open the same Goal and read its current task/source state. Revalidate enabled extensions after update. A version string alone does not establish active host/runtime source or optional provider readiness.

Contributors: @LIHUA919 contributed scoped Explore evidence and working-directory review; @songoow contributed governed Goal/material direction and request recovery. Full community and agent-authored attribution follows the product groups.

loopx --version
loopx doctor
loopx --format json status --goal-id "$GOAL_ID"
loopx extension doctor --all-enabled --execute

State Kernel & Control Plane

• Preserve original settlement/retry identity across dependency waits, notice-only actions, monitor deadlines, ambiguous writeback and quota CAS failure. A later wait cannot replace the original effect or falsely settle a failed spend. #6243, #6245, #6265
• Original Goal commitments, current material revisions and one governed local work-item amendment now have explicit review, source fencing, peer adoption and original-operation recovery. This is a bounded File/SQLite source-session stage, not protected or acceptance rewriting. #6323, #6338, #6343
• Authority recovery preserves published backups, paged File history, source outbox/lease evidence and retirement isolation; concrete Todo read owners avoid loading mutation-only code. #6295, #6288, #6303

Capabilities & Workflows

• The Goal workspace puts current work, shared task inspection, meaningful progress, working-directory review and actual source freshness together. Accepted downstream reports can be read and explicitly adopted; lost acknowledgements reconcile and changed sources withdraw adoption. #6191, #6255, #5890
• Opt-in evidence-conditioned stride makes grow/switch/verify, complete experiment cost and adoption criteria explicit. Independent work-size/route-distance/evidence-breadth choices return with subsequent observations; Explore composition enforces exact research lineage. #6298, #6321, #6377
• Chat completes ACP initialization and model discovery before sessions, persists steering retry identity, retries only verified effect-free transport/overload failures once, and avoids idle recovery of completed commands. Ordinary project Codex Chat preserves the existing workspace grant and adds its physical Git metadata root with strict native readback; no full-access escalation. #6314, #6315, #6254

Quality & Testing

• Local and hosted test tooling now pins canonical fixtures, actual Python/backend origins, UTF-8 subprocesses, exact digest/import owners and bounded stall/failure evidence. Closed-PR cleanup reduces obsolete runner demand. These are engineering changes, not a claim that every historical test passed. #6274, #6318, #6336
• The public-safe outbound scanner checks credential dotfiles and extensionless inputs, fails closed on unreadable policy/content/directories or missing roots, and keeps surrounding content out of receipts. Same-origin Chat and observer redaction harden public/private boundaries. #6428, #6476, #6438

Benchmarks & Integrations

• Explicit research checkpoint and warm receiver paths preserve native opportunities, original remaining deadlines, current preparation/replan and retained research; bounded CLI/provider composition is shipped with its frontend/scientific/long-run gaps disclosed. No matched benchmark uplift or isolated model-causal improvement is claimed. #6305, #6415, #6442
• Managed ZCode CLI Goals now expose bind/model/lifecycle/status with current-basis resume refusal. Windows paths and four stdio hosts use truthful UTF-8 transport. Desktop recovery preserves the original page after WebContent loss and avoids downgrading a newer bundled Core. #5752, #6432, #6430
• Optional Finance providers add source-bound quantity/funding research, order-summary/result reconciliation and declared cash/loss-bound calculations. These separately packaged stages do not prove source truth, live venue execution, genuine confirmation or operation-card delivery. #6246, #6275, #6283
• Extension snapshots recheck current API/declaration compatibility, disabled status survives upgrade/rollback, and unverified legacy status stays readable. DSH runtime readiness and native operation transport remain distinct; Lark reports surface failed declared adapters. #6418, #6425, #6463

Documentation & Compatibility

• Package/manual/book identity is1.4.1. Existing new-Goal SQLite/hard-lease defaults are retained; upgrades do not silently migrate old Goals or activate optional research/host providers. Versioned lifecycle guidance below states activation, readback and recovery for every highlighted optional surface. #6492, #6330, #6300
• The updated engineering article explains LoopX development practices across its evolution; code contribution, source attribution and license-preserving reuse remain explicit. #6444, #6380

Community Contributors

• @BigDataDZ: Unicode Windows installation and writeback recovery — #6430.
• @Duang777: retained project state and safe non-recursive uninstall archives — #6202, #6205, #6206.
• @JasonBuildAI (external contributor): Windows/UTF-8 host portability, quoted Claude launchers, peer-directory pagination and canonical canary paths — #6229, #6250, #6354, #6422, #6464, #6472.
• @Jim-jimu: Todo dependency-cycle diagnosis, revision-bound Reward Memory and source-verified Explore pagination — #6237, #6308, #6379.
• @LIHUA919 (external contributor): scoped Explore evidence, reviewed working-directory selection, bounded CLI input and actual browser-backend provenance — #5280, #6209, #6297, #6336.
• @NIU-123370: ACP initialization ordering and pre-session Agent model discovery — #6314, #6315.
• @cocolord: one typed snapshot for quota fallback advice — #4061.
• @hhyykk (external contributor): bounded Explore hashing, listener permission diagnostics, scheduler-independent recovery and current task status — #6373, #6392, #6394, #6481.
• @hzeroyuke (external contributor; first merged contribution in this range): offline-friendly benchmark installation and complete Codex toolface for model aliases — #6364.
• @ikunikun1145 (external contributor; first merged contribution in this range): private-path protection in Finance/DSH projections and truthful failed quota-CAS settlement — #6446, #6449, #6467.
• @jackie-cqz (external contributor): ZCode diagnostics and managed native CLI Goal control — #5752.
• @songoow: governed Goal/material direction, typed selection/settlement, reusable IO owners and fail-closed extension/public-safe contracts — #6251, #6268, #6323, #6338, #6409, #6413, #6418, #6425, #6428, #6432, #6476, #6357.
• @wchwawa: NoKV SDK/helper qualification diagnostics and same-origin Chat admission — #4726, #6438.
• @yilin-succeed (external contributor): objective-isolation regressions on current Goal writers — #4418.
• @loopx-agent (agent-authored): agent-authored workspace/adoption, settlement recovery, stride/checkpoint, desktop and release-integration work — #5890, #6191, #6243, #6298, #6305, #6385, #6442, #6462, #6492, #6499.

Optional Capability Activation & Use

Replace Goal/Agent/instance variables with a currently authorized target. Read-only previews do not authorize execution. Source-checkout examples require the matching v1.4.1 checkout; optional providers are independently installed.

Goal workspace and result adoption

Activation: Open an existing Goal in the workspace; select its current task or accepted report. Read the exact source, then explicitly adopt. A view or accepted report alone does not adopt it.

Validation: Use status for the same Goal and read the current task/source in the packaged workspace. Reload after correction; changed source must withdraw stale adoption. Missing host/source remains unavailable.

Disable / rollback: Withdraw adoption through the same report action, or stop only the owned session. Preserve original operation receipts; reconcile lost acknowledgement instead of issuing a new effect.

Authority boundary: Viewing/adopting context grants no execution, acceptance rewrite, new Goal/Agent, provider, filesystem or cross-project access. Current claim, lease and source authority still apply.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/apps/presentation/dashboard/README.md

loopx chat --goal-id "$GOAL_ID"
# In another terminal, read the same Goal:
loopx --format json status --goal-id "$GOAL_ID"

Governed local work amendments

Activation: Explicitly use a supported active source_session_v1 Goal with original File/SQLite creation commitments. In Goal settings → Task ownership → Goal commitments and governed changes, review and freeze the permitted work template. The CLI previews writes until --execute; inspect current instance/revision before adopt/propose/commit.

Validation: Read --governed inspect and the original --governed receipt after a response loss. A commit must preserve previous Todos/leases, create unclaimed work and advance one canonical revision; peer adoption and replan settlement are separate current observations.

Disable / rollback: Stop new operations by leaving this entrypoint unused and retire through the existing Goal lifecycle when appropriate. Preserve immutable intent, historical readers and fences; a committed amendment is not undone by disabling its UI. Never resume an old writer by deleting intent.

Authority boundary: Only the declared work-item class is delegated. No protected permission, acceptance rewrite, arbitrary Todo text, automatic Turn adoption, request reservation or whole-Goal completion is granted. Existing Goals without original commitments need the source owner migration.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/architecture/rfcs/shared-goal-alignment-and-governed-amendment-v0.md

loopx --format json goal-amendment-proposal --governed inspect \
  --goal-id "$GOAL_ID" --goal-instance-id "$GOAL_INSTANCE_ID"
loopx goal-amendment-proposal --help

Goal material direction basis

Activation: In Goal settings → Task ownership → Direction materials and revisions, declare local material and apply its reviewed provider revision/authority digest. Each registered Agent reads and acknowledges its own exact returned offer.

Validation: Run goal-direction inspect for the exact Goal instance/Agent and read the original receipt. Read alone is not acknowledgement; changed or unavailable bytes invalidate an old passing projection.

Disable / rollback: Use the same editor disable action, or goal-direction --operation disable with current provider revision/new operation ID in --request-json; preview, then --execute. Retain generation-aware history/readers.

Authority boundary: Local material access proves neither remote discovery, comprehension, route adoption, Vision nor acceptance. Another Agent receipt cannot replace this Agent; replay adds no source access.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/architecture/rfcs/goal-direction-baseline-v0.md

loopx --format json goal-direction --operation inspect --goal-id "$GOAL_ID" \
  --goal-instance-id "$GOAL_INSTANCE_ID" --agent-id "$AGENT_ID"
loopx goal-direction --help

Evidence-conditioned stride reviews

Activation: Default off. Select evidence_conditioned_v1 in Goal review/stride, preview and apply. A CLI strategy-only update retains cadence; strategy-only capability JSON clears its Goal override and inherits live device cadence (ordinary default: six settled work Turns).

Validation: Read configure-goal and the next fresh review packet. Propose source-bound grow/switch/verify, separately justifying work size, route distance, evidence breadth, full cost and adoption criteria; subsequent evidence returns with the decision.

Disable / rollback: Set --execution-stride-strategy off --execute to stop the strategy while retaining cadence. Clear strategy and both cadence options for full device inheritance. Preserve original-version receipts.

Authority boundary: Configuration grants no experiment, quota, clone, weaker acceptance or automatic pivot. A recorded decision does not prove adoption, utility or benchmark gain.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/protocols/goal-vision-replan-contract-v0.md

loopx configure-goal --goal-id "$GOAL_ID" --execution-stride-strategy evidence_conditioned_v1
# Apply only the reviewed preview:
loopx configure-goal --goal-id "$GOAL_ID" --execution-stride-strategy evidence_conditioned_v1 --execute
loopx --format json configure-goal --goal-id "$GOAL_ID"
loopx configure-goal --goal-id "$GOAL_ID" --execution-stride-strategy off --execute

Explore scoped composition

Activation: Explicitly select composition mode explicit_only and its scope in the existing Explore Goal capability editor; the ordinary disabled mode remains unchanged. Review and apply the existing configuration owner.

Validation: Read effective configure-goal and the current obligation. Closing a live research duty requires exact gap/obligation/Todo/experiment lineage. Oversized active graphs remain an explicit refusal rather than evidence of missing work.

Disable / rollback: Preview/apply --explore-composition-mode disabled through configure-goal or the same editor. Preserve findings, current scope and historical experiment evidence; do not rewrite a receipt to make it current.

Authority boundary: Composition adds evidence qualification only: no new model, scoring, tool, remote write, experiment budget or acceptance authority.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/explore/README.md

loopx --format json configure-goal --goal-id "$GOAL_ID"
loopx configure-goal --goal-id "$GOAL_ID" --explore-composition-mode disabled
# Add --execute only after reviewing the preview.

Research checkpoint and warm continuation

Activation: Explicit developer/provider use: ordinary Codex Goals use research-checkpoint capture with reviewed request JSON and a quiescent completed source; verify before separate admit-session/adopt-context. Benchmark workers declare --checkpoint-request and protected --checkpoint-controller-file; warm receivers pass --restore-checkpoint/--restore-sha256/--restore-config and fresh Goal/Agent IDs. Pin the provider revision and original remaining deadline/opportunities.

Validation: Verify immutable artifacts and original capture/restore receipts. Read retained research/native cursor and current preparation/replan before normal cadence; reconcile lost replies by original operation ID. Help launches no benchmark.

Disable / rollback: Omit capture declarations or restore flags for a cold worker. Stop only the owned receiver/controller through its original runner. Never copy a live source or rewrite deadlines/receipts; any phase extension must be equal across compared arms.

Authority boundary: This is CLI/provider composition, not complete packaged frontend restore, scientific-profile, sustained-continuation or matched-gain certification. Archive verification grants no scoring, submission, registration, remote write or budget.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/research-checkpoint.md

loopx research-checkpoint --help
loopx research-checkpoint capture --help
loopx research-checkpoint verify --help
loopx research-checkpoint admit-session --help
loopx research-checkpoint adopt-context --help

Managed ZCode native Goal

Activation: Bind a registered Agent/current Goal through the detail drawer or zcode-goal bind; select an advertised model if needed, then explicitly start. Binding is idle and sends no model request; skill installation is separate.

Validation: Read zcode-goal status. Changed/missing source or acceptance basis refuses plain resume and stays paused; inspect, stop, then explicitly start a fresh objective. Disconnection stays unknown.

Disable / rollback: zcode-goal stop closes the owned process/target and retains history/binding. Do not start/resume while disabled. slash-commands --uninstall --surface zcode removes managed facades only, not native sessions.

Authority boundary: Core quota is not a native per-token hard budget; native completion does not settle work or prove acceptance. No Desktop attachment, credentials, extra tools/permissions/MCP/Hooks/Automations are granted.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/zcode_goal_mode/README.md

loopx doctor --agent-type zcode
loopx --format json zcode-goal status --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"
loopx --format json zcode-goal stop --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"

Claude Goal host and UTF-8 transport

Activation: Explicitly refresh managed Claude facades with slash-commands --install --surface claude-code, then use the existing connected-Goal workflow. The launcher quotes its interpreter and pins UTF-8 stdio.

Validation: Read doctor and existing host status. Unreadable PreToolUse input refuses; authenticated Claude model execution is not certified here.

Disable / rollback: Stop the owned loop via its controller, then slash-commands --uninstall --surface claude-code to remove only managed facades. Retain user settings/history.

Authority boundary: No new shell, file, tool, Goal, quota or account permission. UTF-8 is transport compatibility, not execution authority.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/claude_goal_mode/README.md

loopx slash-commands --install --surface claude-code
loopx doctor
# Stop the owned host before uninstalling its facades:
loopx slash-commands --uninstall --surface claude-code

DSH host and operation transport

Activation: Explicitly select deepseek-harness in the existing connected-Goal host setup; configure its already authorized model/provider. --host dsh selects the existing in-process adapter. Core updates do not independently publish or upgrade the separate DSH plugin.

Validation: Read the original Turn journal and transport-specific preflight result. A healthy DSH model runtime does not prove native operation-card delivery; preserve UTF-8 and source-redacted observer diagnostics.

Disable / rollback: Stop the owned loop via its controller and omit this host/profile on a new run. Recover an unknown operation using its original intent/receipt, never blind redelivery. Plugin disable/uninstall belongs to its separately versioned lifecycle.

Authority boundary: Host execution and native operation transport are distinct. Neither configuration grants model credentials, external channel access, confirmation, payment/order authority or an automatic schedule.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/dsh_goal_mode/README.md

python -m loopx.dsh_goal_mode --help
loopx turn run-once --help
loopx doctor

Extension lifecycle fail-closed readback

Activation: Install/enable an explicitly chosen compatible extension through the existing extension lifecycle; inspect requires_loopx_api, capability/provider declarations and owned runtime first. Legacy unverified status remains readable without making the extension available.

Validation: Run extension doctor --all-enabled --execute after core update. Saved snapshots are rechecked against the current API; duplicate/conflicting declarations refuse. A disabled extension must remain disabled across upgrade and rollback.

Disable / rollback: extension disable ID --execute stops admission; extension rollback ID selects a retained compatible revision. Follow its doctor before re-enabling. Do not edit a snapshot or cached readiness to bypass compatibility.

Authority boundary: Installation/readability is not activation, grant inheritance or validation. An extension cannot mint Core acceptance, permission, scheduler or external-source authority.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/extensions.md

loopx extension list --format json
loopx extension doctor --all-enabled --execute --format json
# For an explicitly selected installed extension:
loopx extension disable "$EXTENSION_ID" --execute --format json

Public-safe outbound scans

Activation: The builtin capability is available by default; scanning is an explicit per-command operation. Supply a readable PUBLIC_SAFE_CONFIG only when using local private markers. The selector now scans credential dotfiles and extensionless files while excluding documented VCS/vendor/binary entries.

Validation: Run scan_cli against the actual intended root or staged diff. Unreadable config/content/directories, missing roots and oversized inputs fail closed; receipts retain bounded labels/line numbers without nearby content.

Disable / rollback: Stop invoking the scanner to stop scans. Unset PUBLIC_SAFE_CONFIG only to deliberately select generic-rule-only mode; it is not a repair for a missing required private policy. Restore the readable original policy/root and rerun before submission.

Authority boundary: A clean scan is not complete confidentiality proof or external publication authority. No file mutation, upload, permission or repository write is granted.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/public_safe_outbound/README.md

python -m loopx.capabilities.public_safe_outbound.scan_cli --scan-root . --format json
# Alternatively, inspect only the actual staged diff:
git diff --cached | python -m loopx.capabilities.public_safe_outbound.scan_cli --diff --format json

Finance research and quantity reconciliation

Activation: Install/enable the optional package/manifest from the matching checkout and doctor its runtime; select reviewed quantity/funding/stock input. Core release does not independently publish this provider.

Validation: Run the versioned synthetic quantity example. Read source-bound stock instants, signed flows, unit/period/version and unresolved coverage; private file references must refuse.

Disable / rollback: Disable loopx-finance-value-discovery or roll back to a retained qualified provider. Preserve periods/sources/receipts; never invent cash flow from a stock change.

Authority boundary: Arithmetic proves neither source truth, suitability nor permission to transact. No corpus, venue account, confirmation, order, transfer, signing or acceptance authority.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/packages/loopx-finance-value-discovery/README.md

python -m pip install ./packages/loopx-finance-value-discovery
loopx extension install --manifest packages/loopx-finance-value-discovery/extension.toml --execute --format json
loopx extension doctor loopx-finance-value-discovery --execute --format json
loopx extension run loopx-finance-value-discovery --input-json packages/loopx-finance-value-discovery/examples/quantity-research-v1.json --execute --format json
loopx extension disable loopx-finance-value-discovery --execute --format json

Finance execution bounded readback

Activation: Install/enable the optional package/manifest and doctor it. --check-order-summary, --reconcile-execution and --check-risk-bound take exact reviewed JSON; default M1 simulation stays unchanged.

Validation: Compare an unexpired order/retained venue bounds to declared cash/loss budgets. Missing amount, unsupported/cross-margin basis, stale/drifted evidence or excess stays unknown/held. Calculation does not authenticate evidence.

Disable / rollback: Disable loopx-finance-execution or roll back to a retained qualified provider. Preserve confirmation/order/result IDs; recover uncertain venue results by original ID, never resubmit automatically.

Authority boundary: Checks certify no real venue order, genuine confirmation, live DSH worker/card return, funds reservation or transaction. Core publishes neither this separate provider nor financial authority.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/packages/loopx-finance-execution/README.md

python -m pip install ./packages/loopx-finance-execution
loopx extension install --manifest packages/loopx-finance-execution/extension.toml --execute --format json
loopx extension doctor loopx-finance-execution --execute --format json
loopx-finance-execution --help
loopx extension disable loopx-finance-execution --execute --format json

Revision-bound Reward Memory recall

Activation: Default off, scoped to named Agents in one Goal. Preview configure-goal --reward-memory-config CONFIG --reward-memory-agent AGENT; apply only with corpus-owner authority and provider canary write/exact readback.

Validation: Read experiment-status for the same Goal/Agent. Invocation revision must match; stale config/source cannot reuse an available receipt. Repair via the original registry/config owner, never copied history digests.

Disable / rollback: Preview/apply configure-goal --clear-reward-memory-config for this Goal. Retain corpus/receipt history; provider deletion is separately authorized.

Authority boundary: Recall is neither universal personal policy nor an execution grant. No cross-Agent/Goal corpus, credentials, escalation, background learner or source truth is inferred.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/reward_memory/README.md

loopx reward-memory experiment-status --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --format json
loopx configure-goal --goal-id "$GOAL_ID" --clear-reward-memory-config
# Add --execute only after reviewing this Goal-scoped disable preview.

Periodic report profiles and declared adapters

Activation: inspect-profile --preset weekly reads the built-in profile without scheduling or provider calls. Custom --profile-json defaults omitted enabled to false; enabled profiles require sources/renderers. Configure declared adapters before formal generation/delivery.

Validation: Read the activation receipt, then separate generation from delivery. Artifact, sink identity, subscription grant and exact result must agree. A failed declared adapter must remain visible.

Disable / rollback: Set custom enabled=false and disable the original subscription to revoke pending delivery. Remove optional archive/provider declarations to stop those paths; reconcile uncertain delivery using the original report key/receipt.

Authority boundary: Inspection grants no provider execution, publication, messaging or access. Generation is not delivery; archive, destination credentials and subscriptions remain explicit.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/periodic_report/README.md

loopx periodic-report inspect-profile --preset weekly --format json
loopx periodic-report --help

Desktop publication preflight

Activation: Maintainer-only per-dispatch opt-in: run desktop-release-artifacts.yml at a reviewed workflow ref with an existing admitted tag and publication_probe=true. This selects the read-only publication probe instead of asset upload.

Validation: Inspect that specific Actions run and download its preflight evidence. Verify builder/tag identity, receipt and access results; do not infer a successful hosted probe from local unit tests.

Disable / rollback: Omit publication_probe or set it to false on a later dispatch to restore the normal publisher path. This neither retracts existing assets nor authorizes publication; keep the original run and evidence.

Authority boundary: The probe grants no release creation, asset overwrite, signing key or unrelated repository access. Authenticated upload and installer qualification remain separate operations.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/product/release-readiness.md

gh workflow run desktop-release-artifacts.yml --repo loopx-project/loopx \
  --ref "$REVIEWED_WORKFLOW_REF" -f tag="$EXISTING_ADMITTED_TAG" -f publication_probe=true
gh run view "$PROBE_RUN_ID" --repo loopx-project/loopx
gh run download "$PROBE_RUN_ID" --repo loopx-project/loopx

Managed pull request review workflow

Activation: Use project-skill install for loopx-pr-review and the chosen host; inspect the preview, then --execute. Invoke review for exact NUMBER@HEAD; a new head needs a new review.

Validation: Read project-skill status and the preserved capability packet. Run pr-review --check-result for the exact result/packet, publish the full bilingual review and read it back. Diagrams aid explanation only.

Disable / rollback: Preview project-skill uninstall for the same project/skill/surface, then --execute to remove only managed files. Retain reviews/user instructions; never reuse approval after a head change.

Authority boundary: Installation or review grants no merge/admin bypass, first-screen approval, benchmark launch or private evidence publication. Exact-head readiness and explicit maintainer authority remain required.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/skills/loopx-pr-review/SKILL.md

loopx project-skill status --project . --skill loopx-pr-review --surface codex
loopx project-skill install --project . --skill loopx-pr-review --surface codex
loopx pr-review --target-exact-head "$PR_NUMBER@$HEAD_OID" --format json
loopx project-skill uninstall --project . --skill loopx-pr-review --surface codex

Original request correction and effect recovery

Activation: Use manager-inbox control on an existing exact request: Goal instance, original receiver, request ID and inspect/revise/cancel/recover. Corrections need reviewed control JSON/stable operation ID; preview before --execute. Original conversation control needs its persisted route/grant.

Validation: Inspect the original request/effect. Recover an ambiguous reserved Goal effect by its original operation before another revision; only independently read terminal Goal receipt releases the reservation. CLI/MCP/conversation use the same scope checks.

Disable / rollback: Omit control to retain the request. Cancel that exact request with --control-operation cancel, new stable operation ID and reviewed preview; this never undoes committed Goal work. Preserve receipts; recover uncertainty.

Authority boundary: Only the original authorized requester/receiver or verified original conversation principal may act. No impersonation, cross-Goal request, protected action, acceptance rewrite or completed-effect rollback.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/architecture/rfcs/shared-goal-alignment-and-governed-amendment-v0.md

loopx manager-inbox control --goal-id "$GOAL_ID" --goal-instance-id "$GOAL_INSTANCE_ID" \
  --peer-agent-id "$RECEIVER_AGENT_ID" --request-id "$REQUEST_ID" --control-operation inspect --format json
loopx manager-inbox --help

Ordinary project Codex Git writes

Activation: Uses the existing ordinary-project workspace_write grant (chat default) and host_default scope; no new switch. --project-workspace-grant workspace_read disables App writes. Start/resume adds only the physical project .git to native write roots.

Validation: Read chat --help, Codex doctor and pre-dispatch native permission readback; missing/different Git root refuses. Authorized disposable native fetch/branch/commit and same-thread resume were checked without a model request.

Disable / rollback: Stop the owned server/session, restart with --project-workspace-grant workspace_read and use a newly scoped thread. workspace_only preserves community isolation without this root. Retain history and native settings.

Authority boundary: No publication, destructive Git, account, extra network, external worktree/symlink or other-project access is granted. Existing network/tmp/approved roots remain; active Bot deployment still needs its normal update/restart.

Docs: https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/protocols/host-integration-surface-v0.md

loopx chat --help
loopx doctor --agent-type codex
# Existing server activation/restriction choices (stop only your owned server first):
# loopx chat --project-workspace-grant workspace_write --project-filesystem-scope host_default
# loopx chat --project-workspace-grant workspace_read

Install / Update

For a fresh PyPI install:

python3 -m pip install --upgrade loopx==1.4.1
loopx workflow-skills --install
loopx slash-commands --install
loopx doctor

Restart the host after refreshing its managed material. Existing pip/pipx/archive users use update check → plan → apply through their existing owner. Desktop users review the target in the existing update panel. Do not confuse Core package acquisition, host material, runtime activation and enabled-extension qualification.

Validation & Limits

Qualified source 4f27801c4e26fd51246d352eada72c98f8e7e657: 2,213 affected tests passed (5 explicit native-stage skips),515 public smokes passed; Ruff, mypy, typecheck, isolated upgrade/managed-host/UI and real PostgreSQL checks passed. Evolving/high (single-tool API profile) completed22 scenarios ×2 actor invocations and6 contrasts with zero failures/skips; only bounded connection/incomplete-stream retries, no semantic reroll. Earlier failures remain recorded. This is affected-source closure, not a repeated full-tree pytest pass or matched long-run/scientific gain.

GitHub/PyPI packages, four desktop installers, signed Mac updates, desktop-stable and source stable verified at1.4.1; actual PyPI1.4.0→1.4.1 upgrade and doctor passed. Fixture-only recovery retained product source4f27801c4. Mac: ad-hoc, no Apple notarization; Windows client not executed here.

中文摘要

图文发布指南： 升级与使用。个人用户所有权、两张正式版截图及原生渲染已核实。阅读仍受文档现有分享权限约束。

工作区更清晰，研究决定更明确，本地 host 执行与恢复更可靠。升级沿用原安装 owner 的 check → plan → apply，再核对1.4.1、doctor、当前 Goal 和启用扩展。主要更新覆盖工作区/结果采用、来源与受治理工作、可选 stride/Explore、host/桌面恢复和公开安全边界；每项用法、停用与权限限制见下方。重规划修复保留原结算 effect 和独立等待时钟，并按目标余量与完整成本判断研究路线。

升级决策

谁需要升级： 使用1.4.0工作区、协作结果、结算/重试恢复、扩展生命周期及 Windows/本地 host 的用户；冻结实验继续固定原源码，只有明确新开运行才改变。

解决了什么： 修复任务上下文过期、原操作恢复缺口、隐藏 transport 失败和 basis 检查缺失；可选研究决定明确版本和证据。

是否有破坏性变更： 无。沿用新 Goal SQLite/hard-lease 默认；不可读/无效策略、来源或 API 状态会更严格拒绝。stride、材料/受治理工作、checkpoint 与 native host 仍显式启用；旧 Goal 不自动迁移，独立 provider 与实验资格单独处理。

如何验证： 版本应为 loopx1.4.1，核对 doctor、同 Goal 当前任务/来源，更新后重新 doctor 已启用扩展。版本号不单独证明 host/runtime 已激活或可选 provider 合格；命令见上方。

贡献者： @LIHUA919 完善 Explore 范围证据与工作目录审阅；@songoow 完善 Goal/材料方向和请求恢复。完整社区与 Agent 作者贡献位于产品分组之后。

状态内核与控制面

• 依赖等待、notice、monitor deadline、写回不明和 quota CAS 失败均保留原结算/重试身份；新等待不能取代原 effect，失败扣费不能被记为成功。 #6243, #6245, #6265
• 原始 Goal 承诺、当前材料版本与一种受治理本地工作变更支持明确审阅、source fence、同伴采纳和原操作恢复；这是有界 File/SQLite source-session 阶段，不包括保护操作或改写验收。 #6323, #6338, #6343
• 权威恢复保留已发布备份、分页 File 历史、source outbox/租约证据与退役隔离；Todo 读取使用具体 owner，减少不必要的 mutation 导入。 #6295, #6288, #6303

能力与工作流

• Goal 工作区集中呈现当前工作、共享任务、有效进展、工作目录与来源新鲜度；已验收后续报告可阅读并显式采用，丢失确认先对账，来源变化撤回采用。 #6191, #6255, #5890
• 可选 evidence-conditioned stride 明确 grow/switch/verify、完整实验成本与采纳标准；工作规模/路线距离/证据广度分别论证，原决定随后续观察返回；Explore 组合要求精确研究 lineage。 #6298, #6321, #6377
• Chat 在会话前完成 ACP 初始化与模型发现，持久化 steering 重试身份，只对已证实无 effect 的 transport/overload 失败有界重试一次，并避免恢复已完成命令；普通项目 Codex Chat 沿用原 workspace grant，以严格 native 读回支持项目真实 Git 元数据写入，不切换 full access。 #6314, #6315, #6254

质量与测试

• 测试工具对齐 canonical fixture、实际 Python/backend 来源、UTF-8 子进程、精确 digest/import owner 和有界阻塞/失败证据；关闭 PR 清理减少无效 runner 占用，不代表历史测试全部通过。 #6274, #6318, #6336
• 公开出站扫描覆盖凭据点文件和无扩展输入，对策略/内容/目录不可读或 root 缺失失败关闭，回执不带邻近正文；Chat 同源约束与观察脱敏强化边界。 #6428, #6476, #6438

基准与集成

• 显式研究 checkpoint 与 warm receiver 保留 native 机会、原剩余 deadline、当前 preparation/replan 与累积研究；交付有界 CLI/provider 组合并保留前端、科学与长程资格缺口，不声称配对 benchmark 增益或单模型因果提升。 #6305, #6415, #6442
• 托管 ZCode CLI Goal 提供绑定、模型、生命周期与状态，并在 basis 变化时拒绝旧恢复；Windows 路径与四类 stdio host 对齐 UTF-8；桌面 WebContent 丢失后恢复原页面，并正确识别较新 bundled Core。 #5752, #6432, #6430
• 可选 Finance provider 增加来源绑定的数量/资金研究、订单摘要/结果对账和声明现金/损失边界计算；独立包阶段不证明来源真实、live venue 执行、真实确认或操作卡交付。 #6246, #6275, #6283
• 扩展快照重新校验当前 API/声明，停用状态跨升级/rollback 保留，旧 unverified 状态仍可读；DSH runtime 与 native 操作 transport 分开，Lark 报告明确已声明 adapter 的激活失败。 #6418, #6425, #6463

文档与兼容性

• 包、手册和双语书籍统一到1.4.1；沿用新 Goal SQLite/hard-lease 默认，升级不静默迁移旧 Goal 或启用可选研究/host provider。下方逐项给出启用、读回与恢复。 #6492, #6330, #6300
• 新的工程文章解释 LoopX 演进中的开发实践，并明确代码贡献、来源署名与保留许可证的复用边界。 #6444, #6380

社区贡献者

• @BigDataDZ: Windows Unicode 安装与写回恢复 — #6430.
• @Duang777: 保留项目状态及安全、非递归卸载归档 — #6202, #6205, #6206.
• @JasonBuildAI（外部贡献者）: Windows/UTF-8 host 可移植性、Claude 解释器引用、同伴目录分页及 canonical canary 路径 — #6229, #6250, #6354, #6422, #6464, #6472.
• @Jim-jimu: Todo 依赖环诊断、revision-bound Reward Memory 与来源验证的 Explore 分页 — #6237, #6308, #6379.
• @LIHUA919（外部贡献者）: Explore 范围证据、审阅工作目录、有界 CLI 输入与真实浏览器 backend 来源 — #5280, #6209, #6297, #6336.
• @NIU-123370: ACP 初始化顺序及会话前 Agent 模型发现 — #6314, #6315.
• @cocolord: quota fallback 建议复用同一 typed snapshot — #4061.
• @hhyykk（外部贡献者）: Explore 有界哈希、listener 权限诊断、独立于 scheduler 的恢复与当前任务状态 — #6373, #6392, #6394, #6481.
• @hzeroyuke（外部贡献者；本区间首次合并贡献）: benchmark 离线安装与模型别名完整 Codex 工具面 — #6364.
• @ikunikun1145（外部贡献者；本区间首次合并贡献）: Finance/DSH 私有路径保护及 quota CAS 失败的真实结算状态 — #6446, #6449, #6467.
• @jackie-cqz（外部贡献者）: ZCode 诊断与托管 native CLI Goal 控制 — #5752.
• @songoow: 受治理 Goal/材料方向、typed 选择/结算、复用 IO owner 与扩展/公开安全失败关闭合同 — #6251, #6268, #6323, #6338, #6409, #6413, #6418, #6425, #6428, #6432, #6476, #6357.
• @wchwawa: NoKV SDK/helper 资格诊断与 Chat 同源接纳 — #4726, #6438.
• @yilin-succeed（外部贡献者）: 当前 Goal writer 的 objective 隔离回归 — #4418.
• @loopx-agent（Agent 作者）: Agent 作者的工作区/采用、结算恢复、stride/checkpoint、桌面与发布集成工作 — #5890, #6191, #6243, #6298, #6305, #6385, #6442, #6462, #6492, #6499.

可选能力启用与使用

用当前授权目标填写 Goal/Agent/instance；预览不授权执行。source 示例需要匹配 v1.4.1 checkout，独立 provider 须分别安装。

Goal workspace and result adoption

**启用：**使用 loopx chat --goal-id GOAL 打开既有打包工作区，查看 Goal 工作、共享任务、工作目录及已验收的后续结果。先发现并阅读精确版本，再显式确认采用；发现不会写入。

**验证：**重载后核对保留任务、当前会话与原来源/版本；来源过期或不可用必须撤回采用，确认响应不明先对账原回执。

**停用 / 回退：**关闭详情，或 Ctrl+C 停止自己启动的 Chat 服务。采用撤回沿用原来源/版本流程，不删除执行历史或重启工作来清理显示。

**权限边界：**界面不授予执行、同伴租约、验收、仓库或模型权限；采用与独立验收分别呈现，选择目录不增加目录访问授权。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/apps/presentation/dashboard/README.md

loopx chat --goal-id "$GOAL_ID"
# In another terminal, read the same Goal:
loopx --format json status --goal-id "$GOAL_ID"

Governed local work amendments

**启用：**显式使用支持的 active source_session_v1 Goal 与原始 File/SQLite 创建承诺。在 Goal 设置→任务所有权→Goal 承诺与受治理变更中审阅并冻结工作模板；CLI 写入默认预览，--execute 才应用。

**验证：**先 --governed inspect，回复丢失按原 operation 查询 --governed receipt；提交保留既有 Todo/租约、创建未认领工作并推进一个 canonical revision。同伴采纳、replan 结算另行读回。

**停用 / 回退：**停止使用入口即停止新操作，必要时由既有 Goal 生命周期前向退役；保留不可变 intent、历史 reader 与 fence。停用界面不撤销已提交工作，不能删除 intent 来恢复旧 writer。

**权限边界：**只委托声明的 work-item 类别，不授予保护操作、改写验收、任意 Todo 内容、自动 Turn 采用、请求预留或整 Goal 完成；缺少原始承诺的旧 Goal 需要原 owner 迁移。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/architecture/rfcs/shared-goal-alignment-and-governed-amendment-v0.md

loopx --format json goal-amendment-proposal --governed inspect \
  --goal-id "$GOAL_ID" --goal-instance-id "$GOAL_INSTANCE_ID"
loopx goal-amendment-proposal --help

Goal material direction basis

**启用：**在 Goal 设置→任务所有权→方向材料与版本中显式声明本地材料，按审阅过的 provider revision 与 authority digest 应用。每个已注册 Agent 分别读取，并确认返回的原 offer/digest。

**验证：**对精确 Goal instance/Agent 执行 goal-direction inspect，再查原操作回执；读取不等于确认，字节变化或后续不可用必须使旧通过投影失效。

**停用 / 回退：**使用同一设置的停用入口，或 goal-direction --operation disable，--request-json 带当前 provider revision 与新的 operation ID；先预览再 --execute。保留回执及支持 generation 的兼容 reader。

**权限边界：**本地有界读取不证明私有远程发现、模型理解、路线采用、Vision 写入或验收；不能借同伴回执代替本 Agent，历史重放不授予新来源访问。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/architecture/rfcs/goal-direction-baseline-v0.md

loopx --format json goal-direction --operation inspect --goal-id "$GOAL_ID" \
  --goal-instance-id "$GOAL_INSTANCE_ID" --agent-id "$AGENT_ID"
loopx goal-direction --help

Evidence-conditioned stride reviews

**启用：**默认 off。在既有“Goal 复盘与步幅”编辑器选择 evidence_conditioned_v1，预览后应用。CLI 单字段策略保留已有周期；能力 JSON 仅含策略会清除 Goal 周期覆盖并继承设备实时周期，普通默认仍为六个已结算工作 Turn。

**验证：**通过 configure-goal 和下一次新鲜复盘读回。启用后要求来源绑定的 grow/switch/verify 提案，分别论证工作规模、路线距离、证据广度、成本与采纳标准；后续证据与原决定一起返回。

停用 / 回退：--execution-stride-strategy off --execute 关闭策略并保留周期；同时清除策略与两个周期字段才恢复设备继承。保留历史，不把冻结旧回执重解释为新策略。

**权限边界：**配置不授予实验启动、quota、checkpoint 复制、放宽验收或自动转向；决定与读回不证明模型采用、实际效用或 benchmark 增益。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/protocols/goal-vision-replan-contract-v0.md

loopx configure-goal --goal-id "$GOAL_ID" --execution-stride-strategy evidence_conditioned_v1
# Apply only the reviewed preview:
loopx configure-goal --goal-id "$GOAL_ID" --execution-stride-strategy evidence_conditioned_v1 --execute
loopx --format json configure-goal --goal-id "$GOAL_ID"
loopx configure-goal --goal-id "$GOAL_ID" --execution-stride-strategy off --execute

Explore scoped composition

**启用：**在既有 Explore Goal 能力编辑器显式选择 explicit_only 与 scope，审阅后应用；disabled 普通路径保持不变。

**验证：**读回有效 configure-goal 与当前义务；关闭研究义务必须匹配 gap/obligation/Todo/experiment 的精确 lineage。活动图过大应明确拒绝，不能据此认定没有工作。

**停用 / 回退：**经 configure-goal 或同一编辑器预览并应用 --explore-composition-mode disabled；保留发现、scope 与历史实验，不改回执伪造当前性。

**权限边界：**组合只增加证据资格，不授予模型、评分、工具、远程写入、实验预算或验收权限。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/explore/README.md

loopx --format json configure-goal --goal-id "$GOAL_ID"
loopx configure-goal --goal-id "$GOAL_ID" --explore-composition-mode disabled
# Add --execute only after reviewing the preview.

Research checkpoint and warm continuation

**启用：**仅开发者/provider 显式启用：声明 --checkpoint-request，另行保护 --checkpoint-controller-file。warm receiver 显式提供 --restore-checkpoint/--restore-sha256/--restore-config 及新的 Goal/Agent；固定文档 provider revision，保留原剩余 deadline 与评估机会。 普通 Codex Goal 显式使用 research-checkpoint capture、审阅过的 request JSON 和停止工作的 completed native source；先校验精确 artifact，再分别 admit-session、adopt-context。

**验证：**接纳 receiver 前先验证不可变制品及原 capture/restore 回执；正常 cadence 前核对保留研究、native cursor 与当前 preparation/replan。响应丢失按原 operation 对账；以下 help 不启动 benchmark。

**停用 / 回退：**新 worker 不声明 checkpoint 即关闭捕获，不传 restore 参数即冷启动。只经原 runner 停止拥有的 receiver/controller；不复制活跃源、不改旧 deadline/回执。显式阶段延长须在比较两臂同等声明。

**权限边界：**归档验证不等于真实研究续接；当前是有界 CLI/provider 组合阶段，完整前端恢复、实际科学 profile、持续续接与配对增益尚未认证。不增加评分、注册、提交、远程写入或预算权限。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/research-checkpoint.md

loopx research-checkpoint --help
loopx research-checkpoint capture --help
loopx research-checkpoint verify --help
loopx research-checkpoint admit-session --help
loopx research-checkpoint adopt-context --help

Managed ZCode native Goal

**启用：**在既有 Goal 详情或 zcode-goal bind 中显式绑定当前 Goal 与已注册 Agent，需要时选已公布模型，再明确 start；bind 只创建空闲会话，不发模型请求，技能安装另行处理。

**验证：**读回 zcode-goal status。验收/来源 basis 改变或缺失时，普通 resume 拒绝并保持暂停；检查当前 basis 后先 stop，再显式 start 交付新目标。断联观察保持 unknown。

**停用 / 回退：**zcode-goal stop 关闭拥有的进程、清除 target 并保留历史/binding；不 start/resume 即保持停用。slash-commands --uninstall --surface zcode 只移除拥有的技能，不停止原会话。

**权限边界：**Core quota 接纳/撤销不是 native 按 token 硬预算，native 完成不结算 LoopX 工作或证明验收；不增加 Desktop attachment、账号/模型凭据、工具、权限、MCP/Hooks 或 Automations。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/zcode_goal_mode/README.md

loopx doctor --agent-type zcode
loopx --format json zcode-goal status --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"
loopx --format json zcode-goal stop --goal-id "$GOAL_ID" --agent-id "$AGENT_ID"

Claude Goal host and UTF-8 transport

**启用：**通过 slash-commands --install --surface claude-code 显式刷新拥有的 Claude 技能/adapter，再启动既有连接 Goal 流程；launcher 保留解释器路径引用并固定 UTF-8。

**验证：**读回 doctor 与原 host 状态；PreToolUse 输入读不到须拒绝，不能静默不作决定。本发布不认证 Claude 真实登录模型执行。

**停用 / 回退：**先经原 controller 停止拥有的循环，需要移除入口时 slash-commands --uninstall --surface claude-code；保留用户配置与历史。

**权限边界：**交付 host 材料不增加 shell、文件、工具、Goal、quota 或账号权限；UTF-8 修复只解决兼容性。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/claude_goal_mode/README.md

loopx slash-commands --install --surface claude-code
loopx doctor
# Stop the owned host before uninstalling its facades:
loopx slash-commands --uninstall --surface claude-code

DSH host and operation transport

**启用：**在已连接 Goal 的 host 配置中显式选择 deepseek-harness，并使用原授权模型/provider；--host dsh 选择既有进程内 adapter。Core 更新不单独发布或升级 DSH plugin。

**验证：**读回原 Turn journal 与 transport-specific preflight；DSH 模型 runtime 可用不证明 native 操作卡已投递。保留 UTF-8 与移除本地来源的观察诊断。

**停用 / 回退：**经原 controller 停止拥有的循环，新运行不选择该 host/profile 即停用。未知操作按原 intent/receipt 恢复，不盲目重投；plugin 停用/卸载遵循独立版本生命周期。

**权限边界：**Host 执行与 native 操作 transport 分开；配置不授予模型凭据、外部频道、确认、支付/下单或自动计划权限。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/dsh_goal_mode/README.md

python -m loopx.dsh_goal_mode --help
loopx turn run-once --help
loopx doctor

Extension lifecycle fail-closed readback

**启用：**通过既有 extension 生命周期安装/启用明确选择的兼容扩展；先核对 requires_loopx_api、capability/provider 声明和拥有的 runtime。旧 unverified 状态可读不等于可用。

**验证：**Core 更新后运行 extension doctor --all-enabled --execute；保存的快照重新校验当前 API，重复或冲突声明拒绝，已停用扩展升级/rollback 后必须仍停用。

**停用 / 回退：**extension disable ID --execute 停止接纳；extension rollback ID 选择保留的兼容 revision，再 doctor 后才启用。不改快照或缓存 readiness 绕过兼容检查。

**权限边界：**安装/可读不等于启用、继承授权或验证通过；扩展不增加 Core 验收、权限、调度或外部来源访问。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/extensions.md

loopx extension list --format json
loopx extension doctor --all-enabled --execute --format json
# For an explicitly selected installed extension:
loopx extension disable "$EXTENSION_ID" --execute --format json

Public-safe outbound scans

**启用：**builtin 默认可用，扫描须每次显式调用。需要本地私有标记时才配置可读 PUBLIC_SAFE_CONFIG；现在扫描凭据点文件/无扩展名文件，并按文档排除 VCS/vendor/binary。

**验证：**对实际目标 root 或 staged diff 运行 scan_cli；配置、内容、目录不可读、root 缺失或输入过大均失败关闭。回执仅保留有界标签/行号，不带邻近正文。

**停用 / 回退：**不再调用即停止扫描。只有明确选择通用规则模式才 unset PUBLIC_SAFE_CONFIG，不能当作修复必需私有策略缺失；恢复原策略/root 后再提交前重扫。

**权限边界：**干净扫描不证明保密完备，也不授予对外发布；不增加文件修改、上传、权限或仓库写入。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/public_safe_outbound/README.md

python -m loopx.capabilities.public_safe_outbound.scan_cli --scan-root . --format json
# Alternatively, inspect only the actual staged diff:
git diff --cached | python -m loopx.capabilities.public_safe_outbound.scan_cli --diff --format json

Finance research and quantity reconciliation

**启用：**可选独立 provider；从 v1.4.1 checkout 显式安装 package/manifest，启用后 doctor。选择审阅过的 quantity/funding/stock 输入；Core 发布不单独发布该 distribution。

**验证：**运行版本化合成 quantity-research 示例，在既有研究审阅面读回来源绑定的存量时点、有符号流量、单位/期间/版本及缺失覆盖；私有文件引用必须拒绝。

**停用 / 回退：**disable loopx-finance-value-discovery，或 rollback 到保留且合格的 provider；保留原期间、来源和回执，不把存量变化虚构为现金流观测。

**权限边界：**研究计算不证明来源真实性、投资适用性或交易许可；不增加 corpus、场所账号、确认、下单、转账、签署或验收权限。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/packages/loopx-finance-value-discovery/README.md

python -m pip install ./packages/loopx-finance-value-discovery
loopx extension install --manifest packages/loopx-finance-value-discovery/extension.toml --execute --format json
loopx extension doctor loopx-finance-value-discovery --execute --format json
loopx extension run loopx-finance-value-discovery --input-json packages/loopx-finance-value-discovery/examples/quantity-research-v1.json --execute --format json
loopx extension disable loopx-finance-value-discovery --execute --format json

Finance execution bounded readback

**启用：**可选独立 provider；显式安装/启用版本化 package/manifest 并 doctor。CLI 接受 --check-order-summary、--reconcile-execution、--check-risk-bound 的精确审阅 JSON；默认 M1 模拟不变。

**验证：**比较未过期精确订单、保留 venue bounds 与调用者现金/损失预算；金额缺失、不支持/cross-margin、证据过期/漂移或超预算仍 unknown/held。计算读回不认证 venue 证据。

**停用 / 回退：**disable loopx-finance-execution 或 rollback 到保留合格 revision；保留原确认和订单/结果 ID，未知 venue 结果只按原 ID 恢复，不自动重投。

**权限边界：**计算检查不认证真实 venue 订单、真实确认、live DSH worker/卡片返回、资金预留或交易权；Core 不单独发布该 provider 或授予金融执行。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/packages/loopx-finance-execution/README.md

python -m pip install ./packages/loopx-finance-execution
loopx extension install --manifest packages/loopx-finance-execution/extension.toml --execute --format json
loopx extension doctor loopx-finance-execution --execute --format json
loopx-finance-execution --help
loopx extension disable loopx-finance-execution --execute --format json

Revision-bound Reward Memory recall

**启用：**默认 off，只作用于单 Goal 的显式 Agent。通过既有 configure-goal --reward-memory-config CONFIG --reward-memory-agent AGENT 预览，再在 corpus owner 授权及 provider canary 写入/精确读回后 --execute。

**验证：**对同 Goal/Agent 读 experiment-status；recall 必须绑定实际 invocation revision，配置/来源漂移不能复用旧 available 回执。按原 registry/config owner 修复，不向历史复制 digest。

**停用 / 回退：**预览并应用 configure-goal --clear-reward-memory-config 停用本 Goal binding，保留 corpus 与回执；provider 删除另需独立授权。

**权限边界：**可召回证据不等于通用个人策略或执行授权；不推导跨 Agent/Goal corpus、凭据、提权、后台学习或来源真实性。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/reward_memory/README.md

loopx reward-memory experiment-status --goal-id "$GOAL_ID" --agent-id "$AGENT_ID" --format json
loopx configure-goal --goal-id "$GOAL_ID" --clear-reward-memory-config
# Add --execute only after reviewing this Goal-scoped disable preview.

Periodic report profiles and declared adapters

**启用：**逐命令选择 profile：inspect-profile --preset weekly 只读内置配置，不启动 scheduler/provider。自定义用 --profile-json；省略 enabled 即 false，启用项必须声明来源和 renderer。正式生成/交付前由原 owner 配置声明的 adapter。

**验证：**执行 periodic-report inspect-profile --preset weekly，检查激活回执。生成和交付分别读回：规范化 artifact、sink 身份、订阅权限与精确结果均须一致；声明 adapter 失败不能静默省略。

**停用 / 回退：**将自定义 profile 的 enabled 设为 false，并由原 owner 停用订阅，使待交付项失去该授权；去掉可选 archive/provider 声明即关闭对应路径。交付不明按原 report key/回执对账。

**权限边界：**检查配置不启动 provider、发布 artifact、发送消息或授予访问。生成不等于正式交付，归档、目标凭据与有效订阅仍须显式配置。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/loopx/capabilities/periodic_report/README.md

loopx periodic-report inspect-profile --preset weekly --format json
loopx periodic-report --help

Desktop publication preflight

**启用：**维护者逐次 dispatch 启用：在审阅过的 workflow ref 上运行 desktop-release-artifacts.yml，提供已接纳 tag 与 publication_probe=true；选择只读发布探针，关闭该次资产上传。

**验证：**检查该次 Actions run 并下载 preflight 证据，核对 builder/tag、回执和访问结果；本机测试不证明托管探针成功。

**停用 / 回退：**后续 dispatch 省略 publication_probe 或设为 false 恢复正常发布路径；这不撤回已有资产，也不授予发布权限。保留原 run 和证据。

**权限边界：**探针不授予创建 release、覆盖资产、签名密钥或其他仓库访问；认证上传与 installer 资格验证另行完成。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/product/release-readiness.md

gh workflow run desktop-release-artifacts.yml --repo loopx-project/loopx \
  --ref "$REVIEWED_WORKFLOW_REF" -f tag="$EXISTING_ADMITTED_TAG" -f publication_probe=true
gh run view "$PROBE_RUN_ID" --repo loopx-project/loopx
gh run download "$PROBE_RUN_ID" --repo loopx-project/loopx

Managed pull request review workflow

**启用：**用 project-skill install 为选定 host 显式安装 loopx-pr-review，先检查预览再 --execute。调用技能或 pr-review 时固定 NUMBER@HEAD；head 变化须重新审阅。

**验证：**读 project-skill status 与保留的 capability-owned review packet；对精确 result/packet 执行 pr-review --check-result，发布完整双语评审并读回。可选解释图只辅助理解。

**停用 / 回退：**对同 project/skill/surface 预览 project-skill uninstall，再 --execute 移除其托管文件；保留评审及用户指令，不复用旧 head 的通过结论。

**权限边界：**安装技能或评审通过不授予 approve/merge、admin bypass、首屏批准、benchmark 启动或私有证据发布权；合并仍需精确 head readiness 与维护者明确授权。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/skills/loopx-pr-review/SKILL.md

loopx project-skill status --project . --skill loopx-pr-review --surface codex
loopx project-skill install --project . --skill loopx-pr-review --surface codex
loopx pr-review --target-exact-head "$PR_NUMBER@$HEAD_OID" --format json
loopx project-skill uninstall --project . --skill loopx-pr-review --surface codex

Original request correction and effect recovery

**启用：**对已有精确请求逐操作启用：manager-inbox control 带 Goal instance、原 receiver、request ID 与 inspect/revise/cancel/recover。修正带审阅过的 control JSON 和稳定 operation ID，先预览再 --execute。原对话 owner 必须有持久化 route/grant，不能猜 thread。

**验证：**读回原请求与 effect 结果。预留 Goal effect 不明时，先按原 operation 恢复再开启修订；只有独立读回的终态 Goal receipt 才能释放预留。CLI、MCP 和原对话操作沿用同一范围校验。

**停用 / 回退：**不调用 control 即保持原请求；--control-operation cancel 带新的稳定 operation ID，经预览后取消精确请求。取消不回滚已提交 Goal 工作；保留原回执，不明结果用 recover。

**权限边界：**仅原授权 requester/receiver 或已验证的原对话主体可操作；不允许冒充、跨 Goal 请求、新保护操作、改写验收或回滚已完成 effect。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/architecture/rfcs/shared-goal-alignment-and-governed-amendment-v0.md

loopx manager-inbox control --goal-id "$GOAL_ID" --goal-instance-id "$GOAL_INSTANCE_ID" \
  --peer-agent-id "$RECEIVER_AGENT_ID" --request-id "$REQUEST_ID" --control-operation inspect --format json
loopx manager-inbox --help

Ordinary project Codex Git writes

**启用：**沿用普通项目既有 workspace_write grant（chat 既有默认）与 host_default 文件系统范围，不新增能力开关；显式 --project-workspace-grant workspace_read 关闭 App 写入 grant。接纳 start/resume 时仅把项目真实 .git 目录加入 native 可写 roots。

**验证：**先看 chat --help 与选定 Codex host 的 doctor，再核对 dispatch 前既有 native thread 权限读回；Git root 缺失或不同即拒绝。已在授权隔离项目验证 native fetch/branch/commit 与同 thread resume，无模型请求。

**停用 / 回退：**停止拥有的 Chat server/session，以 --project-workspace-grant workspace_read 重启并创建新范围 thread 来去掉写权。--project-filesystem-scope workspace_only 保持社区隔离，不增加普通 host-default Git root。保留历史和 native 配置。

**权限边界：**文件系统支持不授予发布、破坏性 Git、远程账号、额外网络、外部 worktree 元数据、symlink 目标或其他项目访问。保留原 network/tmp 与已有 roots；运行中的 Bot 仍需按原流程更新/重启。

文档： https://github.com/loopx-project/loopx/blob/v1.4.1/docs/reference/protocols/host-integration-surface-v0.md

loopx chat --help
loopx doctor --agent-type codex
# Existing server activation/restriction choices (stop only your owned server first):
# loopx chat --project-workspace-grant workspace_write --project-filesystem-scope host_default
# loopx chat --project-workspace-grant workspace_read

发布验证

精确源码 4f27801c4e26fd51246d352eada72c98f8e7e657：2,213项定向测试通过，5项 native阶段显式跳过；515项公开smoke及静态、隔离升级/技能/UI、真实PostgreSQL验证通过。Evolving/high（单工具 API profile）完成22场景×2次actor调用、6项对照，零失败/跳过；仅有界连接/不完整流重试，无语义重投。保留旧失败；不声称完整最终树pytest、配对长程或科学增益。

GitHub/PyPI包、四项桌面制品、签名Mac更新、desktop-stable与源码stable均已验证1.4.1；真实PyPI1.4.0升级及doctor通过。恢复仅修测试，产品仍为4f27801c4。Mac为ad-hoc、无Apple公证；未运行Windows客户端。

Compare: v1.4.0...v1.4.1