@scotthuang/agent-knock-knock 0.3.2 → 0.4.0

package/CHANGELOG.md

@@ -1,581 +1,19 @@
 # Changelog
 
-## 0.3.2 - 2026-07-28
+## 0.4.0 - 2026-07-29
 
 ### Changed
 
-- Build against the verified OpenClaw `2026.6.5` installation floor so ClawHub reports the package without a minimum-host-version drift warning.
-
-## 0.3.1 - 2026-07-28
-
-### Changed
-
-- Lower the OpenClaw plugin API, Gateway, and npm peer floor to the first verified stable API release, `2026.5.12`, and the normal installation host floor to `2026.6.5`, while continuing to build against `2026.7.1-2`.
-- Add isolated compatibility coverage for the minimum and build-time OpenClaw hosts, plus an explicit failing boundary at `2026.5.10-beta.2`.
-
-## 0.3.0 - 2026-07-28
-
-### Added
-
-- Publish the complete OpenClaw plugin, bundled AKK skill, and package-local relay through ClawHub, with stable npm and ClawHub installation paths.
-- Add bounded runtime diagnostics, fail-closed session selectors, stable short references, five-minute tmux and Managed ACPX quickstarts, and opt-in live smoke scripts.
-
-### Changed
-
-- Make installation configure and verify the OpenClaw runtime while preserving unrelated plugin settings.
-- Confirm asynchronous Gateway callback runs before settling delivery, instead of treating a `started` acknowledgement as delivered.
-
-### Fixed
-
-- Find the matching Codex completion across plausible same-workspace rollouts and reconcile durable completion before a newer same-pane task can supersede it.
-- Keep uncertain terminal completion and callback delivery states retryable without duplicating callbacks or attributing a newer task's screen output to an older monitor.
-
-## 0.3.0-beta.2 - 2026-07-28
-
-### Added
-
-- Add bounded real probes and OpenClaw config, plugin runtime, skill, workspace, and Gateway diagnostics to `doctor`.
-- Add fail-closed session selectors, stable short references, and independent five-minute tmux and Managed ACPX quickstarts.
-- Add deterministic onboarding coverage and explicitly opt-in live ACPX and tmux smoke scripts.
-
-### Changed
-
-- Make `install-openclaw` configure workspace, default agent, and mode atomically, preserve approval policy, restart the Gateway at most once, and optionally verify readiness.
-- Verify the packaged README after every ClawHub publish and the default README after stable releases.
-
-## 0.3.0-beta.1 - 2026-07-28
-
-### Added
-
-- Add ClawHub package metadata, bundled skill discovery, a square catalog icon, and a package-local OpenClaw runtime entry.
-- Add manifest/runtime contract tests and a ClawHub validation and dry-run release path.
-
-### Changed
-
-- Compile against the tested OpenClaw `2026.7.1-2` SDK instead of a handwritten type shim.
-- Separate ClawHub and npm installation guidance, including the explicit optional-tool grant required by a clean OpenClaw profile.
-- Publish prerelease builds to npm's `next` dist-tag and mark their GitHub releases as prereleases.
-
-### Fixed
-
-- Align the declared `/akk` command, bundled skill, and tool inventory with the capabilities registered at runtime.
-- Resolve the bundled CLI relative to the loaded plugin entry so ClawHub installs do not depend on a global executable.
-
-## 0.2.51 - 2026-07-25
-
-### Added
-
-- Add disabled-by-default exact command and workspace auto-approval for hookless Claude Code tmux turns by intersecting the current one-time Bash dialog with owner-private local transcript evidence.
-
-### Changed
-
-- Allow trusted `autoApprove` rules to target Claude as well as Codex, while leaving unmatched, unsupported, or ambiguous requests on the manual approval path.
-
-### Fixed
-
-- Deliver Gateway callback attempts outside conversation locks and recover interrupted approval-notification outboxes with stable message identities without duplicating message events.
-- Preserve terminal monitor ownership across inline auto-approval, callback retries, and post-delivery crashes, and track consumed Claude dialogs by transcript request identity plus an observed prompt-generation boundary so redraws cannot replay Enter while later identical requests remain approvable.
-
-### Security
-
-- Keep raw Claude commands executor-local and expose only bounded hashes and opaque identities through callbacks, state, events, logs, and screen summaries.
-- Re-evaluate the exact policy and recapture the prompt, selected one-time action, transcript evidence, and process identity after the one-shot dispatch reservation before sending one Enter; stale, changed, replayed, or uncertain requests fail closed.
-- Restrict pending-approval transcript evidence to compatible Claude Code `2.1.x` releases at `2.1.198` or later while retaining forward-compatible durable completion detection.
-
-## 0.2.50 - 2026-07-24
-
-### Changed
-
-- Verify hook-free Claude tmux completion and approval on Claude Code `2.1.218`, and accept newer versions when their transcripts preserve the required identity and completion structure instead of requiring an exact release allowlist.
-
-### Fixed
-
-- Recognize Claude Code `2.1.218`'s `ctrl+e to explain` Bash approval footer while continuing to approve only the highlighted one-time `Yes` choice.
-
-### Security
-
-- Reject mixed transcript versions within one turn and keep nonzero or malformed Claude pending-agent and pending-workflow counts from completing early.
-
-## 0.2.49 - 2026-07-24
-
-### Added
-
-- Detect hook-free Claude Code tmux completion from the owner-private local JSONL transcript by binding the exact process, session, prompt, file boundary, `end_turn`, and `turn_duration`, then reusing the existing deterministic, idempotent terminal callback flow.
-
-### Changed
-
-- Make tmux and ACPX/ACP independent installation choices, with tmux recommended for shared human/agent control.
-- Make Claude Code tmux control and durable completion hook-free by default.
-- Disable the legacy `install-claude-hooks` CLI so AKK no longer writes Claude Code settings.
-- Recover persisted callback outbox entries after process or Gateway restart, with bounded delivery timeouts and idempotent retries.
-
-### Security
-
-- Limit Claude screen approval to explicit user confirmation for a current managed turn and an exact one-time Bash prompt.
-- Bind approval fingerprints to the raw screen digest, terminal/process identity, conversation, and message; revalidate before input and reject expired, reused, stale, or automatic Claude approvals.
-- Read only the post-send Claude transcript range from a non-symlink, owner-private file; reject unknown schemas, partial writes, rotation, truncation, PID reuse, ambiguous prompts, unresolved tools, and background or sidechain work.
-
-## 0.2.48 - 2026-07-23
-
-### Added
-
-- Add Node 22/24 CI, Dependabot coverage, contribution templates, and a social preview asset.
-
-### Changed
-
-- Refresh the README and project metadata around OpenClaw orchestration, ACPX delegation, and shared Codex/Claude tmux control.
-- Pin the default Codex ACP adapter to the tested `@agentclientprotocol/codex-acp@1.1.7` release and require Node.js 22.14+.
-- Keep the installed OpenClaw skill on token-free plugin callbacks instead of recommending the legacy direct-token script.
-
-### Fixed
-
-- Respect the configured default agent and custom store for `/akk` commands, preserve callback delivery for slash-command tmux sends, and keep agent-specific settings out of unrelated follow-ups.
-- Include native and terminal sessions in `/akk list`, resolve relative plugin stores from the configured workspace, and recognize Claude's native versioned executable path.
-- Reconcile eligible tmux monitors after Gateway restarts, prevent duplicate monitors, preserve wrapper process ancestry across list and takeover, and safely migrate exact legacy Codex terminal identities.
-- Make CLI help, version, doctor requirements, and failure exit status predictable.
-
-### Security
-
-- Require an exact live agent identity before tmux input, revalidate auto approvals inside the executor against the observed prompt, and avoid persisting unmanaged Claude hook payloads.
-- Recheck keys-mode approval fingerprints immediately before input and use tmux bracketed paste for multiline tasks.
-- Keep Gateway tokens out of normal plugin callbacks, coding-agent environments, CLI output, and child-process arguments.
-- Store local task data with private permissions, atomic state replacement, tokenized control locks, serialized stale-lock recovery and event writes, path-containment checks, and symlink rejection.
-- Refuse shared runtime-log directories without changing their permissions and append through no-follow file descriptors.
-
-## 0.2.47 - 2026-07-23
-
-### Added
-
-- Add tmux control for Claude Code, including discovery, safe idle-only sends, Hook-backed monitoring, structured allow/deny, and exact-rule auto approval.
-- Add an idempotent `install-claude-hooks` command that preserves existing Claude settings and hooks.
-
-### Fixed
-
-- Bind Claude permission and Stop/StopFailure events to strong runtime identity, terminal, conversation, message, and active lease; keep background work from completing early and deliver each terminal result once.
-- Show a redacted permission target summary, preserve conservative manual fallback for unverifiable dialogs, and normalize only the exact source-tagged Tokenjuice Claude wrapper for policy matching.
-
-## 0.2.46 - 2026-07-22
-
-### Added
-
-- Add a registry-driven terminal-agent adapter layer, neutral process discovery, agent-aware terminal IDs, and a documented extension path while keeping legacy Codex IDs compatible.
-
-### Security
-
-- Fail closed on missing capabilities and revalidate the full ordered approval-key action against a freshly captured prompt before sending terminal input.
-
-## 0.2.45 - 2026-07-20
-
-### Added
-
-- Persist terminal callback delivery state and add bounded automatic retries plus `AKK retry-callback` / `agent_knock_knock_retry_callback` for idempotent manual recovery after delivery failures.
-
-### Fixed
-
-- Require a request-hash-matched Codex rollout `task_complete` (or terminal-screen completion fallback) before closing terminal bridge tasks, preventing same-workspace assistant progress from another or unfinished turn from being reported as done.
-- Keep terminal bridge tasks in `callback_pending` or `callback_failed` until their OpenClaw callback is actually delivered, and close them only after successful delivery.
-
-## 0.2.44 - 2026-07-19
-
-### Fixed
-
-- Retry local OpenClaw plugin installation with the supported `--force` flow when OpenClaw explicitly requires trust confirmation for a reviewed local source.
-
-## 0.2.43 - 2026-07-19
-
-### Added
-
-- Make terminal bridge inactivity timeouts activity-aware, retain a separate 12-hour hard lifetime, and add `AKK renew` plus matching OpenClaw skill and tool support for explicitly extending long-running tasks.
-- Persist timeout deadlines and activity evidence so renewals and stalls remain auditable without sending any input to the coding agent.
-
-### Fixed
-
-- Bind Codex terminal callbacks to a matching rollout `task_complete` turn and request hash so final-answer prose cannot be mistaken for a working marker or a result from another task.
-- Atomically supersede older bridge tasks when a tmux pane is reused, while preserving the previous callback path if replacement delivery fails and serializing concurrent sends.
-- Preserve terminal-screen completion fallback for wrapped or scrolled prompts by comparing the completed screen with a pre-send fingerprint.
-
-## 0.2.42 - 2026-07-13
-
-### Fixed
-
-- Detect idle Codex terminal completion from the `Worked for ...` divider when rollout history is unavailable and the visible final answer contains no progress bullet markers.
-- Exclude the completion divider, input placeholder, and model footer from terminal-screen fallback callbacks.
-
-## 0.2.41 - 2026-07-11
-
-### Added
-
-- Add deterministic, default-disabled auto-approval policies for terminal-controlled Codex command prompts, with exact command and workspace rules configured through the trusted OpenClaw plugin config.
-
-### Security
-
-- Revalidate approval fingerprints immediately before automatic approval, reject shell composition and workspace escapes, and audit the matched rule and policy fingerprint without involving an LLM.
-
-## 0.2.40 - 2026-07-10
-
-### Fixed
-
-- Reinstall existing OpenClaw plugins with a supported force-replace command instead of combining incompatible `--force` and `--link` flags.
-
-## 0.2.39 - 2026-07-10
-
-### Fixed
-
-- Prevent terminal bridge monitors from treating intermediate Codex progress as completion while terminal activity is unknown, and redact environment-style secrets from terminal excerpts.
-
-## 0.2.38 - 2026-07-10
-
-### Fixed
-
-- Make `install-openclaw` reusable for existing plugin installations and add `--skill-only` for synchronizing the OpenClaw skill without reinstalling the plugin.
-
-## 0.2.37 - 2026-07-10
-
-### Fixed
-
-- Close completed terminal bridge callback conversations so one-shot tmux sends do not remain in the default delegated list.
-
-## 0.2.36 - 2026-07-10
-
-### Added
-
-- Notify OpenClaw when a terminal bridge Codex session is waiting for approval, including explicit approve and deny instructions.
-
-### Fixed
-
-- Resume terminal bridge monitoring after `AKK approve` succeeds so approved Codex work can still deliver its final callback.
-
-## 0.2.35 - 2026-07-09
-
-### Added
-
-- Add a README cover illustration for Agent Knock Knock and include the optimized cover asset in the npm package.
-
-## 0.2.34 - 2026-07-09
-
-### Fixed
-
-- Replace the README architecture image with an inline Mermaid diagram so the diagram renders without depending on raw GitHub image delivery.
-
-## 0.2.33 - 2026-07-09
-
-### Fixed
-
-- Include the README architecture diagram asset in the published npm package so relative image links resolve on package readme surfaces.
-
-## 0.2.32 - 2026-07-09
-
-### Fixed
-
-- Added an idle terminal-screen fallback for terminal bridge callbacks when Codex rollout history is unavailable or does not expose the current tmux task result.
-
-## 0.2.31 - 2026-07-09
-
-### Fixed
-
-- Fixed terminal bridge callback delivery by avoiding unauthenticated gateway URL overrides and persisting the bridge send baseline so stale cwd-matched Codex rollout messages are not treated as the current task result.
-
-## 0.2.30 - 2026-07-09
-
-### Added
-
-- Added terminal bridge monitoring for tmux-controlled Codex conversations so AKK can send only the user-facing task text, observe Codex rollout/terminal state, and deliver the OpenClaw callback itself.
-
-## 0.2.29 - 2026-07-08
-
-### Fixed
-
-- Clarified AKK skill and `agent_knock_knock_send` tool routing so messages to listed AKK/Codex sessions, terminal-controlled rows, or tmux targets such as `my-work:0.1` use `send` instead of starting a new delegation.
-
-## 0.2.28 - 2026-07-07
-
-### Fixed
-
-- Trimmed trailing newlines before sending terminal-controlled Codex messages through tmux and submitted with `C-m`, preventing payload newlines from being left in the Codex input box without starting the task.
-
-## 0.2.27 - 2026-07-06
-
-### Fixed
-
-- Fixed Codex prompt delivery when using an explicit ACPX `--agent` adapter by invoking the top-level `prompt` subcommand before `-s <session>`.
-
-## 0.2.26 - 2026-07-06
-
-### Fixed
-
-- Routed AKK Codex ACPX delegation through the supported `@agentclientprotocol/codex-acp` adapter with an explicit `--agent` command, avoiding OpenClaw ACPX installs that still bundle deprecated `@zed-industries/codex-acp`.
-- Kept `AKK_CODEX_ACPX_AGENT_COMMAND` as an override for custom Codex ACP adapters while refusing explicit `@zed-industries/codex-acp` overrides.
-
-## 0.2.25 - 2026-07-06
-
-### Fixed
-
-- Refused to start Codex through ACPX installations that still reference deprecated `@zed-industries/codex-acp`, preventing AKK from repeatedly launching stale or quarantined Codex adapter binaries.
-
-## 0.2.24 - 2026-07-04
-
-### Fixed
-
-- Ignored injected Codex `<environment_context>` messages when choosing the initial request for `AKK describe`, falling back to the Codex thread title when no real user message is available.
-
-## 0.2.23 - 2026-07-04
-
-### Fixed
-
-- Declared `agent_knock_knock_describe` in the OpenClaw plugin contract and metadata so OpenClaw exposes the new describe tool.
-- Updated the installed AKK skill routing rules to use `agent_knock_knock_describe` for session-content summary requests instead of falling back to direct terminal inspection.
-
-## 0.2.22 - 2026-07-04
-
-### Added
-
-- Added `AKK describe` / `AKK summary` to summarize what an AKK-managed, native Codex, or terminal-controlled Codex session is about.
-- Added the `agent_knock_knock_describe` OpenClaw tool and `/akk describe <conversation-id>` command.
-- Reused Codex rollout history for exact session matches, added cwd-based fallback matching, and returned screen-only summaries with explicit low-confidence limitations when no history can be found.
-
-## 0.2.21 - 2026-07-03
-
-### Added
-
-- Added structured tmux Codex activity detection for terminal-controlled sessions, exposing `activity_state` and `activity_reason` in `AKK status` and `AKK list`.
-- Classified terminal-controlled Codex panes as `awaiting_approval`, `working`, `idle`, or `unknown` using conservative current-screen heuristics.
-- Added regression coverage for approval, stale approval scrollback, working, idle, unknown, and list activity-state output.
-
-## 0.2.20 - 2026-07-02
-
-### Fixed
-
-- Ignored stale Codex approval prompts left in tmux scrollback after later terminal activity, preventing `AKK list`, `AKK status`, and `AKK approve` from treating already-approved prompts as pending.
-- Added regression coverage to ensure stale approval scrollback does not mark terminal-controlled sessions as blocked or send an approval key.
-
-## 0.2.19 - 2026-06-30
-
-### Fixed
-
-- Added a guarded cwd-based tmux pane fallback so wrapper-launched Codex sessions, such as sidecar-managed panes, are detected as `terminal_controlled` when their process cwd uniquely matches a tmux pane path.
-- Preserved existing pid ancestry matching and avoided cwd fallback when multiple panes share the same cwd, preventing ambiguous terminal-control attachment.
-
-## 0.2.18 - 2026-06-30
-
-### Fixed
-
-- Classified Codex remote compact stream disconnects as recoverable executor failures instead of leaving AKK tasks stuck as generic `stalled` conversations.
-- Allowed `AKK recover` to replay saved AKK history and the pending OpenClaw message after a recognized monitor-time executor failure.
-
-## 0.2.13 - 2026-06-25
-
-### Fixed
-
-- Allowed terminal-controlled `AKK status` and `AKK approve` to use `terminal_controlled` ids from `AKK list` directly without requiring an AKK delegated state file.
-- Revalidated the current tmux target and pane pid before reading the terminal screen or sending the approval key.
-
-## 0.2.11 - 2026-06-25
-
-### Added
-
-- Added terminal-controlled `AKK cancel` support by sending Control-C to the controlled tmux pane.
-- Added structured terminal reachability, screen, and approval state to `AKK status` for terminal-controlled sessions.
+- Focus Agent Knock Knock on one transparent execution path: shared tmux terminals for Codex and Claude Code.
+- Reuse exactly one idle agent pane in the configured workspace and fail closed when no unique safe target exists.
+- Prevent a new terminal generation from replacing an active task, persist exact dispatch receipts across stores, and fence uncertain submissions until their owner is explicitly closed.
+- Simplify installation, diagnostics, plugin tools, the bundled skill, and documentation around the tmux handoff workflow.
 
 ### Removed
 
-- Removed the public `agent discover` CLI path and `agent_knock_knock_agent_discover` OpenClaw tool; use `AKK list` for active/local session visibility.
-- Removed the legacy `safe_resume` takeover strategy; native takeover now uses active process-based terminate, terminal-control, or fork paths.
-
-## 0.2.10 - 2026-06-25
-
-### Fixed
-
-- Added parsing for underscore-delimited tmux pane output observed in OpenClaw Gateway terminal diagnostics, allowing those panes to be matched back to Codex processes.
-
-## 0.2.9 - 2026-06-25
-
-### Fixed
-
-- Added a whitespace-delimited tmux pane parser fallback for environments where `tmux list-panes` output is not parsed by the tab-delimited path, and included a bounded `stdoutPreview` in terminal diagnostics.
-
-## 0.2.8 - 2026-06-25
-
-### Added
-
-- Added `AKK list --terminal-debug` and the OpenClaw `terminalDebug` list parameter to expose tmux command, socket, pane, and spawn-attempt diagnostics when terminal-controlled sessions are not detected through the Gateway.
-- Added `scripts/verify-terminal-control.mjs` to validate local Codex/tmux PID mapping against either the source build or an installed package before publishing.
-
-## 0.2.7 - 2026-06-25
-
-### Fixed
-
-- Added tmux socket discovery across `/private/tmp/tmux-*` and `/tmp/tmux-*` directories so OpenClaw Gateway environments do not need to infer the interactive terminal user's tmux socket from their own process UID.
-
-## 0.2.6 - 2026-06-25
-
-### Fixed
-
-- Added tmux executable path fallback for OpenClaw Gateway environments whose service `PATH` cannot resolve `tmux`, allowing `AKK list` and terminal-control actions to discover and operate tmux panes via common absolute install paths.
-
-## 0.2.5 - 2026-06-25
-
-### Added
-
-- Added grouped `AKK list` output for `delegated`, `native`, and `terminal_controlled` sessions while preserving the legacy `tasks` field.
-- Added terminal-controlled approval state to `AKK list`, allowing tmux-backed Codex sessions to show when an approval prompt is visible and approvable.
-
-### Changed
-
-- Updated OpenClaw routing guidance so active/native Codex session questions use `AKK list` instead of the separate native discover tool.
-
-## 0.2.4 - 2026-06-24
-
-### Fixed
-
-- Added tmux socket fallback discovery for OpenClaw Gateway environments whose service `TMPDIR` differs from the user's interactive terminal, allowing active discovery to mark tmux-controlled Codex panes reliably.
-- Preserved the discovered tmux socket path through terminal-control takeover, status, send, and approve operations so follow-up terminal actions use the same tmux server.
-
-## 0.2.3 - 2026-06-24
-
-### Fixed
-
-- Added the tmux takeover tools to the OpenClaw plugin manifest contracts, allowing OpenClaw to expose `agent_knock_knock_agent_discover`, `agent_knock_knock_agent_takeover`, and `agent_knock_knock_approve` after installation.
-
-## 0.2.2 - 2026-06-24
-
-### Added
-
-- Added tmux-backed terminal-control takeover for native Codex sessions, including tmux pane metadata in active discovery, confirmed `terminal_control` attachment, direct follow-up sends to the existing pane, and a conservative `approve` command for the currently visible Codex approval prompt.
-
-## 0.2.1 - 2026-06-22
-
-### Changed
-
-- Deprecated the standalone restart recovery path from the OpenClaw-facing tool, slash command, skill, and docs. Conversations that need recovery now present `recover`, `close`, or starting a new independent delegation.
-- Changed `send` to automatically fall back to AKK replay recovery when the previous ACPX session is unavailable. Explicit recovery decisions remain available only through `--recovery-policy explicit`.
-
-## 0.2.0 - 2026-06-22
-
-### Added
-
-- Added experimental native Codex session discovery and takeover flows, including safe resume, OpenClaw-summarized fork, and confirmed terminate-then-resume adoption for active Codex CLI sessions.
-- Added an explicit `allowCwdOnly` terminate-then-resume fallback for Codex TUI processes that do not expose a session id in argv; it still requires a user-confirmed pid and cwd re-scan before termination.
-- Added tests for Codex native session takeover planning, confirmed termination safeguards, fork summary confirmation, and native `codex exec resume` follow-up delivery.
-
-### Changed
-
-- Updated Codex takeover sends to use native `codex exec resume` for adopted Codex sessions, avoiding ACPX session binding issues when resuming existing terminal Codex sessions.
-
-## 0.1.2 - 2026-06-21
-
-### Fixed
-
-- Fixed stalled-task notifications from the background monitor when no explicit Gateway token is stored, allowing local OpenClaw CLI credentials to deliver the stalled callback and including a short executor-output trace in the notification.
-
-## 0.1.1 - 2026-06-21
-
-### Fixed
-
-- Normalized Codex ACPX slash-style model shorthands such as `gpt-5.5/medium` to ACPX bracket model ids such as `gpt-5.5[medium]`, preventing Codex tasks from stalling before callback when a configured model uses the older shorthand.
-
-## 0.1.0 - 2026-06-21
-
-First public npm release for Agent Knock Knock, including OpenClaw plugin integration, local coding-agent delegation, task management, Cursor support, diagnostics, and packaged installation.
-
-### Added
-
-- Added ACPX executor metadata so delegations can target Claude Code, Codex, or Cursor.
-- Added CLI task management commands: `list`, `status`, `send`, and `close`.
-- Added OpenClaw plugin tools for task listing, task status, follow-up sends, and local task close.
-- Added the `/akk` OpenClaw command for direct delegation, task listing, status, follow-up sends, and task close from command-capable chat surfaces.
-- Added manifest contracts and tool metadata for the new OpenClaw plugin tools.
-- Added cooperative AKK cancellation through `agent-knock-knock cancel`, `agent_knock_knock_cancel`, and `/akk cancel <conversation-id>` for Codex, Claude, and Cursor ACPX sessions.
-- Added Codex ACPX proxy and model configuration support for environments that require `ALL_PROXY` or a ChatGPT-compatible model id.
-- Added lazy idle-timeout cleanup for idle AKK sessions, configurable with `idleTimeoutMinutes` and defaulting to 10080 minutes.
-- Added runtime diagnostics logs under `~/.agent-knock-knock/logs`, with local timestamps, daily NDJSON files, secret redaction, log-level filtering, and retention cleanup.
-- Added safe executor trace support through `agent-knock-knock status --trace` and the OpenClaw status tool's `trace` parameter, summarizing tool calls, permission requests, client events, monitor events, and redacted thinking markers.
-- Added a coding-agent executor registry that centralizes Codex, Claude Code, and Cursor ACPX metadata, aliases, protocol actors, session naming, and proxy/model configuration.
-- Added an explicit recovery-decision foundation with `needs_recovery`, `agent-knock-knock recover`, `agent-knock-knock restart`, `agent_knock_knock_recover`, `agent_knock_knock_restart`, `/akk recover`, and `/akk restart`.
-- Added Cursor as a supported ACPX coding-agent executor with `AKK Cursor`, `/akk cursor`, `agent="cursor"`, Cursor session/model/proxy config keys, and conservative explicit recovery decisions.
-- Added npm package metadata, public package file allowlist, and a tag-driven GitHub Actions release workflow for publishing `@scotthuang/agent-knock-knock`.
-- Added `agent-knock-knock install-openclaw` for installing the packaged OpenClaw plugin and skill template locally.
-- Added `agent-knock-knock doctor` for checking required local commands, available coding agents, and packaged runtime files.
-- Added tests for Codex-backed delegation and task management flows.
-- Added tests for the executor registry used by Codex, Claude Code, and Cursor.
-- Added tests for explicit recovery decisions, AKK history replay recovery, and restart without history replay.
-- Added tests for Cursor delegate, send, cancel, route metadata, and unavailable-session recovery decision behavior.
-- Added tests for runtime diagnostics logging, redaction, retention cleanup, and CLI log emission.
-
-### Changed
-
-- Changed actionable callback delivery from Gateway `sessions.send` to `chat.send` with `deliver: true`, so callbacks injected into channel-scoped OpenClaw sessions can use the original outbound channel route.
-- Changed coding-agent `done` callbacks to move AKK conversations to `idle` instead of terminal `done`, allowing follow-up sends to continue the same session until manual close or idle timeout.
-- Updated `done` callback messages to include concise AKK convenience commands such as `AKK list`, `AKK send <conversation-id>: <message>`, `AKK status`, and `AKK close`.
-- Updated OpenClaw plugin routing descriptions so `AKK` and `akk` both indicate Agent Knock Knock, and unspecified AKK delegations default to Codex.
-- Updated the OpenClaw skill template so follow-up language such as "continue", "再让它", or "给刚才那个" defaults to reusing the most recent matching open AKK session through `agent_knock_knock_send`.
-- Updated the OpenClaw skill template with AKK cancel routing for stopping current in-flight work without closing the reusable AKK session.
-- Updated the OpenClaw skill template with AKK/akk routing, configurable default-agent delegation, explicit agent routing, and task management tool usage.
-- Updated follow-up sends from OpenClaw to launch the coding agent in the background so OpenClaw can continue to subsequent tool calls.
-- Migrated the implementation source to TypeScript and moved runtime entrypoints to compiled `dist/src` output.
-- Migrated the Node test suite to TypeScript and updated `npm test` to build before running compiled tests from `dist/test`.
-- Updated `npm run build` to clean stale compiled output before TypeScript compilation, and added `prepack` so npm publish/pack builds fresh runtime files.
-- Changed `delegate` to generate a unique ACPX session for each new coding-agent task unless a session is explicitly configured, preventing concurrent AKK tasks from sharing the same default Codex, Claude, or Cursor session.
-- Added a background executor monitor that marks agent-waiting tasks `stalled` when the launched process exits without a callback or when the configurable `agentTimeoutMinutes` callback timeout is reached, then attempts to notify the original OpenClaw session.
-- Renamed the OpenClaw skill template from `bidirectional-chat` to `agent-knock-knock` so the installed skill matches the project and plugin name.
-- Updated documentation to describe local coding agents, task management, cooperative cancellation, and the home-directory conversation store.
-- Updated documentation with the local update flow required to rebuild and reload the linked OpenClaw plugin after project changes.
-- Updated the OpenClaw skill template and README with conservative recovery guidance for coding agents whose native session resume is unreliable.
-- Updated the OpenClaw skill template, README, plugin manifest, and protocol docs for Cursor routing and task management.
-- Updated README, plugin metadata, and skill guidance to document the configurable `defaultAgent` flow for unspecified AKK delegations.
-- Documented ACPX approval behavior: Claude Code permission requests work with `--approve-all`, while some Codex sensitive operations can fail directly under AKK's non-interactive/background path instead of surfacing an approvable ACPX permission request.
-
-### Fixed
-
-- Fixed OpenClaw plugin delegations generating `--record-only` callback commands even when Gateway callback delivery was configured, which caused completed Codex tasks to be logged locally without returning results to OpenClaw.
-- Fixed chat-routed `/akk cursor ...`, `/akk claude ...`, and `/akk codex ...` requests being treated as unspecified delegations when OpenClaw called the delegate tool without an explicit `agent` parameter.
-
-### Verified
-
-- `npm test` passes 53 tests.
-- `npm --cache /private/tmp/akk-npm-cache pack --dry-run` verifies the scoped npm package contents.
-- Local OpenClaw installation validated with the npm-installed Agent Knock Knock plugin loaded, the updated `agent-knock-knock` skill installed, and the gateway restarted successfully.
-- Live OpenClaw validation created a Claude task, listed Claude tasks, sent a follow-up message, and closed the task through plugin tools.
-- Live OpenClaw validation created a Codex task with a configured `ALL_PROXY` value and `model=gpt-5.5/medium`, listed Codex tasks, sent a follow-up message, received Codex `done`, and closed the task through plugin tools.
-- Live ACPX validation created smoke sessions for Codex and Claude Code, called AKK cancel for each, observed `executor_cancel_requested` events with status 0, and closed the smoke ACPX sessions.
-- Published `@scotthuang/agent-knock-knock@0.1.0` to npm.
-- Created GitHub Release `v0.1.0`.
-
-### Initial MVP Baseline
-
-Initial MVP for managed bidirectional agent delegation.
-
-#### Added
-
-- Added a minimal Node.js project with no runtime dependencies.
-- Added structured bidirectional message protocol for OpenClaw manager and Claude Code developer roles.
-- Added conversation state tracking with response-round budget support.
-- Added budget thresholds:
-  - 50 response-round soft limit
-  - 100 response-round hard limit
-  - 30-round convergence warning
-  - 40-round completion/degrade/failure warning
-- Added NDJSON event logging for conversations, messages, raw exchanges, and closure events.
-- Added Claude Code bootstrap prompt generation.
-- Added CLI commands for creating conversations, recording messages, generating bootstrap prompts, and delegating tasks.
-- Added shell wrapper for delegation startup.
-- Added OpenClaw skill template for bidirectional Claude Code delegation.
-- Added protocol documentation.
-- Added two-Claude simulation script for weather lookup and multi-round architecture decision scenarios.
-- Added tests for response-round accounting, budget escalation, and done-message closure behavior.
-
-#### Verified
-
-- `npm test` passes.
-- Conversation creation writes state and NDJSON logs.
-- Message recording updates conversation state.
-- Two-Claude architecture scenario completes with correct round accounting:
-  - `task`, `question`, `answer`, `question`, `answer` count as response rounds.
-  - `done` does not increment response rounds.
-- Two-Claude weather scenario records blocked/failure flows when live weather lookup is unavailable.
+- Remove the alternative managed execution stack, blanket background approval, non-tmux agent integration, model/proxy overrides, and callback bootstrap contract.
+- Remove obsolete quickstarts, smoke tests, simulations, modules, and maintenance assets for unsupported execution paths.
 
-#### Known Issues
+## Earlier releases
 
-- `blocked` handling was corrected after the first weather test; older logs may show `blocked` as `requires_response=false`.
-- The two-Claude simulation currently stores full raw `acpx` output in logs, including client/tool status lines.
-- The script does not yet parse the model's returned JSON into canonical message body fields.
-- Real OpenClaw Gateway integration is not implemented yet; current implementation simulates OpenClaw with Manager Claude.
+The pre-0.4 history remains available in [GitHub Releases](https://github.com/scotthuang/agent-knock-knock/releases).