@jawcode-dev/coding-agent 1.1.0

package/CHANGELOG.md

@@ -0,0 +1,376 @@
+# Changelog
+
+## [Unreleased]
+
+## [1.0.0] - 2026-06-16
+
+### Changed
+
+- Improved the Bun runtime version guard diagnostic: when the Bun running `gjc` is older than the required version, the error now names the exact detected Bun runtime path and prints a platform-specific upgrade and PATH fix (Windows gets the `irm bun.sh/install.ps1|iex` reinstall plus a `%USERPROFILE%\.bun\bin` PATH hint) instead of a bare `bun upgrade` (#525).
+
+- Aligned the `codex-standard` and `codex-pro` model profiles on the `openai-codex/gpt-5.5` baseline so they no longer default to stale mixed model generations (`gpt-5.4`, `gpt-5.2-codex`, `gpt-5.1-codex-max`, `gpt-5.3-codex-spark`); the profiles now differentiate purely by per-role reasoning effort (#532).
+- Reduced the default RPC `get_state` payload by omitting static `dumpTools` and `systemPrompt` fields unless requested via `include: ["tools", "systemPrompt"]` (#539).
+- Changed the default automatic compaction threshold to 73% of the active context window so Codex sessions compact before the previous ~85% large-window reserve threshold could overflow.
+- Removed `cua-driver` from managed default MCP autoload and added a discoverable `computer_use` tool that lazily starts the cua-driver backend only on first use.
+
+### Fixed
+
+- Tightened the Windows/psmux tmux provider boundary: `gjc team` now honors `GJC_TMUX_COMMAND` (not just `GJC_TEAM_TMUX_COMMAND`) so the team leader resolves the same multiplexer as `gjc session`/`gjc --tmux`; and when a multiplexer lists a session that lacks GJC's `@gjc-profile` ownership tag, `gjc session status` now returns `gjc_tmux_session_untagged` with a `detail` hint and `gjc team` reports the same cause, instead of a bare `gjc_tmux_session_not_found` / `unmanaged_tmux_session`. Documented that alternative multiplexers such as psmux on Windows are not fully supported because they do not round-trip tmux user options (#531).
+- Hardened RPC stdio lifecycle behavior: `gjc --mode rpc` now reports malformed JSONL frames as parse-error responses without killing the session, flushes durable session state before exiting on EOF/shutdown, and has red-team coverage for attached persistence, reload, malformed-frame recovery, and concurrent child-session isolation.
+### Fixed
+
+- Hardened the harness RPC submit/router contract so `submit` is no longer advertised or accepted during finalizing/non-idle lifecycle windows, non-idle RPC state reports `submitted:false` with a retryable gate, and degraded owner endpoints fall back to `owner-not-live` without false acceptance (#544).
+
+- Ran estimated context maintenance before sending a new prompt, including tool-output pruning and threshold compaction, so large tool results appended after the last assistant turn cannot push the next model request over the context window.
+- `gjc team` now self-heals a missing `@gjc-profile` ownership tag when the current leader pane was genuinely launched by `gjc --tmux` (detected via `GJC_TMUX_LAUNCHED=1`): the session is re-tagged with `set-option` and startup proceeds, instead of hard-failing with `unmanaged_tmux_session` after a mid-startup attach failure or registry race stripped the tag. Sessions without the GJC launch marker are still rejected unchanged, so foreign tmux sessions cannot be hijacked.
+
+## [0.4.5] - 2026-06-12
+
+### Added
+
+- Added a dim `(ctrl+s to observe sessions)` discoverability hint under the `subagent` await panel header while any awaited subagent is still running, pointing to the full session observer overlay; the hint shows in both collapsed and expanded states and disappears once no subagent is running.
+- Added a `phase-rollup` receipt family (receipt-of-receipts) to the harness control plane: a hash-sealed rollup that supersedes N child task receipts at a lifecycle boundary, preserving per-child `{id, status, outputUri, outputSha256, receiptSha256, tokens, costTotal, clonedTokens, lowRoi}` pointers plus aggregate ROI totals, with a pure deterministic builder (`buildPhaseRollupReceipt`) and fail-closed semantic validation.
+- Added a pure receipt-ingestion fast path (`ingestReceipts`): fail-closed batch validation + lifecycle transition computation via the existing state machine, plus a deterministic model-facing digest hard-capped at 280 chars — groundwork for LLM-free receipt routing.
+- Added advisory spawn-ROI reconciliation (`reconcileSpawnRoi`) and deterministic fork-context mode advice (`adviseForkContextMode`) surfaced in task receipts without changing task success semantics.
+- Added the Grok Build provider contract design document.
+
+### Changed
+
+- Reduced compiled CLI startup and native bundle pressure with default-small grammar loading, tokenizer tiering, and compiled fast-help paths.
+- Preserved dev/main release metadata and changelog consistency for the 0.4.5 lockstep release.
+
+- Added native `gjc ultragoal steer --kind` support for documented steering mutations beyond `add_subgoal`, including split, reorder, wording revision, ledger annotation, and blocked-goal supersession contracts with structured audit expectations.
+### Fixed
+
+- Kept the unified `goal` tool registered and active by default whenever `goal.enabled` is true, including explicit tool subsets and `gjc ultragoal create-goals` arming flows.
+- Restored no-argument `gjc` interactive startup instead of launching help.
+- Rendered and executed Cursor-native tool calls, including detached/native handler paths and empty-pattern composer grep guards.
+- Tool-output pruning no longer rewrites already-sent provider-facing history mid prompt-cache epoch and now persists pruned message updates back into canonical session storage.
+- Preserved provider abort root causes in the final TUI abort label, kept replay rendering idempotent, and added a `PI_STREAM_IDLE_TIMEOUT_MS` remediation hint when stream idle watchdogs fire.
+- Hardened harness owner recovery/finalize paths and submit-prompt-file handling.
+### Added
+
+- Added FTS5 semantic (full-text) search for local memory stage1 rows (`src/memories/memory-fts.ts`): BM25-ranked queries with synonym-expanded terms, LIKE fallback, manual upsert FTS sync, and `memory reindex` support; artifact search remains line-based.
+- Added per-turn local memory Task Snapshot prompt injection (99.01 M6): `buildLocalTaskSnapshot` / `buildMemoryTaskSnapshot` diversify top hits (episode cap 2, one per ref) and wrap them in `<memories>` via `localBackend.buildTaskSnapshot`.
+- Added `memory list`, `memory status`, and `memory reindex` for the local backend (99.01 M8), with FTS5 stage1 search and manual-save FTS sync via `syncStage1FtsRow`.
+- Added unified local memory CLI surface (99.01 M7): `memory browse [--limit N]`, shared `loadMemoryConfig` in `src/memories/memory-config.ts`, lexicographic search tie-breaks, and ref alias normalization for read/context.
+- Added `memory search --cloud` to route a single search through Hindsight recall even when `memory.backend` is `local` or `off` (requires `hindsight.apiUrl`).
+- Added local memory search path scoping: `searchMemoriesCore` / `filterScopePaths` restrict episode and manual hits by thread rollout path (profile `memory` and shared `summary` artifacts always pass); CLI `memory search <query> [--scope path[,path...]]`.
+
+### Changed
+
+- Local memory search: artifact `MEMORY.md` / `memory_summary.md` bodies are indexed in FTS5 (`memory_artifacts_fts`), synced on consolidation and `memory reindex`; new setting `memories.searchMode` (`hybrid` | `fts` | `like`) controls stage1 and artifact query backends.
+
+- **jaw-interview ask (TUI, 082.3 v2)**: options `1…N` only in the list; **`N+1. 출력창`** is a separate bordered composer panel below (not a list row). ↑↓ includes the output panel; trimmed text-only submit; empty/whitespace Enter skips. Interview question scroll uses a bounded title viewport (fixes `MAX_SAFE_INTEGER` starving scroll rows).
+
+## [0.4.4] - 2026-06-10
+
+### Added
+
+- Made coordinator turns event-backed, adding an MCP coordinator server and the `gjc coordinator` / `gjc mcp-serve` commands backed by durable turn/session state (#479).
+- Surfaced awaited sub-agent live streaming status in the await panel (#475).
+
+### Changed
+
+- Refreshed the README brand images (#477).
+
+### Fixed
+
+- Persisted ralplan role-agent artifacts via the CLI, returning receipt-only output to the caller (#474).
+- Collected the `ask` tool "Other" custom input inline below the option list (#476).
+
+## [0.4.3] - 2026-06-10
+
+### Added
+
+- Added a `busyPromptMode` setting (`steer` | `queue`, default `steer`) so a prompt submitted while the agent is busy can either steer the active turn or be queued to run after it completes, keeping steering and queued-next-turn semantics distinct. Ctrl+Enter still always queues as a follow-up (#434).
+### Fixed
+
+- Fixed a persistent `monitor` notification flood where a cancelled or evicted monitor kept delivering queued `task-notification` follow-ups (surviving process death, log deletion, and `job cancel` returning not-found). Monitors now purge their queued notifications on cancel/terminal/eviction, retain a short tombstone so post-eviction `job cancel` still purges, coalesce rapid duplicate output to the latest state, and close a cancel/trailing-flush race.
+- Fixed `ultragoal` execution leaking across concurrent independent GJC sessions. The runtime `goal-mode-request.json` is now stamped with the producing session's `GJC_SESSION_ID`, and the consumer only activates a request that belongs to the current session (another session's request is left intact instead of being consumed/deleted). Legacy unscoped requests remain consumable for single-session compatibility ([#457](https://github.com/lidge-jun/jawcode/issues/457)).
+
+## [0.4.2] - 2026-06-09
+
+### Changed
+
+- Added conservative `timeout-minutes` values to all CI workflow jobs to prevent indefinite hangs.
+- Made coordinator MCP turn waiting state-backed by durable turn/session files, with runtime session sidecar updates for running/completed/error states and Meeseeks guidance that avoids fixed sleep/capture-pane loops.
+
+### Fixed
+
+- Failed stale coordinator turns quickly when their recorded tmux session is gone, clearing active-turn state instead of burning await timeouts.
+- Improved the grep limit-reached message to show the current limit value and suggest using `--limit` for more results.
+- Passed the active model's `maxTokens` (reserved completion budget) into the auto-compaction threshold and context-usage reserve so prompt packing reserves output for large-window models, keeping the safe input budget below the total context window (e.g. ~272K for a 400K/128K model) instead of filling the whole window ([#442](https://github.com/lidge-jun/jawcode/issues/442)).
+- Fixed a `gjc harness` recovery deadlock where a session created by `start` without `--detach` (persisted as `started` with no owner lease/endpoint) could never get a live owner: `recover` refused to spawn one because no prior endpoint existed, while `start` reported `session-already-exists`. `recover` now bootstraps a fresh owner for a never-started session (no lease, no endpoint, no owner-run evidence) without writing a misleading `vanish` receipt, reported via `bootstrappedOwner: true`. Bootstrap is independent of the vanish classifier's `ownerRequired` verdict (nothing has vanished), so a session started in a non-git workspace (git delta `unknown`) is recovered too, while a deleted worktree is still refused (#421).
+
+## [0.4.1] - 2026-06-07
+
+### Changed
+
+- Hardened the default system prompt with a `<skill-discipline>` block (never ignore skill text, keep read-only/interview skills from mutating, recommend and invoke the matching `/skill` on approval) and tightened `<communication>` to ban permission-begging/deferral phrasing and never announce remaining work instead of doing it (#392).
+- Cleaned up the bundled GJC workflow skill docs and defaulted execution handoff to ultragoal while prioritizing ralplan refinement (#395, #396).
+
+## [0.4.0] - 2026-06-06
+
+### Added
+
+- Added an agent-driven RPC workflow lifecycle control plane and a `workflow_gate` contract: `negotiate_unattended`, `workflow_gate`, and `workflow_gate_response` frames are validated by an answer-schema validator, persisted through a durable gate broker, and wired into live session dispatch (#314).
+- Added a binding-only `gjc-plugins` sub-skill plugin framework that loads, validates, and activates declarative plugin packs without granting implicit arbitrary execution surface (#347).
+- Added generated JSON Schemas for the config and models files (`schemas/config.schema.json`, `schemas/models.schema.json`) with a `check:schemas` drift gate (#377).
+- Added `cacheRetention` support in the models config (#381).
+- Added an Opus max reasoning preset (#372).
+
+### Changed
+
+- Improved slash command and `skill:*` suggestion ranking so `/team` surfaces the matching skill before weaker fallback candidates.
+
+### Removed
+
+- Removed the optional `@jawcode-dev/swarm-extension` package and its `gjc-swarm` CLI. The YAML/DAG swarm orchestration extension was a standalone optional feature not imported by any other package; it is no longer built or published.
+
+### Fixed
+
+- Routed unattended deep-interview ask-tool questions through `workflow_gate { kind: "question" }` events, including Round 0 topology and challenge-mode metadata, free-text option/schema shape, headless RPC answers, and synchronous response race handling (#316).
+- Preserved harness owner-vanish evidence after prompt acceptance: no-owner `recover` now either restores a detached owner when a prior endpoint exists or returns a public-safe concrete owner-exit reason plus a vanish receipt, and no-owner `observe`/`events` expose the preserved owner-exit summary.
+- Fixed LSP and MCP server lifecycle cleanup so clients/servers and their child processes are torn down on abort/exit instead of leaking (#389).
+- Preserved session retention across resident session rewrites.
+- Fixed harness session lookup testability without changing runtime owner-routing behavior.
+- Prevented release catalog file specs from recursing during catalog resolution (#351).
+- Reconciled the ultragoal skill mode-state and HUD with the plan/ledger so status reflects real goal progress (#342/#346).
+
+## [0.3.2] - 2026-06-05
+
+### Added
+
+- Added model profiles with a `--mpreset <profile>` CLI flag and a `/model` selector "Profiles" section that activate a named profile's default model plus per-agent-role model overrides in one step, validating required-provider credentials before applying and surfacing a custom provider onboarding wizard for missing API-compatible providers.
+- Integrated `ai-slop-cleaner` as an internal Ultragoal sub-skill fragment that runs as the mandatory completion-gate cleanup sweep over a story's changed files, reporting blocking and advisory findings without editing code or mutating `.jwc/` state.
+### Changed
+
+- Edit tool diff generation (`generateDiffString`) now uses the native `diffLines` from `@jawcode-dev/natives` (a byte-identical Rust port of jsdiff) instead of the pure-JS implementation, removing the multi-second Myers blowup on large-file edits (~16x faster on ~1MB files) with identical diff output.
+
+### Fixed
+
+- Reconciled native Ultragoal commands with workflow mode-state and the HUD: `gjc ultragoal create-goals`, `complete-goals`, `checkpoint`, steering, review-blocker recording, and status now sync `.jwc/state/ultragoal-state.json` plus `skill-active-state.json` from the durable `.jwc/ultragoal` plan/ledger, clearing stale active HUD chips after all goals complete.
+- Forwarded the parent session id when task subagents validate configured role-agent model overrides, preventing session-scoped OAuth providers from being misread as unauthenticated and falling back to the parent chat model.
+- Removed unintended public memory-tool guidance and registration: Hindsight retain/recall/reflect helpers are now compatibility-only, local memory prompt injection no longer advertises `memory://` reads, and regression tests guard the public tool surface.
+- Fixed `read` hashline anchors drifting on truncated reads so the `line+hash` anchors consumed by `edit`/`apply_patch` stay correct when a file is read past the truncation boundary.
+- Reconciled the ultragoal skill mode-state (`current_phase`/`active`) and HUD chip with the `.jwc/ultragoal` plan/ledger on every `gjc ultragoal` command (`create-goals`/`complete-goals`/`checkpoint`/`steer`/`record-review-blockers`/`status`), so `gjc state ultragoal read`, the skill-tool chain guard, and the HUD no longer sit at a stale `active:true`/`goal-planning` after a run completes (#342). A new `reconcileWorkflowSkillState` performs a session-scoped (`GJC_SESSION_ID`) derived write that bypasses only transition-edge validation while preserving schema/unknown-phase validation, version/checksum stamping, and audit provenance (`owner: gjc-runtime`, `verb: reconcile`); reconciliation is best-effort and surfaces failures via stderr and a `reconcile_failed` ledger event without changing command status/stdout. Removed the duplicate sessionless active-state sync from the `gjc ultragoal` command wrapper.
+
+## [0.3.1] - 2026-06-05
+### Added
+
+- Added opt-in crash diagnostics for subprocess failures, with a shared crash taxonomy/report writer, bash/Python/LSP/DAP crash notices, and a native Rust panic-report hook gated by `GJC_NATIVE_CRASH_DIAGNOSTICS` / `GJC_CRASH_DIAGNOSTICS`.
+- Started the GJC backend bridge foundation with a shared agent-wire protocol module, event envelopes, RPC command scope matrix, UI request broker, typed unsupported UI results, a guarded `--mode bridge` handshake surface, and RPC mode dispatch refactored onto the shared command dispatcher.
+- Documented the experimental `--mode bridge` protocol in `docs/bridge.md` and the `GJC_BRIDGE_*` environment variables in `docs/environment-variables.md` (TLS-mandatory startup, bearer auth, coarse command scopes with a `prompt` floor, single live `AgentSession` per process, bounded event-stream replay with `reset`, and the semantic-not-pixel UI capability matrix), and added bridge event-stream/idempotency regression tests plus a docs-conformance check that pins the docs against the protocol version, scope/command catalog, negotiated capabilities/frame types, and unsupported UI surfaces. The bridge protocol/SDK are experimental (`BRIDGE_PROTOCOL_VERSION` 1) and may change in additive, version-negotiated ways.
+
+### Fixed
+
+- Made opt-in crash diagnostics create/chmod report directories to `0700` and report files to `0600` so captured command, cwd, and stderr previews are private even under a permissive umask.
+- Scoped `agent://` and `artifact://` resolution to the caller's artifacts directory plus explicitly authorized parent/child tree directories, removed registry-wide live-session lookup/enumeration, and made missing agent-output metadata sidecars fail closed.
+- Fail-closed experimental bridge session endpoints by default for 0.3.1: events, commands, controller ownership, UI responses, host tool results, and host URI results are disabled unless an internal endpoint matrix explicitly enables them; only health/help and the authenticated handshake remain available, with the handshake advertising no enabled session surface by default.
+- Render terminal-pasted clipboard image temp paths as compact `[image N]` prompt placeholders while attaching the image payload, instead of inserting raw `/var/folders/.../clipboard-*.png` path text.
+- Preserved `gjc harness` `owner-vanished:*` blockers when a replacement owner becomes live so unrecovered prior-owner evidence remains visible until explicit recovery or terminal completion evidence; only safe startup liveness false-negatives such as `detached-owner-not-live` are auto-cleared.
+- Fixed the interactive agent unexpectedly stopping after automatic context maintenance instead of resuming the in-flight task. Post-compaction continuation now schedules exactly one source per completion (overflow retry → queued messages → synthetic auto-continue prompt), the threshold/handoff auto-continue prompt skips a redundant pre-send compaction check, overflow retry strips only the context-overflow failed turn (never normal/aborted/silent-abort tails), and non-resumable or superseded continuations log a structured reason instead of stranding the session.
+- Fixed the native Stop skill-state hook letting active GJC workflow skills stop prematurely. The Stop hook no longer treats a missing/unreadable mode-state file as terminal for handoff workflows, and handoff skills (`deep-interview`, `ralplan`) now keep blocking Stop even in the `handoff` phase until they are demoted (`active:false`) or cleared, so they always end by offering the next handoff step via the ask tool. Non-handoff skills (`team`, `ultragoal`) retain the fail-open safety valve when their mode-state file is corrupt or invalid.
+
+## [0.3.0] - 2026-06-03
+
+### Added
+
+- Added runtime-enforced Ultragoal executor QA/red-team evidence matrices for completion checkpoints, with plan-first contract coverage, user-surface evidence, adversarial cases, artifact references, scoped Executor red-team guidance, and focused rejection tests for shallow or contradictory QA evidence.
+### Fixed
+
+- Made `gjc harness observe` preserve completed RPC owner evidence after the owner exits, including a `completedOwnerExited` diagnostic and durable terminal-result cursor.
+- Clarified that `gjc team` requires an existing tmux-backed leader session from `gjc --tmux`, with actionable help, docs, and failure text.
+- Kept deep-interview ask options visible for long prompts by adding an opt-in scrollable selector title panel with selector-local `PageUp`/`PageDown` prompt scrolling, while leaving normal ask dialogs and global keybinding configuration unchanged.
+
+## [0.2.5] - 2026-06-02
+
+### Added
+
+- Added Claude Code parity `monitor` plus `CronCreate`, `CronList`, and `CronDelete` tools with frozen upstream schema fixtures, inline scheduling, background output monitoring, and ACP permission coverage.
+
+### Changed
+
+- Pruned bundled built-in themes to `red-claw` and `blue-crab`, with `blue-crab` now the default light-appearance theme.
+- Clarified ralplan role-agent handoff guidance so Planner/Architect/Critic return compact artifact receipts after `gjc ralplan --write --json` instead of duplicating full persisted verdict markdown into the parent context.
+- Made `web_search` permissionless by default with a keyless DuckDuckGo fallback, active-model-gated native provider selection, and explicit-only legacy provider selection so custom providers no longer auto-hit stray OpenAI/Codex OAuth credentials.
+
+### Fixed
+
+- Fixed the skill HUD rail showing already-handed-off planning workflows so it renders only the currently-active stage. Handoffs now supersede every same-session-scope row of the caller and callee skills (not just the exact `skill::session_id` key), the visible-state read collapses duplicate same-skill rows to the most-recent one (so a handoff demotion drops a stale `active:true` row and on-disk state self-heals), and the HUD renderer collapses the `deep-interview → ralplan → ultragoal` pipeline to its most-recent stage. Activating a later stage (e.g. `gjc ultragoal` after ralplan) now supersedes the earlier one even when the activation path does not run the `handoff` verb, while `team` still coexists with ultragoal.
+
+## [0.2.4] - 2026-06-02
+
+### Added
+
+- Added the bundled `blue-crab` dark theme and a TUI-only `/theme` selector that persists the selected appearance slot while keeping `red-claw` as the default dark theme.
+- Added `retry.requestMaxRetries` and `retry.streamMaxRetries` settings plus docs for codex-cli-style provider retry budgets ([#157](https://github.com/lidge-jun/jawcode/pull/157)).
+
+### Fixed
+
+- Forwarded a caller-provided `User-Agent` on non-Anthropic proxy base URLs so the anthropic-messages provider no longer strips it, fixing HTTP 403 "request was blocked" rejections from WAF-fronted proxy gateways.
+- Restored Settings theme live preview without persisting browse state so confirm/cancel governs the rendered theme and in-flight previews cannot overwrite an explicit choice ([#166](https://github.com/lidge-jun/jawcode/pull/166)).
+- Preserved the deep-interview session language so interviews continue in the language they started in ([#168](https://github.com/lidge-jun/jawcode/pull/168)).
+- Purged the inherited OMP/pi-mono history from the embedded coding-agent CHANGELOG so post-update notifications and `/changelog` only surface jawcode releases ([#161](https://github.com/lidge-jun/jawcode/pull/161)).
+- Wired `goal` tool `drop`/`complete` semantics so completing or dropping a goal restores the pre-goal toolset in the same session, with refreshed ultragoal docs ([#152](https://github.com/lidge-jun/jawcode/pull/152)).
+- Added local-memory empty-payload guidance so the agent does not claim facts were saved or remembered when local memory has no confirmed payload/readback ([#165](https://github.com/lidge-jun/jawcode/pull/165)).
+- Fixed `gjc update` binary fallback downloads to use the current owner release repository and report actionable manual update commands for unsupported fallback targets ([#164](https://github.com/lidge-jun/jawcode/pull/164)).
+
+## [0.2.3] - 2026-06-01
+
+### Added
+
+- Added `/provider add --preset minimax|minimax-cn|glm` and matching `gjc setup provider --preset ...` shortcuts for OpenAI-compatible MiniMax and GLM/zAI custom-provider onboarding.
+- Added a built-in `skill` tool so the agent can chain into another loaded skill on its next turn. Mirrors `/skill:<name>` typing and subagent `autoloadSkills` by dispatching the chained skill's SKILL.md as a user-attribution custom message; controlled by the new `skill.enabled` setting (default true).
+- Added explicit fork-context task subagents with sanitized bounded parent-history seeds, global `task.forkContext.enabled`, per-agent `forkContext: allowed`, per-task `inheritContext: true`, audit-visible seed metadata, and fresh provider transport state by default.
+- Defaulted the bundled `executor` and `architect` agents to `forkContext: allowed` so callers can opt them into a sanitized parent-context seed via `task({ inheritContext: true })`; per-task selection remains explicit (default false).
+- Added a `/context` slash command that surfaces the active session's token usage breakdown via `buildSessionContext` without forking prompt assembly, so users can see active context before sessions overflow ([#150](https://github.com/lidge-jun/jawcode/pull/150)).
+- Added multi-line focused-option rendering to the `ask` tool's `HookSelectorComponent` via an opt-in `wrapFocused` flag so long option labels stay readable in plan-mode and other shared consumers ([#148](https://github.com/lidge-jun/jawcode/pull/148)).
+- Added a sanctioned native deep-interview spec persistence bridge so deep-interview can save final specs under `.jwc/specs/` before deliberate ralplan chaining without caller-side `.jwc` writes ([#134](https://github.com/lidge-jun/jawcode/pull/134)).
+- Added the skill chaining lifecycle's `handoff` state verb with same-turn dispatch and atomic HUD-truth synchronization across callee mode-state, caller mode-state, session active-state, and root active-state ([#132](https://github.com/lidge-jun/jawcode/pull/132)).
+
+### Changed
+
+- Clarified the README to position team as optional and ultragoal as implementation-through-evidence, so docs readers see workflow handoffs without spelunking into skill sources ([#145](https://github.com/lidge-jun/jawcode/pull/145)).
+- Migrated CI to the gajae self-hosted Linux runner with fork-PR guards, idempotent `fd` symlink setup, and Node 24 pinning for self-hosted jobs ([#124](https://github.com/lidge-jun/jawcode/pull/124)).
+
+### Fixed
+
+- Enforced the deep-interview phase boundary so active interviews block mutation tools until a handoff/spec is produced.
+- Allowed read-only `architect`, `planner`, and `critic` role agents to persist ralplan/state workflow receipts through a restricted `bash` allowlist while blocking general shell and product-file mutations.
+- Made settings theme browsing confirm-only so arrowing through themes no longer changes the rendered theme before the displayed/persisted theme name changes.
+- Made startup CHANGELOG display deterministic by embedding `packages/coding-agent/CHANGELOG.md` into the binary so post-update launches show the shipped history regardless of cwd or `GJC_PACKAGE_DIR`/`PI_PACKAGE_DIR` overrides.
+- Registered `gjc update` as a public root subcommand so it invokes the bundled updater instead of routing into the interactive launcher.
+- Fixed local memory backend persistence so manual enqueue/rebuild starts maintenance immediately and prompt injection reads the active session's memory root.
+- Aligned release-bump CI tests with the highest bundled upstream Claude family and stopped cancelling TTSR retries when an aborted partial was not persisted, so dev PRs no longer regress on v0.2.2-style release assumptions ([#142](https://github.com/lidge-jun/jawcode/pull/142)).
+- Added a runtime guard in the built-in `skill` tool that rejects re-entry into the currently active skill via a typed active-skill bridge from agent sessions into tool sessions, preventing recursive skill handoffs that prompt-only guidance could not stop ([#129](https://github.com/lidge-jun/jawcode/pull/129)).
+
+## [0.2.2] - 2026-05-31
+
+### Added
+
+- Added `gjc session` for listing, inspecting, removing, and attaching GJC-managed tmux sessions ([#105](https://github.com/lidge-jun/jawcode/pull/105)).
+- Added a managed tmux session surface so `gjc --tmux` and `gjc team` reuse a scoped GJC session instead of mutating the global tmux environment.
+- Added a detached contribution prep workflow surface ([#122](https://github.com/lidge-jun/jawcode/pull/122)).
+- Added stricter `models.yml` validation and docs for OpenAI-compatible proxy providers, bridge-command requirements, and team dry-run state behavior ([#121](https://github.com/lidge-jun/jawcode/pull/121)).
+- Added first-class Azure OpenAI and Amazon Bedrock providers ([#119](https://github.com/lidge-jun/jawcode/pull/119)).
+- Added workflow state receipt routing so skill state writes emit a structured `WorkflowStateReceipt` envelope and sync skill-active HUD chips under `.jwc/state/skill-active-state.json` ([#118](https://github.com/lidge-jun/jawcode/pull/118)).
+- Made `gjc state`, `gjc ralplan`, and `gjc deep-interview` work natively as documented in their SKILL.md files. `gjc state read|write|clear|contract` operates directly on `.jwc/state/` receipts (accepts `--input '<json>'` with `@file` shorthand, `--mode <skill>`, positional `<skill>`, `--session-id`, `--thread-id`, `--turn-id`, `--json`, `--replace`); writes emit a structured `WorkflowStateReceipt` envelope and sync skill-active HUD chips. `gjc ralplan` accepts the documented `--interactive`/`--deliberate`/`--architect`/`--critic` flags and the `--write --stage --stage_n --artifact` artifact persistence shape under `.jwc/plans/ralplan/<run-id>/`. `gjc deep-interview` accepts the documented `--quick`/`--standard`/`--deep` resolution flags ([#125](https://github.com/lidge-jun/jawcode/pull/125)).
+- Added a GJC dogfood skill template ([#114](https://github.com/lidge-jun/jawcode/pull/114)).
+
+### Changed
+
+- Changed `gjc --tmux` startup to create a fresh GJC-managed tmux session instead of attaching to an existing default session.
+- Hardened team message delivery and worker enforcement ([#112](https://github.com/lidge-jun/jawcode/pull/112)).
+- Hardened team worker startup and evidence invariants ([#123](https://github.com/lidge-jun/jawcode/pull/123)).
+- Clarified team state while integration is pending ([#109](https://github.com/lidge-jun/jawcode/pull/109)).
+- Removed goal mode budget semantics ([#106](https://github.com/lidge-jun/jawcode/pull/106)).
+- Removed the unused `gjc question` CLI; the `ask` tool covers blocking-question prompts.
+- Dropped the obsolete `loop` and `orchestrate` skills and pinned GJC skills in autocomplete ([#126](https://github.com/lidge-jun/jawcode/pull/126)).
+- Enforced CLI-managed workflow artifacts so planning skills persist via `gjc state` and `gjc ralplan --write` instead of editing `.jwc/` files directly.
+
+### Fixed
+
+- Fixed Anthropic extended-thinking replay recovery after aborted turns ([#115](https://github.com/lidge-jun/jawcode/pull/115)).
+- Fixed `gjc deep-interview` ambiguity threshold default so workflow gating matches the documented contract ([#116](https://github.com/lidge-jun/jawcode/pull/116)).
+- Fixed duplicate PR creation in the team integration path ([#117](https://github.com/lidge-jun/jawcode/pull/117)).
+- Fixed package asset resolution to prefer `GJC_PACKAGE_DIR` over the legacy `PI_PACKAGE_DIR`, preventing update-launch changelog displays from reading stale OMP package assets ([#111](https://github.com/lidge-jun/jawcode/pull/111)).
+## [0.2.1] - 2026-05-30
+
+### Fixed
+
+- Added a `gjc skills` inspection command so installed binaries can list and read embedded workflow skills from any project without relying on source-tree `.jwc` files.
+- Fixed first-run API provider onboarding so `models.yml` parent directories are created before writing, and malformed `/provicer` startup invocations now report the intended `/provider add` spelling instead of falling through to model bootstrap.
+
+## [0.2.0] - 2026-05-28
+
+### Added
+
+- Added scoped GJC tmux profile handling for `gjc --tmux` and `gjc team` sessions without mutating global tmux configuration.
+- Added GJC team integration hardening for worker turn-end integration requests, auto-rebase/auto-merge conflict surfacing, protected checkpoint classification, and leader/worker-visible integration summaries.
+- Added Node 20 release baseline validation to the release/check surface.
+
+### Changed
+
+- Clarified the public workflow contract so `deep-interview` and `ralplan` are invoked through `/skill:<name>`, while `gjc ultragoal` and `gjc team` remain native runtime commands.
+- Updated the README hero image and Discord community invite.
+
+### Fixed
+
+- Restored Ultragoal completion receipt export/generation validation and completion gates.
+- Fixed workflow bridge guidance and tests so private compatibility bridge commands are not advertised as public skill-loading paths.
+
+## [0.1.3] - 2026-05-28
+
+### Changed
+
+- Released the current dev branch fixes with refreshed 0.1.3 package metadata.
+
+## [0.1.2] - 2026-05-28
+
+### Changed
+
+- Updated package metadata for the Gajae Code npm publication.
+
+### Fixed
+
+- Fixed slash-command autocomplete so skill command matches no longer hide built-in fuzzy candidates like `/model` while typing `/mode`.
+
+## [0.1.1] - 2026-05-28
+
+### Changed
+
+- Restored `gjc team` multi-worker GJC-team parity orchestration with current-window worker panes, GJC-scoped state/API semantics, and `N:agent-type` launches.
+- Ported GJC team worker-worktree integration parity so `status`/`resume` auto-checkpoint dirty workers, merge or cherry-pick worker commits, cross-rebase idle workers, and record conflicts under `.jwc` integration artifacts.
+
+### Added
+
+- Added a detached `subagent` control tool for task subagents, with list, inspect, await-with-timeout, and cancel actions.
+- Added shared provider onboarding for OpenAI-compatible and Anthropic-compatible API providers through `gjc setup provider` and `/provider add`, with model-list configuration and redacted setup feedback.
+- Added shared `/model` onboarding guidance and an interactive `/provider` onboarding chooser so first launch, slash commands, and TUI no-model states point at the same provider setup flows.
+- Added a native in-TUI skill HUD rail backed by `.jwc/state/skill-active-state.json`, so active GJC workflow skills are visible without a separate tmux pane.
+- Added bundled `executor`, `architect`, `planner`, and `critic` role agents for task delegation, including source-defined prompt files and role-agent discovery coverage.
+- Added a native `gjc team` runtime that writes GJC-scoped state, mailboxes, task lifecycle files, and telemetry without delegating to an external team binary
+- Added `openai-code` and `gemini` to the web search provider settings so users can configure OpenAI and Gemini web search directly from provider selection
+- Added OpenAI (`openai-code`) and Gemini web search options with updated setup descriptions for `gjc /login openai-code` and Gemini OAuth login
+
+### Changed
+
+- Changed normal `task` subagent launches to return immediately as detached background work while keeping generic `job` controls available.
+- Changed default interactive `gjc` startup to enter a `gajae_code` tmux session before launching the Gajae Code TUI, with non-interactive modes continuing to run directly.
+- Changed `/skill:<name>` handling so canonical skill invocations can be chained in one prompt across interactive and ACP sessions, with autocomplete-only `/name` and `/skill-name` normalization back to the public canonical form.
+- Changed interactive `gjc` startup to launch tmux only when `--tmux` is provided, with direct startup as the default.
+- Changed GJC default definitions so workflow skills remain source-bundled while repo-visible `.jwc` default artifacts are no longer the source of truth; updated system and Ultragoal guidance to use role-agent delegation and ralplan-first planning when needed.
+- Changed bare `gjc setup` to install the normal default workflow skills, while keeping hooks, provider, Python, and speech-to-text setup as explicit optional components.
+- Changed `gjc team` startup to use tmux worker panes backed by dedicated detached git worktrees by default, while keeping `--worktree` as a backward-compatible launch override.
+- Constrained the visible GJC utility surface to the retained workflow/runtime endpoints and four bundled task agents, with MCP, arbitrary skill, plugin, extension, marketplace, and custom discovery surfaces quarantined from default public use.
+- Redesigned the interactive TUI chrome with a minimal opencode-style prompt composer, simple user/gajae transcript labels, a forge-style welcome surface, and compact cwd/pulse indicators tuned for terminal coding-agent ergonomics.
+- Changed web search provider credential lookup to use the shared `AuthStorage` pipeline (`getApiKey`/`getOAuthAccess`) for API-key and OAuth auth instead of direct `AgentStorage` access
+- Changed the `openai-code` web search provider display label from `OpenAI code` to `OpenAI`
+- Updated `anthropic` and `openai`/`gemini` web search option descriptions to reflect their native `web_search`/OAuth requirements
+- Changed `/model` selection to a canonical single default-model action, removing the redundant role assignment menu for smol/slow/vision/plan/designer/commit/task models.
+- Changed public API-compatible provider setup to require `--api-key-env` and reject raw `--api-key` values.
+
+### Removed
+
+- Removed approved non-critical slash-command handlers for plan, share, browser, copy, todo, changelog, context, branch, fork, handoff, force, and quit while keeping /loop, provider setup/login/logout/model selection, and SSH intact.
+- Removed redundant model-selector role assignment options for smol, slow, vision, plan, designer, commit, task, and custom roles so selection uses one canonical default model.
+- Removed obvious non-critical plugin, marketplace, extension, and reload-plugin slash-command handlers from the built-in registry while preserving ambiguous slash-command utilities for a later approval pass.
+- Removed the auto-QA grievance reporting feature, including the `report_tool_issue` tool, `gjc grievances` command, auto-QA settings/env flags, sharing consent prompt, bundled push endpoint, and persistent install ID correlation path.
+- Removed standalone utility feature documentation for plugins, extensions, hooks, marketplace, arbitrary skills, custom tools, task-agent discovery, and TUI/config utility internals from the generated docs index.
+
+### Fixed
+
+- Fixed `gjc ultragoal create-goals` native goal activation so live sessions receive a pending reconciliation request even when the session file already contains an active goal.
+- Made `gjc ultragoal` run natively, preserving active goal state across interrupted turns.
+- Fixed interactive Escape/interrupt recovery so abort cleanup is bounded and forces the session back to idle when a provider stream, tool, or post-turn task ignores cooperative cancellation.
+- Fixed root `gjc --worktree` / `gjc -w` startup so the launch command actually creates and enters the sibling `<repo>.jawcode-worktrees/<branch-slug>` git worktree before starting the session, using collision-resistant branch slugs and avoiding worktree side effects for help/version launches.
+- Fixed root `gjc --worktree <branch>` / `gjc -w <branch>` parsing so named branch worktrees create their own `<branch-slug>` directory instead of reusing the dirty detached worktree for the current branch.
+- Wired GJC native UserPromptSubmit/Stop skill-state hooks, including `gjc setup hooks`, so public workflow keywords activate `.jwc/state`, active skill state can block premature Stop events, and active Ultragoal sessions remind steering prompts to use `gjc ultragoal steer`.
+- Fixed `gjc ultragoal create-goals` to seed GJC goal mode runtime state automatically, avoiding a separate manual `/goal` setup step.
+- Fixed legacy Pi plugin import remapping and stale GJC config-path tests so rebranded `.jwc` discovery contracts pass while preserving legacy compatibility.
+- Fixed web search OAuth-backed providers (including OpenAI code and Gemini) to use broker-managed token retrieval and account metadata, avoiding direct token-store refresh behavior that could cause search authentication failures
+- Updated Tavily missing-credential feedback to prompt users to configure an API-key provider setting instead of referencing `agent.db` directly
+- Refreshed expired OpenAI code provider OAuth tokens during `web_search` execution and persisted the updated credentials so searches continue working after token expiry
+- Wired `/login`, `/logout`, `/model`, and `/provider` TUI slash commands through interactive provider/model selectors and existing OAuth flows.

package/README.md

@@ -0,0 +1,37 @@
+# @jawcode-dev/coding-agent
+
+Core implementation package for the `gjc` coding agent in the `jawcode` monorepo.
+
+For installation, setup, provider configuration, model roles, slash commands, and full CLI reference, see:
+- [Monorepo README (local)](../../README.md)
+- [Monorepo README (GitHub)](https://github.com/lidge-jun/jawcode#readme)
+
+Package-specific references:
+- [CHANGELOG](./CHANGELOG.md)
+- [DEVELOPMENT](./DEVELOPMENT.md)
+- [RenderMermaid guide](../../docs/render-mermaid.md)
+
+## Memory backends
+
+The agent supports three mutually-exclusive memory backends, selected via the `memory.backend` setting (Settings → Memory tab, or `~/.jwc/config.yml`):
+
+- `off` (default) — no memory subsystem runs.
+- `local` — existing rollout-summarisation pipeline; writes `memory_summary.md` and consolidated artifacts under the agent dir.
+- `hindsight` — talks to a [Hindsight](https://hindsight.vectorize.io) server (Cloud or self-hosted Docker). Hindsight uses private backend lifecycle hooks to retain transcripts and recall context; compatibility-only internals remain for legacy backend calls, but GJC does not expose public coding-harness memory tools such as `retain`, `recall`, or `reflect`.
+
+### Hindsight quickstart
+
+1. Run a Hindsight server (Cloud or `docker run -p 8888:8888 ghcr.io/vectorize-io/hindsight:latest`).
+2. Set `memory.backend = "hindsight"` and `hindsight.apiUrl = "http://localhost:8888"` (or your Cloud URL).
+3. Optional environment overrides (env wins over settings):
+   - `HINDSIGHT_API_URL`, `HINDSIGHT_API_TOKEN` — connection
+   - `HINDSIGHT_BANK_ID`, `HINDSIGHT_DYNAMIC_BANK_ID`, `HINDSIGHT_AGENT_NAME` — bank addressing
+   - `HINDSIGHT_AUTO_RECALL`, `HINDSIGHT_AUTO_RETAIN`, `HINDSIGHT_RETAIN_MODE` — lifecycle
+   - `HINDSIGHT_RECALL_BUDGET`, `HINDSIGHT_RECALL_MAX_TOKENS` — recall sizing
+   - `HINDSIGHT_BANK_MISSION`, `HINDSIGHT_DEBUG`
+
+Switching backends mid-session is honoured on the next system-prompt rebuild and the next `/memory` slash command. Existing users with `memories.enabled = true|false` are migrated to `memory.backend = "local"|"off"` exactly once on first launch.
+
+## Abyss Bite TUI theme
+
+The interactive TUI defaults to the bundled `abyss-bite` theme for dark terminals and the bundled `abyss-bite-light` theme for light-appearance terminals. Explicit user theme settings still win; set `theme.dark: abyss-bite` and `theme.light: abyss-bite-light` in `~/.jwc/agent/config.yml` to pin them.

package/dist/types/async/index.d.ts

@@ -0,0 +1,2 @@
+export * from "./job-manager";
+export * from "./support";