open-claude-p 1.1.2 → 1.2.1

package/CHANGELOG.md

@@ -7,6 +7,339 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ---
 
+## [1.2.1] — 2026-09-13
+
+Re-sync with Claude Code 2.1.270 and the current model lineup, and fix a
+permission regression introduced in 1.2.0.
+
+### Fixed
+
+- **`--dangerously-skip-permissions=false` and `OCP_NO_SKIP_PERMS=1` were
+  being ignored.** Two faults stacked up. `forwardedRequestFields` omits
+  `false` values (there is no argv to emit for them), and `runOneShot`
+  resolves a missing field with `?? true` — so an opt-out reached the
+  driver as "unset" and came back out as "enabled". Underneath that sat
+  an older, latent bug: the CLI's opt-out test read
+  `options['dangerously-skip-permissions'] === false`, which is *also*
+  true for the pre-seeded default, so it fired on every invocation and
+  quietly turned the whole default-assignment block into dead code — the
+  default-on behaviour had been coming from the driver's `?? true` all
+  along. `parseArgv` now reports which options argv actually supplied
+  (`result.supplied`), the opt-out keys off that, and the CLI always
+  sends a concrete boolean. Net effect on defaults: none — a plain
+  `ocp "…"` still runs with permissions bypassed — but asking for the
+  bypass back now works.
+
+### Added
+
+- **Claude Code 2.1.270 flags.**
+
+  | Flag | Purpose |
+  |------|---------|
+  | `--permission-prompts <host\|none>` | Who answers permission prompts; `none` denies rather than waiting |
+  | `--restricted` | Drop command/code-running tools and WebFetch, ignore user/project/local settings, confine file tools, refuse `bypassPermissions` |
+  | `--system-prompt-snapshot <on\|off>` | Record the system prompt once per conversation and reuse it verbatim, or re-render each request |
+
+- **`--restricted` is wired to actually work headlessly.** Upstream
+  rejects it outright when `bypassPermissions` is in play ("Error:
+  bypassPermissions not supported in restricted mode") and exits, and
+  ocp's CLI turns `--dangerously-skip-permissions` on by default — so a
+  bare `ocp --restricted "…"` would have failed every time, silently,
+  with no output. Asking for restricted mode is an explicit choice to
+  keep the permission gates, so it now stands the permissive default
+  down and defaults `--permission-prompts` to `none`, which is the only
+  setting that keeps the run from waiting on a prompt no PTY can answer.
+  Passing `--restricted` together with an explicit
+  `--dangerously-skip-permissions` is reported as a validation error
+  instead of being handed to upstream to reject.
+
+  One caveat worth knowing: a restricted run writes no session JSONL, so
+  ocp cannot use its usual clean-transcript path and reconstructs the
+  reply from the TUI frames instead. The answer is correct, but it can
+  carry leftover box-border chrome — observed on roughly two runs in
+  three, against none for the same prompts without `--restricted`.
+
+- **`parseArgv` returns `supplied`**, the set of canonical option names
+  present in argv. Defaults are pre-seeded into `options`, so this is the
+  only way to tell `--flag=false` apart from an unset flag.
+
+### Changed
+
+- **Model references refreshed to the current lineup.** Claude Fable 5.1
+  (`claude-fable-5-1`) supersedes Claude Fable 5 in the documented
+  examples; `claude-opus-5` remains the general-purpose recommendation.
+  Aliases (`fable`, `opus`, `sonnet`, `haiku`) resolve to the latest
+  model in each family and are unaffected. Verified against the upstream
+  model list and the 2.1.270 binary, which ships both `claude-fable-5-1`
+  and `claude-mythos-5-1`.
+
+### Verified
+
+- The 2.1.270 option surface was re-diffed against `OPTION_SPEC`: no
+  previously-forwarded flag was removed, and the remaining unmapped
+  flags (`--bg`, `--cloud`, `--environment`, `--from-pr`,
+  `--remote-control`, `--teleport`, `--tmux`, `--worktree`, `--chrome`)
+  are interactive or cloud-session features outside a headless shim's
+  scope.
+
+---
+
+## [1.2.0] — 2026-08-21
+
+Realign the option surface with Claude Code 2.1.x. The headline is a
+forwarding bug: roughly thirty documented flags parsed cleanly, passed
+validation, appeared in `ocp --help`, and were then silently discarded
+before the upstream `claude` process ever saw them.
+
+### Fixed
+
+- **Flags declared in `OPTION_SPEC` are no longer dropped on the way to
+  the upstream process.** `buildSpawnArgs()` has always been
+  spec-driven — it walks `OPTION_SPEC` and emits `--flag value` for
+  every entry whose `forward.type` is `'argv'`, reading
+  `req[camelCase(spec.name)]`. The other half of that contract was
+  hand-maintained: each call site in `bin/cli.js` listed the request
+  fields it cared about. The two halves drifted, and everything the
+  hand-written list had never picked up went nowhere:
+
+  `--effort`, `--thinking`, `--max-thinking-tokens`, `--fallback-model`,
+  `--tools`, `--add-dir`, `--mcp-config`, `--strict-mcp-config`,
+  `--permission-prompt-tool`, `--system-prompt-file`,
+  `--append-system-prompt-file`, `--settings`, `--setting-sources`,
+  `--agents`, `--agent`, `--plugin-dir`, `--disable-slash-commands`,
+  `--file`, `--ide`, `--betas`, `--bare`, `--init`, `--init-only`,
+  `--maintenance`, `--debug-file`, `--workload`, `--enable-auth-status`,
+  `--allow-dangerously-skip-permissions`, `--include-hook-events`, and
+  `--include-partial-messages`.
+
+  There was no error and no warning — `ocp --effort high "…"` ran at
+  the upstream default effort and reported success. The mapping now
+  comes from the same spec that produces the argv
+  (`src/options/forward.js`), across all three request paths (daemon,
+  print-mode, and the PTY path), so appending an entry to `OPTION_SPEC`
+  is once again the only edit a newly-supported flag needs.
+
+- **A concurrent session in the same directory could answer for this
+  one.** Two independent defects lined up to produce it, and both are
+  fixed:
+
+  `findRecentSessionId` is the filesystem fallback that recovers a
+  session id when none was scraped from the PTY. It accepted two classes
+  of candidate — files created during the request, and pre-existing files
+  whose mtime moved during it (the `--resume` append) — then ranked them
+  together by mtime and took the newest. A background daemon holding a
+  warm PTY in the same cwd appends to its own JSONL continuously, so it
+  reliably out-raced the file the current spawn had just created. The
+  wrong id then flowed into `readSessionText`, whose strict-by-id path
+  faithfully returned that other session's transcript as this request's
+  answer. Newly-created files now win outright; the append class is
+  consulted only when nothing new appeared, which is exactly the
+  `--resume` case it exists for.
+
+  `readSessionText` has a second path, used when no session id is known
+  at all, that scans the project directory and takes the most recently
+  modified transcript. Its own source comments describe why that is
+  unsafe — it is why the by-id path refuses to fall back to a scan — but
+  the CLI reached it on every run where the id was missing. It now
+  accepts an `expectPrompt` option and requires a candidate transcript to
+  actually contain the prompt that was sent before returning its text;
+  the CLI always supplies it. Ownership beats recency, so our own
+  transcript is still found even when a neighbour is newer, and when
+  nothing matches the caller falls back to the PTY-extracted text as
+  before. Callers that omit the option keep the previous behaviour.
+
+- **`--include-partial-messages` is no longer forwarded on the PTY
+  path.** Upstream refuses it without `--print` ("Error:
+  --include-partial-messages requires --print and
+  --output-format=stream-json") and exits immediately, so forwarding it
+  to the interactive spawn failed the whole run. OPTION_SPEC gains a
+  `printModeOnly` field for this class of flag; ocp already synthesises
+  `assistant-partial` events from the TUI frames, so nothing is lost.
+
+- **`--agents` reaches upstream as JSON again.** It is a `json`-kind
+  option, so `parseArgv` stores it already parsed and `buildSpawnArgs`'s
+  `String(value)` turned it into the literal `[object Object]`. Upstream
+  accepts that token without complaint and simply defines no agents, so
+  the flag failed silently. `json`-kind values are now re-serialised, and
+  a value supplied as raw JSON text is passed through unchanged rather
+  than double-encoded.
+
+- **`--effort xhigh` is accepted.** Upstream's levels are `low`,
+  `medium`, `high`, `xhigh`, `max`; ocp's enum was missing `xhigh`, so
+  the value most appropriate for coding and agentic work was rejected
+  at parse time.
+
+- **`--permission-mode manual` is accepted.** Upstream renamed the
+  prompt-on-every-tool mode from `default` to `manual`; ocp's validator
+  still had the old name only, so the current spelling was refused. Both
+  are accepted now — upstream continues to take `default`, so removing
+  it would have broken existing callers for no gain.
+
+- **The `env:` field in `OPTION_SPEC` does something.** It has been part
+  of the documented entry shape since 1.0 but nothing ever read it; the
+  two options that used it were wired up by hand at their call sites, so
+  any new entry declaring `env` silently had no effect. `parseArgv` now
+  applies it, with the precedence the hand-rolled versions already used:
+  **explicit argv > environment variable > spec default**. An explicit
+  `--print-mode=false` beats a set `OCP_PRINT_MODE=1`. Booleans read
+  `1` / `true` / `yes` / `on` as on; everything else, including the
+  empty string, leaves the default in place.
+
+### Added
+
+- **Upstream flags introduced since the spec was last synced.** All
+  forwarded verbatim:
+
+  | Flag | Purpose |
+  |------|---------|
+  | `--autocompact <auto\|tokens>` | Auto-compact window size (`auto`, or 100k–1M) |
+  | `--exclude-dynamic-system-prompt-sections` | Move per-machine sections out of the system prompt so the cached prefix is stable across machines |
+  | `--forward-subagent-text` | Forward subagent text/thinking as messages with `parent_tool_use_id` |
+  | `--prompt-suggestions` | Emit a `prompt_suggestion` event carrying a predicted next prompt |
+  | `--plugin-url <url…>` | Fetch a plugin `.zip` from a URL for this session |
+  | `--safe-mode` | Start with all customizations disabled |
+  | `--ax-screen-reader` | Screen-reader friendly TUI output |
+  | `--brief` | Enable the SendUserMessage tool |
+
+- **`--ax-screen-reader` is worth knowing about even if you do not need
+  accessibility.** It asks the upstream TUI for flat text with no
+  decorative borders and no animations. ocp reconstructs its answer by
+  scraping TUI frames, so there is simply less chrome for the parsers to
+  strip. It stays **opt-in** (flag, or `OCP_AX_SCREEN_READER=1`) because
+  it changes the captured frame shape, and this release does not alter
+  any default.
+
+- **New cross-flag validation.** `--forward-subagent-text` and
+  `--prompt-suggestions` both require `--output-format=stream-json`
+  (upstream only emits them on that channel, and the text/json adapters
+  have nowhere to put them). `--autocompact` is checked for `auto` or a
+  token budget in the 100k–1M range, accepting a `k`/`m` suffix.
+
+- **`buildSpawnArgs` is exported from `open-claude-p`.** It was already
+  a pure function; exporting it lets the spec → request-field → argv
+  round-trip be asserted directly in tests.
+
+### Security
+
+- **`--plugin-url` added to the `passThroughArgv` deny-list.** It loads
+  untrusted plugin code exactly like `--plugin-dir`, which was already
+  listed, so it belongs in the same group.
+
+- **`redactArgvForLog` now redacts every value of a variadic sensitive
+  flag, not just the first.** `--mcp-config` takes multiple tokens and
+  inline MCP JSON routinely carries server credentials; only the first
+  token was masked, so a `--debug` run could print a secret to stderr
+  and from there into a CI log or a pasted bug report. Previously
+  unreachable from the CLI, because `--mcp-config` was one of the flags
+  being dropped.
+
+- **Wiring the dropped flags up widens what an argv-splicing wrapper
+  exposes, and that is now documented.** ocp treats its own command line
+  as trusted; the deny-list guards only *unrecognised* argv forwarded
+  through `passThroughArgv` and has never covered ocp's own spec flags.
+  A wrapper that splices untrusted text into ocp's argv was already
+  unsafe — `--system-prompt`, `--append-system-prompt` and
+  `--allowed-tools` have always been forwarded and the CLI defaults
+  `--dangerously-skip-permissions` to on — but the newly-wired flags
+  turn that from model-mediated influence into direct process
+  execution. `docs/cli-reference.md` gains a **Never splice untrusted
+  input into ocp's argv** section with the safe pattern. The flags are
+  deliberately *not* gated behind an env var: they are documented,
+  first-class options, and hiding them would re-break exactly what this
+  release fixes.
+
+### Tests
+
+- `test/forward.test.js` (39 new tests). The central one walks
+  `OPTION_SPEC` itself and asserts that every argv-forwarding entry
+  survives the trip to the spawn argv, so a flag added to the spec is
+  covered the moment it is declared rather than whenever someone
+  remembers to extend a hand-written list. The rest pin the specific
+  flags that regressed, the refreshed `--effort` / `--permission-mode`
+  surfaces, the new validation rules, and env-var precedence.
+
+### Documentation
+
+- `docs/cli-reference.md` gains the model & behaviour flags
+  (`--effort`, `--thinking`, `--fallback-model`, `--autocompact`,
+  `--exclude-dynamic-system-prompt-sections`), a **Config, plugins &
+  MCP** section covering the flags that were being dropped, the
+  permission-mode values, the new lifecycle flags, and the env-var
+  precedence rule.
+- Model examples across `README.md`, `README.ko.md`, `README.ja.md`,
+  `README.zh.md`, `docs/cli-reference.md`, and the spec's own help text
+  moved off retired ids onto current ones (`claude-opus-5`,
+  `claude-fable-5`) and the current aliases (`fable`, `opus`, `sonnet`,
+  `haiku`).
+
+### Internal
+
+- Two comments in `bin/cli.js` used a non-English example prompt, which
+  the project's contribution rules forbid in committed files. Replaced
+  with English equivalents.
+
+---
+
+## [1.1.3] — 2026-05-19
+
+Fix the `/compact` 24-hour hang and tighten the slash-command path so
+skill invocations are not collateral damage. Driven by a captured
+session (`4af68584-…`) where a manually-triggered `/compact` ran for
+57 s of compaction activity, returned to the input box with no
+assistant turn, and left `runOneShot` waiting for a sentinel that — by
+the upstream's design — could never arrive.
+
+### Fixed
+
+- **`/compact` and other local-builtin slash commands no longer block
+  for `maxResponseMs`.** Claude TUI splits `/`-prefixed prompts into
+  two classes: *local builtins* (`/compact`, `/clear`, `/help`,
+  `/exit`, `/quit`, `/login`, `/logout`, `/cost`, `/status`, `/model`,
+  `/permissions`, `/config`) which run a local handler and never open
+  an `⏺` region, and *LLM-bearing slash invocations* (skills like
+  `/init`, `/review`, `/security-review`, plus every user-installed
+  skill) which DO go through the model. The driver previously
+  appended the OCP_END marker instruction to both and required
+  `hadAssistantText` before the completion detector's idle fallback
+  could fire — fine for the LLM-bearing class, fatal for builtins,
+  which would wait for an assistant region that never opened until
+  the 24 h hard timeout.
+
+  Now `runOneShot` matches the prompt against a narrow whitelist of
+  local builtins. On a match it (a) skips the OCP_END instruction
+  append (the TUI's command parser drops it as junk args anyway, and
+  appending it can pollute free-form-arg commands like `/bug`) and
+  (b) sets the detector's new `allowIdleWithoutResponse` flag so the
+  pre-sentinel idle path is reachable without a prior region entry.
+  Skills and unknown `/<name>` prompts are NOT in the whitelist and
+  keep the existing instruction + strict-idle-gate behaviour, so
+  their LLM responses still complete cleanly via the sentinel.
+
+### Added
+
+- **`CompletionDetector.allowIdleWithoutResponse` (default false).**
+  When true, `_onTick`'s pre-sentinel idle fallback fires after
+  `preIdleMs` of silence even without a prior `assistant-region-
+  entered`. This is the policy switch the driver uses for local
+  builtins; library callers that drive prompts known to produce no
+  assistant turn can opt in directly.
+
+### Tests
+
+- **End-to-end integration coverage for the slash-command path.**
+  `test/driver-slash-command.test.js` spawns the real driver against
+  `test/fixtures/fake-claude-tui.mjs` — a minimal node-pty fixture
+  that reproduces the captured `/compact` shape (spinner activity,
+  `Compacted` stdout, chevron return, no region, no sentinel) — and
+  asserts: `/compact` completes via `reason='idle'` in well under
+  `maxResponseMs`; a plain prompt still completes via `sentinel` with
+  no degraded-capture notice prefix; `/init` (a skill, not a builtin)
+  also completes via `sentinel`, proving the whitelist isn't too
+  greedy. Three new unit tests in `test/detector.test.js` cover the
+  bare `allowIdleWithoutResponse` flag.
+
+---
+
 ## [1.1.2] — 2026-05-19
 
 Recovery-path fixes for the sentinel-missing case. When the end-of-reply

package/README.ja.md

@@ -174,7 +174,7 @@ ocp --output-format stream-json "こんにちは"
 
 # モデルを指定
 ocp --model sonnet "複雑な質問..."
-ocp --model claude-opus-4-7 "アーキテクチャレビュー..."
+ocp --model claude-opus-5 "アーキテクチャレビュー..."
 
 # システムプロンプトを追加
 ocp --append-system-prompt "常に日本語で答えてください" "what's the weather?"
@@ -290,7 +290,7 @@ const result = await driver.runOneShot({
 
   // ── モデル / 動作 ────────────────────────────
   model: 'sonnet',                        // モデル名
-  effort: 'high',                         // 'low' | 'medium' | 'high' | 'max'
+  effort: 'high',                         // 'low' | 'medium' | 'high' | 'xhigh' | 'max'
   thinking: 'adaptive',                   // 'enabled' | 'adaptive' | 'disabled'
   maxTurns: 5,                            // 最大エージェントターン数（shim 強制）
 
@@ -634,7 +634,7 @@ OCP_NO_DAEMON=1 ocp "一度だけ実行"
 
 | req フィールド | CLI フラグ | 型 | 説明 |
 |-------------|---------|---|------|
-| `model` | `--model` | string | モデル名（例：`sonnet`、`claude-sonnet-4-6`） |
+| `model` | `--model` | string | モデル名（例：`sonnet`、`claude-opus-5`） |
 | `systemPrompt` | `--system-prompt` | string | システムプロンプト全体を置換 |
 | `appendSystemPrompt` | `--append-system-prompt` | string | デフォルトシステムプロンプトに追加 |
 | `dangerouslySkipPermissions` | `--dangerously-skip-permissions` | boolean | 権限チェックをスキップ |
@@ -645,11 +645,11 @@ OCP_NO_DAEMON=1 ocp "一度だけ実行"
 | `forkSession` | `--fork-session` | boolean | 再開時に新しいセッション ID を作成 |
 | `sessionId` | `--session-id` | string | 新しいセッションに特定の UUID を指定 |
 | `noSessionPersistence` | `--no-session-persistence` | boolean | セッション保存を無効化 |
-| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `max` |
+| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `xhigh` \| `max` |
 | `thinking` | `--thinking` | enum | `enabled` \| `adaptive` \| `disabled` |
 | `maxTurns` | `--max-turns` | number | 最大エージェントターン数 |
 | `fallbackModel` | `--fallback-model` | string | プライマリモデル過負荷時のフォールバック |
-| `permissionMode` | `--permission-mode` | string | `default` \| `plan` \| `acceptEdits` \| `bypassPermissions` |
+| `permissionMode` | `--permission-mode` | string | `acceptEdits` \| `auto` \| `bypassPermissions` \| `manual` \| `dontAsk` \| `plan` |
 | `mcpConfig` | `--mcp-config` | string[] | MCP 設定パス |
 | `addDir` | `--add-dir` | string[] | ツールがアクセスできる追加ディレクトリ |
 | `bare` | `--bare` | boolean | 最小モード（hooks、LSP、プラグインなどを無効化） |

package/README.ko.md

@@ -169,7 +169,7 @@ ocp --output-format stream-json "안녕"
 
 # 모델 지정
 ocp --model sonnet "복잡한 질문..."
-ocp --model claude-opus-4-7 "설계 리뷰..."
+ocp --model claude-opus-5 "설계 리뷰..."
 
 # 시스템 프롬프트 추가
 ocp --append-system-prompt "항상 한국어로 답변하세요" "what's the weather?"
@@ -285,7 +285,7 @@ const result = await driver.runOneShot({
 
   // ── 모델 / 동작 ──────────────────────────
   model: 'sonnet',                        // 모델 지정
-  effort: 'high',                         // 'low' | 'medium' | 'high' | 'max'
+  effort: 'high',                         // 'low' | 'medium' | 'high' | 'xhigh' | 'max'
   thinking: 'adaptive',                   // 'enabled' | 'adaptive' | 'disabled'
   maxTurns: 5,                            // 최대 에이전트 턴 수 (shim 강제)
 
@@ -634,7 +634,7 @@ OCP_NO_DAEMON=1 ocp "한 번만 실행"
 
 | req 필드 | CLI 플래그 | 타입 | 설명 |
 |----------|-----------|------|------|
-| `model` | `--model` | string | 모델 이름 (예: `sonnet`, `claude-sonnet-4-6`) |
+| `model` | `--model` | string | 모델 이름 (예: `sonnet`, `claude-opus-5`) |
 | `systemPrompt` | `--system-prompt` | string | 시스템 프롬프트 전체 대체 |
 | `appendSystemPrompt` | `--append-system-prompt` | string | 기본 시스템 프롬프트에 추가 |
 | `dangerouslySkipPermissions` | `--dangerously-skip-permissions` | boolean | 권한 검사 건너뜀 |
@@ -645,11 +645,11 @@ OCP_NO_DAEMON=1 ocp "한 번만 실행"
 | `forkSession` | `--fork-session` | boolean | resume 시 새 세션 ID 생성 |
 | `sessionId` | `--session-id` | string | 새 세션에 특정 UUID 지정 |
 | `noSessionPersistence` | `--no-session-persistence` | boolean | 세션 저장 비활성화 |
-| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `max` |
+| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `xhigh` \| `max` |
 | `thinking` | `--thinking` | enum | `enabled` \| `adaptive` \| `disabled` |
 | `maxTurns` | `--max-turns` | number | 최대 에이전트 턴 수 |
 | `fallbackModel` | `--fallback-model` | string | 기본 모델 과부하 시 폴백 |
-| `permissionMode` | `--permission-mode` | string | `default` \| `plan` \| `acceptEdits` \| `bypassPermissions` |
+| `permissionMode` | `--permission-mode` | string | `acceptEdits` \| `auto` \| `bypassPermissions` \| `manual` \| `dontAsk` \| `plan` |
 | `mcpConfig` | `--mcp-config` | string[] | MCP 설정 경로 |
 | `addDir` | `--add-dir` | string[] | 도구가 접근할 추가 디렉토리 |
 | `bare` | `--bare` | boolean | 최소 모드 (hooks, LSP, 플러그인 등 비활성) |

package/README.md

@@ -192,7 +192,7 @@ ocp --output-format stream-json "Hi"
 
 # Specify model
 ocp --model sonnet "Complex question..."
-ocp --model claude-opus-4-7 "Architecture review..."
+ocp --model claude-opus-5 "Architecture review..."
 
 # Append system prompt
 ocp --append-system-prompt "Always reply in English" "what's the weather?"
@@ -413,7 +413,7 @@ const result = await driver.runOneShot({
 
   // ── Model / Behavior ──────────────────────────
   model: 'sonnet',                        // model name
-  effort: 'high',                         // 'low' | 'medium' | 'high' | 'max'
+  effort: 'high',                         // 'low' | 'medium' | 'high' | 'xhigh' | 'max'
   thinking: 'adaptive',                   // 'enabled' | 'adaptive' | 'disabled'
   maxTurns: 5,                            // max agent turns (shim-enforced)
 
@@ -747,7 +747,7 @@ appears below; the canonical, fully-described table is in
 
 | req field | CLI flag | Type | Description |
 |-----------|----------|------|-------------|
-| `model` | `--model` | string | Model name (e.g. `sonnet`, `claude-sonnet-4-6`) |
+| `model` | `--model` | string | Model name (e.g. `sonnet`, `claude-opus-5`) |
 | `systemPrompt` | `--system-prompt` | string | Replace entire system prompt |
 | `appendSystemPrompt` | `--append-system-prompt` | string | Append to default system prompt |
 | `dangerouslySkipPermissions` | `--dangerously-skip-permissions` | boolean | Skip permission checks |
@@ -758,14 +758,19 @@ appears below; the canonical, fully-described table is in
 | `forkSession` | `--fork-session` | boolean | Create new session ID on resume |
 | `sessionId` | `--session-id` | string | Assign a specific UUID to the new session |
 | `noSessionPersistence` | `--no-session-persistence` | boolean | Disable session saving |
-| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `max` |
+| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `xhigh` \| `max` |
 | `thinking` | `--thinking` | enum | `enabled` \| `adaptive` \| `disabled` |
 | `maxTurns` | `--max-turns` | number | Maximum agent turns |
 | `fallbackModel` | `--fallback-model` | string | Fallback when primary model is overloaded |
-| `permissionMode` | `--permission-mode` | string | `default` \| `plan` \| `acceptEdits` \| `bypassPermissions` |
+| `permissionMode` | `--permission-mode` | string | `acceptEdits` \| `auto` \| `bypassPermissions` \| `manual` \| `dontAsk` \| `plan` |
 | `mcpConfig` | `--mcp-config` | string[] | MCP config paths |
 | `addDir` | `--add-dir` | string[] | Additional directories tools can access |
 | `bare` | `--bare` | boolean | Minimal mode (disables hooks, LSP, plugins, etc.) |
+| `autocompact` | `--autocompact` | string | Auto-compact window: `auto`, or 100k–1M tokens |
+| `safeMode` | `--safe-mode` | boolean | Start with all customizations disabled |
+| `axScreenReader` | `--ax-screen-reader` | boolean | Flat, animation-free TUI output — less chrome for the parsers to strip |
+| `forwardSubagentText` | `--forward-subagent-text` | boolean | Forward subagent text/thinking as messages (stream-json only) |
+| `excludeDynamicSystemPromptSections` | `--exclude-dynamic-system-prompt-sections` | boolean | Keep the cached prompt prefix stable across machines |
 | `debug` | `--debug` | boolean | Print debug logs to stderr |
 | `verbose` | `--verbose` | boolean | Verbose output |
 | `cwd` | `--cwd` | string | PTY process working directory |

package/README.zh.md

@@ -164,7 +164,7 @@ ocp --output-format stream-json "你好"
 
 # 指定模型
 ocp --model sonnet "复杂问题..."
-ocp --model claude-opus-4-7 "架构评审..."
+ocp --model claude-opus-5 "架构评审..."
 
 # 添加系统提示
 ocp --append-system-prompt "始终用中文回答" "what's the weather?"
@@ -280,7 +280,7 @@ const result = await driver.runOneShot({
 
   // ── 模型 / 行为 ──────────────────────────────
   model: 'sonnet',                        // 模型名称
-  effort: 'high',                         // 'low' | 'medium' | 'high' | 'max'
+  effort: 'high',                         // 'low' | 'medium' | 'high' | 'xhigh' | 'max'
   thinking: 'adaptive',                   // 'enabled' | 'adaptive' | 'disabled'
   maxTurns: 5,                            // 最大代理轮次（shim 强制）
 
@@ -624,7 +624,7 @@ OCP_NO_DAEMON=1 ocp "只运行一次"
 
 | req 字段 | CLI 标志 | 类型 | 说明 |
 |---------|---------|------|------|
-| `model` | `--model` | string | 模型名称（如 `sonnet`、`claude-sonnet-4-6`） |
+| `model` | `--model` | string | 模型名称（如 `sonnet`、`claude-opus-5`） |
 | `systemPrompt` | `--system-prompt` | string | 替换整个系统提示 |
 | `appendSystemPrompt` | `--append-system-prompt` | string | 追加到默认系统提示 |
 | `dangerouslySkipPermissions` | `--dangerously-skip-permissions` | boolean | 跳过权限检查 |
@@ -635,11 +635,11 @@ OCP_NO_DAEMON=1 ocp "只运行一次"
 | `forkSession` | `--fork-session` | boolean | 恢复时创建新的会话 ID |
 | `sessionId` | `--session-id` | string | 为新会话指定特定 UUID |
 | `noSessionPersistence` | `--no-session-persistence` | boolean | 禁用会话保存 |
-| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `max` |
+| `effort` | `--effort` | enum | `low` \| `medium` \| `high` \| `xhigh` \| `max` |
 | `thinking` | `--thinking` | enum | `enabled` \| `adaptive` \| `disabled` |
 | `maxTurns` | `--max-turns` | number | 最大代理轮次 |
 | `fallbackModel` | `--fallback-model` | string | 主模型过载时的备用模型 |
-| `permissionMode` | `--permission-mode` | string | `default` \| `plan` \| `acceptEdits` \| `bypassPermissions` |
+| `permissionMode` | `--permission-mode` | string | `acceptEdits` \| `auto` \| `bypassPermissions` \| `manual` \| `dontAsk` \| `plan` |
 | `mcpConfig` | `--mcp-config` | string[] | MCP 配置路径 |
 | `addDir` | `--add-dir` | string[] | 工具可访问的附加目录 |
 | `bare` | `--bare` | boolean | 最小模式（禁用 hooks、LSP、插件等） |