Merge dev into main#43
Merged
Merged
Conversation
… Skill (#40) * docs(clones): 토스 이승건(sglee) 빌트인 클론 추가 (#38) PO SESSION 7편 전체, EO Korea 인터뷰 2편, 퓨처앤잡 인터뷰 1편, Square of Toss / 창업 현실 영상 2편, Fortune Korea Visionary 기사 2편을 바탕으로 구성. 총 14개 dated knowledge 파일. categories: founder (primary), expert (PO/Growth 프레임). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(clones): 센드버드 김동신(johnkim) 빌트인 클론 추가 (#39) * docs(clones): 토스 이승건(sglee) 빌트인 클론 추가 PO SESSION 7편 전체, EO Korea 인터뷰 2편, 퓨처앤잡 인터뷰 1편, Square of Toss / 창업 현실 영상 2편, Fortune Korea Visionary 기사 2편을 바탕으로 구성. 총 14개 dated knowledge 파일. categories: founder (primary), expert (PO/Growth 프레임). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * docs(clones): 센드버드 김동신(johnkim) 빌트인 클론 추가 @doshkim 존잡생각 채널 10편, 외부 인터뷰·강연 4편, 서울경제 CEO&STORY 기사 1편을 기반으로 구성. 총 15개 dated knowledge 파일. 프레임 중심 정리: PPTM(글로벌 진출), Art of Pivot, Capacity/People Scaling, RAMP(내적 동기), 2x2 행복 매트릭스, 지속가능한 헌신의 구간, 긍정적 집요함 등. categories: founder (primary), expert (B2B SaaS/피봇/글로벌). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * Add standalone openclone CLI without replacing Claude Code The CLI reuses existing persona and knowledge markdown as the source of truth while keeping the Claude Code skill, hooks, setup, and statusline paths intact. Provider routing starts with OpenAI-compatible APIs and adds explicit Codex OAuth and Ollama paths for local/API-based chat. Constraint: Existing Claude Code skill behavior must remain compatible Constraint: Markdown clone files remain the canonical data model Rejected: Treat Codex OAuth token as a plain api.openai.com/v1 key | Codex OAuth requires the Codex backend transport Rejected: Trigger built-in knowledge fetch during list | list must be read-only Confidence: high Scope-risk: moderate Directive: Do not route Codex OAuth through api.openai.com/v1; use the Codex backend/provider path Tested: npm run validate Tested: markdownlint-cli2 "**/*.md" Tested: shellcheck -S error hooks/*.sh scripts/*.sh setup uninstall Tested: npm audit --audit-level=moderate Tested: npm run build Tested: live Codex OAuth smoke with gpt-5.5 Tested: live Ollama smoke with ministral-3:8b Not-tested: Native provider adapters beyond OpenAI-compatible, Codex OAuth, and Ollama * Bring CLI closer to Claude Code persona fidelity The standalone CLI now uses AI SDK tool calling to let the model inspect local markdown knowledge on demand, fetch/search external facts, and keep an Ollama-run-like in-memory conversation when no prompt is supplied. The package is scoped for the new npm org and release publishing is tied to GitHub Release tags so package versions are not manually drift-prone. Constraint: Existing Claude Code skill, hooks, setup, and statusline behavior must remain intact Constraint: Markdown persona and knowledge files remain the canonical data model Constraint: npm package should publish under the newly created @OpenClone scope, not the unrelated unscoped openclone name Rejected: Always stuff every knowledge file into the prompt | too expensive and weaker than dynamic file reads Rejected: Publish on every main merge | npm release should require an explicit GitHub Release event Confidence: high Scope-risk: moderate Directive: Do not make main merges publish npm automatically unless release policy changes; publish is Release-published driven Directive: Keep Codex OAuth explicit and routed through the Codex backend/provider path Tested: npm run validate Tested: npm run build Tested: markdownlint-cli2 "**/*.md" Tested: shellcheck -S error hooks/*.sh scripts/*.sh setup uninstall Tested: npm audit --audit-level=moderate Tested: npm pack --dry-run confirms @openclone/openclone and no stale codex-auth artifact Not-tested: Actual npm publish because it requires a real GitHub Release and npm token * Prevent runaway interactive CLI context Long persona chats previously kept every turn verbatim until the provider failed or degraded from an oversized request. The CLI now compacts older turns into a continuity summary while preserving recent turns, and exposes /compact for explicit user control. Constraint: Keep the feature provider-agnostic and dependency-free Constraint: Preserve one-shot chat, Claude Code support, and existing CLI tool calling Rejected: Exact tokenizer-based budgeting | provider/model tokenizers differ and are unnecessary for a first safety guard Rejected: Drop old turns without summary | loses user goals and decisions that persona continuity needs Confidence: high Scope-risk: narrow Directive: Compaction is approximate char-based; revisit only if provider-specific token windows become first-class config Tested: npm run validate Tested: npm run build Tested: markdownlint-cli2 "**/*.md" Tested: shellcheck -S error hooks/*.sh scripts/*.sh setup uninstall Tested: npm audit --audit-level=moderate Tested: npm pack --dry-run Tested: Architect review APPROVE * Teach agents how to guide openclone CLI users The CLI now has enough provider and runtime surface area that agents need a separate, progressively disclosed guide instead of overloading the Claude Code /openclone dispatcher skill. The new openclone-cli skill keeps the main skill small and splits local Ollama, Codex OAuth, OpenAI-compatible, conversation, knowledge, and troubleshooting details into reference files. Constraint: Root SKILL.md remains the Claude Code dispatcher and must not absorb CLI help details Constraint: Ollama guidance must state that openclone is only the client and does not start the Ollama server Rejected: Put all CLI provider instructions in one SKILL.md | too much context and poor progressive disclosure Rejected: Reuse the root openclone skill for CLI help | it is tied to Claude Code slash-command behavior Confidence: high Scope-risk: narrow Directive: Keep provider-specific CLI guidance in skills/openclone-cli/references rather than bloating SKILL.md Tested: npm run validate Tested: npm run build Tested: markdownlint-cli2 "**/*.md" Tested: shellcheck -S error hooks/*.sh scripts/*.sh setup uninstall Tested: npm audit --audit-level=moderate Tested: npm pack --dry-run includes skills/openclone-cli files Tested: Architect review APPROVE * Keep Codex OAuth response references resolvable Codex OAuth runs through the Responses API, and AI SDK tool or conversation steps can send item references back to the backend. The oauth provider defaulted to non-persistent response items, which made those references disappear between requests and produced Item not found errors.\n\nDefault Codex OAuth to stored response items while keeping an explicit opt-out for privacy-sensitive local experiments.\n\nConstraint: Codex backend rejects AI SDK item references when response items are not persisted\nRejected: Disable tools for Codex OAuth | would break dynamic knowledge and web lookup behavior\nRejected: Force users to set an env var for the working path | the default path should not fail at runtime\nConfidence: high\nScope-risk: narrow\nDirective: Do not default Codex OAuth store back to false without re-testing multi-step tool calls\nTested: npm run build\nTested: npm test\nTested: npm run validate\nTested: npx markdownlint-cli2 README.md skills/openclone-cli/references/codex-oauth.md * Persist CLI conversations and add cross-clone history view Conversation history persistence - Add HistoryStore that saves each interactive session to ~/.openclone/conversations/<slug>/<sessionId>.json on every turn and on exit, with a stable sessionId that is reused across resumes so a single session is one JSON file. - runConversation now accepts initialMessages, initialSummary, and an onPersist callback; the library itself does no I/O so callers can swap the store. CLI wires HistoryStore via chatCommand. - New chat flags: --resume (latest session for that clone), --resume=<id> (specific session), --no-persist (ephemeral run). - On resume the prior summary and every restored message are replayed to stdout in chronological order before the live prompt loop, so scrolling up shows the full prior dialogue. History command - New 'openclone history' subcommand. Bare 'openclone history' now prints help instead of silently falling back to the active clone. - 'openclone history <slug>' lists sessions for one clone with a table header, a per-session 'resume:' hint, and a footer tip. - 'openclone history --all' walks every directory under ~/.openclone/conversations/, groups sessions by clone, and tags any group whose slug is missing from CloneLoader.listClones() with '[orphan: clone not found]'. - HistoryStore.listClonesWithSessions and HistoryStore.listAllSessions back the cross-clone view. - --quiet suppresses the column header and per-session hints for piping. Codex OAuth default - Default codexStore to false so the ChatGPT backend (chatgpt.com/backend-api/codex) accepts requests; OPENCLONE_CODEX_STORE=1 remains as an explicit opt-in for environments that allow store=true. npm publish payload - Trim package.json files array to dist/, clones/, README.md, LICENSE. references/, scripts/, hooks/, skills/, assets/, and SKILL.md are Claude Code skill assets and are not loaded by the npm CLI binary. - Build script now chmods dist/cli/index.js so global installs work. - Bump README status badge to v0.3.0. - Ignore *.tgz so npm pack output does not get committed by accident. Tests - 53 tests pass (was 20). New coverage for history store round-trips, cross-clone listing, orphan classification, persist callbacks, resume replay output, history command formatting, and the codex store default flip. * Restructure README install paths and refresh openclone-cli Agent Skill README - Split the install section into three clearly separated entry points: A. Claude Code skill install, B. Standalone CLI install (the Vercel Agent Skill is now documented inside the CLI section as the recommended way for any AI agent to teach users how to use the CLI), C. Codex CLI experimental install. The mixed 'Node.js CLI' + 'Agent skill for CLI help' text was reorganized into B1 install, B2 Vercel Agent Skill install with example agent prompts, B3 direct usage, B4 provider setup, B5 interactive mode, B6 dev checkout. - Document the canonical install command for the agent skill: npx skills add open-clone/openclone --skill openclone-cli. Vercel Agent Skill (skills/openclone-cli) - Delete skills/openclone-cli/agents/openai.yaml. agents/ is not part of the agentskills.io / vercel-labs/skills directory layout, and the YAML duplicated frontmatter that already lives in SKILL.md. - SKILL.md description now lists the v0.3.0 surface (npm install, history, --resume, --resume=<id>, --no-persist, conversation persistence path, knowledge lookup, errors) so the skill activates on the right user phrases. Body adds 'session persistence and resuming prior conversations' as an explicit dispatch reason for conversation-and-knowledge.md. - references/conversation-and-knowledge.md no longer claims chat history is dropped on exit. New sections cover persistence path, --no-persist, --resume / --resume=<id>, the resumed banner, replay to stdout, openclone history (single + --all + --quiet), and how the compaction summary survives across resumes. - references/quickstart.md gains a 'Sessions and resuming' section covering history listing, resume by latest or by id, --no-persist, and --quiet. Notes that bare 'openclone history' prints help. - references/codex-oauth.md updated for the flipped default: the CLI now sends store=false because the ChatGPT backend rejects store=true for ChatGPT-tier OAuth tokens. OPENCLONE_CODEX_STORE=1 is documented as opt-in for hypothetical future backend behavior, with the expected 'Store must be set to false' 400 noted as a troubleshooting case. * Make CI green: drop legacy openai.yaml check, sync i18n README rows validate-skill.ts - Drop the 'agents/openai.yaml' presence check. The agents/ directory is not part of the agentskills.io / vercel-labs/skills directory layout; the canonical layout is SKILL.md + scripts/ + references/ + assets/. Requiring openai.yaml duplicates SKILL.md frontmatter and tied us to a non-spec format. SKILL.md frontmatter remains the single source of skill metadata. README_en.md, README_zh.md - Add johnkim and sglee rows so validate-readme-i18n no longer flags drift. johnkim and sglee landed on dev via #38 and #39 after the last translation regen, leaving the i18n validator failing on dev. English translations are direct; Chinese rows follow the existing REVIEW NEEDED machine-translation convention. --------- Co-authored-by: Hayun Song <100770062+4entrepreneur@users.noreply.github.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Korean README is the single source of truth and was significantly expanded in #40 (added Standalone CLI section B1-B6, Vercel Agent Skill install instructions, two-path A/B/C install table, npm release section, session persistence and resume documentation, and the johnkim / sglee clone rows). The English and Chinese mirrors were left at the v0.2 era structure, missing all of the above. This commit retranslates both files from scratch against README.md at commit a5072ed so the three READMEs match section-for-section. README_en.md - Sync header bumped to a5072ed. - Status badge bumped from v0.2 to v0.3.0. - Hero tagline updated to mention both the Claude Code skill and the standalone CLI. - New install split: A (Claude Code skill), B (Standalone CLI with B1 install / B2 Vercel Agent Skill / B3 direct usage / B4 provider setup / B5 interactive mode / B6 dev checkout), C (Codex CLI experimental). - B2 documents the canonical "npx skills add open-clone/openclone --skill openclone-cli" install with example natural-language prompts. - B5 documents on-disk session persistence at ~/.openclone/conversations/<slug>/<sessionId>.json, --resume / --resume=<id>, the [resumed: N message(s)] banner, the in-terminal replay of prior dialogue, and the compaction summary surviving across resumes. - B4 explains the Codex OAuth store=false default and why previous_response_id is unnecessary. - New "npm release" section describes the publish-npm.yml workflow and dist-tag rules. README_zh.md - Same restructure mirrored into Simplified Chinese. - Sync header bumped to a5072ed; REVIEW NEEDED marker preserved because no native zh-CN reviewer has signed off on the new sections. - Status badge bumped to v0.3.0. - Same A/B/C install split, B1-B6 substructure, and additional paragraphs covering Vercel Agent Skill install, conversation persistence, --resume / --resume=<id>, replay-on-resume, and the Codex OAuth store=false default. Verification - node .github/scripts/validate-readme-i18n.ts now reports '3 files, 14 clone row(s) cross-referenced, picker + sync headers + install fragments verified'. - markdownlint-cli2 across all three READMEs returns 0 errors. - npm run validate (typecheck + 53 unit tests + 5 hook smoke tests) passes.
…ns (#42) * chore(deps): add Ink 7 + React 19 + marked stack for TUI chat Wires up the runtime dependency tree and tsconfig JSX support that the forthcoming Ink-based interactive chat path requires: - Add ink@^7.0.1, react@^19.2.0, react-reconciler@^0.33.0 as runtime deps so the published CLI ships with everything needed for the TUI. - Add marked@^14 + marked-terminal@^7 for ANSI-rendered markdown of assistant responses. - Add @types/react@^19 to devDependencies. - tsconfig: "jsx": "react" (legacy transform — emits createElement calls and works with tsc-only ESM output without needing the react/jsx-runtime resolution dance). - tsconfig include extended to src/**/*.tsx so the new src/ui/ tree is picked up. No behavior change yet; existing tests still pass. * refactor(conversation): export compactMessages + CONVERSATION_DEFAULTS Lifts the previously-private compactMessages helper to a public export and surfaces a CONVERSATION_DEFAULTS object with the env-driven compaction tunables (compactMaxChars, compactKeepTurns, compactSummaryMaxChars). Both will be reused by the new Ink TUI conversation driver to share the exact same compaction behavior as the readline path. No callers change; all existing tests still pass. * feat(ui): Ink-based TUI chat with markdown rendering Adds an Ink (React 19) TUI front-end for the interactive openclone chat REPL while keeping the existing readline path intact for non-TTY/test scenarios. New components (src/ui/): - App.tsx: root component owning the full conversation state machine. Replicates the readline runConversation behavior 1:1: boot banner, resume banner + prior-summary block + replayed messages + continuation marker, slash commands (/help, /clear, /compact, /bye+/exit+/quit), auto-compaction, persistence after every turn and on exit, persist-failure recovery. - Markdown.tsx: marked + marked-terminal renderer for assistant responses. Headers, bullets, fenced code blocks, inline code, links all render with ANSI styling. Pure renderMarkdown(text) helper is exported for testing. - MessageView.tsx: renders MessageItem (user / assistant / system). User shows '>>> {text}', assistant goes through Markdown, system renders as plain Text. - PromptInput.tsx: single-line input via Ink's useInput. Handles typed input, backspace, Enter (submits), Ctrl+C (abort). Also handles paste-with-newline edge case so pasted commands like '/bye\r' submit cleanly. - hooks/useStateAndRef.ts: vendored verbatim from google-gemini/gemini-cli (Apache-2.0). Provides stale-closure-safe state for streaming event handlers. - runInkConversation.tsx: the entry point that mirrors runConversation options. Mounts App via ink.render, awaits exit, defensively restores stdin (raw mode off, unref, pause) on the way out. Layout uses Ink's <Static> for committed history (no re-render on streaming) plus a separate live-updating region for the current streaming response. Streaming text renders raw while in flight and gets the full markdown pass once the response is committed post-stream — this mirrors the readline path's character-by-character write-then-newline rhythm. Tests (all green, 76 total): - markdown-render.test.mjs (8 cases): pure renderMarkdown unit tests. - ink-render.mjs: thin custom render harness for Ink under node --test. FakeStdin/FakeStdout that emit 'readable' events the way Ink expects, inline 10-line ANSI stripper, no extra deps. - ink-render-smoke.test.mjs: harness sanity check. - ink-conversation.test.mjs (14 cases): full port of the conversation.test.mjs runConversation suite to runInkConversation, asserting on captured frames. Original readline tests (16) still pass unchanged. * feat(cli): dispatch chat to Ink TUI on TTY, readline elsewhere Adds the integration point that decides which conversation backend to use: - If both stdout and stdin are TTYs, lazy-import and call runInkConversation. This avoids pulling Ink/React into the single- shot path (--prompt, piped stdin) and keeps non-interactive tests hitting the unchanged readline runConversation. - After Ink runs to completion, process.exit(0) explicitly so any leftover stdin event listeners or raw-mode artifacts can't keep the process alive after /bye. - Single-shot path (lines below the if) is unchanged. - All existing CLI commands (list, status, history, --dry-run, --prompt, piped stdin) remain bit-for-bit identical. * feat(ui): proper TUI layout with header bar, input box, role chrome The previous Ink port was visually indistinguishable from the ollama-run-style readline REPL because every output line was a 1:1 textual replica. This commit replaces the textual chrome with real TUI primitives so the interactive chat experience now has visible structure (boxes, borders, role colors, spinner) and clearly looks different from a plain readline loop. What changed: - HeaderBar.tsx: cyan rounded box at the top showing 'openclone · {clone} · {model} · ses HH:MM'. - InputBox.tsx: magenta rounded box with '›' prompt + cursor. When streaming is in flight the box turns gray, swaps the prompt for a dot spinner + 'thinking…', and ignores keys. A dim 'enter to send · /help · /bye to exit · ctrl+c' hint sits below the box. - MessageView.tsx: 'you ›' / '{cloneName} ›' role headers with colored, bold prefixes; bodies indented two columns. Markdown rendering still applies to assistant bodies. System lines split into 'system-banner' (yellow dim, like '[resumed: ...]', '[compacted ...]', '--- prior summary ---') and 'system' (plain dim text) so banners stand out from inline help/error text. - App.tsx: re-renders with the new chrome and a streaming region that shows the speaker label + spinner + raw tokens while a turn is in flight. - runInkConversation: accepts speakerLabel / modelLabel / sessionLabel pass-through. - chatCommand: derives speakerLabel from clone.displayName, modelLabel from provider.modelId, sessionLabel from a short slice of sessionId. Tests: still 76/76 green. Three Ink-conversation assertions loosened from '>>> {text}' / 'openclone conversation: {label}' matchers to plain-text matchers, since the new chrome no longer mirrors the readline output byte-for-byte. The behavioral guarantees (turn-by-turn persistence, /clear, /compact, auto-compact, resume, persist-failure resilience, Ctrl+C exit) are unchanged. Deps: + ink-spinner@^5 (runtime). * fix(ui): render markdown with marked lexer + Ink JSX (color-coded) The previous Markdown component piped marked-terminal's string output into a single Ink <Text> node. With default options marked-terminal emits little to no styling — '# H1' came through as the literal string '# H1', '**bold**' came through as 'bold' with no decoration, '`code`' came through as plain 'code'. Users saw raw markers and a visually flat block. Switched to walking marked's token tree directly and emitting Ink <Text> nodes with explicit colors and modifiers per token type: - Headings: marker stripped, depth maps to color (H1 magenta+bold+ underline, H2 cyan+bold+underline, H3 yellow+bold, H4 green+bold, H5/H6 blue/red). - Inline emphasis: **bold** -> yellow + bold; *italic* -> cyan + italic; ~~del~~ -> gray + strikethrough. - Inline code: green text on gray background, padded. - Links: blueBright + underline; the href trails in dim gray. - Lists: cyan bullets ('•') for unordered, cyan numbers for ordered. Nested block items recurse correctly. - Code blocks: rounded gray border with the language label dim at the top and green body lines. - Blockquotes: gray '│' rail on the left, content rendered as nested blocks. - Horizontal rule: dim gray '─' x 40. - HTML / unknown blocks: dim gray fallback. - Decodes the standard HTML entities (& < > etc.) that marked emits inside text tokens so they render readably. Tests: markdown-render.test.mjs is rewritten to mount Markdown via the ink-render harness and assert on stripped frame content. Added explicit checks that '#', '**', '~~', backtick, and other markers do NOT appear in the rendered output, plus per-feature presence checks for headings, lists (both kinds), code, links. 11 cases, all green. Suite total now 79/79. Streaming raw text continues to render through the unchanged App 'streaming' branch (plain <Text>), so the post-stream commit pass through Markdown is what gets the styling — same visual contract as before, just with actual styling applied. * feat(conversation): raise default compaction thresholds for 250K-token contexts Old defaults were sized very conservatively (compacted at the 5–10% mark of any modern context window): OPENCLONE_COMPACT_MAX_CHARS 24000 OPENCLONE_COMPACT_KEEP_TURNS 6 OPENCLONE_COMPACT_SUMMARY_MAX_CHARS 6000 That made openclone fire auto-compaction in dialogues most users would still consider 'just getting started', wasted a per-trigger summarization round-trip, and lost more raw turns than necessary on flagship-context models. New defaults assume a 250K-token context window at ~70% utilization, with Korean-leaning ~2 char/token estimate to stay on the safe side: OPENCLONE_COMPACT_MAX_CHARS 350000 (~70% of 250K tok @ 2 chars/tok) OPENCLONE_COMPACT_KEEP_TURNS 8 (more raw turns kept post-compact) OPENCLONE_COMPACT_SUMMARY_MAX_CHARS 20000 (proportional to new threshold) Smaller-context users (Ollama 8B, gpt-4o-mini etc.) or anyone optimizing for per-turn token cost can still drop the env vars back to the old values explicitly. The README (Korean / English / Chinese) and skills/openclone-cli/references/conversation-and-knowledge.md are updated to document the new defaults plus the downgrade hint. * feat(cli): agent-friendly multi-turn via --resume + --prompt single-shot Previously the single-shot path (--prompt, piped stdin) was stateless: it called streamChat once, wrote the response to stdout, and exited without saving anything. --resume hard-routed to interactive mode, ignoring any --prompt flag. That meant agents had two bad options: 1. Drive an interactive child process over pipes (fragile, requires parsing TUI output). 2. Re-send the entire conversation history themselves on every call (no shared state with the persistence layer). This commit makes single-shot stateful, mirroring what runConversation already does for interactive mode: - --prompt without --resume: starts a new session, persists the turn (user + assistant messages) to ~/.openclone/conversations/<slug>/<sessionId>.json, prints '[session: <id>]' to stderr. - --resume --prompt 'X': loads newest session, appends the new turn, persists, prints '[session: <id>]' to stderr (same id). - --resume=<id> --prompt 'X': loads that specific session. - --no-persist --prompt 'X': stateless behavior, identical to pre-change. No disk write, no stderr session line. stdout receives only the streamed response (terminated with \n), stderr receives the session marker. This split lets agents capture either independently: RESPONSE=$(openclone chat douglas --prompt 'q1' 2>session.log) SESSION=$(grep -oE '\[session: [^]]+\]' session.log | sed 's/\[session: //;s/\]//') openclone chat douglas --resume=$SESSION --prompt 'q2' Auto-compaction also applies in single-shot: if the loaded history plus the new prompt exceeds OPENCLONE_COMPACT_MAX_CHARS, older turns are summarized before the LLM call and '[auto-compacted N older message(s)]' is emitted to stderr. Implementation: - src/lib/single-shot.ts: new runSingleShot() helper. Takes cloneSlug, model, system, prompt, resume flags, and stream injection point. Returns { sessionId, response, messages, conversationSummary, persisted, ... }. Stdout/stderr writers are injectable so unit tests can capture both streams. - src/cli/index.ts: chatCommand single-shot block delegates to runSingleShot. Interactive vs single-shot decision now treats any explicit prompt (--prompt / positional / piped stdin) as single-shot, even when --resume is set. Drops dead streamChat import. - test/single-shot.test.mjs: 7 cases covering new-session create, --resume newest, --resume=<id>, missing-session error, --no-persist, stdout/stderr split, and conversationSummary carry-over. Tests: 79 -> 86 passing. Interactive behavior unchanged: every existing conversation.test.mjs and ink-conversation.test.mjs case still passes byte-for-byte. Docs: - README.md / README_en.md / README_zh.md: B3 'Use it directly' table updated with the new --resume + --prompt rows and a worked agent-style multi-turn example. - skills/openclone-cli/references/conversation-and-knowledge.md: new 'Single-shot mode for agents' section with stream-split contract, capture patterns, and persistence semantics. * docs(skill): update openclone-cli description for guru-consultation triggers Added explicit consultation-oriented triggers (advice, strategy, analysis, domain expertise) and listed the 10 official clones with their categories so that agents can recognize when to invoke openclone for expert perspectives. The installation/troubleshooting triggers remain.
Collaborator
Author
dev → main 변경점 점검 체크리스트
변경 범위 요약
검증 상태
참고 수치
|
Single-shot sessions can start multiple persisted threads in the same millisecond, and timestamp-only IDs could overwrite or collapse distinct sessions. Generate default session IDs monotonically while preserving deterministic timestamp formatting for explicit test dates.\n\nConstraint: CI validate failed on rapid --resume=<id> single-shot test due to millisecond ID collision\nRejected: Add sleeps in tests | would hide the persistence collision and slow CI\nConfidence: high\nScope-risk: narrow\nDirective: Keep explicit Date-based newSessionId calls deterministic for fixture readability\nTested: npm run build\nTested: npm run validate\nNot-tested: Remote GitHub Actions after push
Collaborator
Author
|
CI validate 실패 원인 수정했습니다.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
devintomain.Diff scope
main.Validation
npm run typecheck: passed.npm test: failed:test/single-shot.test.mjssubtestsingle-shot: --resume=<id> targets a specific sessionasserts two generated timestamp session IDs differ, but both were equal in the same millisecond.node --test test/single-shot.test.mjs: same failure reproduced.