mirror of
https://github.com/mksglu/context-mode.git
synced 2026-09-19 03:27:16 +08:00
main
11 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Mert Koseoglu
|
7dc26dd4a8 |
feat(cursor): route external MCP tools through PreToolUse (closes #529 partial — cursor)
Extends PR #532's external-MCP routing protection to Cursor. Adds `MCP:(?!ctx_)` to the preToolUse matcher (Cursor's MCP wire shape is `MCP:<tool>`, with own tools at `MCP:ctx_*` — verified via tests/fixtures/cursor/pretooluse-mcp.json and hooks/cursor/posttooluse.mjs). Also extends `routing.mjs` `isExternalMcpTool` to recognise the `MCP:` prefix shape so the routing branch returns external-MCP guidance instead of silently passing through. The shared detection function now classifies Cursor `MCP:*`, Claude `mcp__*`, and Kiro `@*/` tools uniformly. - src/adapters/cursor/hooks.ts: export EXTERNAL_MCP_MATCHER_PATTERN, append to PRE_TOOL_USE_MATCHERS - hooks/core/routing.mjs: extend isExternalMcpTool to detect `MCP:` (cursor) prefix - configs/cursor/hooks.json, hooks/cursor/hooks.json: mirror runtime matcher - tests/adapters/cursor-external-mcp-routing.test.ts: 6-test vertical slice covering matcher regex, adapter config, and routing.mjs integration |
||
|
Mert Koseoglu
|
d05a37584e |
feat(prose): retire prose-style enforcement entirely (#482)
Issue #482 (makoMakoGo) reported that context-mode's caveman/terse injection pressures the model toward brevity on its FINAL ANSWER, not just on tool-output reporting. Cited evidence: Moonshot AI on kimi-k2.5 (anomalyco/opencode#20258, PR #20259) — aggressive brevity prompts measurably degrade coding/reasoning benchmarks because the model drops assumptions, caveats, verification evidence, failure modes, and security warnings the user actually needs. Considered: A — config switch ("injectCommunicationStyle: false"). Rejected: switches default-on become dead code. B — close FR with rationale. Rejected: ignores valid evidence. C — refine wording ("compress when reporting raw tool output, be complete for technical answers"). Rejected: still text injection, model-dependent, half-measure. D — full strip everywhere. Adopted. The decision after grilling: context-mode's value is data routing (sandbox, FTS5, session continuity), not prose styling. The brevity injection conflated three goals — keeping raw data out of context (real, hard-enforced), summarizing tool output compactly (LLMs auto- calibrate), and final-answer prose style (the wrong target). Strip all 22 sites where prose-style language landed. Sites stripped (A-Z): hooks/routing-block.mjs - <communication_style> block (Terse like caveman, fragments OK, auto-expand for security warnings) - <response_format> block (Concise summary, 2-3 bullets) src/server.ts (5 MCP tool descriptions + 2 cosmetic comments) - ctx_execute "When reporting results — terse..." - ctx_execute_file same - ctx_search same - ctx_fetch_and_index same (URL + commands shapes, both) - cosmetic comment "Caveman style — terse status line" - rewrote concurrency note: "Indexing is serial regardless of concurrency" → "Fetches parallelize up to your concurrency setting; FTS5 indexing serializes the writes after (SQLite single-writer rule)." — same fact, less jargon. configs/ (15 adapter MD files — every shipped system prompt) antigravity/GEMINI.md, claude-code/CLAUDE.md, codex/AGENTS.md, cursor/context-mode.mdc, gemini-cli/GEMINI.md, jetbrains-copilot/ copilot-instructions.md, kilo/AGENTS.md, kiro/KIRO.md, omp/SYSTEM.md, openclaw/AGENTS.md, opencode/AGENTS.md, pi/AGENTS.md, qwen-code/ QWEN.md, vscode-copilot/copilot-instructions.md, zed/AGENTS.md All had identical "## Output" block: 3 caveman lines stripped, workflow lines ("Write artifacts to FILES", "Descriptive source labels") kept. CLAUDE.md (repo root — internal dev instructions) Same caveman block stripped. We don't ship this file but we do eat our own dog food. README.md Pillar 4 ("Output Compression — Terse like caveman...") rewritten to "No prose-style enforcement" — explicitly cites the kimi-k2.5 benchmark evidence as the rationale. web/index.html Removed Ch 4b entirely (the "Output compression" chapter with before/after example pushing terse style on the model). Tests: - tests/session/continuity.test.ts: SessionStart routing-block assertion flipped from "must include 'Terse like caveman'" to "must NOT include caveman/terse-style directive". - tests/core/server.test.ts: Task hook injection assertion same flip. Two cosmetic comment renames ("Caveman style — terse status line" → "Status line: counts + sections + size"), test name rename ("caveman style" → "compact format"). Added new "prose-style policy (#482)" describe block at end of file with 3 negative-pin tests covering server.ts MCP descriptions, routing-block, and README. Full suite: 82/82 files passed, 2645 passed, 20 skipped, 0 failed. Net +3 new tests (the policy describe block). CONTRIBUTING.md New "Prose-style policy (#482)" section documents the decision so future contributors don't re-add the injection. Cites the Moonshot benchmark evidence + the regression test that pins the deletion. This addresses #482 in full. Closing the issue with a comment that walks the requester through the decision and links the policy section. |
||
|
Mert Köseoğlu
|
734bdbc2b7 |
v1.0.104 (#401)
* fix(insight): move showAllInsights useState before early return (React #310) * 1.0.102 * fix(test): normalize pluginRoot path separators for Windows (#369) buildNodeCommand() converts backslashes to forward slashes to prevent MSYS path mangling on Windows. The test assertion was comparing against raw pluginRoot (backslashes from mkdtempSync) causing CI failure on windows-latest while macOS and Ubuntu passed. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix(stats): persist counter + show lifetime + auto-memory + business value framing - Add tool_calls table to SessionDB — counter survives upgrades and --continue - Show persistent memory totals (events across all sessions) - Show auto-memory count from ~/.claude/projects/*/memory/ - Replace hardcoded '9 more' with actual category count - Use Opus pricing ($15/M) for cost calculations - Replace '3.0x' with '3x longer sessions' phrasing - Add 'Bottom line' footer with session/lifetime cost summary Closes the upgrade-resets-stats bug. ctx_stats now correctly shows that data persists across compaction, restart, and upgrade. * fix(windows): normalize hooks.json placeholders on startup (#378) The committed hooks/hooks.json and .claude-plugin/plugin.json use ${CLAUDE_PLUGIN_ROOT} placeholders + bare 'node' command. On Windows + Claude Code, this causes runtime loader failures (cjs/loader:1479) because: 1. bare 'node' may not resolve via PATH (Git Bash issue, see #369) 2. ${CLAUDE_PLUGIN_ROOT} resolution can hit MSYS path mangling (see #372) 3. backslash paths get corrupted in shell quoting Fix: start.mjs detects Windows on every MCP server boot. If hooks.json or plugin.json contain unresolved placeholders, rewrites them with: - process.execPath (absolute, quoted) instead of bare 'node' - Forward-slash paths (prevent MSYS translation) - Double-quoted paths (handle spaces) Idempotent — only rewrites when placeholder pattern is detected. Survives upgrades — runs at every start. Closes #378 * fix(cache-heal): use shebang on Unix, self-heal stale node paths After Brew updates Node, the versioned Cellar path written to ~/.claude/settings.json becomes stale, causing 'session start' errors: /opt/homebrew/Cellar/node/25.9.0_2/bin/node (gone after upgrade) vs the stable symlink: /opt/homebrew/bin/node (always current) Root cause: start.mjs wrote `process.execPath` directly, which on Brew returns the versioned path snapshot. Fix (2 layers): 1. New installs on Unix: write cache-heal script with shebang (#!/usr/bin/env node) + chmod +x, register hook as bare script path. `env` resolves node from PATH at runtime — survives any Node upgrade. 2. Self-heal: every MCP boot, check if existing hook command references a node path that no longer exists. If stale, rewrite using current pattern. Windows unchanged (no shebang support) — uses process.execPath + buildHookCommand pattern, plus self-heal for any breakage. Reported by @vigo on Discord. * fix(lifecycle): trigger isParentAlive re-check on stdin EOF to close 30s CPU-spin window (#388) (#389) * fix(test): normalize pluginRoot path separators for Windows (#369) buildNodeCommand() converts backslashes to forward slashes to prevent MSYS path mangling on Windows. The test assertion was comparing against raw pluginRoot (backslashes from mkdtempSync) causing CI failure on windows-latest while macOS and Ubuntu passed. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * ci: update install stats * ci: update install stats * ci: update install stats * ci: update install stats * ci: update install stats * fix(lifecycle): trigger isParentAlive re-check on stdin EOF to close 30s CPU-spin window (#311, #388) The vendored MCP SDK's StdioServerTransport only registers 'data' and 'error' listeners on process.stdin. When the parent (e.g. Claude Code) dies abruptly without sending SIGTERM, the server keeps reading from a half-closed pipe and CPU-spins until the 30s ppid poll catches up. In practice this manifests as orphaned context-mode processes accumulating ~80h of CPU time before being SIGKILL'd manually (#388). The fix adds a single 'end' listener on process.stdin inside the lifecycle guard. It does NOT shut down on 'end' alone — that's the false-positive behavior #236 tore out. Instead, 'end' triggers the same isParentAlive() probe the periodic timer runs, just earlier: - parent alive → no-op (#236 regression test still passes) - parent dead → 30s detection window collapses to ~0 Skipped on TTY (OpenCode ts-plugin), where stdin is not the MCP channel. Tests: added a unit test that emits stdin 'end' under both alive and dead parent conditions, and updated the existing listener-invariance test to pin the new contract (only 'end' touched, restored on cleanup). All existing tests still pass. --------- Co-authored-by: Mert Koseoglu <bm.ksglu@gmail.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: Ray <cho_meiko@okuribito-funeral.jp> * fix(executor): hide Windows console + drop .sh extension for shell exec (#384) On Windows, ctx_execute(language: 'shell', ...) had two problems: 1. Silent output - child_process.spawn without windowsHide:true creates a visible console window that intercepts stdout, leaving the MCP response empty. 2. Git Bash popup - temp script written as 'script.sh' triggers Windows file association for .sh files. bash.exe opens a visible window over the user's IDE. Fix (minimal, two surgical changes): - spawn(..., { windowsHide: isWin }) via buildSpawnOptions(platform) - isWin && language === 'shell' ? 'script' : 'script.{ext}' via buildScriptFilename(language, platform) Both changes are Windows-gated. Linux/macOS behavior unchanged. Does NOT change shell invocation semantics (no bash -c wrapper). Does NOT add SHELL env override. Both deferred - separate features. Helpers exposed as pure functions for unit testing without mocking spawn or filesystem. Closes #384. Supersedes #385 with smaller surface area. * fix(executor): full Windows shell coverage — bash -c source + SHELL override (#384) Builds on commit |
||
|
Mert Köseoğlu
|
e7dbf587b7 |
feat: unified persistent memory & timeline search (#367)
* feat: unified persistent memory & timeline search (#363) Two pillars for session continuity after compaction: Pillar 1 — Unified Timeline Search: - ctx_search gains sort="timeline" mode searching 3 sources chronologically - ContentStore (current session) + SessionDB (prior sessions) + auto-memory - Default sort="relevance" preserves current behavior exactly - New searchAllSources() in src/search/unified.ts - SessionDB.searchEvents() with project_dir filter + LIKE escaping - searchAutoMemory() with adapter-aware dash-separated paths - FTS5 schema migration: 4 new UNINDEXED columns (source_category, session_id, event_id, timestamp) Pillar 2 — Auto-Injection on Compaction: - buildAutoInjection() fires ONLY on SessionStart(source="compact") - <behavioral_directive> for roles, <rules> for decisions, <active_skills> for skills - 500 token hard cap with priority-based overflow strategy - Never fires mid-session — context-mode does not pollute the context it protects Static Reinforcement (~99 tokens, session start only): - 12 adapter configs: Session Continuity + MEMORY sections (4 tiers by hook support) - routing-block.mjs: <session_continuity> tag + tool hierarchy 0. MEMORY - ctx_search tool description: SESSION STATE reminder Bug Fixes: - snapshot.ts: add missing case "role" + buildRolesSection() - extract.ts: skill priority P3 → P2 - pi-extension.ts: minPriority 2 → 3 12 New Categories: Phase 1: user-prompt, compaction, rejected-approach, session-resume, constraint, knowledge-reuse Phase 2: agent-finding, error-resolution, external-ref, blocked-on, iteration-loop, latency 8 review rounds, 16+ engineers per round, TDD-first implementation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: move rejected-approach from pretooluse to posttooluse via marker file PreToolUse cannot safely load SessionDB (native module loading breaks hook stdout on CI). Use tmpdir marker file pattern (same as latency) instead: pretooluse writes marker, posttooluse reads and writes event. Also implements knowledge-reuse event (search_hit_prior) with sessionId wiring and 3 new tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: wire searchAllSources into ctx_search handler + skip empty guard for timeline Two critical wiring bugs found by final review: 1. searchAllSources was imported but never called — sort="timeline" silently fell back to ContentStore-only search 2. Empty-index guard blocked timeline mode even when SessionDB had data Now sort="timeline" creates a read-only SessionDB connection and calls searchAllSources() for unified 3-source search. Default sort="relevance" unchanged — calls store.searchWithFallback() directly. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: correct tool name prefixes in qwen-code, openclaw, kilo adapter configs - qwen-code: bare ctx_* → mcp__context-mode__ctx_* (22 refs) - openclaw: bare ctx_search → context-mode__ctx_search (line 49) - kilo: bare ctx_search → context-mode_ctx_search (line 49) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: address PR review — 6 blockers resolved 1. Timeline sort: ContentStore results get timestamp fallback (new Date()) 2. Read path no longer mutates state: knowledge-reuse write removed from ctx_search (read tool should not write) 3. configDir: use CLAUDE_CONFIG_DIR env var + absolute path (fixes #290 regression) 4. SessionDB: open once before query loop, close in finally (was per-query) 5. Auto-memory: 1MB file size guard on readFileSync 6. Debug logging: all catch blocks log with DEBUG=context-mode env var Removes 3 knowledge-reuse tests that validated deleted write-on-read behavior. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: timeline timestamp correctness + output format + CI fix 1. auto-memory: set timestamp from file mtime (was missing entirely) 2. ContentStore: use sessionStartTime instead of new Date() per-result 3. Normalize SQLite datetime format to ISO before sort 4. Output headers show origin + timestamp: [prior-session | 2026-04-27 14:30 | source] 5. pretooluse.mjs: move marker writes BEFORE stdout (fixes macOS CI timeout) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: macOS CI + auto-memory match quality + snippet boundaries 1. macOS CI: test sentinel path uses /tmp (matching sentinelDir()) instead of tmpdir() which returns /var/folders/... on macOS — fixes isMCPReady() returning false in CI, causing all redirect assertions to fail 2. Auto-memory: raise minimum term length from 2 to 3 chars, use word boundary regex instead of String.includes() (prevents "if" matching "specific", "definition", etc.) 3. Snippet extraction: walk to nearest paragraph boundary (\n\n) instead of cutting mid-word at arbitrary byte offset Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: surface real timestamps from FTS5 chunks for correct timeline sort ContentStore search results now carry actual indexing timestamps instead of falling back to Date.now(). Changes: 1. store.ts: add chunks.timestamp to all 12 search SQL SELECT statements 2. store.ts: populate timestamp on INSERT with new Date().toISOString() 3. types.ts: add timestamp to SearchResult interface 4. unified.ts: use r.timestamp from ContentStore, fallback sessionStartTime Timeline sort now produces correct chronological ordering across all 3 sources (ContentStore indexed_at, SessionDB created_at, auto-memory mtime). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: remove session bundle files from PR Session bundles (session-db, session-extract, session-snapshot) are CI-generated artifacts. Only server.bundle.mjs and cli.bundle.mjs belong in commits. CI will regenerate session bundles on merge. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: remove PRD from PR — belongs in issues, not in repo Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Mert Koseoglu
|
7e35caf43e |
feat: output compression + caveman-style response rules across all 14 adapters
Replace hard "500 words" limit with linguistic compression rules: - Terse like caveman. Technical substance exact. Only fluff die. - Drop: articles, filler (just/really/basically), pleasantries, hedging. - Fragments OK. Short synonyms. Code unchanged. - Pattern: [thing] [action] [reason]. [next step]. - Auto-expand for: security warnings, irreversible actions, user confusion. Changes: - routing-block.mjs: <word_limit> → <communication_style> with soft rules - 14 adapter config MDs: compressed, 500-word limit removed - 6 MCP tool descriptions: added response style tip - README.md: "The Problem" → "How Context Mode Solves It" subheading, output compression as 4th pillar, platform count 12→14 - web/index.html: new "Output compression" section with before/after, platform count 12→14, JetBrains + Qwen Code tags added - CLAUDE.md: updated output section - 3 test files: assertions updated for new routing block text ~65-75% output token reduction with full technical accuracy. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Mert Koseoglu
|
c678fe3750 |
fix(cursor): add stop hook, sync .mdc with routing block, clean HOOK_SCRIPTS
- Create hooks/cursor/stop.mjs — session-end event recording - Remove afterAgentResponse from HOOK_SCRIPTS (no script exists) - Change HOOK_SCRIPTS type to Partial<Record> for incomplete mappings - Sync context-mode.mdc with routing-block.mjs: ctx_purge, /clear preservation, tool hierarchy, forbidden actions, ctx_commands - Add 3 stop hook tests to cursor-hooks.test.ts Addresses @murataslan1 beta test report (issue #45 comment). |
||
|
Mert Koseoglu
|
071ded7a88 |
feat: Think in Code — MANDATORY paradigm across all 12 platforms
Add "Think in Code" manifesto to all instruction files, tool descriptions, and routing messages. LLMs are now instructed to PROGRAM analysis via ctx_execute instead of reading raw data into context to process mentally. Changes: - ctx_execute, ctx_execute_file, ctx_batch_execute tool descriptions updated - 4 routing block messages (curl/wget, inline HTTP, build tools, WebFetch) - 12 platform instruction files with platform-correct tool prefixes - Root CLAUDE.md manifesto Rules enforced: - Pure JavaScript only — no npm dependencies - Node.js built-ins only (fs, path, child_process) - Always try/catch, null-safe - Node.js + Bun compatible - console.log() only the answer Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Mert Koseoglu
|
725951a24a |
1.0.59
fix: add PreToolUse to Codex hooks.json + stop to Cursor hooks.json README: - Codex: hooks no longer optional, PreToolUse added for routing enforcement - Codex: removed from "no hook support" references (lines 747, 830) - Codex: continuity table updated (PostToolUse/SessionStart: Yes) - Cursor: stop hook in hooks.json example Configs: - configs/codex/hooks.json: created with PreToolUse+PostToolUse+SessionStart - configs/cursor/hooks.json: added stop hook entry Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Mert Koseoglu
|
4bbbd46a29 |
fix: platform-specific tool prefixes in configs, Kiro mcp.json rename, Cursor .mdc, README rewrite
## Config tool prefix fixes (evidence-based from official docs) - configs/opencode/AGENTS.md: mcp__context-mode__ → context-mode_ (OpenCode convention) - configs/zed/AGENTS.md: mcp__context-mode__ → mcp:context-mode: (Zed convention) - configs/codex/AGENTS.md: mcp__context-mode__ → bare ctx_ (Codex convention) - configs/openclaw/AGENTS.md: mcp__context-mode__ → bare ctx_ (OpenClaw convention) - configs/pi/AGENTS.md: mcp__context-mode__ → bare ctx_ (Pi convention) ## Kiro adapter fix - Renamed configs/kiro/mcp_config.json → mcp.json (matches official Kiro docs) - Updated src/adapters/kiro/index.ts: 5 references fixed - Updated tests/adapters/kiro.test.ts ## Cursor routing rules - Created configs/cursor/context-mode.mdc with alwaysApply: true - Uses bare tool names per Cursor convention ## README rewrite (DX Engineer) - All 12 platform install sections rewritten in consistent book format - Every section: Prerequisites → Install → Verify → Routing - Platform-correct tool names in all examples - No references to auto-created files (disabled) - Collapsible sections for advanced config Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|
Mert Koseoglu
|
eae8dfcf95 |
fix(cursor): use verified env vars, remove broken sessionStart hook
Based on source-verified audit against Cursor forum, GitButler deep dive, and macos-notify-mcp: - Detection: replace fabricated CURSOR_CWD/CURSOR_SESSION_ID with verified CURSOR_TRACE_ID (MCP context) and CURSOR_CLI (terminal context) - sessionStart: set capabilities.sessionStart=false (Cursor rejects it with "Unknown hook type" per forum.cursor.com/t/149566). Routing instructions now delivered via server.ts startup (same as Codex CLI) - hooks.json: remove undocumented type/loop_limit/failClosed fields, remove sessionStart entry - hooks.ts: remove sessionStart from REQUIRED_HOOKS - README + docs/platform-support.md: update Cursor sections to reflect verified capabilities Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |
||
|
Ryne Andal
|
e29a51ec94 |
feat: add native Cursor support via event hooks (#89)
* feat: add native Cursor support with event hooks - Introduced native Cursor hooks: preToolUse, postToolUse, and sessionStart. - Added configuration files for Cursor in and . - Updated documentation to include Cursor platform support and hook details. - Implemented routing and formatting for Cursor-specific events. - Updated session management to add Cursor-specific session handling. * fix: update test assertions for CursorAdapter plugin registration, make paths platform-agnostic |