feat(continue): add Continue.dev integration for VS Code + JetBrains

[prooshani]

Summary

Adds Continue (the open-source AI assistant for VS Code + JetBrains) as a first-class caveman target, alongside the existing Cursor/Windsurf/Cline/Copilot integrations. While verifying the new integration end-to-end I discovered a latent tool-calling bug that affects every existing agent-mode integration too, and bundled the fix here so the Continue integration ships working from the first merged commit.

Two commits on this branch:

1. feat(continue) — new .continue/rules/caveman.md, CI sync, README/CLAUDE updates
2. fix(rules) — fence tool calls from caveman compression (edits sources only, CI regenerates agent copies post-merge)

Commit 1 — Continue.dev integration

• New .continue/rules/caveman.md with alwaysApply: true frontmatter — auto-loaded by Continue in any workspace containing the file (format per the Continue rules docs)
• CI sync: sync-skill.yml now rebuilds the Continue rule from rules/caveman-activate.md on every change, same pattern already used for Cursor/Windsurf
• README: new install-table row, What You Get column, and a detail block that explains the distinction between .continue/skills/ (on-demand, what npx skills add -a continue installs) and .continue/rules/ (always-on, what this PR adds). Per-project + global install snippets and reload instructions for both VS Code and JetBrains included.
• CLAUDE.md + CLAUDE.original.md: synced-files table, CI frontmatter note, and Agent distribution table updated for reviewers/future contributors

Commit 2 — tool-call fence (bug fix)

The bug

When caveman is active in an agent mode (Continue, Cursor, Windsurf, Cline, Codex, Claude Code — any agent that does function/tool calling), tool calls fail with Tool not found. The compression directive in the rule (short synonyms, abbreviate, drop articles) bleeds into the tool_use JSON, so the model emits read instead of read_file, grep instead of grep_search, {"fp": ...} instead of {"filepath": ...}, and the tool registry rejects the call.

The fix

Added an explicit "tool calls stay exact" fence to both source files:

• rules/caveman-activate.md — short form (mirrors the terse style of that file)
• skills/caveman/SKILL.md — long form with examples, under its own ## Tool Calls section

Per CONTRIBUTING.md, I edited sources only. Post-merge CI will regenerate the agent-specific copies (.cursor/rules/*.mdc, .windsurf/rules/*.md, .continue/rules/*.md, .clinerules/*.md, .github/copilot-instructions.md, all SKILL.md copies) on its next run.

Reproduction

1. Enable caveman rule in Continue.dev (VS Code, agent mode) against a model that does tool calling
2. Ask the agent to read any file
3. Model emits read → Continue resolver: Tool not found

Verification

I ran an 11-step tool-call test (file reads, codebase grep, glob, terminal, edits, multi-tool chain, intentional-error path) in a separate project (haevn admin-panel) under three configurations:

Config	Result
Caveman disabled	10/10 pass (step 10 errors exactly as designed) — baseline
Caveman enabled, no fence	Multiple Tool not found failures on abbreviated tool names
Caveman enabled, with fence (this PR)	10/10 pass, exact tool names: read_file, grep_search, file_glob_search, run_terminal_command, edit_existing_file. Explanation prose still caveman-style (fragments, dropped articles). Tool_use JSON untouched.

Files changed

.continue/rules/caveman.md         | new (22 lines)
.github/workflows/sync-skill.yml   | +8 / -1
CLAUDE.md                          | +3 / -1
CLAUDE.original.md                 | +3 / -1
README.md                          | +29 / -13
rules/caveman-activate.md          | +2 / -0
skills/caveman/SKILL.md            | +12 / -1

No existing agent behavior changed (the fence makes tool calls work where they were silently failing). No secrets, no destructive edits.

Test plan

• Rule file frontmatter validated against Continue docs (name, description, alwaysApply: true)
• CI-simulated file generation byte-matches the committed .continue/rules/caveman.md
• Workflow YAML parses clean
• Rule loads in Continue (VS Code) inside the caveman repo
• Rule loads in Continue (VS Code) inside an external project — always-on confirmed across arbitrary workspaces
• stop caveman / normal mode reverts to normal prose (auto-clarity preserved)
• Code blocks in responses remain unmodified (boundary rule preserved)
• Agent-mode tool calling works with caveman active after fence — 11-step diagnostic passes
• Reviewer: verify in JetBrains Continue extension (I only tested VS Code; rule format is workspace-agnostic and documented as cross-IDE by Continue)
• Reviewer: verify fence helps on other agents with tool calling (Cursor agent, Cline, etc.) — repro is agent-mode-general, not Continue-specific

Notes for reviewer

• The Continue detail block was intentionally placed in the existing shared <details> for Cursor/Windsurf/Cline/Copilot rather than creating a new top-level section, since the mechanism (workspace rule file with always-on frontmatter) is identical to Cursor/Windsurf.
• npx skills add -a continue is already supported upstream by vercel-labs/skills — no coordination needed there. That command installs into .continue/skills/ (on-demand). This PR complements it with the always-on rule path.
• I kept the commits separate so the bug fix can be cherry-picked independently if you'd prefer to ship the Continue integration and the tool-call fence as two PRs. Happy to split if that's cleaner for review.