Changelog

All notable changes to this plugin are documented here.

Format follows Keep a Changelog; versioning follows SemVer.

---

[0.4.3] - 2026-08-08

The accumulator guidance was the inverse of the bug

A user hit a defect the whole stack missed: a var total re-accumulated by a loop on every bar, which grows for the life of the chart. The skills' advice — "declare accumulators var so they persist" — is correct for a running total and is exactly what produces this. Run the old checklist against the broken code and every item ticks.

No skill mentioned for or while anywhere, so a reader had nothing to catch the loop shapes with. All three skills now cover both directions and ask the question that decides it: does the accumulator's lifetime match its meaning? "Total so far" wants var; "total for this bar" must not have it.

Now detected as S3 in pinescript-v6-validator@0.3.0, pinned here.

Honest omissions added to "What the validator CANNOT see"

That table implied name-checking was covered. It is not:

• Constant and built-in-variable names are never checked. shape.trianglup, color.grene and plot.style_circlez all validate clean and all fail on TradingView. Only parameter names are checked, never the values. lookup_pine_reference covers functions only, so found: false for barstate.islast means "not a function", not "not real".
• Re-declaring a name with = is a TradingView error and passes here.
• ta.* on the right of and/or escapes S2, because v6 short-circuits.

Gate hardening

• validate_skill_examples.py returned 0 when the engine was absent, so a local make gate printed "Gate passed" over zero validated examples. It now fails unless --allow-skip is passed deliberately.
• The five scaffolds advertised as "validated in CI" were opened by no script at all — true by luck, not enforcement. Now validated as whole files: 31 examples, 5 of them scaffolds.
• New make anchors resolves every semantic check's docAnchor against real skill headings, using the installed engine. Four of nine were dead when it was written, including S1's — the most-cited defect in Pine Script. The engine's own test asserted docAnchor.includes('#'), which all four dead links passed.

Skills now reference each other

Previously none did, so an agent that loaded pinescript-v6 never learned the validated scaffolds existed two directories away. The PostToolUse hook is now documented, so agents stop redundantly re-validating after every write.

Added a screenshot of Antigravity auditing a script through the plugin.

---

[0.4.2] - 2026-08-08

Changed

• Requires pinescript-v6-validator@^0.2.1, which repairs three checks that each worked on their happy path and silently missed a shape occurring constantly in real code:

  • S1 skipped every multi-line request.security(). Wrapping is the normal formatting for that function, so most real repainting went undetected.
  • S2 checked only the true branch of a ternary.
  • S9 counted strategy.cancel as an exit. Cancel withdraws a pending order; it does not close a position.

  v0.4.1 was tagged before that engine fix landed, so hub installs picked up the broken checks. This release exists to correct that.

[0.4.1] - 2026-08-08

Fixed — found by cross-harness UAT and an adversarial coherence review

• Codex discovered zero skills while reporting a successful install. .agents/skills was a symlink; installers copy rather than clone, so it arrived empty. Now real files, kept in step by scripts/sync_agents_skills.py and gated.
• The hook never tried the npm engine. It resolved only local checkouts of the extension repo, so on any normal install it found nothing and exited silently — while the README promised "an actual control".
• PLUGIN.md still demanded a built checkout of the extension, contradicting the README. It listed one skill when four ship, and "both diagnostic paths" when there are three.
• pinescript-strategy told agents the validator "cannot catch any of the traps below" — two of which are now S1 and S2 and are caught. An agent would have discounted a live finding as out of scope.
• pinescript-validation documented S3 and S4 as shipping checks with remedies. Neither is implemented; both are now struck through and flagged as check-by-eye.
• Version was 0.4.0 in five manifests but 0.1.0 in package.json, PLUGIN.md, the MCP server and three skill files, with no 0.4.0 changelog entry at all.

[0.4.0] - 2026-08-07

Added

• MCP server surfaces semantic findings from engine 0.2.0 with // pine-ignore honoured. Each carries a check field; its absence marks a syntactic diagnostic, which cannot be suppressed.
• pinescript-validation documents every check and its remedy.

[0.3.0] - 2026-08-07

Added

• skills/pinescript-strategy — entries, exits, position sizing, and the five traps that make a backtest lie. Built from user evidence rather than assumption: practitioners report that "one overlooked mistake — like a repainting signal or scope error — can invalidate months of backtesting" (PickMyTrade). Covers repainting, unconfirmed-bar entries, ta.* in conditionals, the v6 lazy-evaluation trap, and optimistic fill assumptions — every one of which compiles cleanly and is still wrong.
• Two CI-validated scaffolds: basic-strategy and risk-managed-strategy, the latter with ATR stops, risk-based sizing and webhook JSON payloads for broker automation (TradersPost / PickMyTrade shape).

Notable

Competitive research (_plugin/research/COMPETITIVE-LANDSCAPE.md) found that every comparable AI tool is prose-only — none can check the code it emits. Two failure modes exist and the industry conflates them: syntactic errors, which a validator solves, and semantic errors that compile perfectly and are still wrong. The expensive failures are all semantic.

Embedded Pine examples validated in CI: 14 -> 22.

---

[0.2.0] - 2026-08-07

Added

• skills/pinescript-validation — every diagnostic class the validator emits, with its deterministic fix. Includes an explicit what the validator cannot see section (type compatibility, runtime logic, repainting, platform limits), because a clean run is not proof a script is correct, and a workflow for auditing an existing codebase of broken .pine files.
• skills/pinescript-indicator — three scaffolds validated in CI (overlay-indicator, oscillator, drawing-objects) covering plotting, drawing objects with cleanup, tables, alerts and the plot/object caps. Copy a scaffold rather than composing from memory: composing from recollection is how wrong parameter names get in.
• hooks/validate-pine.sh — PostToolUse hook validating every .pine file the agent edits. Real installs into Claude Code and Antigravity both reported Hooks: 0; the plugin told agents to validate and shipped nothing to enforce it.

Changed

• The engine now comes from npm. pinescript-v6-validator@0.1.0 is published, so the plugin no longer requires a local checkout of the VS Code extension. The on-disk fallbacks remain for developing against unreleased engine changes.
• Corrected the description in all five manifests. It advertised skills that did not exist — promising components the package does not ship.

Verified

Installed for real into Claude Code (Skills 3 · Hooks 1 · MCP 1) and Antigravity. Packaging consistency across all four harness manifests is now enforced by make gate. 14 embedded Pine examples validated in CI, up from 4.

---

[0.1.0] - 2026-08-07

Initial release. Scaffold plus the first skill and a working MCP server.

Added

• Agent Plugins 1.0.0 conformance — plugin.json and mcp.json at the plugin root, closed schema, MCP declared out of band. .claude-plugin/ manifests for Claude Code.
• skills/pinescript-v6 — execution model (var, :=, why ta.* must run unconditionally), the two call forms of line.new / label.new / box.new, anti-repainting with request.security, platform limits, and the API TradingView shipped through July 2026 — including rules it removed.
• MCP server with two tools:
  • validate_pine_script — runs both diagnostic paths, so its verdict matches what the VS Code extension shows
  • lookup_pine_reference — real signatures with overloads surfaced explicitly, and near-miss suggestions when a symbol is not found
• make gate — spec conformance, skill contracts, reference-URL resolution, embedded Pine example validation, and MCP behaviour tests. Wired as a pre-commit hook via make hooks, and run in CI.
• 14 MCP behaviour tests, including one asserting both diagnostic sources are wired in and one covering alertcondition messages containing commas.

Notable

scripts/validate_skill_examples.py extracts every ```pine block from every skill and runs it through the real validator. Contrast blocks are split at the WRONG/RIGHT boundary so the correct half is validated rather than the whole block being skipped. When the engine cannot be located it skips loudly — an unverified example must never look verified.

Known limitations

• One skill. pinescript-validation, pinescript-strategy and pinescript-indicator are planned.
• The engine is not published to a package registry. The plugin locates it on disk, so it requires a local checkout of pinescript-vscode-extension. Until that is resolved this is not ready for a plugin-hub listing.
• No mascot or illustrative assets yet.