Repository navigation
Conversation
An opt-in plugin that reviews a prompt after it is sent and stays quiet unless it has a concrete objection. This slice is the gate: pass, attach, or hold, with one Windows card that can add a guardrail to the outgoing model call. prompt.text is never mutated and the send is never held. The plugin half is TypeScript against @opencode/plugin: it builds a bounded digest, writes a transient request, and injects earned context through the context hook. The daemon half is Windows PowerShell 5.1 + WPF: two deterministic rules, one card, a pid guard, and a stale-request pass. This is the first slice requested in anomalyco#59587. Triage, learning, autostart, and the doctor are later pull requests. Proof: 4 Bun tests against a fake daemon, 38 live checks against the real daemon and card (driven through UI Automation), clean typecheck and oxlint, and 2 PowerShell files with no parse errors. docs/PROOF.md carries the receipts.
Core awaits the prompt hook inside SessionPrompt.prepare, before the message is admitted, so the previous revision was a pre-send gate: it polled 1.5 s for a verdict and 5 s for a card decision on the send path. That contradicted the post-send design. The hook now does one small file write and returns the event untouched. There is no verdict file, no decision file, no poll, no timeout, and no context hook, so the model request is byte-for-byte what opencode would have sent. The daemon reads the request, deletes it, and shows at most one card whose actions are clipboard copy and dismiss. The attach rule is gone with the context hook. It attached a fixed Target: sentence to any change request without a file extension, which changed the system prompt, cost prompt-cache reuse, and told the model the user had established text they never wrote. Tests no longer race a fake daemon: 7 deterministic Bun tests (19 expects) prove one request, an unmutated event, a sub-100 ms return with no daemon, no text on disk when the coach is dead, and the filters. The live test drives the real card and reads the real clipboard: 35 checks, 0 failed. Also fixed: the plugin never created its log directory, so it logged nothing when it loaded before the daemon had ever run.
The card now has two shapes. A destructive ask with no recovery point keeps the guardrail copy. A credential ask that would produce plaintext password handling or has no rate limiting shows 'N things before you accept this' with Explain why, Fix my prompt, and Ignore. Explain why swaps the card in place to the first-principles lesson and offers Back. Fix my prompt puts a rewritten ask on the clipboard. Ignore closes it and leaves the clipboard alone. The rules move to src/Coach.Rules.ps1 so they can be tested headless. tests/Rules.ps1 asserts all 38 branches with no desktop. tests/Run-Minimal.ps1 grows to 66 live checks and clicks every button, including the Explain -> Back loop and the clipboard contents. The plugin test's timing bound is now a generous regression tripwire (< 1 s) instead of a 100 ms latency claim; the same test proves non-blocking structurally by asserting no reply channel exists.
…-only Seven findings now share one card: plaintext passwords, missing rate limits, secrets in the repo, SQL built from input, shell commands built from input, user content rendered as HTML, and unvalidated input. Each has its own why and its own line in the fixed prompt; the card keeps the two most severe. Rules live in src/Coach.Rules.ps1 so tests/Rules.ps1 can assert every branch with no desktop. 65 rule checks, 76 live checks. Drops docs/PROOF.md; the receipts are in the PR.
Wire the extension card to render the triage findings instead of a static message, so every security lesson (SQL, shell, XSS, secrets, error handling) is actually visible. Drop a dead pattern helper and a broken regex while aligning the tests with the rule engine, and point the lint script at the project-wide oxlint. Co-Authored-By: Solar Mini4 <nous>
github-actions
Bot
added
needs:compliance
Contributor
|
Hey! Your PR title Please update it to start with one of:
Where See CONTRIBUTING.md for details. |
Contributor
|
This PR doesn't fully meet our contributing guidelines and PR template. What needs to be fixed:
Please edit this PR description to address the above within 72 hours, or it will be automatically closed. If you believe this was flagged incorrectly, please let a maintainer know. |
Contributor
|
Thanks for working on Code Coach, @0xhzsa! This pull request is too large to review as one change: about 1,900 changed lines, not counting tests. It also mixes two separate things:
Please split it into focused pull requests that can each be reviewed on their own, for example:
You're welcome to open these. |
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
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.
$I added the Code Coach extension as a new card in the extension list. The card shows a quick triage verdict on the learner\u2019s code after they send it, using the same deterministic rule engine we already have \u2014 no model calls, no network, silent by default.\n\nChanges:\n- Extension card in the extension list that renders the triage findings\n- Engine fixes: SQL-concat regex, aligned XSS/shell/secret conditions\n- Tests: profile.record() append-only, rewriter asserts the full intent string\n- lint script: oxlint src\n\nVerified locally: bun test 20 pass, tsc --noEmit clean, oxlint clean.