From 8554016c0eaa685688904fab0bc19407be971c56 Mon Sep 17 00:00:00 2001 From: Devin Oldenburg Date: Sun, 21 Jun 2026 15:27:56 +0200 Subject: [PATCH] docs(changelog): complete the v0.7.0 entry for release Add the non-code agentic goals feature (task-aware gates + requireCodeReview) and a documentation note (the rewrite + corrected benchmark figures) to the v0.7.0 section ahead of cutting the release. --- CHANGELOG.md | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/CHANGELOG.md b/CHANGELOG.md index f4316f8..2bea539 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -36,6 +36,34 @@ with a one-switch escape hatch: - `CONFIG_DOCS` is the single doc source for every key; a test keeps it in lockstep with `DEFAULT_CONFIG` so docs can't drift. +### Feature: non-code agentic goals + +Goal Mode is no longer only for writing code. A research, analysis, explanation, +or planning goal produces a written answer plus recorded evidence rather than a +diff, and that is now a first-class path to completion. + +- The review gates are task-aware: the code-only gates (`goal-diff-reviewer`, + `goal-verifier`) are required only once a goal is code-bearing — it has edited a + file, recorded changed files, or gone dirty. A non-code goal is gated on its + recorded evidence by the always-on reviewers (`goal-prompt-auditor`, + `goal-reviewer`, `goal-final-auditor`), so an empty `git diff` no longer blocks a + correct text deliverable. The edit signal is monotonic, so a code goal cannot + shed those gates by reverting or committing. +- New `requireCodeReview` option (`auto`, `always`, or `never`; default `auto`) and + `GOAL_GUARD_REQUIRE_CODE_REVIEW` to override the detection. +- The reviewer prompts and the Goal agent were updated to record and evaluate a + text/evidence deliverable instead of treating a clean working tree as a failure. + +### Documentation + +- The user-facing documentation was rewritten to be precise, current, and free of + decorative emoji, in a professional voice. Verifying every claim corrected the + headline benchmark figures to the current run — 92.3% detection and 0.8% false + positives on 704 real third-party commands, at roughly 1.35 microseconds per + command — and the canonical `opencode.json` plugin-options form + (`["./plugins/goal-guard.js", { … }]`) is now used consistently across the README, + the customization skill, the `/goal-mode-customize` command, and `goal-config`. + ### Hardening since v0.6.12 (≈190 issues) A large triage-and-fix sweep landed: deeper shell-analyzer coverage and bypass