chore: v0.3.0

knowledge/decisions/2026-04-10-1042-expand-getreviewcomments-to-fetch-all-three-github-pr.md

@@ -0,0 +1,38 @@
+---
+title: "Expand get_review_comments to fetch all three GitHub PR comment types"
+date: 2026-04-10
+author: "bennisjonas@gmail.com"
+source: "https://github.com/Jonasb8/memex/commit/cb591c4516921463b79e30f770b26cf27340515d"
+repo: "Jonasb8/memex"
+confidence: 0.55
+tags: []
+---
+
+# Expand get_review_comments to fetch all three GitHub PR comment types
+
+> ⚠️ **Low confidence** — limited rationale present in source. Verify before relying on this record.
+
+## Context
+
+The existing `get_review_comments` function only retrieved top-level review bodies submitted via GitHub's 'Review changes' flow. This meant inline line-level code comments and general PR thread (issue-level) comments were silently ignored, leaving potentially significant reviewer feedback uncaptured.
+
+## Decision
+
+Refactor `get_review_comments` to make three separate GitHub API calls — one for top-level review bodies (`gh pr view --json reviews`), one for inline review comments (`repos/{repo}/pulls/{pr_number}/comments`), and one for general PR thread comments (`repos/{repo}/issues/{pr_number}/comments`) — aggregating all results into a single deduplicated list.
+
+## Alternatives considered
+
+_None recorded_
+
+## Constraints
+
+- GitHub exposes PR comments across three distinct API endpoints with no single endpoint returning all comment types
+- Must use GitHub CLI (`gh`) subprocess calls rather than a direct API client
+
+## Revisit signals
+
+_None_
+
+---
+
+_Extracted by Memex from [commit cb591c45](https://github.com/Jonasb8/memex/commit/cb591c4516921463b79e30f770b26cf27340515d) · 2026-04-10_

knowledge/decisions/2026-04-10-1042-use-structured-embed-text-title-context-decision-alternatives.md

@@ -0,0 +1,38 @@
+---
+title: "Use structured embed text (title + context + decision + alternatives + constraints) instead of raw markdown for semantic search"
+date: 2026-04-10
+author: "Jonasb8"
+source: "https://github.com/Jonasb8/memex/pull/11"
+pr: 11
+repo: "Jonasb8/memex"
+confidence: 0.55
+tags: []
+---
+
+# Use structured embed text (title + context + decision + alternatives + constraints) instead of raw markdown for semantic search
+
+> ⚠️ **Low confidence** — limited rationale present in source. Verify before relying on this record.
+
+## Context
+
+The memex system was previously embedding raw markdown content for knowledge records, which reduced semantic search quality. The PR introduces a cleaned, structured representation for embedding to improve retrieval relevance.
+
+## Decision
+
+Embedding now uses a structured representation composed of title, context, decision, alternatives, and constraints fields rather than raw markdown, and query excerpts are drawn from structured sections (## Context and ## Decision) rather than the first raw paragraph.
+
+## Alternatives considered
+
+_None recorded_
+
+## Constraints
+
+_None recorded_
+
+## Revisit signals
+
+_None_
+
+---
+
+_Extracted by Memex from [PR #11](https://github.com/Jonasb8/memex/pull/11) · 2026-04-10_

knowledge/decisions/2026-04-10-1043-rerun-extraction-with-author-reply-on-lowconfidence-nudge.md

@@ -0,0 +1,39 @@
+---
+title: "Re-run extraction with author reply on low-confidence nudge comments"
+date: 2026-04-10
+author: "bennisjonas@gmail.com"
+source: "https://github.com/Jonasb8/memex/commit/cd30d56bbfda29c9f4157d686627d439dc04c870"
+repo: "Jonasb8/memex"
+confidence: 0.60
+tags: []
+---
+
+# Re-run extraction with author reply on low-confidence nudge comments
+
+> ⚠️ **Low confidence** — limited rationale present in source. Verify before relying on this record.
+
+## Context
+
+When Memex detects a PR that looks like it contains a decision but lacks sufficient rationale, it posts a comment asking the author for one sentence of context. Previously, that reply was not acted upon — extraction was only triggered on merged PRs.
+
+## Decision
+
+Extend the GitHub Actions workflow to also trigger on `issue_comment` events, passing the comment body and author as environment variables so that when the PR author replies to a low-confidence nudge, Memex re-runs extraction with the reply and writes the knowledge record.
+
+## Alternatives considered
+
+_None recorded_
+
+## Constraints
+
+- The workflow must support two distinct trigger paths (merged PR and issue comment) using the same extraction job
+- PR number must be resolved from either the pull_request or issue event context depending on trigger type
+- The git branch checkout step must handle the case where the branch already exists (from the initial low-confidence pass)
+
+## Revisit signals
+
+_None_
+
+---
+
+_Extracted by Memex from [commit cd30d56b](https://github.com/Jonasb8/memex/commit/cd30d56bbfda29c9f4157d686627d439dc04c870) · 2026-04-10_

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "memex-oss"
-version = "0.2.0"
+version = "0.3.0"
 description = "Institutional knowledge extraction for engineering teams"
 readme = "README.md"
 license = { text = "AGPL-3.0-or-later" }