fix(cli): post-upgrade What's-new parses the modern changelog format

[muqsitnawaz]

Problem

The post-upgrade "What's new" summary (showWhatsNew → renderWhatsNew) only recognizes the old changelog format — standalone **Heading** lines with - sub-bullets. The changelog moved to single-line - **Title.** verbose prose… entries several releases ago (all of 1.20.48–1.20.50 + Unreleased), so line.startsWith('**') matches nothing and every recent upgrade prints a bare version header with zero bullets — releases look like they shipped nothing.

Fix

renderWhatsNew extracts the bold heading from both formats: ^- \*\*(.+?)\*\* for modern single-line entries (heading kept, prose dropped) and the existing standalone-line match for old sections. Range semantics (from, to] and the Unreleased exclusion are unchanged and now pinned by tests.

Verification

• Unit tests cover both formats, prose-dropping, and Unreleased exclusion (6 tests green).
• Against the real CHANGELOG.md, renderWhatsNew('1.20.49', '1.20.50') renders:

v1.20.50
  • Distributed agent teams: teammates can now run on different machines across your fleet, not just the box running `teams start`.
  • NEW: `agents doctor --devices` shows a cross-device agent-readiness matrix.
  • `agents run codex` / `agents teams` now honor your configured Codex model instead of silently defaulting to `gpt-5.3-codex`.
  • Fix: `agents add claude@<version>` now produces a runnable install — it no longer ships a half-built binary that dies with "claude native binary not installed."

Previously that range rendered v1.20.50 with no bullets.

[prix-cloud]

Code Reviewer

Verdict: Changes requested

Build: ✅ tsc clean (bun run build from apps/cli, exit 0, no diagnostics)
Tests: ✅ bun test src/lib/whats-new.test.ts — 6 pass, 0 fail, 17 expect() calls. Full local bun test hung indefinitely on an unrelated test (no output after 3+ min, unrelated to this diff's files); killed it and relied on gh pr checks 880 instead, which shows all shards (test-shard 1/2/3, typecheck, test-windows, bench, gitleaks) green.

Read before reviewing: root CLAUDE.md, apps/cli/AGENTS.md (no nested AGENTS.md/CLAUDE.md under apps/cli/src or apps/cli/src/lib). Build/test commands and CI scoping (apps/cli/AGENTS.md "Build, test, dev") matched what I ran.

Issues that need attention

The modern-format regex over-matches old-format sections, reintroducing the exact "noisy/wrong summary" failure mode for any upgrade range that crosses a pre-1.20.36 version. whats-new.ts:39 — const entryBullet = line.match(/^- \*\*(.+?)\*\*/); — matches any line starting with - **, not just top-level entries in the modern format. But older changelog sections use the format the PR says is "unchanged": a standalone **Heading** line followed by - sub-bullets, and several of those sub-bullets are themselves bold-led (- **Some claim.** detail…), e.g. CHANGELOG.md:268 under the 1.20.30 → **agents doctor --fix...** heading (CHANGELOG.md:261). I ran renderWhatsNew(changelog, '1.20.7', '1.20.30') against the real CHANGELOG.md (via the built dist/lib/whats-new.js) and got, among the v1.20.30 bullets:

  • `agents doctor --fix` + a daemon safety check: heal the gap between defined and installed
  • `agents doctor --fix`
  • Daemon safety check:
  • `agents doctor` overview now covers every installed version, not just defaults.
  • Says exactly WHAT is out of sync — plugins first.
  • Fixed a false "drift" that could never be reconciled:
  • Corrected a false promise in the sync-status readout.
  ...
  • Auto-cache is on by default.
  • Configurable, still flexible.
  • Explicit `always` now persists
  • Upgrade self-heal:
  • Disclosure fixed.
  • Back-compat:
  • Heal-on-upgrade:
  • Broker version-skew self-heal:
  • Opt-in by construction:
  • Security trade-off (documented in `docs/secrets.md`):
  • Release note:

The correct heading (`agents doctor --fix` + a daemon safety check: heal the gap between defined and installed) is now followed by three of its own sub-bullets rendered as if they were independent top-level entries — several as bare sentence fragments ("Daemon safety check:", "Upgrade self-heal:", "Back-compat:", "Release note:") that mean nothing out of context. This isn't a hypothetical: 1.20.30 is a real, reachable fromVersion for anyone upgrading from an old install, and the PR's own regression tests (whats-new.test.ts) never exercise a range that mixes the old standalone-heading format with a bold-led sub-bullet, so nothing caught it.

Before this PR, the old-format matcher (line.startsWith('**') && line.endsWith('**')) only matched a whole line wrapped in **…**, so a - **Bold.** prose sub-bullet never qualified — old-format ranges rendered cleanly (one bullet per heading). This PR fixes modern-format ranges but regresses old-format ranges that happen to contain bold-led sub-bullets, trading "recent upgrades show nothing" for "old upgrades show garbage" on a different set of inputs.

A fix likely needs the entry-bullet match to only fire when NOT already inside an old-style heading's sub-bullet list — e.g. track whether the current version section is using the standalone-heading format (seen a **Heading**-only line since the last ## version) and skip the - ** match in that mode, or only recognize - ** as a top-level entry when it appears immediately after the version header / a blank line, not mid-section.

Things to verify manually

None beyond the above — the described 1.20.49→1.20.50 verification in the PR body reproduces correctly locally against the real changelog.

---

Reviewed by Code Reviewer — actually ran the build and tests on this branch.

[muqsitnawaz]

Addressed in 13fe819: the - ** entry match is now suppressed once a standalone **Heading** line is seen in the current version section (flag resets per ## version), so old-format bold-led sub-bullets render as details again, not entries.

Re-ran your exact repro against the real CHANGELOG.md via the built dist/lib/whats-new.js: renderWhatsNew('1.20.7', '1.20.30') now emits one bullet per heading with zero sentence fragments (Daemon safety check:, Back-compat:, Release note: etc. all gone), and the modern 1.20.49→1.20.50 range still renders all four entry titles. New tests pin both the bold-led sub-bullet suppression and a mixed-format range.

[prix-cloud]

Code Reviewer

Verdict: Ready to merge

Build: Green (tsc clean via bun run build)
Tests: src/lib/whats-new.test.ts — 8 passed, 0 failed (4 pre-existing + 4 new). src/lib/agent-spec/primitives.test.ts (the compareVersions this module depends on) — 7 passed, 0 failed. Full-repo bun test / bun test src/lib stalled with no CPU progress in this sandbox (looks tmux/network/device-dependent, unrelated to this diff) and was killed after ~60–90s — see "Things to verify manually."

Read AGENTS.md (root) and apps/cli/AGENTS.md before running anything — the latter documents the authoritative bun install && bun run build && bun test flow and confirms whats-new.test.ts sits correctly next to its source (1:1 convention).

Changes that work well

• sectionUsesStandaloneHeadings is scoped and reset per version section (## 1.2.3 boundary), so a range spanning both changelog eras parses each section by its own format instead of a single global mode. I confirmed this against the real CHANGELOG.md: a 1.20.35 → 1.20.42 range correctly renders modern single-line entries for 1.20.42/1.20.41 (dropping their prose) and correctly renders old-format **Heading** entries for 1.20.36 while suppressing its bold-led sub-bullets (- **A bold-led sub-bullet.**-style lines) — matching the intent described in the new tests.
• Ran renderWhatsNew('1.20.49', '1.20.50') against the actual CHANGELOG.md in the tree — output matches the PR description's claimed output exactly (4 bullets, prose dropped, headings intact).
• Tests pin the exact failure mode being fixed (modern format previously matched nothing) plus the trickier regression risk (old-format bold-led sub-bullets must still be suppressed) and the Unreleased exclusion in both formats. Good coverage for a 19-line logic change.
• No signature change to renderWhatsNew; both call sites (apps/cli/src/index.ts:317, :501, :710) are unaffected.

Things to verify manually

• The full bun test run (and even the src/lib-scoped run) stalled in this sandbox with no CPU progress across a 5-minute wait, unrelated to any file touched here — likely a test elsewhere needing tmux/network/a registered device that isn't available in this environment. Worth confirming CI (tests.yml, the required check) is green on this branch, since I could not run the full suite to completion here.

---

Reviewed by Code Reviewer — actually ran the build and tests on this branch.