[General] Proposed v0.8 integration branch for review

[xsyetopz]

Area

Coding agent and CLI

Topic

I have working integration branch for large part of the currently open v0.8.x work in here based on upstream commit 97b994c3.

It combines fixes for the open v0.8.x Issues, selected open PRs, CI repairs, and several bugs I found (and locally fixed before forking) while using the result locally.

If maintainers want this work, I can open one integration PR or split it into smaller PRs in the order they prefer.

By the way, it SHOULD NOT be reviewed as one 318-commit block. The commits preserve the source branches and integration boundaries, so that reviewers can inspect one area at a time.

---

These issues were directly addressed:

• Custom endpoint documentation gaps for new open-source users #534
• Complete CI and release compatibility hardening #1379 Make persisted session and configuration state crash-safe #1380 Fence daemon worker identity and recover supervisor ownership safely #1381 Complete queued, interrupted, and archived session lifecycle recovery #1382 Harden MCP OAuth discovery and Codex transports #1383 Bound transcript repair, compaction, and autonomous recovery #1384

There're some partial/substantial bits of #1182 . Also, #1251 said they fixed #1120 which was closed by then. I listed this one specifically for provenance.

My branch ALSO includes reviewed and,in many cases, rebased OR hardened versions of work from these pull requests:

• fix anthropic record-valued tool arguments #367 fix no-env prime authentication isolation #374
• Disable Ghostty inline image escapes by default #413 feat(coding-agent): surface refinement status and outcomes #447 fix(coding-agent): make subagent cancellation explicit #464
• Fix model search ranking by intent #539 Add distinguishing JSON mode types for autonomous mode for --json mode #569
• fix(agent): fail closed on provider quota exhaustion #795
• Smooth fullscreen transcript wheel scrolling #887
• prompt: forbid blocking time.sleep() polling in the RLM control loop #1034
• test: add deterministic local swarm evidence substrate #1106 fix(coding-agent): passivate quiescent roots after restart #1107 test(coding-agent): prove B00B production-path swarm gate #1115 feat(daemon): fence lifecycle operations by durable identity #1123 ENG-5106 Create new agents at the current view's depth with ctrl+n #1136 feat: add /exit as an alias for /quit #1142 [v0.8 MCP 1/4] establish generic MCP transport foundation #1175 [v0.8 Release 1/3] bind artifacts to immutable source commits #1176 [v0.8 Prompts P01] preserve typed system-prompt provenance #1177 feat(coding-agent): add tokens/sec to the working loader #1179 feat(coding-agent): add async bash() to IPython kernel #1187 prompt: make long-running RLM work nonblocking, visible, and clear #1188
• [v0.8 Prompts P02] add clear direct prose guidance #1237 feat(kernel): dispatch host requests as capabilities #1243 fix(security): contain private session artifacts #1249 fix(security): keep no-session descendants ephemeral #1251 fix(coding-agent): handle clipboard helper failures #1252 fix(security): cancel RLM work on host teardown #1253 Resume interrupted work after an automatic compaction #1279
• [v0.8 MCP 5] enforce project MCP trust boundaries #1337 [v0.8 MCP 8] close late probe sessions #1338 Capability cleanup + three admission/revocation fixes for the host-request dispatcher #1357 Agents view: hint 'type to search sessions' instead of 'type to start' #1367 Keep the draft when opening the agents view #1372 Add Ctrl+J to toggle edit diffs independently of tool output #1388 Keep the original session running and listed when forking #1389 refactor(daemon): consolidate rlm subagent metadata onto the spawn ledger #1390 feat(coding-agent): always-visible edit summary with inline Ctrl+J diff #1392

---

Only current heads of PRs #1170 , #1337 , and #1338 are present unchanged. The other ones were integrated from earlier revisions, rebased, combined w/ dependent work, or changed after review. Their commit history, and attribution, were kept where possible.

What have you tried or considered?

• bun run check passes on current branch.
• Latest fork CI run passed on macOS, Linux, and Windows: https://github.com/xsyetopz/prime-agent-next/actions/runs/31928024625
• Specific tests were run while each upstream issue, and pull-request slice, was integrated.
• Chained skill autocomplete test passes with 38 tests in the file. Oh, and user's own testing also worked.
• Local installer dry-run builds a temporary bundle, installs production dependencies, and launches prime-agent --version with Node.
• Local manual checks covered multi-skill completion, /exit, clipboard image markers, and image input to a vision-capable model.

Additional context

These changes came from my own testing rather than an upstream issue or pull request:

• Bun standalone build crashed on macOS when native module called uv_async_init. I let local installer build non-minified ESM bundle, and run it with Node instead.
• First Python installer could pass dry-run, and still fail during real build. Dry-run now checks workspace packages, source types, required assets, prod deps, and temp launch.
• Copying example dirs failed on broken, or external node_modules symlinks. Asset copies now exclude nested node_modules.
• Clipboard image paste blocked while image was resized. Editor now attaches original image immediately, and resizes retained copy in background.
• [image #N] markers were not styled consistently. Marker matching is case-insensitive, and keeps existing lowercase inserted form.
• /skill:<name> was treated like ordinary slash command. Skill invocations now autocomplete and highlight anywhere in promptbox, including several invocations in one line, and a second skill after leading skill.
• /exit was missing even though /quit existed (Hello, Codex CLI users being comfy with that command!). I found this on my own; it overlaps the closed PR feat: add /exit as an alias for /quit #1142 .
• Bare Enter could be lost in Kitty keyboard-protocol terminals--the "Shift+Enter" for newlining issue.
• Multiple skill blocks could fail to expand as separate 1st-class invocations.
• Child agents could not use an explicitly selected thinking level (Codex CLI has same issue they haven't fixed in eons now)
• Goal continuations needed explicit pause, and objective-update controls.

---

Multi-skill hover info:

[image #x] highlighting works, and pasting them is really fast:

/exit alias: