Agentflow v8.4.15

• Fix #27: switch Claude to explicit manual capture of later owner submissions because UserPromptSubmit cannot reliably identify generated messages. Keep startup capture, Stop enforcement and Codex automatic capture; remove owned Claude prompt hooks on setup and silently ignore stale Claude prompt calls. Restart Claude after updating; separately update/migrate old global installations.

• Fix #28: validate established notebooks before init host detection or setup writes, refuse damaged structure with restoration guidance, and preserve existing content.

• Fix #29: restore the missing Traditional Chinese setup fallback examples with exact command parity.

• Affected behavior and terminal journeys pass. Three broader fixtures reproduce on the pre-change source; this release does not claim the entire suite is green.

Agentflow v8.4.14

• Report blocked automatic archiving in startup output and prompt-capture notices, including duplicate input and bare activation. Name the retained round and explain that its live history remains authoritative.

• Require answer recovery and verification of complete saved answers before using the existing explicit answered-round override. Keep unresolved requests and every later round live; do not treat completion or an answered marker alone as proof.

Verification: 54 archive/terminal/release checks passed; additional caller checks passed 285/286. One Stop ownership assertion also fails on the unchanged v8.4.13 source and remains unresolved. This release does not claim a fully passing repository suite.

Agentflow v8.4.13

Agentflow v8.4.13

• Keep possible tasks raised in discussion visible in the existing tracker.

• Discuss first when requested, then ask a specific confirmation question in the saved Reply. Tracking an idea does not authorize implementation.

• Preserve existing authorization and record rejected or deferred ideas so they do not keep returning. Updated both public READMEs, guides and assistant brief.

Agentflow v8.4.12

Agentflow v8.4.12 repairs message controls and clarifies their usage.

• Actual unquoted no-ag bypasses Agentflow capture and completion recovery. Use no-ag, no-ag: task, no-ag, task, or no-ag task for the current message. Quoted examples and conditional descriptions do not activate it.

• Multiline quoted pipeline controls and review waivers stay inactive.

• Removed legacy saved no-ag review waivers and cached bypass markers. Stop uses current prompt or session transcript evidence; missing evidence keeps ordinary checks. The separate Git guard remains active.

• Updated the public guides, command references, README translations, release brief and changelog.

Verification: the initial 276 affected hook/control checks and 32 alignment/release checks passed; after correcting a review finding, all 120 affected lane/collector checks passed (including two new regressions), with unaffected evidence reused; actual interactive Codex and both-host direct-hook terminal journeys passed. These results do not claim a green unrelated full repository suite.

Agentflow v8.4.11

• Fix #25 and #26: hook setup uses the running installation when neither global host script exists. Project-scoped Skills installations and Claude plugin-cache installations no longer receive missing global hook paths; existing global path priority and symlink spelling are preserved.

• Recognized stale hooks are repaired when startup or the hook installer runs after updating. Keep the installation directory available and restart the assistant when prompted.

• Enable Linux terminal checks to recognize /dev/pts devices. Verify common workflows on Ubuntu with Node 24 and 26; five pre-existing failed checks remain in four groups, so the complete suite is not green.

• Include the previously committed inline-request safeguards: recover unanswered owner requests, protect them during archiving, and keep one complete saved answer location per request.

Validation: common workflows passed on Ubuntu 24.04 with Node 24 and 26. Each full run completed with 1,433 passing checks, five pre-existing failures, and 33 skips. The remaining failures concern post-close edit expectations, old settings expectations, a worker-permission test fixture without an installed worker CLI, and the skill's size budget. This release does not have an entirely green contributor suite.

Agentflow v8.4.10

• Raise automatic notebook archiving from 500 to 750 lines. Keep the 768 KiB threshold, exact-byte verification, active-round protection and pending inline-answer protection.

• Add on-demand guidance for large mixed requests: track each requested outcome, preserve authorization and dependencies, and carry unfinished work forward without claiming that continuation has started.

• Start each numbered devlog final-report item with a verbatim quotation of the original request or question, limited to its first 30 words; quote shorter requests in full and place any omission marker outside the quotation.

• Clarify line-start examples for the existing skip-ag control; normal review, delegation and stream requirements remain in effect.

Agentflow v8.4.9

Agentflow v8.4.9 uses installed skill paths for Codex, Claude, and Git guard hooks, migrates recognized older commands while preserving unrelated settings, and keeps worktree cleanup compatible with separately installed copies. Automatic notebook archiving now triggers at 500 lines or 768 KiB, while preserving the current round and pending inline answers.

This release also includes the earlier branch updates in 8.4.7 and 8.4.8: startup settings audits, verification of completed rounds in the Stop hook, and preservation of existing directory-specific ignore rules and project skill exceptions.

Installed skill symlinks still need their targets updated if the development checkout moves. Personal project hook repairs are not part of the public package. Native Windows looper support remains unavailable.

This publishes the missing GitHub Release and tag for the existing 8.4.9 public package, fixing issue #24. The tag identifies public commit 2007860; no package files were changed for this repair. See CHANGELOG.md for the version-by-version details.

Agentflow v8.4.6

Agentflow v8.4.6 makes the devlog's short [SUMMARY] contain one concise bullet for each numbered [FINAL REPORT] item, in the same order. The English and Traditional Chinese guides and the assistant brief now describe that rule consistently. Runtime behavior is unchanged.

The release uses the existing publisher. All 18 focused release tests passed, the assembled public package matched the published checkout, and an independent reviewer passed Outcome, Minimality, and Conformance on the exact source commit before publication.

Agentflow v8.4.5

Agentflow v8.4.5 adds the optional away-gates: on|off project setting. When enabled, it supplies Design Go and Result Go after the required evidence passes. Owner-only choices, Stop decisions, and failed checks remain in place.

The release also publishes the updated assistant brief at skills/agentflow/docs/agent-brief.md and links it from both public READMEs. The release tooling copies its exact committed bytes and refuses a missing source brief before changing the public checkout.

Release tests passed (16 development cases and 2 public cases). The public package was assembled from a clean source commit, checked for version and brief-byte equality, and passed independent review after a changelog correction.

Agentflow v8.4.4

Clarify saved replies so they retain requested answers and essential results without repeating content already saved in a report. Includes the owner’s skill wording edits and updated version headings and changelog.

Release and document checks pass after a small equivalent wording compression to preserve the existing skill size limit. Independent review passed. No executable code or configuration changes.