Skip to content

aigent-os v1.2.0: Native skill matching and Caddy retirement

Latest

Choose a tag to compare

@wrg32786 wrg32786 released this 10 Nov 16:36
· 62 commits to master since this release
d1caeae

aigent-os v1.2.0: Native skill matching and Caddy retirement

Stable release. Source commit d1caeae (master). This replaces v1.1.0 as Latest. Managed Auto-Refresh remains the product; this cut retires Caddy, switches skill matching to Claude Code's native descriptions, and hardens the refresh path.

What it is

aigent-OS is a Claude Code harness that refreshes its own context instead of compacting. At a context threshold (80% by default) it asks Claude to write a capsule, waits for the exact line Capsule Complete, Ready For Clear, sends one /clear, checks through the SessionStart hook that a fresh session started, and resumes from the capsule. No ack, no clear. Memory is plain markdown in git.

Highlights since v1.1.0

  • Caddy is retired; skill matching is native (#160, #161, #162). The caddy.sh router was taking 13 to 22 s per prompt and was cancelled on most of them, so its hints never reached the agent. Claude Code now lists each installed skill's description, and every product skill has a short trigger-led description (180 characters at most) so that matching works. A new daemons/prompt-context.sh hook keeps the [CLOCK] line and starts memory capture detached, returning in milliseconds. The installer retires Caddy hooks from existing installs. #162 finishes the retirement (system-check, boot banner, capture notices, narrower hook retirement).
  • A transient file lock no longer turns automatic refresh off (#164). On Windows, a scanner or indexer can hold a file for a moment while the runner renames its state file into place. That used to throw EPERM, fail the transport tick, and turn automation off. The rename is retried, and automation turns off only after 5 failed ticks in a row.
  • context-capsule is a standalone Claude Code plugin (#103, #117, #121). /capsule, /refresh and /resume, plus SessionStart and PreCompact hooks, with zero dependencies. On a full aigent-OS install the plugin hooks stay inert so capsules are not written twice.

Other notable changes

  • Semantic search reads the seat's own notes and no longer lets one file take over the corpus (#149, issue #148). An exact title or file-name match now leads the results (#155, issue #150).
  • A busy seat no longer waits forever for a pre-capsule checkpoint, which had been holding ordinary delivery for hours (#154, issue #151).
  • Capsules record outside side effects across a refresh, and unresolved rows can survive resume across capsules (#135, #142, issues #81, #83, #134). Done history renders operator rows above claims (#131, issue #80).
  • The managed runner no longer inherits CLAUDE_CODE_CHILD_SESSION from a parent Claude Code session, which used to turn the seat's transcript off (#125, issue #98). An empty transcript is not-ready at the capsule, clear and control boundaries (issue #111).
  • fleet-delivery 0.4.x: native room_send / room_drain, a hold from capsule request through resume, and fixes for unresolved claims, pointer fences, and a submit that straddles /clear (#109, #123, #128, #133, #137, issues #105, #124, #126, #136). fleet-band can show a pending resume (#158).
  • Install stamps the source commit and doctor reports it; stale installed docs are named and never overwritten (#156, #159, issue #147). A failed dependency update can no longer destroy a working tree (#143, issue #138).
  • A bounded pre-permission interlock stands between automatic keystrokes and the permission dialog (#114). This reduces, and does not close, issue #104. Automated input acceptance remains NOT PASSED.

Full list: CHANGELOG.md and the compare link below.

Known limits

Read these before you rely on it.

  • Evidence is still single-seat. The public result remains 11 of 11 refresh cycles on one Windows seat, in one live run during v1.1.0-rc.1 testing. This is not a fleet or reliability claim, and not a claim that refresh beats compaction.
  • The managed runner can still miss a permission dialog and answer it (#104). #114 adds a bounded pre-permission interlock, but automated input acceptance remains NOT PASSED. Until that is closed, don't run the managed runner unattended on a seat that can hit permission prompts, and watch the seat during refresh cycles. Automatic input is enabled by default. You can turn it off per seat with AUTO_CLEAR_OFF=1 or the runtime/auto-clear-off marker.
  • Scrolling the seat window sends mouse escape sequences to the child as operator input (#99). They show up as stray characters, flood the error log, and count as operator input to the composer guard. Until it's fixed, avoid scrolling inside the seat terminal while a refresh cycle is in progress.
  • Done items are still unverified claims (#80). #131 improves how operator history and claims render, but a Done bullet can still hold the agent's own last message, e.g. "it's deployed". A verified tag needs a product source of command exit status (#132).
  • The Stop hook advises, it doesn't block. discipline-check.mjs flags confident "fixed / deployed / verified" claims with no check and exits 0. An opt-in blocking mode is #82.
  • The recollection benchmark has no passing result. Don't read evals/recollection as a recall-quality claim.
  • Vault sync only pushes where you approved. Memory is committed and pushed only to a destination you've set in the repo-local aigent.vaultSyncPushUrl, and it has to match every push URL git would actually use. Setup and revocation are in docs/vault-sync.md.
  • macOS has the least live mileage with the managed runner and node-pty. On Windows, install.sh needs Git Bash; PowerShell alone can't run it.

Closed since v1.1.0, so they are no longer listed as limits: a parent Claude Code session turning the seat's transcript off (#98), fleet-delivery typing into a seat during a refresh (#105), and the missing record of outside side effects across a refresh (#81, with a late-correction fixture in #83).

Install

Requires Claude Code, Node.js 18+ with npm, and bash (Git Bash on Windows).

git clone https://github.com/wrg32786/aigent-os.git
cd aigent-os
git checkout v1.2.0
bash install.sh --dry-run

The dry run prints every planned change and writes nothing. When it looks right, run bash install.sh, open a new terminal and run aigent. bash scripts/doctor.sh checks hooks, settings, search and runtime state, and --attest compares the install to the pinned baseline.

To use only the standalone plugin, from a Claude Code session: /plugin marketplace add wrg32786/aigent-os then /plugin install context-capsule@aigent-os.

Upgrade

From v1.1.0: check out v1.2.0 and rerun bash install.sh in place (dry run first). The installer retires Caddy hooks from the existing install. Files you declare in .aigent/operator-owned.json are kept byte for byte. Installed docs under docs/ are add-only: copies that differ from this source are named as stale and are not overwritten. The installer also writes .aigent/install-source.json with the source commit.

If a refresh is in flight, let it finish (or check that the latest capsule is saved) before you upgrade.

Rollback

git checkout v1.1.0
bash install.sh

Memory, capsules and vault notes are data, and the installer doesn't touch them. The safe rollback is a verified snapshot of the install and its data. After rolling back, Caddy hooks are not automatically restored; put back any hook wiring you still want, or take it from the v1.1.0 tree.

Source