From 6cb2f04f63dc0f1599d2a8c17db78f26dc40f5f6 Mon Sep 17 00:00:00 2001 From: dancingteeth Date: Wed, 21 Oct 2026 19:24:42 +0400 Subject: [PATCH] chore(release): 0.8.3 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - package.json 0.8.3 (lockfile has no root version; unchanged). - Version labels kept in sync: README current line, CONSUMERS.md, docs/embed-api.md, llms.txt, site/ pages + site/llms.txt, and the site version test. - CHANGELOG: Unreleased -> 0.8.3 — 2026-10-21 with PR links (#44 #45 #46 #47); adds the parked-runs-exit-2 and deliverables-confined entries. --- CHANGELOG.md | 34 ++++++++++++++------------ CONSUMERS.md | 2 +- README.md | 2 +- docs/embed-api.md | 2 +- llms.txt | 4 +-- package.json | 2 +- site/about/index.html | 2 +- site/about/index.md | 2 +- site/docs/index.html | 2 +- site/docs/index.md | 2 +- site/index.md | 2 +- site/llms.txt | 4 +-- src/site/landingAgentReadiness.test.ts | 12 ++++----- 13 files changed, 38 insertions(+), 34 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 81af251..e0993ba 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -10,28 +10,32 @@ tags: Notable changes to `@dancingteeth/agent-looper`. Dates are UTC. -## Unreleased +## 0.8.3 — 2026-10-21 + +npm 0.8.2 no longer crashes without `@cursor/sdk`; verify output goes to a sidecar log by default; judge and worker faults park instead of crashing; Harness² ideas: `## Facts`, deliverables, opt-in lessons and stall probes. ### Headline -- **Behaviour change: verify sidecar by default** — `verifyLogMode` now defaults to `sidecar` (was `inline`). Full verify output is written under `/verify-logs/` (gitignore `verify-logs/` in consumer repos); the worker prompt gets a tail-biased preview plus path. Opt into `inline` to paste capture only. In-memory capture uses head+tail with a middle elision marker (not head-only truncation). -- **Billing token breakdown on green runs** — run log and `run-report.md` sum uncached input, output, and cache read/write across worker and judge sessions, with per-role rows. List vs billed cost semantics unchanged. Cursor SDK runs mark cache splits as **unknown** (not zero). -- **Judge failures park the run** — when the review gate is on and the primary judge errors, the harness classifies the fault, retries once, then tries `reviewSecondary` as fallback primary (merge secondary skipped for that verdict). If both fail, the run parks (`status: waiting`, CLI exit `2`, failure domain `review_judge_failed`) instead of hard-failing like a verify miss. -- **Expected tool-error stats** — `log.ndjson` and `run-report.md` aggregate `tool_end` errors and plain-text tool-call stalls per runtime/model for workers and judges. -- **Transient worker SDK errors** — classify worker faults (connection stalled, `status=error`, timeouts, 5xx/rate-limit vs fatal auth/config). Retry on the next iteration with session recycle and backoff; preserve the working tree; log failure context; retain partial Cursor usage when the SDK reported it. Cap consecutive transients (`maxConsecutiveTransientAgentErrors`, default 3) then park (`agent_transient_exhausted`). -- **Secondary review errors** — a failed optional `reviewSecondary` merge session is logged and noted in `review.md`; the primary verdict alone decides completion. -- **Lazy optional peers** — the CLI no longer statically imports `@cursor/sdk` (or other runtime SDKs) at startup. `agent-loop --version`, `--help`, `agent-loop-init`, and `agent-check` for a non-Cursor runtime work with only `@dancingteeth/agent-looper` installed. Selecting a runtime whose SDK is missing fails with an actionable install line (`Runtime "cursor" needs @cursor/sdk: pnpm add -D @cursor/sdk`) instead of `ERR_MODULE_NOT_FOUND`. -- **CI: npm pack smoke** — packs the tarball, installs it in a temp project with zero optional peers, and asserts `--version` / `--help` plus the friendly missing-SDK error; also checks OpenCode-only peers without `@cursor/sdk`. +- **Lazy optional peers** — the CLI no longer statically imports `@cursor/sdk` (or other runtime SDKs) at startup. `agent-loop --version`, `--help`, `agent-loop-init`, and `agent-check` for a non-Cursor runtime work with only `@dancingteeth/agent-looper` installed. Selecting a runtime whose SDK is missing fails with an actionable install line (`Runtime "cursor" needs @cursor/sdk: pnpm add -D @cursor/sdk`) instead of `ERR_MODULE_NOT_FOUND`. This fixes the 0.8.2 crash when `@cursor/sdk` is not installed. ([#44](https://github.com/dancingteeth/agent-looper/pull/44)) +- **Behaviour change: verify sidecar by default** — `verifyLogMode` now defaults to `sidecar` (was `inline`). Full verify output is written under `/verify-logs/` (gitignore `verify-logs/` in consumer repos); the worker prompt gets a tail-biased preview plus path. Opt into `inline` to paste capture only. In-memory capture uses head+tail with a middle elision marker (not head-only truncation). ([#45](https://github.com/dancingteeth/agent-looper/pull/45)) +- **Parked runs keep exit `2`** — every waiting park (budget, verify environment, judge down, transient worker cap, stall-probe restore failure) exits `2` in the CLI, the same as an incomplete run (per Paul); no separate exit code. ([#45](https://github.com/dancingteeth/agent-looper/pull/45)) +- **Judge failures park the run** — when the review gate is on and the primary judge errors, the harness classifies the fault, retries once, then tries `reviewSecondary` as fallback primary (merge secondary skipped for that verdict). If both fail, the run parks (`status: waiting`, CLI exit `2`, failure domain `review_judge_failed`) instead of hard-failing like a verify miss. ([#45](https://github.com/dancingteeth/agent-looper/pull/45)) +- **Transient worker SDK errors** — classify worker faults (connection stalled, `status=error`, timeouts, 5xx/rate-limit vs fatal auth/config). Retry on the next iteration with session recycle and backoff; preserve the working tree; log failure context; retain partial Cursor usage when the SDK reported it. Cap consecutive transients (`maxConsecutiveTransientAgentErrors`, default 3) then park (`agent_transient_exhausted`). ([#45](https://github.com/dancingteeth/agent-looper/pull/45)) +- **Billing token breakdown on green runs** — run log and `run-report.md` sum uncached input, output, and cache read/write across worker and judge sessions, with per-role rows. List vs billed cost semantics unchanged. Cursor SDK runs mark cache splits as **unknown** (not zero). ([#45](https://github.com/dancingteeth/agent-looper/pull/45)) +- **Expected tool-error stats** — `log.ndjson` and `run-report.md` aggregate `tool_end` errors and plain-text tool-call stalls per runtime/model for workers and judges. ([#45](https://github.com/dancingteeth/agent-looper/pull/45)) +- **Secondary review errors** — a failed optional `reviewSecondary` merge session is logged and noted in `review.md`; the primary verdict alone decides completion. ([#45](https://github.com/dancingteeth/agent-looper/pull/45)) +- **CI: npm pack smoke** — packs the tarball, installs it in a temp project with zero optional peers, and asserts `--version` / `--help` plus the friendly missing-SDK error; also checks OpenCode-only peers without `@cursor/sdk`. ([#44](https://github.com/dancingteeth/agent-looper/pull/44)) ### Harness² ideas -- **`stallProbes`** — opt-in (`false` by default): on the first hard stall at `escalateAfterStagnation`, run two sequential probe workers with different prompt sections, snapshot/restore the git workspace between them, keep whichever verify result goes green or moves the failure fingerprint, log `probe: { variant, kept }`, then allow model escalation on later stalls. Disabled when `stagnationThreshold` is `0`. Restore puts back the branch (a probe's `git checkout -b` / detached HEAD) and HEAD (a probe's own commits are soft-reset; an unrelated HEAD is refused), restores `.gitignore` files and `info/exclude` first, and never deletes a path the snapshot's ignore rules ignore. A failed restore parks the run (`status: waiting`, failure domain `stall_probe_restore_failed`) with the restore error and the probe's original error. +- **`## Facts` block** — the stable prompt head states the verifier command exactly as configured (not trimmed) and the deliverables, on its own line after the rules; without either, the head is byte-identical to before. ([#46](https://github.com/dancingteeth/agent-looper/pull/46)) +- **Plain-text tool-call warning + safety rule 7** — when the previous worker reply wrote tool calls as plain text instead of calling tools, the next prompt's volatile tail warns about it; safety rule 7 tells workers to use the tool interface. ([#46](https://github.com/dancingteeth/agent-looper/pull/46)) +- **`deliverables`** — optional repo-relative paths in `loop.json`; worker prompts state them in `## Facts`, and the harness nudges the worker before verify when files are missing or empty (`deliverableNudges`, default `2`). ([#46](https://github.com/dancingteeth/agent-looper/pull/46)) +- **Deliverables confined to the repo** — `deliverables` must be repo-relative: absolute paths and `..` escapes are rejected when `loop.json` is parsed, so the deliverable check and nudges never look outside the repo. Worker faults during a deliverable nudge go through the same fault path as the main worker call (transient classification, escalation, parking). ([#46](https://github.com/dancingteeth/agent-looper/pull/46)) +- **Fix: HTTP/2 `CANCEL` crash** — a Connect-node `http/2 stream closed with error code CANCEL (0x8)` rejection nobody awaited (Cursor runtime) is now treated as a detached HTTP/2 transport error instead of crashing the process. ([#46](https://github.com/dancingteeth/agent-looper/pull/46)) +- **`lessons.md`** — **opt-in** (`"lessons": true`; default `false`). After a green verifier run, the harness writes deduped verify gotchas to the loop bundle; the next run injects them before `## Workspace (git)`. Gotchas are the most informative lines of each failure overcome (failing `[verify]` step + `FAIL`/error lines, stderr first; banner, progress and summary lines skipped; at most 3 lines of 240 chars per failure, 6000 chars per file). Verifier output is stored and injected as a fenced block labelled as data, not instructions (the fence is longer than any backtick run in the output). Workers cannot edit the file (frozen). ([#47](https://github.com/dancingteeth/agent-looper/pull/47)) +- **`stallProbes`** — opt-in (`false` by default): on the first hard stall at `escalateAfterStagnation`, run two sequential probe workers with different prompt sections, snapshot/restore the git workspace between them, keep whichever verify result goes green or moves the failure fingerprint, log `probe: { variant, kept }`, then allow model escalation on later stalls. Disabled when `stagnationThreshold` is `0`. Restore puts back the branch (a probe's `git checkout -b` / detached HEAD) and HEAD (a probe's own commits are soft-reset; an unrelated HEAD is refused), restores `.gitignore` files and `info/exclude` first, and never deletes a path the snapshot's ignore rules ignore. A failed restore parks the run (`status: waiting`, failure domain `stall_probe_restore_failed`) with the restore error and the probe's original error. ([#47](https://github.com/dancingteeth/agent-looper/pull/47)) - **Snapshot limits:** ignored files (`.env`, `node_modules/`, local data) are **not** saved in the snapshot — restore leaves them as they are rather than restoring them. Not safe with concurrent writers: restore reverts *every* change made in the worktree during the probe round (your editor, another loop, another process's commits), not only the probe's. Run one loop per worktree and don't edit it while probes run. -- **`lessons.md`** — **opt-in** (`"lessons": true`; default `false`). After a green verifier run, the harness writes deduped verify gotchas to the loop bundle; the next run injects them before `## Workspace (git)`. Gotchas are the most informative lines of each failure overcome (failing `[verify]` step + `FAIL`/error lines, stderr first; banner, progress and summary lines skipped; at most 3 lines of 240 chars per failure, 6000 chars per file). Verifier output is stored and injected as a fenced block labelled as data, not instructions (the fence is longer than any backtick run in the output). Workers cannot edit the file (frozen). -- **`deliverables`** — optional repo-relative paths in `loop.json`; worker prompts state them in `## Facts`, and the harness nudges the worker before verify when files are missing or empty (`deliverableNudges`, default `2`). -- **`## Facts` block** (#46) — the stable prompt head states the verifier command exactly as configured (not trimmed) and the deliverables, on its own line after the rules; without either, the head is byte-identical to before. -- **Plain-text tool-call warning + safety rule 7** (#46) — when the previous worker reply wrote tool calls as plain text instead of calling tools, the next prompt's volatile tail warns about it; safety rule 7 tells workers to use the tool interface. -- **Fix: HTTP/2 `CANCEL` crash** (#46) — a Connect-node `http/2 stream closed with error code CANCEL (0x8)` rejection nobody awaited (Cursor runtime) is now treated as a detached HTTP/2 transport error instead of crashing the process. ## 0.8.2 — 2026-10-20 diff --git a/CONSUMERS.md b/CONSUMERS.md index 8a9a370..cf1b0ff 100644 --- a/CONSUMERS.md +++ b/CONSUMERS.md @@ -29,7 +29,7 @@ Or in `package.json`: ```json { "devDependencies": { - "@dancingteeth/agent-looper": "^0.8.2", + "@dancingteeth/agent-looper": "^0.8.3", "@cursor/sdk": "^1.0.18" } } diff --git a/README.md b/README.md index 51df831..5ca2708 100644 --- a/README.md +++ b/README.md @@ -27,7 +27,7 @@ pnpm exec agent-loop run .cursor/loops/my-task Each runtime also needs its own SDK or CLI and key, e.g. `@cursor/sdk` + `CURSOR_API_KEY` for Cursor, `claude` on PATH for Claude Code, `opencode` for OpenCode. [`docs/runtime-map.md`](./docs/runtime-map.md) lists them all; `pnpm exec agent-check ` confirms one is wired. Flags and the rest are below. You do not need them for a first green run. -How the loop is shaped: [`README.intro.md`](./README.intro.md). Technical deep dive: [`ARCHITECTURE.md`](./ARCHITECTURE.md) (including §1.1 — the harness is a small control-flow graph; the Ralph loop lives inside the worker node). npm releases: [`docs/releasing.md`](./docs/releasing.md). What shipped: [`CHANGELOG.md`](./CHANGELOG.md) (current **0.8.2**). Embedding this package as a library dependency in another ADE: [`docs/embed-api.md`](./docs/embed-api.md). Reporting a vulnerability: [`SECURITY.md`](./SECURITY.md). +How the loop is shaped: [`README.intro.md`](./README.intro.md). Technical deep dive: [`ARCHITECTURE.md`](./ARCHITECTURE.md) (including §1.1 — the harness is a small control-flow graph; the Ralph loop lives inside the worker node). npm releases: [`docs/releasing.md`](./docs/releasing.md). What shipped: [`CHANGELOG.md`](./CHANGELOG.md) (current **0.8.3**). Embedding this package as a library dependency in another ADE: [`docs/embed-api.md`](./docs/embed-api.md). Reporting a vulnerability: [`SECURITY.md`](./SECURITY.md). Supports pluggable **agent SDK** workers (`runtime`) and judges (`reviewRuntime`). Shipped today: **Cursor**, **Cline** (Pass / Credits), **OpenCode** (Go + BYOK), **Pi**, **Codex**, **DSH** (PATH `dsh`), **Muse** (PATH `muse`), **Claude** (PATH `claude`, `--safe-mode`). Defaults and cost notes: [`docs/runtime-map.md`](./docs/runtime-map.md). DSH companion for `dsh web` (skills + scaffold, not a second harness): [`docs/dsh-plugin.md`](./docs/dsh-plugin.md). OpenCode companion plugin (1.x + 2.x: skills, detached `agent_loop_start`, shell guard): [`plugins/opencode-agent-looper/`](./plugins/opencode-agent-looper/). Muse Code: [`docs/muse-runtime.md`](./docs/muse-runtime.md). Claude Code: [`docs/claude-runtime.md`](./docs/claude-runtime.md). To measure cheap-worker claims on a frozen loop: [`docs/runtime-cost-bench.md`](./docs/runtime-cost-bench.md). The judge runs on the worker's runtime when `reviewRuntime` is unset (each runtime has its own judge model, e.g. OpenCode → DeepSeek V4 Pro, Claude → Opus); set `reviewRuntime` + `reviewModel` to pair a cheap worker with a different judge. diff --git a/docs/embed-api.md b/docs/embed-api.md index d6f7e8b..af34cae 100644 --- a/docs/embed-api.md +++ b/docs/embed-api.md @@ -114,7 +114,7 @@ the whole file. ## Public API and what breaks a major -Current package version: **0.8.2**. Until 1.0, minor bumps may still add fields to `Stable` +Current package version: **0.8.3**. Until 1.0, minor bumps may still add fields to `Stable` types (additive, non-breaking for a host that ignores unknown fields) — a major bump is reserved for anything that removes or renames a field, changes a field's type, or removes an export from the [Supported surface](#supported-surface-stable-vs-experimental) table above. diff --git a/llms.txt b/llms.txt index fbb1f69..6338417 100644 --- a/llms.txt +++ b/llms.txt @@ -9,7 +9,7 @@ Start here: - README.intro.md — product spine (worker / verify / judge / human) - ARCHITECTURE.md — loop engine, escalation, cache-friendly prompts - AGENTS.md — working agreements for this repo -- CHANGELOG.md — shipped versions (current: 0.8.2) +- CHANGELOG.md — shipped versions (current: 0.8.3) - SECURITY.md — vulnerability reporting - docs/embed-api.md — library embed contract (stable vs experimental, frozen shapes) - docs/claude-runtime.md — PATH `claude -p --safe-mode` worker/judge @@ -54,4 +54,4 @@ Worker default: composer-2.5. Judge default: grok-4.6 (`reviewModel`). `runtime: ## Package -npm: `@dancingteeth/agent-looper` **0.8.2**. Layout: src/loop, src/review, src/agents, templates, .cursor/loops. Embed: docs/embed-api.md. Security: SECURITY.md. +npm: `@dancingteeth/agent-looper` **0.8.3**. Layout: src/loop, src/review, src/agents, templates, .cursor/loops. Embed: docs/embed-api.md. Security: SECURITY.md. diff --git a/package.json b/package.json index 415c5f7..72dd9db 100644 --- a/package.json +++ b/package.json @@ -1,6 +1,6 @@ { "name": "@dancingteeth/agent-looper", - "version": "0.8.2", + "version": "0.8.3", "description": "Fix-until-green harness for coding agents: frozen goal + verify.sh, a fresh worker each round, optional judge review. Works with Cursor, Claude Code, Codex, OpenCode, Cline, Pi, DSH, and Muse.", "keywords": [ "agent", diff --git a/site/about/index.html b/site/about/index.html index de30772..9657bd9 100644 --- a/site/about/index.html +++ b/site/about/index.html @@ -44,7 +44,7 @@

About Agent Looper

  • Product site
  • Source
  • npm
  • -
  • Changelog — 0.8.2
  • +
  • Changelog — 0.8.3
  • Author
  • diff --git a/site/about/index.md b/site/about/index.md index 9ee0efa..2ab3ca6 100644 --- a/site/about/index.md +++ b/site/about/index.md @@ -9,5 +9,5 @@ Agent Looper is made by Paul Zgordan (dancingteeth), who also writes at The Vacu - Product: https://looper.dancingteeth.net/ - Source: https://github.com/dancingteeth/agent-looper - npm: https://www.npmjs.com/package/@dancingteeth/agent-looper -- Changelog (0.8.2): https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md +- Changelog (0.8.3): https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md - Author: https://dancingteeth.net/ diff --git a/site/docs/index.html b/site/docs/index.html index 225b648..1977a0f 100644 --- a/site/docs/index.html +++ b/site/docs/index.html @@ -52,7 +52,7 @@

    Agent Looper developer resources

  • Claude runtime — Claude Code CLI
  • Embed API — contract for embedding Agent Looper
  • Security — reporting vulnerabilities
  • -
  • Changelog — 0.8.2
  • +
  • Changelog — 0.8.3
  • Agent Looper llms.txt
  • Agent Looper homepage (markdown)
  • diff --git a/site/docs/index.md b/site/docs/index.md index ef73225..326ecfb 100644 --- a/site/docs/index.md +++ b/site/docs/index.md @@ -16,7 +16,7 @@ Agent Looper is the npm package `@dancingteeth/agent-looper` and the CLI `agent- - Claude runtime (Claude Code CLI): https://github.com/dancingteeth/agent-looper/blob/main/docs/claude-runtime.md - Embed API (contract for embedding Agent Looper): https://github.com/dancingteeth/agent-looper/blob/main/docs/embed-api.md - Security (reporting vulnerabilities): https://github.com/dancingteeth/agent-looper/blob/main/SECURITY.md -- Changelog (0.8.2): https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md +- Changelog (0.8.3): https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md - Agent Looper llms.txt: https://looper.dancingteeth.net/llms.txt - Agent Looper homepage (markdown): https://looper.dancingteeth.net/index.md diff --git a/site/index.md b/site/index.md index f18d33a..ba9395d 100644 --- a/site/index.md +++ b/site/index.md @@ -105,7 +105,7 @@ Agent Looper is free and MIT-licensed. If it saves you time: - [llms.txt](https://looper.dancingteeth.net/llms.txt) - [Harnesses](https://looper.dancingteeth.net/harnesses/) - [Developer resources](https://looper.dancingteeth.net/docs/) -- [Changelog (0.8.2)](https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md) +- [Changelog (0.8.3)](https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md) - [Embed API](https://github.com/dancingteeth/agent-looper/blob/main/docs/embed-api.md) - [Security](https://github.com/dancingteeth/agent-looper/blob/main/SECURITY.md) - [GitHub](https://github.com/dancingteeth/agent-looper) diff --git a/site/llms.txt b/site/llms.txt index b0df42a..fefc583 100644 --- a/site/llms.txt +++ b/site/llms.txt @@ -6,7 +6,7 @@ When to use Agent Looper: a human wants a feature finished without babysitting C When not to use Agent Looper: there is no deterministic check (pure taste with no script or vision gate); you need a team-ops factory, shared org cloud, or multi-repo orchestration; you want the model to grade itself as “done.” -How an agent should call it: install `@dancingteeth/agent-looper`, ask the human what they want, then implement the feature with Agent Looper (set up the loop, write GOAL.md and verify.sh, run until green). Or from the terminal: `pnpm exec agent-loop-setup` (one-screen setup wizard — recap plus current question) then `pnpm exec agent-loop-prompt --out .cursor/loops/` — the scaffold judge stops as soon as frozen GOAL.md and verify.sh pass freeze (10-minute wall, not a long grind) — then `pnpm exec agent-loop run .cursor/loops/ --runtime `. Leave `reviewRuntime` unset and the optional judge runs on the worker's own runtime with that runtime's judge model (OpenCode → `opencode-go/deepseek-v4-pro`, Claude → `opus`, Codex → `gpt-5.6-sol`, DSH → V4 Pro, Muse → PAYG Spark, Cursor → `grok-4.7`; Cline and Pi use the worker model unless `reviewModel` is set) — there is no Cursor default judge. Set `reviewRuntime` to `cursor` to keep a Cursor judge. On `runtime: cursor`, repo defaults that pin OpenCode Go `escalateModel` are not applied. DeepSeek Flash defaults are 4.1: DSH `deepseek-official/deepseek-flash`, OpenCode Go `opencode-go/deepseek-v4.1-flash`, Cline Pass `cline-pass/deepseek-v4.1-flash`; older `deepseek-v4-flash` slugs still parse. DSH headless: `~/.dsh/.credentials.yaml` must be a flat `KEY: "string"` map — a wrapped store (`version` / `refs` / `records`) or YAML integer fails `agent-check dsh` and `agent-loop run` before spawning the worker; quoting `version` alone is not enough (the harness does not rewrite that file). Optional `--review-gate` re-opens the loop only on blocking review findings. If verify fails because the environment is broken, the loop waits instead of burning another worker. After each visit the harness restores frozen specs and removes any frozen basename a worker planted (e.g. `setup.sh`) so it cannot persist. Before tightening `verify.sh` or `REVIEWS.md` between runs, label a diverse sample of `.cursor/loop-exports/` (see docs/unknowns-preflight.md). `runtime: opencode` works with the OpenCode 1.x CLI (`@opencode-ai/sdk`) and 2.x CLI (`@opencode/client`); without the matching client package the loop fails fast with an install hint instead of timing out. `agent-check opencode` handshakes the server so a client/CLI mismatch fails the check, not the first loop. OpenCode 2.x provider keys live in OpenCode’s credential store (add once). Companion plugin `@dancingteeth/opencode-agent-looper` (0.1.6) teaches Agent Looper skills and starts loops in the background via `agent_loop_start` / `agent_loop_status` (so a shell timeout can’t kill them); `agent_loop_start` can launch through the repo’s `agent:loop` script or a `launchPrefix` when keys live in a secrets wrapper; on 2.x install with `opencode plugin add @dancingteeth/opencode-agent-looper`, on 1.x add `"plugin": ["@dancingteeth/opencode-agent-looper"]` to `opencode.json` — alongside the CLI, not instead of it. OpenCode Go and Cline Pass model lists and list prices come from models.dev (sync picks up Gemini 4 Argon (contributor) and the hosted-$0 Cat Jam row, drops deprecated DeepSeek V4 Flash slugs, and again offers priced `opencode-go/kimi-k3`); unknown well-formed slugs warn at start and read `$0` until priced — budget caps still reject unpriced worker or escalate models. CLI: `agent-loop`, `agent-loop-prompt`, `agent-loop --version`; unknown flags fail with an error. Current npm: **0.8.2** (supported line **0.8.x**). +How an agent should call it: install `@dancingteeth/agent-looper`, ask the human what they want, then implement the feature with Agent Looper (set up the loop, write GOAL.md and verify.sh, run until green). Or from the terminal: `pnpm exec agent-loop-setup` (one-screen setup wizard — recap plus current question) then `pnpm exec agent-loop-prompt --out .cursor/loops/` — the scaffold judge stops as soon as frozen GOAL.md and verify.sh pass freeze (10-minute wall, not a long grind) — then `pnpm exec agent-loop run .cursor/loops/ --runtime `. Leave `reviewRuntime` unset and the optional judge runs on the worker's own runtime with that runtime's judge model (OpenCode → `opencode-go/deepseek-v4-pro`, Claude → `opus`, Codex → `gpt-5.6-sol`, DSH → V4 Pro, Muse → PAYG Spark, Cursor → `grok-4.7`; Cline and Pi use the worker model unless `reviewModel` is set) — there is no Cursor default judge. Set `reviewRuntime` to `cursor` to keep a Cursor judge. On `runtime: cursor`, repo defaults that pin OpenCode Go `escalateModel` are not applied. DeepSeek Flash defaults are 4.1: DSH `deepseek-official/deepseek-flash`, OpenCode Go `opencode-go/deepseek-v4.1-flash`, Cline Pass `cline-pass/deepseek-v4.1-flash`; older `deepseek-v4-flash` slugs still parse. DSH headless: `~/.dsh/.credentials.yaml` must be a flat `KEY: "string"` map — a wrapped store (`version` / `refs` / `records`) or YAML integer fails `agent-check dsh` and `agent-loop run` before spawning the worker; quoting `version` alone is not enough (the harness does not rewrite that file). Optional `--review-gate` re-opens the loop only on blocking review findings. If verify fails because the environment is broken, the loop waits instead of burning another worker. After each visit the harness restores frozen specs and removes any frozen basename a worker planted (e.g. `setup.sh`) so it cannot persist. Before tightening `verify.sh` or `REVIEWS.md` between runs, label a diverse sample of `.cursor/loop-exports/` (see docs/unknowns-preflight.md). `runtime: opencode` works with the OpenCode 1.x CLI (`@opencode-ai/sdk`) and 2.x CLI (`@opencode/client`); without the matching client package the loop fails fast with an install hint instead of timing out. `agent-check opencode` handshakes the server so a client/CLI mismatch fails the check, not the first loop. OpenCode 2.x provider keys live in OpenCode’s credential store (add once). Companion plugin `@dancingteeth/opencode-agent-looper` (0.1.6) teaches Agent Looper skills and starts loops in the background via `agent_loop_start` / `agent_loop_status` (so a shell timeout can’t kill them); `agent_loop_start` can launch through the repo’s `agent:loop` script or a `launchPrefix` when keys live in a secrets wrapper; on 2.x install with `opencode plugin add @dancingteeth/opencode-agent-looper`, on 1.x add `"plugin": ["@dancingteeth/opencode-agent-looper"]` to `opencode.json` — alongside the CLI, not instead of it. OpenCode Go and Cline Pass model lists and list prices come from models.dev (sync picks up Gemini 4 Argon (contributor) and the hosted-$0 Cat Jam row, drops deprecated DeepSeek V4 Flash slugs, and again offers priced `opencode-go/kimi-k3`); unknown well-formed slugs warn at start and read `$0` until priced — budget caps still reject unpriced worker or escalate models. CLI: `agent-loop`, `agent-loop-prompt`, `agent-loop --version`; unknown flags fail with an error. Current npm: **0.8.3** (supported line **0.8.x**). ## Developer resources @@ -24,7 +24,7 @@ How an agent should call it: install `@dancingteeth/agent-looper`, ask the human - [Claude runtime](https://github.com/dancingteeth/agent-looper/blob/main/docs/claude-runtime.md): Claude Code CLI (`--safe-mode`) - [Embed API](https://github.com/dancingteeth/agent-looper/blob/main/docs/embed-api.md): contract for embedding Agent Looper - [Security](https://github.com/dancingteeth/agent-looper/blob/main/SECURITY.md): reporting vulnerabilities -- [Changelog](https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md): 0.8.2 +- [Changelog](https://github.com/dancingteeth/agent-looper/blob/main/CHANGELOG.md): 0.8.3 - [Harnesses — Agent Looper worker runtimes](https://looper.dancingteeth.net/harnesses/index.md) ## Optional diff --git a/src/site/landingAgentReadiness.test.ts b/src/site/landingAgentReadiness.test.ts index 3abd6f6..0b7b77b 100644 --- a/src/site/landingAgentReadiness.test.ts +++ b/src/site/landingAgentReadiness.test.ts @@ -85,7 +85,7 @@ describe('landing agent readiness', () => { } }) - it('site version labels say 0.8.2 current with no stale 0.7.x as current', () => { + it('site version labels say 0.8.3 current with no stale 0.7.x as current', () => { const versionPaths = [ 'about/index.html', 'about/index.md', @@ -96,14 +96,14 @@ describe('landing agent readiness', () => { ] for (const file of versionPaths) { const body = readSite(file) - expect(body, file).toContain('0.8.2') + expect(body, file).toContain('0.8.3') expect(body, file).not.toMatch(/Current npm: \*\*0\.7\./) } const llms = readSite('llms.txt') - expect(llms).toContain('Current npm: **0.8.2**') + expect(llms).toContain('Current npm: **0.8.3**') expect(llms).toContain('supported line **0.8.x**') expect(llms).not.toContain('supported line **0.7.x**') - expect(llms).toMatch(/Changelog.*0\.8\.2/) + expect(llms).toMatch(/Changelog.*0\.8\.3/) }) it('llms.txt follows llmstxt.org v2 and includes when-to-use guidance', () => { @@ -800,9 +800,9 @@ describe('landing agent readiness', () => { expect(surface).not.toMatch(/supported line \*\*0\.7\.x\*\*/i) } - expect(llms).toContain('Current npm: **0.8.2**') + expect(llms).toContain('Current npm: **0.8.3**') expect(llms).toContain('0.8.x') - expect(llms).not.toContain('Current npm: **0.8.1**') + expect(llms).not.toContain('Current npm: **0.8.2**') expect(llms).not.toContain('Current npm: **0.7.') expect(llms).toContain('@dancingteeth/opencode-agent-looper') expect(llms).toMatch(/no Cursor default judge/i)