cli-agent-runner 0.2.2__tar.gz → 0.2.3__tar.gz

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/CHANGELOG.md

@@ -5,6 +5,18 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.2.3] - 2026-07-26
+
+Details: [docs/migrations/0.2.3.md](docs/migrations/0.2.3.md)
+
+### Added
+- `--preset kimi` — a preset for [Kimi Code CLI](https://github.com/MoonshotAI/kimi-code) (`kimi -p --output-format stream-json`). `docs/recipes/kimi.md` also documents running Kimi K3 through the existing `claude` preset via Moonshot's Anthropic-compatible endpoint.
+- `--preset pi` — a preset for [Pi Coding Agent](https://github.com/earendil-works/pi) (`pi -p -na --mode json --model …`; `-na` pins project trust off for unattended runs). `docs/recipes/pi.md` documents driving Kimi K3 through pi via a Moonshot OpenAI-compatible provider.
+
+### Fixed
+- The agent subprocess now runs in `runtime.work_dir` (`cwd=` on spawn). Previously it inherited the supervisor's cwd and only launch conventions (systemd `WorkingDirectory=`, relative `--config`) kept the two aligned — fatal for agent CLIs with no working-directory flag of their own (e.g. `pi`). The startup check now also validates a relative `agent.command[0]` against `work_dir`, matching where it executes.
+- Plugin round-log tail window widened 50 → 200 lines: a stderr burst after the agent's terminal JSONL event could evict it from the scan window, silently dropping usage/transient classification. The round log's merged stdout+stderr contract is now documented (`HookContext.agent_log_path`, `docs/long-running-agents.md`) — the merge is deliberate; auth/network detection reads stderr text from it.
+
 ## [0.2.2] - 2026-07-18
 
 ### Removed
@@ -386,7 +398,7 @@ See `docs/migrations/0.1.17.md`.
 
 ### Acknowledgements
 
-Argus Gateway's post-Q1-audit feedback (2026-05-14) surfaced the per-phase sub-table need (S3) and the multi-file prompt pattern (S1). Their R721 frontmatter-strip lesson informed making the (already-shipping) strip behavior an explicit opt-out config flag rather than hardcoded.
+A downstream integrator's post-Q1-audit feedback (2026-05-14) surfaced the per-phase sub-table need (S3) and the multi-file prompt pattern (S1). Their R721 frontmatter-strip lesson informed making the (already-shipping) strip behavior an explicit opt-out config flag rather than hardcoded.
 
 ## [0.1.15] - 2026-05-14
 
@@ -418,10 +430,10 @@ independent components bundled into one release because both serve the same
 
 ### Acknowledgements
 
-Two of six nice-to-have items surfaced in Argus Gateway's v0.1.12
+Two of six nice-to-have items surfaced in a downstream integrator's v0.1.12
 production-evaluation report (2026-05-14). An earlier scope also included
 per-phase runtime override; spec review caught that `runtime.round_timeout_per_phase`
-already covers Argus's stated need, so that component was deferred to a future
+already covers the integrator's stated need, so that component was deferred to a future
 release where a second per-phase field surfaces. Other items in the report
 (detector helper, hot-reload, replay) remain intentionally out of scope.
 
@@ -440,7 +452,7 @@ release where a second per-phase field surfaces. Other items in the report
 
 ### Acknowledgements
 
-Thanks to the Argus Gateway team — this release answers their imminent
+Thanks to our downstream integrators — this release answers their imminent
 production-deployment requirement for upgrade-without-disruption. After
 confirming round duration (10-40 min) fits within graceful-stop tolerance,
 this release scopes to round-boundary upgrade UX (Level 1). Mid-round
@@ -465,7 +477,7 @@ real production need.
 
 ### Acknowledgements
 
-Thanks to the Argus Gateway team for the deep v0.1.10 audit-session feedback
+Thanks to our downstream integrators for the deep v0.1.10 audit-session feedback
 (6 items, 3-round real-run testing on ARMv8 Pi). This release reframes those
 items into a coherent "Plugin & Operator Transparency" theme across three
 layers: transparency (see what plugins do), operator override (escape hatches
@@ -487,16 +499,16 @@ for audit/debug), and diagnostic quality (errors point at the next debug step).
 
 ### Migration notes
 
-- `Config.plugins` type change is breaking for any caller reading the field as a dict (`cfg.plugins.get("foo")`). Plugin authors using `[plugins.argus_*]`-style keys: read them from `cfg.plugins.raw.get("argus_*")` instead.
+- `Config.plugins` type change is breaking for any caller reading the field as a dict (`cfg.plugins.get("foo")`). Plugin authors using `[plugins.acme_*]`-style keys: read them from `cfg.plugins.raw.get("acme_*")` instead.
 - `LockHeldError` message format changed (now includes holder info: `"another agent-runner is holding /path (held by PID N, age Ns, cmd: ...)"` or stale/missing variants). Operators grepping the exact format string need to update.
-- For Argus's P5 confusion: see new `docs/architecture.md` section "Plugin injection: two paths" — `inject_context` and `disable_pre_round_hooks` are INDEPENDENT flags. Setting one does not affect the other.
+- For the integrator's P5 confusion: see new `docs/architecture.md` section "Plugin injection: two paths" — `inject_context` and `disable_pre_round_hooks` are INDEPENDENT flags. Setting one does not affect the other.
 - **Known limitation**: `[plugins] disable` removes named plugins from the hook / context-enricher / detector / event-kind registries, but does NOT remove a disabled plugin's owned VCS paths (the `register_plugin_owned_paths` registry has no name attribution today). Mostly inert. If this becomes a real issue, file a GitHub issue.
 
 ## [0.1.11] - 2026-05-13
 
 ### Acknowledgements
 
-Thanks again to the Argus Gateway team — this release closes the
+Thanks again to our downstream integrators — this release closes the
 network-resilience gap that 0.1.10's `MonitorRemoteError` propagation
 exposed, plus adds per-occurrence agent network blip observability requested
 during 0.1.10 handover review.
@@ -525,7 +537,7 @@ For network-related failures, the events index points at the diagnostic body:
 
 ### Acknowledgements
 
-Thanks to the Argus Gateway team for the Phase 4 second-pass production feedback that drove every change in this release. Six audit memos across 50 minutes of validated runtime surfaced four specific gaps; this release closes them.
+Thanks to our downstream integrators for the Phase 4 second-pass production feedback that drove every change in this release. Six audit memos across 50 minutes of validated runtime surfaced four specific gaps; this release closes them.
 
 ### Added
 
@@ -553,7 +565,7 @@ Thanks to the Argus Gateway team for the Phase 4 second-pass production feedback
 
 ### Acknowledgements
 
-Thanks to the argus-gateway team for the dev/qa/product wall-time data
+Thanks to our downstream integrators for the dev/qa/product wall-time data
 (Phase 4 feedback §3.1) that drove this API shape. Their three-role
 distribution made the case for per-phase overrides concrete.
 
@@ -579,7 +591,7 @@ documented plugin-author public surface.
 
 ### Acknowledgements
 
-Thanks to the argus-gateway team for Phase 4 dogfooding feedback that drove
+Thanks to our downstream integrators for Phase 4 dogfooding feedback that drove
 every item in this release. 3 audit memos (~90KB) silently swept into an
 orphan stash is a real-world failure mode; this release closes that loop.
 
@@ -608,7 +620,7 @@ orphan stash is a real-world failure mode; this release closes that loop.
 
 - Plugin outputs in plugin-declared paths (e.g. `proposals/`,
   `logs/plugins/my_plugin/`) no longer silently swept into orphan stashes
-  by `process_orphan_wip`. Previously: 90KB Argus audit memos invisible
+  by `process_orphan_wip`. Previously: 90KB audit memos invisible
   after Phase 4 round; required stash archaeology to recover.
 
 ### Migration
@@ -651,7 +663,7 @@ Or regenerate cleanly:
 agent-runner init --preset claude --force
 ```
 
-Plugin authors (Argus Gateway, etc.): no public API was renamed or removed
+Plugin authors (downstream integrators, etc.): no public API was renamed or removed
 from your import surface. The deleted symbols (`agent_runner.agent_runtime.CRITICAL_ENV_DEFAULTS`,
 `agent_runner.agent_runtime.merge_critical_envs`) were internal — not part of
 the documented plugin API. A new public-API contract test
@@ -882,7 +894,7 @@ Initial public release on PyPI as `cli-agent-runner`.
 
 ### Added
 - Three-layer model: Round / Loop / Witness.
-- 13 CLI verbs: `init`, `install`, `uninstall`, `start`, `stop`, `kill`,
+- 14 CLI verbs: `init`, `install`, `uninstall`, `start`, `stop`, `kill`,
   `cancel`, `restart`, `status`, `round`, `serve`, `peek`, `watch`, `monitor`.
 - 11 named defenses (round timeout, process group isolation, orphan stash
   with SHA lock, set-diff classification, smoke check, flock concurrency,
@@ -909,7 +921,10 @@ Initial public release on PyPI as `cli-agent-runner`.
 - Tag-triggered release publishing to PyPI via Trusted Publishing OIDC,
   gated by a manual approval on the `pypi` GitHub environment.
 
-[Unreleased]: https://github.com/wan9yu/cli-agent-runner/compare/v0.2.0...HEAD
+[Unreleased]: https://github.com/wan9yu/cli-agent-runner/compare/v0.2.3...HEAD
+[0.2.3]: https://github.com/wan9yu/cli-agent-runner/compare/v0.2.2...v0.2.3
+[0.2.2]: https://github.com/wan9yu/cli-agent-runner/compare/v0.2.1...v0.2.2
+[0.2.1]: https://github.com/wan9yu/cli-agent-runner/compare/v0.2.0...v0.2.1
 [0.2.0]: https://github.com/wan9yu/cli-agent-runner/compare/v0.1.42...v0.2.0
 [0.1.26]: https://github.com/wan9yu/cli-agent-runner/compare/v0.1.25...v0.1.26
 [0.1.25]: https://github.com/wan9yu/cli-agent-runner/compare/v0.1.24...v0.1.25

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cli-agent-runner
-Version: 0.2.2
+Version: 0.2.3
 Summary: Restart-on-exit supervisor for autonomous CLI agents
 Project-URL: Homepage, https://github.com/wan9yu/cli-agent-runner
 Project-URL: Documentation, https://github.com/wan9yu/cli-agent-runner#readme
@@ -41,7 +41,7 @@ Description-Content-Type: text/markdown
 # agent-runner
 
 A restart-on-exit supervisor for autonomous coding CLIs. Ships presets for
-Claude Code, aider, gemini and codewhale; any prompt-arg CLI via custom
+Claude Code, aider, gemini, codewhale, kimi and pi; any prompt-arg CLI via custom
 config. Spawn the agent round-after-round under defenses that prevent the
 failure modes
 that bite in production: stuck rounds, orphan commits, OAuth burn loops,

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/README.md

@@ -5,7 +5,7 @@
 # agent-runner
 
 A restart-on-exit supervisor for autonomous coding CLIs. Ships presets for
-Claude Code, aider, gemini and codewhale; any prompt-arg CLI via custom
+Claude Code, aider, gemini, codewhale, kimi and pi; any prompt-arg CLI via custom
 config. Spawn the agent round-after-round under defenses that prevent the
 failure modes
 that bite in production: stuck rounds, orphan commits, OAuth burn loops,

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/agent_runner/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.2.2'
-__version_tuple__ = version_tuple = (0, 2, 2)
+__version__ = version = '0.2.3'
+__version_tuple__ = version_tuple = (0, 2, 3)
 
 __commit_id__ = commit_id = None

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/agent_runner/agent_runtime.py

@@ -137,6 +137,7 @@ def run(
     prompt: str,
     prompt_delivery: str = "argv",
     timeout_s: int,
+    work_dir: Path,
     log_path: Path,
     env_extra: dict[str, str],
     max_grace_after_result_s: int = 0,
@@ -149,6 +150,12 @@ def run(
 
     Wall-clock timeout (R1128). On timeout: SIGTERM pgroup → REAP_GRACE_S → SIGKILL.
 
+    work_dir: the agent child's working directory (required). Before 0.2.3 the
+    child inherited the supervisor's cwd and only launch conventions (systemd
+    WorkingDirectory=, relative --config) kept the two aligned — fatal for CLIs
+    with no --cwd flag of their own (e.g. pi). Callers pass the already-absolute
+    cfg.runtime.work_dir.
+
     max_grace_after_result_s: when > 0, start a countdown after the first
     type=result event is detected in the log. After it elapses, reap the
     process group only if the agent has no live worker processes left (a
@@ -175,16 +182,20 @@ def run(
         if stdin_mode
         else _build_argv(command, prompt_arg_template, prompt)
     )
-    env = {**os.environ, **env_extra}
+    # PWD pinned alongside cwd= so shell-reported paths agree; env_extra wins.
+    env = {**os.environ, "PWD": str(work_dir), **env_extra}
     log_path.parent.mkdir(parents=True, exist_ok=True)
     log_file = log_path.open("w", encoding="utf-8")
     start = time.time()
     last_progress_at = start
     proc = subprocess.Popen(
         argv,
+        cwd=work_dir,
         env=env,
         stdin=subprocess.PIPE if stdin_mode else subprocess.DEVNULL,
         stdout=log_file,
+        # Merged on purpose: oauth_fail / network_fail / network-blip detection
+        # regex-scan stderr text out of this log (see hooks.agent_log_path).
         stderr=subprocess.STDOUT,
         start_new_session=True,
     )

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/agent_runner/builtin_plugins/_constants.py

@@ -6,8 +6,13 @@ same window size, raw-text caps, and transient-error back-off defaults.
 
 from __future__ import annotations
 
-_TAIL_LINES: int = 50
-"""Number of log lines to scan from the end of round-N.log."""
+_TAIL_LINES: int = 200
+"""Number of log lines to scan from the end of round-N.log.
+
+200, not 50: the round log is merged stdout+stderr, and a CLI that bursts
+stderr text after its terminal JSONL event (observed with stderr-chatty
+agents like pi) could evict that event from a 50-line window — silently
+dropping usage/transient classification for the round."""
 
 _RAW_CAP: int = 200
 """Maximum length for ``raw`` field in transient_error_detected payload."""

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/agent_runner/hooks.py

@@ -75,16 +75,20 @@ class HookContext:
     """Basename of ``cfg.agent.command[0]`` (e.g. ``'claude'``, ``'gemini'``).
     Distinct from ``agent_name`` (the user-cosmetic ``[agent] name`` field).
     Plugins should guard on ``agent_binary``, not ``agent_name``, so a custom
-    ``[agent] name = "argus_dev"`` does not suppress events.
+    ``[agent] name = "acme_dev"`` does not suppress events.
     Populated from 0.1.30+; defaults to ``None`` for manually-constructed
     HookContext instances (rare; tests set it via ``make_hook_context``).
     """
     agent_log_path: Path | None = None
-    """Path to the agent's stdout JSONL log for this round (e.g.
-    ``log_dir/rounds/R<N>-<timestamp>.log``). Plugins parse this to
-    extract usage / classify errors. Default ``None`` for backward
-    compatibility with manually-constructed HookContext instances
-    (rare; the supervisor always populates this from 0.1.25+).
+    """Path to the agent's round log (e.g.
+    ``log_dir/rounds/R<N>-<timestamp>.log``). Contract: the file is the
+    agent's MERGED stdout+stderr — the merge is deliberate (oauth_fail /
+    network_fail detection regex-scans stderr text out of it), so JSONL
+    consumers must skip non-JSON lines (per-line ``json.loads`` in
+    try/except). Plugins parse this to extract usage / classify errors.
+    Default ``None`` for backward compatibility with manually-constructed
+    HookContext instances (rare; the supervisor always populates this
+    from 0.1.25+).
     """
     dry_run: bool = False
     """When True, plugins should skip side-effect actions (e.g. git ops,

cli_agent_runner-0.2.3/agent_runner/presets/kimi.toml

@@ -0,0 +1,38 @@
+# agent-runner.toml — generated by `agent-runner init --preset kimi`.
+#
+# Prereqs (full setup, exact auth values, and the claude-via-Kimi-K3
+# alternative: docs/recipes/kimi.md):
+#   - Kimi Code CLI on PATH: `brew install kimi-code` (or `npm i -g @moonshot-ai/kimi-code`).
+#   - Host authenticated with a model configured. Cleanest unattended path: set the
+#     KIMI_MODEL_* env vars (a self-contained model spec — no interactive login, no
+#     config.toml); a raw KIMI_API_KEY alone is NOT enough. Or `kimi login` once.
+#   - work_dir is a git repo.
+#
+# `kimi -p` is headless and auto-approves tool calls itself — it REJECTS
+# --yolo / --auto, so the command below must not add them.
+
+[agent]
+command = ["kimi", "--output-format", "stream-json"]
+prompt_arg_template = ["-p", "{prompt}"]
+name = "kimi"
+
+[agent.env]
+KIMI_CODE_NO_AUTO_UPDATE = "1"
+KIMI_DISABLE_TELEMETRY = "1"
+
+[runtime]
+work_dir = "."
+log_dir = "~/.agent-runner/{project}/logs"
+round_timeout_s = 1800
+restart_delay_s = 3
+
+[prompt]
+file = "./prompts/main.md"
+inject_context = true
+
+[vcs]
+dirty_action = "stash"
+stash_idempotency_s = 5
+
+[monitor]
+auth_fail_hint = "Configure Kimi model access on the supervisor host — set the KIMI_MODEL_* env vars (see docs/recipes/kimi.md) or run `kimi login`; `kimi doctor` validates config."

cli_agent_runner-0.2.3/agent_runner/presets/pi.toml

@@ -0,0 +1,42 @@
+# agent-runner.toml — generated by `agent-runner init --preset pi`.
+#
+# Prereqs (full setup + provider auth: docs/recipes/pi.md):
+#   - Pi Coding Agent on PATH: `npm i -g --ignore-scripts @earendil-works/pi-coding-agent`
+#     (or the install script at https://pi.dev). Binary is `pi`.
+#   - A provider + model configured. pi has NO universal default model, so you MUST
+#     set --model below. Our setup: Kimi K3 via a Moonshot provider in
+#     ~/.pi/agent/models.json — see docs/recipes/pi.md — then --model moonshot/kimi-k3.
+#   - work_dir is a git repo. agent-runner starts the agent in it (pi itself
+#     has no --cwd flag).
+#
+# pi runs tools unconfined (full YOLO, no confirmation) and agent-runner does not
+# sandbox it — only point it at a repo/host where unattended automated shell is OK.
+
+[agent]
+# Replace PROVIDER/MODEL — pi has no default. Our default: moonshot/kimi-k3.
+# -na pins project trust off: project-local pi resources (.pi/ settings,
+# extensions, skills, SYSTEM.md; project .agents/skills) never load, instead
+# of depending on the host's saved trust.json decisions or defaultProjectTrust.
+command = ["pi", "-p", "-na", "--mode", "json", "--model", "PROVIDER/MODEL"]
+prompt_arg_template = ["{prompt}"]
+name = "pi"
+
+[agent.env]
+PI_OFFLINE = "1"
+
+[runtime]
+work_dir = "."
+log_dir = "~/.agent-runner/{project}/logs"
+round_timeout_s = 1800
+restart_delay_s = 3
+
+[prompt]
+file = "./prompts/main.md"
+inject_context = true
+
+[vcs]
+dirty_action = "stash"
+stash_idempotency_s = 5
+
+[monitor]
+auth_fail_hint = "Configure a pi provider on the supervisor host — add it to ~/.pi/agent/models.json (see docs/recipes/pi.md) or run `pi /login`, and set --model to a reachable provider/model; `pi -p \"ok\"` validates it."

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/agent_runner/runner.py

@@ -507,6 +507,7 @@ def _run_one_round_inner(cfg: Config, *, phase_override: str | None = None) -> R
         prompt=prompt,
         prompt_delivery=cfg.agent.prompt_delivery,
         timeout_s=timeout_s,
+        work_dir=cfg.runtime.work_dir,
         log_path=log_path,
         env_extra={**framework_env, **dict(cfg.agent.env)},
         max_grace_after_result_s=cfg.runtime.max_grace_after_result_s,

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/agent_runner/startup_check.py

@@ -8,6 +8,7 @@ import os
 import shutil
 from collections.abc import Callable
 from dataclasses import dataclass
+from pathlib import Path
 
 from agent_runner.config import Config
 
@@ -45,6 +46,21 @@ def _check_agent_cli(cfg: Config) -> CheckResult:
     if not cfg.agent.command:
         return CheckResult("agent_cli_in_path", False, "agent.command is empty")
     cli = cfg.agent.command[0]
+    if "/" in cli:
+        # Slash-containing commands are exec'd relative to the agent's cwd
+        # (= runtime.work_dir since the cwd= spawn fix), not PATH — validate
+        # against the same base the exec will use, not the supervisor's cwd.
+        candidate = Path(cli)
+        if not candidate.is_absolute():
+            candidate = cfg.runtime.work_dir / candidate
+        if not (candidate.is_file() and os.access(candidate, os.X_OK)):
+            return CheckResult(
+                "agent_cli_in_path",
+                False,
+                reason=f"{cli!r} not found or not executable under {cfg.runtime.work_dir}",
+                how_to_fix="fix the path relative to runtime.work_dir, or use an absolute path",
+            )
+        return CheckResult("agent_cli_in_path", True)
     if shutil.which(cli) is None:
         return CheckResult(
             "agent_cli_in_path",

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/README.md

@@ -1,7 +1,7 @@
 # agent-runner — Documentation
 
 A restart-on-exit supervisor for autonomous coding CLIs (Claude Code, aider,
-gemini, codewhale, or any prompt-arg CLI). Each round runs the agent once and
+gemini, codewhale, kimi, pi, or any prompt-arg CLI). Each round runs the agent once and
 exits; an external
 service manager (systemd / launchd / bash loop) restarts. State persists
 across restarts via JSON files; defenses (timeout, process-group reap,
@@ -23,10 +23,12 @@ modes.
 | [long-running-agents.md](long-running-agents.md) | Context rot, fresh eyes, long-lineage runs |
 | [recipes/aider.md](recipes/aider.md) | aider walkthrough |
 | [recipes/codewhale.md](recipes/codewhale.md) | codewhale walkthrough |
+| [recipes/kimi.md](recipes/kimi.md) | Kimi Code CLI + running Kimi K3 via the claude preset |
+| [recipes/pi.md](recipes/pi.md) | Pi Coding Agent + Kimi K3 via a Moonshot provider |
 
 ## 中文导读
 
-agent-runner 是一个面向自主编码 CLI（Claude Code、aider、gemini、codewhale，
+agent-runner 是一个面向自主编码 CLI（Claude Code、aider、gemini、codewhale、kimi、pi，
 或任何 prompt-arg CLI）的「跑完即退、自动重启」调度框架。
 进程退出后由外部服务管理器（systemd / launchd / bash loop）立即拉起新一轮，
 状态文件持久化在 JSON 中，关键防御（超时、进程组回收、孤儿 stash、启动 smoke

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/architecture.md

@@ -15,8 +15,8 @@
 Each layer can run without the layer above. The Witness can watch a remote
 Loop.
 
-**Provider-agnostic by design.** Four presets ship (`claude` — the default —
-plus `aider`, `gemini` and `codewhale`) because those are what we run in
+**Provider-agnostic by design.** 6 presets ship (`claude` — the default —
+plus `aider`, `gemini`, `codewhale`, `kimi` and `pi`) because those are what we run in
 production, but the supervisor's defenses, observability, and lifecycle make
 no CLI-specific assumptions in core. Set `[agent].command` to any prompt-arg
 CLI and the same Round / Loop / Witness layers apply.
@@ -97,7 +97,7 @@ API quota / writing to a near-full disk).
 
 ## Monitor: anomaly-only by design
 
-The monitor emits no events during healthy operation — it surfaces alerts only when a detector fires. To verify the monitor process is running, look for the `monitor_started` event in `events-*.jsonl`. Programmatic consumers (e.g. supervisory layers like Argus Gateway) should subscribe to that event kind as the canonical "supervision is up" signal. The event carries `mode: "anomaly-only"` to document the intentional silence.
+The monitor emits no events during healthy operation — it surfaces alerts only when a detector fires. To verify the monitor process is running, look for the `monitor_started` event in `events-*.jsonl`. Programmatic consumers (e.g. an external supervisory layer) should subscribe to that event kind as the canonical "supervision is up" signal. The event carries `mode: "anomaly-only"` to document the intentional silence.
 
 ## Monitor: transient ssh tolerance
 

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/commands.md

@@ -35,7 +35,7 @@ appends `logs/` to `.gitignore`. By default also creates a git commit.
 
 Flags:
 
-- `--preset {claude,aider,gemini,codewhale}` — agent CLI preset to scaffold (default: `claude`)
+- `--preset {claude,aider,gemini,codewhale,kimi,pi}` — agent CLI preset to scaffold (default: `claude`)
 - `--force` — overwrite an existing `agent-runner.toml`
 - `--no-commit` — skip the initial git commit
 

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/configuration.md

@@ -161,6 +161,11 @@ guidance without authoring it themselves:
   credentials, or set `DEEPSEEK_API_KEY` on the supervisor host.
 - `--preset gemini` → verify your API key or check your authentication status
   for Gemini CLI.
+- `--preset kimi` → set the `KIMI_MODEL_*` env vars (see `recipes/kimi.md`), or
+  run `kimi login`; `kimi doctor` validates config.
+- `--preset pi` → configure a pi provider in `~/.pi/agent/models.json` (see
+  `recipes/pi.md`) or run `pi /login`, and set `--model` to a reachable
+  provider/model.
 
 Override in your `agent-runner.toml` if you ship a custom CLI.
 
@@ -176,7 +181,7 @@ concat_separator = "\n\n"        # default; use "\n\n---\n\n" for visible breaks
 strip_yaml_frontmatter = true    # default; set false for non-LLM-CLI agents
 ```
 
-- **Missing `files[0]`** → `ConfigError` (fail-fast; the first file is required).
+- **Missing `files[0]`** → `FileNotFoundError` when the round assembles the prompt (the first file is required; the existence check happens at prompt assembly, not at config load).
 - **Missing `files[n≥1]`** → warning logged, file skipped (supports optional preamble pattern).
 - **Both `prompt.file` and `prompt.files` set** → `ConfigError`.
 - **`prompt.file = "x.md"` shorthand** — single-file back-compat, still works unchanged.

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/long-running-agents.md

@@ -227,10 +227,12 @@ migration recipes.
 ### Reading agent stdout from a plugin
 
 Use `ctx.agent_log_path` (added in 0.1.25). This points to the agent's
-actual JSONL stdout for the current round
-(`log_dir/rounds/R<N>-<timestamp>.log`). Do NOT compute the path from
-`ctx.log_dir + round_num` — historical naming conventions in that directory
-are subject to change.
+round log (`log_dir/rounds/R<N>-<timestamp>.log`) — the agent's **merged
+stdout+stderr**, merged deliberately so auth/network error text emitted on
+stderr stays detectable. Parse it as JSONL that may contain non-JSON lines
+(per-line `json.loads` in try/except, as the built-in plugins do). Do NOT
+compute the path from `ctx.log_dir + round_num` — historical naming
+conventions in that directory are subject to change.
 
 ```python
 def after_round(self, ctx: HookContext, result: Any) -> None:

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/migrations/0.1.20.md

@@ -1,5 +1,15 @@
 # 0.1.20 migration notes
 
+> **Forward note (added 2026-07-18 for 0.2.2)**: the rate-limit surface this page
+> introduces was later replaced — read this page as history, not current behavior:
+> - The three events (`rate_limit_rejected`, `rate_limit_recovered`,
+>   `rate_limit_backoff_capped`) were **removed in 0.1.29** and are no longer
+>   registered kinds, so a subscriber never fires. `transient_error_*` is the
+>   current emission path.
+> - `[runtime] rate_limit_action` was **removed in 0.1.29** and now hard-fails
+>   config load — use `runtime.transient_error_action`.
+> - Events are written to monthly `events-YYYY-MM.jsonl`, not `events.jsonl`.
+
 ## What changed
 
 ### 1. Built-in claude rate-limit detection

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/migrations/0.1.23.md

@@ -35,7 +35,7 @@ transient API errors via a unified classifier:
 Same supervisor mechanism (sleep until `reset_at_epoch + jitter`, 8h cap).
 All 4 classifications honor `[runtime] transient_error_action`.
 
-Motivation: Argus Gateway reported a 15-min claude API 500 outage that caused
+Motivation: a downstream integrator reported a 15-min claude API 500 outage that caused
 16 wasted empty rounds before the operator noticed and intervened. The existing
 rate-limit machinery already had the right back-off design; extending it to
 cover 5xx/timeout was the natural generalization.

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/migrations/0.1.25.md

@@ -86,7 +86,7 @@ should start producing events on the next round.
 - **Auto-detect plugins using old path + warn** — rejected: too magic;
   explicit migration recipe is cleaner
 
-## What argus + other 3rd-party users should do
+## What downstream + other third-party users should do
 
 1. Update to agent-runner 0.1.25
 2. Audit any custom `post_round_hook` plugins for `ctx.log_dir /

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/migrations/0.1.29.md

@@ -70,7 +70,7 @@ from agent_runner.api import emit_transient_error_backoff_capped
 
 The alias surface was introduced in 0.1.20 (`rate_limit_rejected`), generalized
 in 0.1.23 (`transient_error_detected` as canonical), and dual-emitting since.
-Gateway consumers confirmed test-mode integration (not production workloads);
+Downstream consumers confirmed test-mode integration (not production workloads);
 cutting the aliases now avoids carrying back-compat dual-emission indefinitely.
 
 ## Verification recipe

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/migrations/0.1.30.md

@@ -22,7 +22,7 @@ If your config has a custom `[agent] name`:
 ```toml
 [agent]
 command = ["claude", "--some-flag"]
-name = "argus_dev"   # custom cosmetic name
+name = "acme_dev"   # custom cosmetic name
 ```
 
 Previously this silenced all `agent_usage_recorded` and

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/migrations/0.1.35.md

@@ -59,7 +59,7 @@ since 0.1.20.
 ## Why hard-cut (no deprecation cycle)
 
 The 0.1.23 rename is 12 releases / ~6 days old (as measured by the rapid
-0.1.25-0.1.34 ship cadence). The Argus Gateway team — our only known
+0.1.25-0.1.34 ship cadence). A downstream integrator — our only known
 production consumer — explicitly migrated away from the old name during
 their 2026-05-19/20 monitoring overhaul. Carrying an unused back-compat
 entry-point line indefinitely is debt; per project policy

cli_agent_runner-0.2.3/docs/migrations/0.2.3.md

@@ -0,0 +1,47 @@
+# Migration guide: 0.2.3
+
+## TL;DR
+
+No public symbols removed, no TOML changes required. Two additive presets
+(`kimi`, `pi`) and two behavior fixes. Upgrade freely; read on only if you
+launch `serve` from outside `work_dir` or ship plugins that parse round logs.
+
+---
+
+## Fixed: the agent subprocess now runs in `runtime.work_dir`
+
+Before 0.2.3 the agent child inherited the **supervisor's** cwd; only launch
+conventions (the systemd unit's `WorkingDirectory=`, the default relative
+`./agent-runner.toml`) kept the two aligned. Agent CLIs with no working-
+directory flag of their own (e.g. `pi`) had no recourse when the convention
+was violated.
+
+From 0.2.3 the child is spawned with `cwd=work_dir` (and `PWD` pinned to
+match). If you relied on the agent running somewhere other than `work_dir`,
+that setup was already misbehaving — the startup battery usually failed it —
+and now runs where the config says.
+
+Related: a relative, slash-containing `agent.command[0]` (e.g.
+`./run-agent.sh`) is now validated by the startup check against `work_dir` —
+the same base the exec uses — instead of the supervisor's cwd.
+
+## Changed: plugin round-log tail window 50 → 200 lines
+
+The round log is the agent's **merged stdout+stderr** (deliberate: auth- and
+network-error detection regex-scan stderr text out of it). A CLI bursting >50
+stderr lines after its terminal JSONL event could evict that event from the
+50-line scan window, silently dropping usage / transient-error classification
+for the round. The window is now 200 lines. No action needed; the merged-log
+contract is documented at `HookContext.agent_log_path` and
+`docs/long-running-agents.md` — plugin parsers must skip non-JSON lines.
+
+## Added: `--preset kimi` and `--preset pi`
+
+- `kimi` — Kimi Code CLI (`kimi -p --output-format stream-json`); auth via
+  `KIMI_MODEL_*` env vars. See `docs/recipes/kimi.md`.
+- `pi` — Pi Coding Agent (`pi -p -na --mode json --model …`). `-na` pins pi's
+  project trust **off** for unattended rounds: without it, a saved trust
+  decision for `work_dir` or any parent silently loads repo-local `.pi/`
+  resources, including a `SYSTEM.md` that replaces the system prompt. pi has
+  no turn cap or timeout of its own — agent-runner's `round_timeout_s` is the
+  only brake. See `docs/recipes/pi.md`.

{cli_agent_runner-0.2.2 → cli_agent_runner-0.2.3}/docs/plugins.md

@@ -448,7 +448,7 @@ detector. Other plugin detectors and all builtins still run normally.
 
 This example shows the full pattern for a project-specific monitor detector
 that filters out rounds the plugin marks as exempt by some project rule.
-Covers gateway-class needs like "stuck role detection" (count git commits per
+Covers supervisory needs like "stuck role detection" (count git commits per
 round) and "wall-time trend" (compare recent avg vs older avg).
 
 **The shape**: a plugin emits a custom event for exempt rounds; a custom
@@ -684,7 +684,7 @@ from agent_runner.vcs_state import register_plugin_owned_paths
 # Module-top side effect — must register before the first round runs.
 register_plugin_owned_paths([
     "proposals/",                  # trailing slash → prefix match
-    "logs/plugins/my_plugin/**/*", # recursive glob (fnmatch)
+    "logs/plugins/my_plugin/**/*", # recursive glob (globstar)
     "reports/*.md",                # single-segment glob (PurePath.match)
 ])
 ```
@@ -697,7 +697,7 @@ register_plugin_owned_paths([
 | `"proposals"` (no slash) | `proposals` exactly | Single-segment literal. |
 | `"reports/*.md"` | `reports/dev.md` | `*` does not cross slashes. |
 | `"reports/**/*.md"` | `reports/dev.md`, `reports/sub/qa.md` | `**` matches across directory separators. |
-| `"logs/plugins/**/*"` | `logs/plugins/argus/state.json` | Same — `**` covers intermediate dirs. |
+| `"logs/plugins/**/*"` | `logs/plugins/acme/state.json` | Same — `**` covers intermediate dirs. |
 
 ### Caveat — this is NOT a "make work_dir messy" license