openclaw-code-agent 4.3.7 → 4.4.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +14 -9
- package/dist/index.js +207 -196
- package/openclaw.plugin.json +13 -9
- package/package.json +6 -4
- package/skills/code-agent-orchestration/SKILL.md +2 -1
package/README.md
CHANGED
|
@@ -4,7 +4,7 @@
|
|
|
4
4
|
[](https://www.npmjs.com/package/openclaw-code-agent)
|
|
5
5
|
[](LICENSE)
|
|
6
6
|
|
|
7
|
-
`openclaw-code-agent` runs Claude Code and
|
|
7
|
+
`openclaw-code-agent` runs Claude Code, Codex, and experimental OpenCode as managed background coding sessions from OpenClaw chat. It adds plan approval, session lifecycle, wake routing, worktree isolation, merge/PR follow-through, and explicit goal loops on top of the agent backends.
|
|
8
8
|
|
|
9
9
|
Use it when you want to start coding work from Telegram, Discord, or another OpenClaw-supported channel and keep the job observable after the first message.
|
|
10
10
|
|
|
@@ -18,10 +18,10 @@ Use it when you want to start coding work from Telegram, Discord, or another Ope
|
|
|
18
18
|
- **Full session lifecycle**. Suspend, resume, fork, interrupt, and recover sessions across restarts with persisted metadata and output.
|
|
19
19
|
- **Explicit goal-task loops**. Opt into verifier-driven repair loops or Ralph-style completion loops when you need iterative autonomous execution toward a specific goal.
|
|
20
20
|
- **Real operator visibility**. `agent_sessions`, `agent_output`, and `agent_stats` show status, buffered output, duration, and USD cost.
|
|
21
|
-
- **
|
|
21
|
+
- **Multiple harnesses, one control plane**. Claude Code, Codex, and experimental OpenCode share the same tools, routing, notification pipeline, and worktree strategy model while each backend uses its own adapter and resume substrate.
|
|
22
22
|
- **One continuation path**. Follow-ups, approvals, revisions, interrupts, and redirects all continue the existing session instead of launching a duplicate.
|
|
23
23
|
|
|
24
|
-
This plugin is separate from OpenClaw's bundled `acpx` runtime plugin and bundled core `codex` plugin. Those own adjacent OpenClaw runtime/provider surfaces; `openclaw-code-agent` owns chat orchestration and repository follow-through for its own Claude Code and
|
|
24
|
+
This plugin is separate from OpenClaw's bundled `acpx` runtime plugin and bundled core `codex` plugin. Those own adjacent OpenClaw runtime/provider surfaces; `openclaw-code-agent` owns chat orchestration and repository follow-through for its own Claude Code, Codex, and experimental OpenCode harnesses. See [docs/ACP-COMPARISON.md](docs/ACP-COMPARISON.md) for the boundary details.
|
|
25
25
|
|
|
26
26
|
## From Chat To Resolved Work
|
|
27
27
|
|
|
@@ -32,13 +32,13 @@ This plugin is separate from OpenClaw's bundled `acpx` runtime plugin and bundle
|
|
|
32
32
|
|
|
33
33
|
### Direct Completion
|
|
34
34
|
|
|
35
|
-
For small trusted changes, an orchestrator can launch a session, let
|
|
35
|
+
For small trusted changes, an orchestrator can launch a session, let the selected harness finish, and report the verified outcome back to chat. The session stays observable through launch, completion, cost, duration, and commit summary.
|
|
36
36
|
|
|
37
37
|
|
|
38
38
|
|
|
39
39
|
### Plan Review
|
|
40
40
|
|
|
41
|
-
The default review loop is plan-first. Claude Code and
|
|
41
|
+
The default review loop is plan-first. Claude Code, Codex, and experimental OpenCode feed the same approval UX: the plugin blocks implementation until approval, then continues the same session after the plan is approved. Codex can provide structured plan artifacts; OpenCode is currently treated as text-only with plugin-owned plan gating. The user can approve, request a revision, or reject the plan from the originating thread.
|
|
42
42
|
|
|
43
43
|
|
|
44
44
|
|
|
@@ -115,7 +115,7 @@ Add the smallest useful config under `plugins.entries["openclaw-code-agent"]` in
|
|
|
115
115
|
For the first run, choose:
|
|
116
116
|
|
|
117
117
|
- `defaultWorkdir`: a git repository root you expect to use often.
|
|
118
|
-
- `defaultHarness`: `claude-code` or `
|
|
118
|
+
- `defaultHarness`: `claude-code`, `codex`, or `opencode`. Treat `opencode` as experimental.
|
|
119
119
|
|
|
120
120
|
The default policy is intentionally review-first:
|
|
121
121
|
|
|
@@ -127,7 +127,7 @@ Because worktree isolation defaults to `delegate`, `defaultWorkdir` should norma
|
|
|
127
127
|
|
|
128
128
|
Chat-launched sessions route updates back to their originating chat thread. For agent-launched tool sessions without an origin route, configure `fallbackChannel` or `agentChannels` in the reference guide.
|
|
129
129
|
|
|
130
|
-
|
|
130
|
+
The current package targets and validates against the OpenClaw SDK package `openclaw@2026.6.1`, while keeping the plugin peer floor at `>=2026.4.21`. No host upgrade or host config change is required for this plugin compatibility update. OpenClaw `2026.6.1` changes plugin runtime/install lookup, plugin SDK exports, approval normalization, cron/session delivery, Telegram callback and topic delivery, runtime tool allowlists, Codex App Server recovery, and adjacent bundled/external agent plugins, but `openclaw-code-agent` does not need new SDK imports for them: the manifest already declares `contracts.tools`, the plugin still imports only `openclaw/plugin-sdk/plugin-entry`, and its own session store, wake routing, callbacks, worktree flows, and harness model restrictions remain plugin-owned.
|
|
131
131
|
|
|
132
132
|
If you use Codex, make sure the local `codex` command or `OPENCLAW_CODEX_APP_SERVER_COMMAND` override is available and authenticated. Codex-specific defaults live under `harnesses.codex`: `reasoningEffort` is sent as `reasoningEffort`, and `fastMode: true` sends `service_tier: "fast"` on Codex App Server thread, resume, and turn payloads. When Codex auth is inconsistent, this is the recommended `~/.codex/config.toml` setting:
|
|
133
133
|
|
|
@@ -135,6 +135,8 @@ If you use Codex, make sure the local `codex` command or `OPENCLAW_CODEX_APP_SER
|
|
|
135
135
|
forced_login_method = "chatgpt"
|
|
136
136
|
```
|
|
137
137
|
|
|
138
|
+
If you use experimental OpenCode, make sure local `opencode >= 1.16.2` is available and configured with provider auth. The plugin starts `opencode serve` per session on localhost and uses OpenCode's classic session lifecycle routes while v2 session wait remains unavailable. Leave `harnesses.opencode.defaultModel` unset to let OpenCode choose its configured provider default, or pass an explicit `provider/model` string for a launch.
|
|
139
|
+
|
|
138
140
|
## First Session
|
|
139
141
|
|
|
140
142
|
In chat, ask OpenClaw to start work:
|
|
@@ -161,7 +163,7 @@ Stop this session.
|
|
|
161
163
|
|
|
162
164
|
### Plan Review
|
|
163
165
|
|
|
164
|
-
By default, Claude Code and
|
|
166
|
+
By default, Claude Code, Codex, and experimental OpenCode produce a plan before implementation. The plan can be approved, revised, or rejected through buttons when available, or with plain-text `Approve`, `Revise`, or `Reject` in the same thread.
|
|
165
167
|
|
|
166
168
|
Revisions stay attached to the same session, so the newest plan is the actionable one.
|
|
167
169
|
|
|
@@ -182,6 +184,7 @@ Merge, PR, ordinary terminal, and no-change worktree outcomes use a two-step com
|
|
|
182
184
|
### Goal Tasks
|
|
183
185
|
|
|
184
186
|
Goal tasks are explicit autonomous loops for work that should keep iterating toward a defined finish line. They do not replace ordinary coding sessions.
|
|
187
|
+
Goal iteration progress is controller progress: the counter advances only when the goal controller starts another agent turn after a missing completion promise or failed verifier. If the agent performs several review/implementation passes inside one successful turn, those internal passes should appear in the single completion summary rather than as separate goal iteration notifications.
|
|
185
188
|
|
|
186
189
|
Ask in normal chat:
|
|
187
190
|
|
|
@@ -189,6 +192,7 @@ Ask in normal chat:
|
|
|
189
192
|
Start a verifier goal in /repo: fix the failing auth flow and keep running pnpm test until it passes.
|
|
190
193
|
Start a Ralph-style goal for /repo: ship the draft workflow, and consider it complete when the output says DONE.
|
|
191
194
|
Show goal status.
|
|
195
|
+
Change the auth goal to also update the smoke tests.
|
|
192
196
|
Stop the auth goal.
|
|
193
197
|
```
|
|
194
198
|
|
|
@@ -214,9 +218,10 @@ Most users interact in chat. The tool surface is for OpenClaw agents and advance
|
|
|
214
218
|
| `agent_worktree_cleanup` | Clean safe worktrees or dismiss one pending decision |
|
|
215
219
|
| `goal_launch` | Start an explicit verifier or Ralph-style goal loop |
|
|
216
220
|
| `goal_status` | Show one goal task or list all goal tasks |
|
|
221
|
+
| `goal_edit` | Change the goal text for an active goal task |
|
|
217
222
|
| `goal_stop` | Stop a running goal task |
|
|
218
223
|
|
|
219
|
-
Chat commands mirror the common workflows when you want explicit commands instead of natural-language chat, but most human use should start with plain requests like the examples above. Available commands are `/agent`, `/agent_sessions`, `/agent_output`, `/agent_respond`, `/agent_kill`, `/agent_stats`, `/goal`, `/goal_status`, and `/goal_stop`.
|
|
224
|
+
Chat commands mirror the common workflows when you want explicit commands instead of natural-language chat, but most human use should start with plain requests like the examples above. Available commands are `/agent`, `/agent_sessions`, `/agent_output`, `/agent_respond`, `/agent_kill`, `/agent_stats`, `/goal`, `/goal_status`, `/goal_edit`, and `/goal_stop`.
|
|
220
225
|
|
|
221
226
|
## Docs
|
|
222
227
|
|