@higherdev/cli 0.1.1 → 0.1.3

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.
Files changed (3) hide show
  1. package/README.md +30 -14
  2. package/dist/index.js +1014 -697
  3. package/package.json +1 -1
package/README.md CHANGED
@@ -56,15 +56,17 @@ hd ask "why is HD-12 stuck?"
56
56
  hd msg HD-12 --to builder "also rename the old column" --interrupt
57
57
  ```
58
58
 
59
- `hd plan` runs the **architect** here on your machine, on Codex against your
60
- ChatGPT subscription. It reads the repo, decides what should be built, files an
61
- epic, and can tune the platform itself: models, effort, routing notes, prompt
62
- addenda, conventions, budgets. Creating an epic fires `epic.created`, which the
59
+ `hd plan` talks to **your own model**, here on your machine, on your ChatGPT
60
+ subscription. It is not a worker in a workspace and not a row in the agent
61
+ registry: nothing on the runner dispatches to it, and it reaches the whole
62
+ platform. `hd brain` shows which model it is and connects the subscription the
63
+ first time. It reads the repo, decides what should be built, files an epic,
64
+ and can tune the platform itself: models, effort, routing notes, prompt addenda,
65
+ conventions, budgets. Creating an epic fires `epic.created`, which the
63
66
  orchestrator already triggers on, so the handoff needs no glue.
64
67
 
65
68
  ```bash
66
- hd connect # ChatGPT OAuth, once per machine
67
- hd doctor # what a run would cost, before it runs
69
+ hd brain # which model, and connect it the first time
68
70
  hd plan # a conversation: type, it answers, keep going
69
71
  ```
70
72
 
@@ -82,16 +84,27 @@ hd plan --read-only "what would you change about the agent lineup?"
82
84
  hd plan --once "file the epic we discussed"
83
85
  ```
84
86
 
85
- ### The subscription is checked, not assumed
87
+ ### How it reaches the model
86
88
 
87
- The platform runs on subscriptions, never per-token APIs, so `hd plan` **refuses
88
- to start** unless Codex is signed in through the ChatGPT OAuth flow. It reads
89
- `~/.codex/auth.json` and distinguishes four states: connected, signed in with an
90
- API key (which bills per token), signed out, and not installed. `hd connect`
91
- runs `codex login` for you, which opens the browser and writes the token itself.
92
- `hd` never sees or stores a Codex credential.
89
+ `hd` calls the ChatGPT backend directly with your subscription credentials and
90
+ runs its own agent loop over its own tools, in process. It does not spawn
91
+ `codex exec`. That matters: a subprocess made Codex's coding agent the brain, so
92
+ it owned the system prompt, it had a shell, and a greeting took twenty seconds.
93
+ Direct, a turn answers in about two, `hd` owns the prompt, there is no shell in
94
+ the path, and text streams as it is written.
93
95
 
94
- `--allow-api-billing` overrides the refusal when you actually mean to spend.
96
+ The credential comes from `~/.codex/auth.json`, because the Codex CLI is what
97
+ performs the browser login. Nothing else about Codex is involved. `hd connect`
98
+ runs that login for you; `hd` never sees or stores the token, and renews the
99
+ session automatically rather than asking you to sign in again.
100
+
101
+ `hd plan` **refuses to start** unless that session is a real ChatGPT one, since
102
+ the platform runs on subscriptions and never per-token APIs. It distinguishes
103
+ connected, signed in with an API key (which bills per token), signed out, and
104
+ not installed. `--allow-api-billing` overrides the refusal when you mean it.
105
+
106
+ The architect reads code through `list_files`, `read_file`, and `search_code`,
107
+ scoped to the workspace checkout. Read only, no shell, no path escape.
95
108
 
96
109
  `hd mcp` is the same platform as MCP tools on stdio. `hd plan` attaches it to
97
110
  Codex for the invocation, with no global config edit. Attach it to Claude Code or
@@ -102,6 +115,9 @@ hd mcp --list # the 24 tools, with what each does
102
115
  hd mcp --read-only # safe to hand to anything
103
116
  ```
104
117
 
118
+ The architect does not go through MCP: its tools are the same functions, called
119
+ in process. `hd mcp` exists so anything else can reach them.
120
+
105
121
  ## Everything else
106
122
 
107
123
  ```