@dreki-gg/pi-plan-mode 0.2.0 → 0.3.0

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.
Files changed (5) hide show
  1. package/CHANGELOG.md +14 -0
  2. package/README.md +44 -70
  3. package/extensions/plan-mode/index.ts +463 -548
  4. package/extensions/plan-mode/utils.ts +26 -11
  5. package/package.json +4 -4
package/CHANGELOG.md CHANGED
@@ -1,5 +1,19 @@
1
1
  # @dreki-gg/pi-plan-mode
2
2
 
3
+ ## 0.3.0
4
+
5
+ ### Minor Changes
6
+
7
+ - [`5c9d134`](https://github.com/dreki-gg/pi-extensions/commit/5c9d134131599cd102f77d3849660e3e6f885f70) Thanks [@jalbarrang](https://github.com/jalbarrang)! - Redesign plan mode as a two-phase workflow with file-based handoff
8
+
9
+ - Plan phase uses `claude-opus-4-6:medium` with read-only tools + strict bash allowlist
10
+ - Plans are written to `.plans/<kebab-name>/PLAN.md` with a `START-PROMPT.md` for clean context handoff
11
+ - Execute phase uses `gpt-5.5:low` with full tool access, starting from START-PROMPT.md in a clean context
12
+ - Todo extraction is deferred to execution time (extracted from PLAN.md on "Execute Plan")
13
+ - New menu: Execute Plan, Refine Plan (adversarial self-review), Follow up, Exit plan mode
14
+ - Model and thinking level are saved/restored across phase transitions
15
+ - Removed domain-model, plan-files, and autocomplete sub-workflows
16
+
3
17
  ## 0.2.0
4
18
 
5
19
  ### Minor Changes
package/README.md CHANGED
@@ -1,13 +1,8 @@
1
1
  # @dreki-gg/pi-plan-mode
2
2
 
3
- Cursor-like planning workflow for [pi](https://github.com/badlogic/pi-mono).
3
+ Two-phase planning workflow for [pi](https://github.com/badlogic/pi-mono).
4
4
 
5
- It gives pi a dedicated **plan mode** that:
6
- - locks the agent into a read-only planning pass
7
- - nudges it to use `questionnaire` before planning when scope is unclear
8
- - hands off to `/skill:domain-model` when you want terminology/design pressure-testing
9
- - hands off to `/skill:create-implementation-plans` when you want self-contained `*.plan.md` files
10
- - restores full tool access for execution once the plan is approved
5
+ Plan with `claude-opus-4-6:medium`, execute with `gpt-5.5:low`. Plans are persisted as files in `.plans/` for clean context handoff between models.
11
6
 
12
7
  ## Install
13
8
 
@@ -21,89 +16,68 @@ Recommended companions:
21
16
  pi install npm:@dreki-gg/pi-questionnaire
22
17
  ```
23
18
 
24
- And make sure these skills are available globally or project-locally if you want the full workflow:
25
- - `domain-model`
26
- - `create-implementation-plans`
27
-
28
- If those skills are missing, the extension falls back to plain prompts that emulate the same workflow.
29
-
30
19
  ## What it provides
31
20
 
32
- | Feature | Name | Notes |
33
- |---|---|---|
34
- | Flag | `--plan` | Start pi in read-only planning mode |
35
- | Command | `/plan [prompt]` | Enable plan mode or immediately send a planning prompt |
36
- | Command | `/plan-status` | Show current phase + extracted plan steps |
37
- | Command | `/plan-domain [prompt]` | Stress-test the current plan against the domain model |
38
- | Command | `/plan-plans [prompt]` | Generate self-contained implementation plan files |
39
- | Command | `/plan-execute [prompt]` | Restore full tool access and execute the approved plan |
40
- | Shortcut | `Ctrl+Alt+P` | Toggle plan mode |
21
+ | Feature | Name | Notes |
22
+ | -------- | -------------- | ---------------------------------------------- |
23
+ | Flag | `--plan` | Start pi in plan mode |
24
+ | Command | `/plan [prompt]` | Enter plan mode, optionally with a starting prompt |
25
+ | Command | `/todos` | Show current plan progress |
26
+ | Shortcut | `Ctrl+Alt+P` | Toggle plan mode |
41
27
 
42
28
  ## Workflow
43
29
 
44
- ### 1. Start planning
30
+ ### 1. Plan (`claude-opus-4-6:medium`)
45
31
 
46
32
  ```text
47
- /plan add a plan mode similar to Cursor for this repo
33
+ /plan add authentication middleware with JWT support
48
34
  ```
49
35
 
50
- While planning, the extension restricts tools to a read-only set such as:
51
- - `read`
52
- - `bash` (allowlisted read-only commands only)
53
- - `grep`
54
- - `find`
55
- - `ls`
56
- - `questionnaire` (if installed)
57
- - `lsp` / `context7_*` (if installed)
36
+ The planner has access to read-only tools plus `edit`/`write` restricted to `.plans/` files. Bash is locked to a strict allowlist of safe commands.
58
37
 
59
- The system prompt tells pi to:
60
- - inspect the real codebase first
61
- - use `questionnaire` when the task is still underspecified
62
- - respond with a numbered plan under a `Plan:` header
63
- - stay read-only
38
+ The planner:
39
+ - Inspects the codebase using read-only tools
40
+ - Uses `questionnaire` when requirements are underspecified
41
+ - Creates `.plans/<kebab-name>/PLAN.md` with the full numbered plan
42
+ - Creates `.plans/<kebab-name>/START-PROMPT.md` — a self-contained handoff prompt with all context needed to execute without the planning conversation
43
+ - Can add supporting files in the same directory for extra context
64
44
 
65
- ### 2. Stress-test the design
45
+ ### 2. Choose next step
66
46
 
67
- ```text
68
- /plan-domain
69
- ```
47
+ When the planner finishes, a menu appears:
70
48
 
71
- If `domain-model` is installed, the extension invokes:
49
+ | Option | Description |
50
+ | ---------------- | ---------------------------------------------------------------------- |
51
+ | **Execute Plan** | Extract todos from PLAN.md, switch to gpt-5.5:low, start with START-PROMPT.md |
52
+ | **Refine Plan** | Adversarial review — planner critiques its own plan and updates files |
53
+ | **Follow up** | Open an editor for additional instructions to the planner |
54
+ | **Exit plan mode** | Disable plan mode and restore original model |
72
55
 
73
- ```text
74
- /skill:domain-model
75
- ```
56
+ ### 3. Execute (`gpt-5.5:low`)
76
57
 
77
- Otherwise it sends an equivalent fallback prompt.
58
+ When **Execute Plan** is selected:
59
+ 1. Todos are extracted from `PLAN.md`
60
+ 2. Model switches to `gpt-5.5:low` with full tool access
61
+ 3. The executor starts with a **clean context window** using `START-PROMPT.md`
62
+ 4. Each step must be marked with `[DONE:n]` before moving to the next
63
+ 5. Progress is tracked in a widget in the status bar
64
+ 6. When all steps complete, the original model and thinking level are restored
78
65
 
79
- ### 3. Write implementation plan files
66
+ ## Plan directory structure
80
67
 
81
- ```text
82
- /plan-plans
83
68
  ```
84
-
85
- This temporarily enables `edit`/`write`, but only for plan-file authoring. The prompt explicitly forbids product-code changes in this phase.
86
-
87
- If `create-implementation-plans` is installed, the extension invokes:
88
-
89
- ```text
90
- /skill:create-implementation-plans
69
+ .plans/
70
+ └── add-auth-middleware/
71
+ ├── PLAN.md # Numbered plan with context
72
+ ├── START-PROMPT.md # Self-contained executor handoff prompt
73
+ └── ... # Optional supporting files
91
74
  ```
92
75
 
93
- Otherwise it falls back to a plain prompt that asks pi to create grounded `*.plan.md` files.
94
-
95
- ### 4. Execute
96
-
97
- ```text
98
- /plan-execute
99
- ```
76
+ ## Footer indicators
100
77
 
101
- That restores the original tool set and asks pi to execute the approved plan. If the plan has numbered steps, the extension tracks progress via `[DONE:n]` tags.
78
+ - `📝 plan` — plan mode active (opus-4-6:medium, strict bash)
79
+ - `📋 exec 2/5` — executing plan with gpt-5.5:low, 2 of 5 steps done
102
80
 
103
- ## Notes
81
+ ## Bash safety
104
82
 
105
- - Planning mode is **hard enforced** by tool restriction, not just prompt wording.
106
- - `bash` is restricted to read-only commands while planning.
107
- - Plan steps are extracted from `Plan:` sections and shown in a widget/status area.
108
- - State is persisted across session resume and tree navigation.
109
- - The implementation-plan phase is a **controlled write phase** intended for planning docs, not code changes.
83
+ In plan mode, bash is restricted to read-only commands (ls, grep, git status, cat, rg, etc.). Destructive commands (rm, mv, git commit, etc.) are blocked.