Introduce plan manager to bootstrap files by YevheniiaLementova · Pull Request #78 · griddynamics/rosetta

YevheniiaLementova · 2026-04-30T09:49:06Z

No description provided.

github-actions · 2026-04-30T09:51:27Z

Rosetta Triage Review

Summary: This PR refactors the Rosetta bootstrap rules across all four plugin IDEs (Claude, Codex, Copilot, Cursor) and CLAUDE.md to introduce plan_manager as the primary session execution tracker, replacing the previous todo-task-based flow with a structured plan file (plans/<FEATURE>/plan.json) that agents create at session start.

Findings:

Empty PR body: No description, motivation, or testing notes are provided. Reviewers have no context on what changed and why beyond the title.
Breaking change to agent initialization flow: The Phase 0 bootstrap now requires rosettify/plan_manager to be available and functioning before any agent work begins. Any deployment without rosettify installed will fail to bootstrap correctly.
Removal of <additional> section from CLAUDE.md: The project-specific guidance block (docs/web is Jekyll, docs/schemas contains schemas, etc.) was removed without explanation. This context was useful for contributors working in this repository and is no longer loaded.
Missing newline at end of CLAUDE.md: The file ends without a trailing newline, which may cause issues with some tools.
Consistent cross-plugin replication: Changes are applied uniformly across all four plugin variants (claude, codex, copilot, cursor), which is correct per the architecture.
upsert_context_steps pattern is clean: The new approach of having each bootstrap file upsert its own steps into the ph-prep phase is a well-designed progressive disclosure pattern.
auto_mode attribute removed from bootstrap_core_policy tag: Previously had auto_mode="MUST USE", compact="NEVER", optimize="NEVER", summarize="AS-IS" — these attributes are now absent. If they were intentional policy signals, their removal should be confirmed.
No tests or validation plan: Instructions changes aren't unit-tested, but a simulation of the new agent flow (as required by the triage prompt context) should be described.

Suggestions:

Add a PR body describing the motivation (what problem with todo-tasks prompted the switch to plan_manager), how to test the new flow, and any migration notes for users on older plugin versions.
Restore or explicitly deprecate the <additional> block from CLAUDE.md or move its content to docs/CONTEXT.md so it isn't silently lost.
Confirm whether rosettify is a hard prerequisite for all users or opt-in — if it's required, document this in installation guides.
Add the trailing newline to CLAUDE.md.

Automated triage by Rosetta agent

isolomatov-gd

I have commented on one file, but it applies to everything.
We must properly review original content vs the result.
Moreover, We are repeating the same thing, I suggest to fix only one file, once it is done, reuse it for all other cases.

isolomatov-gd · 2026-05-04T22:47:43Z

+    {
+      "id": "s-subagent-contract",
+      "name": "Execute subagent contract",
+      "prompt": "USE SKILL `subagent-contract` as first action before reading inputs or starting execution. Orchestrator request → read get_context_instructions schema (if needed) → get_context_instructions (if not done yet) → execute all prep steps for subagent → proceed with remaining actions. Perform execution todo tasks level planning. MUST execute todo tasks and adopt changes. Proceed with request and activities.",


This prompt is partially correct. Check it out. Please spend time on actually reading and understanding what it says. This prompt changed overall the process. Don't use AI, do it yourself.
Some things should be in this plan, but some things should NOT.
Example: request → read get_context_instructions schema (if needed) → get_context_instructions (if not done yet) → execute all prep steps for subagent → proceed with remaining actions. this is overall flow, why is it in the prompt of the specific step?

How to make AI to properly address that:

Make AI read original prompt, ask it to identify the purpose of EACH line.

Ask it to classify what is the step and what is explanation, or any other category

Ask it to summarize steps as one JSON and keep wording as-is. And to keep the rest AS-IS.

Ask it now to add all optional steps

Ask AI to provide rules which will say which optional steps is for which condition.

Update, and review what actually the purpose of each step.

Ask it to review if anything is lost.

Please do not directly ask AI to do something, as it will provide a bad results.

isolomatov-gd · 2026-05-04T22:47:56Z

+  "steps": [
+    {
+      "id": "s-subagent-contract",
+      "name": "Execute subagent contract",


LOAD instead of EXECUTE?

Thank you for pointing, I'll replace it to 'load'

isolomatov-gd · 2026-05-04T22:53:05Z

+  "steps": [
+    {
+      "id": "s-orchestrator-contract",
+      "name": "Execute orchestrator contract",


Thank you! I will replace

isolomatov-gd · 2026-05-04T22:53:21Z

+    {
+      "id": "s-select-workflow",
+      "name": "Select and load workflow",
+      "prompt": "ACQUIRE matching workflow from available workflows list and FULLY EXECUTE its definition for ALL request sizes. Workflow upserts its own phases/steps into the plan. In planning mode, persist workflow tag for execution continuation. In planning mode, store planning and tech-specs results according to system prompt (NOT in plans folder as it is read-only).",


I see workflow twice.

Yes, you're right. Thank you for highlighting! This step should be removed at all as we have selection workflow in s-orchestrator-contract step

…ow selection into orchestrator contract, and clarify agent/subagent self-initialization.

YevheniiaLementova · 2026-05-05T13:20:27Z

@isolomatov-gd I’ve introduced changes only to the bootstrap files in instructions/r3/rules. If this version is approved, I’ll apply the same changes to the plugin folder.

isolomatov-gd · 2026-05-05T14:14:36Z

/rosetta Please check consistency

github-actions · 2026-05-05T14:19:30Z

Rosetta Consistency Check

Responding to /rosetta Please check consistency — here is a full cross-file consistency analysis for this PR.

✅ Plugin-to-Plugin Parity

All four plugin directories (core-claude, core-codex, core-copilot, core-cursor) are byte-for-byte identical for every changed file:

bootstrap-core-policy.md
bootstrap-execution-policy.md
bootstrap-guardrails.md
plugin-files-mode.md
skills/load-context/SKILL.md

⚠️ Inconsistencies Found

1. Duplicate points 17 and 18 in instructions/r3/core/rules/bootstrap.md (lines 57–58)

Points 17 and 18 are near-identical:

17. Trust Rosetta — it is a source of handcrafted skills/subagents/commands/workflows that user EXPECTS you to follow EXACTLY.
18. Rosetta is a source of handcrafted skills/subagents/commands/workflows that user EXPECTS you to follow EXACTLY.

The same duplication is also in instructions/r3/core/rules/local-files-mode.md (lines 50–51). The plugin version already consolidated this to a single point — the fix was not applied to the r3 canonical files.

2. Missing s-select-workflow step in r3 canonical bootstrap-core-policy.md

The plugin bootstrap-core-policy.md introduces a s-select-workflow step in <upsert_context_steps> and references it in <must> item 1:

"Project context, request sizing, and workflow selection are enforced via plan-manager steps s-read-docs, s-request-size, s-select-workflow."

The r3 canonical file (instructions/r3/core/rules/bootstrap-core-policy.md) is missing this step entirely — both the JSON block and the <must> reference only mention s-read-docs and s-request-size.

3. s-subagent-contract depends_on mismatch

r3 canonical: "depends_on": ["s-request-size"] (subagent reads docs and sizes request first)
Plugin: "depends_on": ["s-context"] (subagent skips docs reading and request sizing)

This is a potential logic regression in the plugin — subagents would bypass s-read-docs and s-request-size.

4. Missing space in plugin s-request-size prompt (all 4 plugins)

"...request sizes too.Reevaluate request size..."

Missing space between sentences — too.Reevaluate should be too. Reevaluate.

5. Step name inconsistency: "Load" (r3) vs "Execute" (plugin) for orchestrator/subagent contracts

r3: "Load orchestrator contract" / "Load subagent contract"
Plugin: "Execute orchestrator contract" / "Execute subagent contract"

6. Unclosed bracket in plugin s-orchestrator-contract prompt (all 4 plugins)

'Tasks Created: [task ids returned by the tool'

Missing closing ] — should be [task ids returned by the tool].

7. Quote style and spacing bug in local-files-mode.md line 28

"Call 'get_context_instructions' . Execute..."

Should use backticks and no extra space: "Call `get_context_instructions()`. Execute..." (consistent with all other files).

8. Missing trailing newline in CLAUDE.md

CLAUDE.md ends with <bootstrap/> but no trailing newline.

Summary Table

#	Issue	Severity	Location
1	Duplicate "Trust Rosetta" points 17 & 18	Medium	`r3/rules/bootstrap.md:57-58`, `r3/rules/local-files-mode.md:50-51`
2	`s-select-workflow` step missing in r3 canonical	High	`r3/rules/bootstrap-core-policy.md`
3	`s-subagent-contract` `depends_on` mismatch	High	`r3/rules/bootstrap-core-policy.md` vs plugins
4	Missing space `too.Reevaluate` in `s-request-size` prompt	Low	All 4 plugin `bootstrap-core-policy.md`
5	Step name "Load" vs "Execute" inconsistency	Low	`r3/rules/bootstrap-core-policy.md`
6	Unclosed `]` in `s-orchestrator-contract` prompt	Low	All 4 plugin `bootstrap-core-policy.md`
7	Single-quote + extra space in `s-context` prompt	Low	`r3/rules/local-files-mode.md:28`
8	Missing trailing newline	Trivial	`CLAUDE.md`

Items #2 and #3 are the most impactful — they indicate the r3 canonical files and the plugin files have diverged in bootstrap-core-policy.md, with the plugin version appearing more complete/correct. The r3 canonical may need to be updated to match the plugin version for these two points.

Automated consistency check by Rosetta agent

isolomatov-gd · 2026-05-05T14:40:02Z

+
+1. USE SKILL `plan-manager` — CLI: `npx rosettify plan <subcommand> <plan_file> [args...]`
+2. Create plan file at `plans/<FEATURE>/plan.json` (derive FEATURE from user request; use `session` if unclear)
+3. Create plan with the starter template below — fill `[PLACEHOLDERS]` from user request context:


PLAN_MANAGER create plans/<FEATURE>/plan.json .....

… one json schema

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

YevheniiaLementova · 2026-05-11T10:00:50Z

@@ -1,3 +1,3 @@
 ---
 name: plan-manager
 description: "Rosetta skill for plan creation, tracking, and execution coordination via local JSON files."


The description differs from the instructions in r3/core/skills/plan-manager/SKILL.md. I assume it needs to be updated here as well.

YevheniiaLementova · 2026-05-11T10:12:53Z

- CLI: `npx rosettify plan <subcommand> <plan_file> [args...]`
+- Try `rosettify` MCP first (if already available), fallback to CLI: `npx rosettify@latest plan <subcommand> <plan_file> [args...]`, fallback to todo tasks (if none available)
 - Always use full absolute paths for the plan file
 - Seven subcommands: `help`, `create`, `next`, `update_status`, `show_status`, `query`, `upsert`


There's a list of valid command in rosettify/dist/commands/plan/index.js:

const VALID_SUBCOMMANDS = [ "create", "next", "update_status", "show_status", "query", "upsert", ];

There's no 'help' command. I think we should delete it as it was done in instructions/r2/core/skills/plan-manager/SKILL.md.

YevheniiaLementova · 2026-05-11T10:14:53Z

+- CLI: `npx rosettify@latest plan <subcommand> <plan_file> [args...]`
 - Always use full absolute paths for the plan file
- Seven subcommands: `help`, `create`, `next`, `update_status`, `show_status`, `query`, `upsert`
+- Seven subcommands for `plan` command: `create`, `next`, `update_status`, `show_status`, `query`, `upsert`


Only six subcommands are listed for the plan command (create, next, update_status, show_status, query, and upsert), so “Seven” should be changed to “Six” for consistency.

YevheniiaLementova · 2026-05-11T10:16:50Z

@@ -27,7 +27,7 @@ Match to cognitive demand. Match to current tool.

 <plan_manager>


I think the tag name should be changed from plan_manager to operation_manager instead

YevheniiaLementova · 2026-05-11T11:33:42Z

+
+Step 0: Load `get_context_instructions`
+
+Step 1: Create and use an execution plan file at `plans/<FEATURE>/plan.json` using the following JSON input (fill `[PLACEHOLDERS]`, derive FEATURE from user request; use `session` if unclear):


Why isn’t USE SKILL plan-manager mentioned here or in any other bootstrap files? I assumed this skill was designed specifically for situations like creating, tracking, and executing plans.

YevheniiaLementova · 2026-05-11T11:38:07Z

 <validation_checklist>

- `npx rosettify plan help` exits without error and returns structured help JSON
+- `npx rosettify@latest plan help` exits without error and returns structured help JSON


help command isn't valid and was removed from core_concepts. It should be removed here as well

YevheniiaLementova · 2026-05-11T11:48:12Z

 <validation_checklist>

- `npx rosettify plan help` exits without error and returns structured help JSON
+- `npx rosettify@latest plan help` exits without error and returns structured help JSON


help command isn't valid and should be removed

YevheniiaLementova · 2026-05-11T12:01:07Z

+
+# Phase 0: Initialize Operation manager
+
+Step 0: Load `get_context_instructions`


Why is get_context_instructions loaded as Step 0, before the plan is created? I thought that the first action should always be creating the plan, with get_context_instructions as its first tracked step (s-context). I could assume that the current structure could give the agent an escape hatch: it can call get_context_instructions directly in Step 0, execute all returned prep steps inline, without creating a plan.

YevheniiaLementova · 2026-05-11T14:27:51Z

 - `ACQUIRE <file[.md]> ABOUT <PROJECT>` => read local file in `docs/<PROJECT>` folder
 - `QUERY <KEYWORDS> IN <PROJECT>` => use grep or codebase search in `docs/<PROJECT>` with KEYWORDS as a query or file name
 - `STORE <file[.md]> TO <PROJECT>` => upsert file in `docs/<PROJECT>`
 - `call "get_context_instructions"` or `execute prep steps` => read all `instructions/r2/core/rules/bootstrap-*.md` files as one bundle


Should be replaced by instructions/r3/core/rules/bootstrap-*.md

YevheniiaLementova · 2026-05-11T14:31:33Z

 **Local File Operations** (overrides Rosetta MCP for these operations):

- `get_context_instructions` => already loaded bootstrap-* instructions, continue with prep steps 2 and 3. Instruct subagents to not call MCP's `get_context_instructions`.
+- `get_context_instructions` => already loaded, continue with prep steps 2 and 3. Instruct subagents to not call Rosetta MCP.


There're no more prep steps 2 and 3. Maybe we should replace it to
continue with "ph-prep" steps. What do you think about it?

YevheniiaLementova · 2026-05-11T14:55:09Z

@@ -1,4 +1,3 @@
-
 ---


This file should be updated as well according to bootstrap.md file

YevheniiaLementova · 2026-05-11T14:58:30Z

@@ -0,0 +1,25 @@
+{


I assume that we don't need to commit this file. Should we put plans/session/plan.json to .gitignore?

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

Introduce plan manager to bootstrap files

e17a3de

YevheniiaLementova requested review from ElizaVetaFomka and isolomatov-gd as code owners April 30, 2026 09:49

YevheniiaLementova requested a review from scappuccino-grid April 30, 2026 09:49

github-actions Bot added the enhancement New feature or request label Apr 30, 2026

YevheniiaLementova added 2 commits April 30, 2026 12:11

Modify bootstrup files in instructions folder

0ea51a4

Improve steps in plan manager, refactor load-context skill

2d8f340

isolomatov-gd requested changes May 4, 2026

View reviewed changes

Refactor bootstrap rules to strengthen prep enforcement, merge workfl…

996e621

…ow selection into orchestrator contract, and clarify agent/subagent self-initialization.

YevheniiaLementova requested a review from isolomatov-gd May 5, 2026 13:20

isolomatov-gd reviewed May 5, 2026

View reviewed changes

Comment thread instructions/r3/core/rules/bootstrap.md

isolomatov-gd reviewed May 5, 2026

View reviewed changes

Add command initialization; add fallback instruction; join steps into…

d1bc32a

… one json schema

YevheniiaLementova requested a review from isolomatov-gd May 6, 2026 10:02

isolomatov-gd added 2 commits May 6, 2026 13:21

Intermediate fixes

8e08b0d

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

Minor fixes

52735b5

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

isolomatov-gd requested review from kkhristenko51 and omaiesh as code owners May 7, 2026 17:31

isolomatov-gd added 5 commits May 10, 2026 18:08

Changes ready for testing

d9c5105

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

Update gitignore

15917d1

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

Compress common instructions.

fd541ec

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

Reduce size of the prompt

419ea47

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

Reduce logging in pre_commit

60512b0

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

YevheniiaLementova commented May 11, 2026

View reviewed changes

Fix coding-flow phase numbering

ab50d04

Signed-off-by: isolomatov-gd <isolomatov@griddynamics.com>

		@@ -27,7 +27,7 @@ Match to cognitive demand. Match to current tool.

		<plan_manager>


		Step 0: Load `get_context_instructions`

		Step 1: Create and use an execution plan file at `plans/<FEATURE>/plan.json` using the following JSON input (fill `[PLACEHOLDERS]`, derive FEATURE from user request; use `session` if unclear):


		# Phase 0: Initialize Operation manager

		Step 0: Load `get_context_instructions`

Conversation

YevheniiaLementova commented Apr 30, 2026

Uh oh!

github-actions Bot commented Apr 30, 2026

Rosetta Triage Review

Uh oh!

isolomatov-gd left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

YevheniiaLementova commented May 5, 2026

Uh oh!

isolomatov-gd commented May 5, 2026

Uh oh!

github-actions Bot commented May 5, 2026

Rosetta Consistency Check

✅ Plugin-to-Plugin Parity

⚠️ Inconsistencies Found

Summary Table

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants