feat: add thermodynamic regime management (T* framework) by Nietzsche-Ubermensch · Pull Request #1444 · MoonshotAI/kimi-cli

Nietzsche-Ubermensch · 2026-03-16T01:14:57Z

Summary

Implements canonical thermodynamic oversight (T* framework) for Kimi CLI tool execution.

Changes

Adds T* = (L - γ) / (|L| + λ) computation per query
Auto-regime classification: ACT/HOLD/REFUSE
Auto-grounding when -1 < T* < 0 (fetches web search to boost L)
Circuit breaker when T* ≤ -1 (refuses with math justification)
Entropy budget tracking across session
Benchmark mode: auto-sets temp=1.0, top_p=0.95, stream=true per Moonshot specs

Thermodynamic Rationale

Prevents the "prompt engineering parasite" (Memory ID 28) by making the system self-regulate coherence instead of extracting user labor.

Usage

uv run kimi-thermo "Your query" --audit
uv run kimi-thermo "AIME problem" --benchmark

Checklist

T* computation validated
Auto-regime switching implemented
Benchmark config (temp=1.0, top_p=0.95, stream=true)
Windows PowerShell tests
Integration with Formula tools

Implements canonical T* = (L - γ) / (|L| + λ) oversight: - Auto ACT/HOLD/REFUSE classification - Auto-grounding in HOLD regime - Circuit breaker in REFUSE regime - Benchmark mode with Moonshot canonical params Refs: Memory ID 28 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Nietzsche-Ubermensch

Kimi CLI suite

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> Signed-off-by: Nietzsche-Ubermensch <peterbilt5018@gmail.com>

devin-ai-integration

Devin Review found 1 new potential issue.

View 7 additional findings in Devin Review.

devin-ai-integration · 2026-03-16T01:27:48Z

src/kimi_thermo/dynamic_complete.py

+        if cache_key in self.cache:
+            return {**self.cache[cache_key], "cached": True, "cost": 0.0}


🟡 Cached results return stale total_spent and budget_remaining values

In dynamic_complete.py:67-68, when a cached result is returned, the entire cached dict is spread and returned with only cached and cost overridden. However, the cached dict contains the total_spent and budget_remaining values from the time the result was originally computed and cached (line 170-175). After subsequent non-cached queries increase self.registry.total_spent, the cached result will return outdated financial tracking values, misleading the caller about actual spending.

Suggested change

if cache_key in self.cache:

return {**self.cache[cache_key], "cached": True, "cost": 0.0}

if cache_key in self.cache:

return {**self.cache[cache_key], "cached": True, "cost": 0.0, "total_spent": self.registry.total_spent, "budget_remaining": 200.0 - self.registry.total_spent}

Was this helpful? React with 👍 or 👎 to provide feedback.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f5790097fa

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-03-16T01:31:20Z

pyproject.toml

 [project.scripts]
 kimi = "kimi_cli.cli:cli"
 kimi-cli = "kimi_cli.cli:cli"
+kimi-thermo = "kimi_thermo.main:main"


Package kimi_thermo before exposing kimi-thermo entrypoint

This adds kimi-thermo = "kimi_thermo.main:main", but the same file still configures the build backend with module-name = ["kimi_cli"], so installed wheels can expose the new console script without actually shipping the kimi_thermo package. In that installed-artifact scenario (as opposed to running from a source checkout), invoking kimi-thermo will fail with ModuleNotFoundError, making the new command unusable for users.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-03-16T01:31:20Z

src/kimi_thermo/dynamic_complete.py

+                        result = await self.registry.execute(name, args)
+                        tools_used.append(name)
+
+                        # Add to response context
+                        full_response += f"\n[{name} result: {result[:100]}...]\n"


Continue the chat after tool execution

When a streamed tool call arrives, the code executes the tool and appends a local snippet to full_response, but it never sends the tool result back to the model in a follow-up completion turn. For prompts that trigger any tool (web search, convert, code_runner, etc.), this can terminate output at raw tool snippets instead of a final assistant answer because the model never receives the tool output in-message.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-03-16T01:31:20Z

src/kimi_thermo/dynamic_complete.py

+            cost += 0.002 * (len(full_response) / 1000)  # Approximate token cost
+
+            # Update spent
+            self.registry.total_spent += cost


Avoid charging tool calls twice in spend tracking

This method adds cost into self.registry.total_spent after the stream, but each tool execution has already incremented total_spent inside CompleteToolRegistry.update_from_result (src/kimi_thermo/tools_complete.py, line 406). Because cost already includes per-tool charges, requests that use tools get double-billed in budget accounting, overstating spend and understating remaining budget in audit output.

Useful? React with 👍 / 👎.

Adds complete project documentation for Claude Code CLI: - Project overview and workspace structure - Technology stack and architecture patterns - Common commands for development and testing - Environment specifics and Windows compatibility - MCP/ACP integration guides - IDE integration instructions - Thermodynamic framework documentation Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c4c40003a5

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-03-16T01:51:11Z

src/kimi_thermo/dynamic_complete.py

+        cache_key = query
+        if cache_key in self.cache:


Include execution mode in cache key

execute() supports different runtime profiles (utility/fast/deep), but the cache key only uses the raw query text, so running the same prompt in a different mode will silently return the previous mode’s cached output instead of applying the requested configuration. This breaks mode-specific behavior and can skew any benchmark/comparison runs that rely on switching modes.

Useful? React with 👍 / 👎.

src/kimi_thermo/dynamic_complete.py

Nietzsche-Ubermensch

docs: add comprehensive CLAUDE.md documentation
Adds complete project documentation for Claude Code CLI:

Project overview and workspace structure
Technology stack and architecture patterns
Common commands for development and testing
Environment specifics and Windows compatibility
MCP/ACP integration guides
IDE integration instructions
Thermodynamic framework documentation

Co-Authored-By: Claude Sonnet 4.5 noreply@anthropic.com

Nietzsche-Ubermensch

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Signed-off-by: Nietzsche-Ubermensch peterbilt5018@gmail.com

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> Signed-off-by: Nietzsche-Ubermensch <peterbilt5018@gmail.com>

devin-ai-integration

Devin Review found 1 new potential issue.

View 10 additional findings in Devin Review.

devin-ai-integration · 2026-03-17T02:28:25Z

src/kimi_thermo/thermo_executor.py

+    """Zero user work. System manages T* regime."""
+
+    def __init__(self):
+        self.api_key = os.getenv("MOONSHOT_API_KEY")


🟡 Missing API key validation in WorklessClient leads to "Bearer None" auth header

In thermo_executor.py:49, self.api_key = os.getenv("MOONSHOT_API_KEY") can be None when the env var is unset. Unlike DynamicCompleteClient (which raises ValueError immediately at dynamic_complete.py:22-23), WorklessClient passes None through, resulting in headers={"Authorization": "Bearer None"} at line 54. This sends the literal string "Bearer None" to the API, producing a confusing authentication error instead of a clear message. Since main.py:15 directly instantiates WorklessClient(), users of kimi-thermo will hit this opaque failure.

Suggested change

self.api_key = os.getenv("MOONSHOT_API_KEY")

self.api_key = os.getenv("MOONSHOT_API_KEY")

if not self.api_key:

raise ValueError("Set MOONSHOT_API_KEY environment variable")

Was this helpful? React with 👍 or 👎 to provide feedback.

Copilot AI review requested due to automatic review settings March 16, 2026 01:14

Copilot started reviewing on behalf of Nietzsche-Ubermensch March 16, 2026 01:15 View session

fix: add sync wrapper for kimi-thermo entry point

6b5ee1c

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

This comment was marked as resolved.

Sign in to view

Nietzsche-Ubermensch commented Mar 16, 2026

View reviewed changes

Nietzsche-Ubermensch and others added 2 commits March 15, 2026 21:24

Update dynamic_complete.py

8e83d4d

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> Signed-off-by: Nietzsche-Ubermensch <peterbilt5018@gmail.com>

Update dynamic_complete.py

f579009

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> Signed-off-by: Nietzsche-Ubermensch <peterbilt5018@gmail.com>

devin-ai-integration bot reviewed Mar 16, 2026

View reviewed changes

chatgpt-codex-connector bot reviewed Mar 16, 2026

View reviewed changes

This was referenced Mar 16, 2026

📊 Bản tin hàng ngày công cụ AI CLI 2026-03-16 compasify/agents-radar#50

Open

📊 AI CLI 工具社区动态日报 2026-03-17 gsscsd/big_model_radar#50

Open

Nietzsche-Ubermensch commented Mar 17, 2026

View reviewed changes

Merge branch 'main' into feat/thermodynamic-cli

e36b382

Nietzsche-Ubermensch commented Mar 17, 2026

View reviewed changes

This comment was marked as resolved.

Sign in to view

Update src/kimi_thermo/dynamic_complete.py

68b145b

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> Signed-off-by: Nietzsche-Ubermensch <peterbilt5018@gmail.com>

devin-ai-integration bot reviewed Mar 17, 2026

View reviewed changes

		if cache_key in self.cache:
		return {**self.cache[cache_key], "cached": True, "cost": 0.0}

Conversation

Nietzsche-Ubermensch commented Mar 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Thermodynamic Rationale

Usage

Checklist

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

This comment was marked as resolved.

Uh oh!

Nietzsche-Ubermensch left a comment

Choose a reason for hiding this comment

Uh oh!

devin-ai-integration bot left a comment

Choose a reason for hiding this comment

Uh oh!

devin-ai-integration bot Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Mar 16, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Nietzsche-Ubermensch left a comment

Choose a reason for hiding this comment

Uh oh!

Nietzsche-Ubermensch left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as resolved.

Uh oh!

devin-ai-integration bot left a comment

Choose a reason for hiding this comment

Uh oh!

devin-ai-integration bot Mar 17, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Nietzsche-Ubermensch commented Mar 16, 2026 •

edited

Loading