PR #54448: fix: preserve Kimi tool call arguments instead of clearing valid JSON

• Repository: openclaw/openclaw
• Author: vinsew
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/54448

Description (problem / solution / changelog)

Summary

Fixes three related bugs in the Kimi (kimi-coding/k2p5) tool call argument repair pipeline that cause tool calls to fail with empty or undefined arguments:

• tryParseMalformedToolCallArguments treats valid JSON as unfixable: When accumulated streaming JSON is already valid, JSON.parse(raw) succeeds but the function returns undefined — the same signal as "unrepairable". The caller then clears arguments, wiping correct tool parameters.
• clearToolCallArgumentsInMessage assigns {} instead of undefined: The empty object pollutes shared references. Downstream validators see a present-but-empty object and report must have required property 'command' instead of recognizing arguments as absent.
• else branch in wrapStreamRepairMalformedToolCallArguments unconditionally clears arguments: When repair returns undefined, arguments may already be correctly set via content_block_start.input (Kimi's non-incremental streaming path). Clearing here wipes them. The fix lets content_block_stop handle final argument resolution.

Root cause

Kimi k2p5 has two streaming behaviors:

1. Incremental: content_block_start with input: {}, then input_json_delta events build up arguments
2. Non-incremental: content_block_start with full input: {path: "..."}, no input_json_delta at all

The repair wrapper assumed path 1 exclusively. In path 2, partialJson stays empty, tryParseMalformedToolCallArguments("") returns undefined, and the else branch wiped the already-correct arguments.

Changes

Reproduction

With kimi-coding/k2p5:

# Before fix:
Validation failed for tool "exec":
  - command: must have required property 'command'
Received arguments: {}

# Or:
read tool called without path: toolCallId=tool_xxx argsType=undefined

Note on remaining pi-ai issues

Two additional Kimi tool call bugs exist in the upstream @mariozechner/pi-ai package (not addressed here):

• content_block_stop overwrites block.arguments with parseStreamingJson("") → {} even when partialJson is empty
• A "Kimi API bug workaround" in content_block_start converts input: {} to undefined

These require a separate fix in pi-ai. This PR addresses all bugs within OpenClaw's own codebase.

Test plan

• All 95 existing tests in attempt.test.ts pass
• New test: valid JSON arguments are preserved through the repair pipeline
• New test: arguments set by content_block_start.input survive incomplete delta attempts
• Updated test: trailing suffix growing too long preserves last valid repair instead of wiping
• TypeScript type check passes (pnpm tsgo)
• Full lint/format check passes (pnpm check)
• Manual verification with kimi-coding/k2p5 on Telegram

Fixes #54442 Fixes #53747 Refs #54366

🤖 Generated with Claude Code

Changed files

• src/agents/pi-embedded-runner/run/attempt.test.ts (modified, +93/-2)
• src/agents/pi-embedded-runner/run/attempt.ts (modified, +13/-23)

PR #56697: fix(tools): include parameter examples in validation errors to help models self-correct

• Repository: openclaw/openclaw
• Author: Co-Messi
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/56697

Description (problem / solution / changelog)

Summary

Include tool-specific parameter examples in validation error messages so models can self-correct on retry.

Problem

Models like Kimi k2p5 intermittently send empty objects ({}) instead of required tool parameters. The current error message says:

Missing required parameter: path alias. Supply correct parameters before retrying.

This is too vague — the model retries with the same empty payload, creating a loop of 5+ failed attempts.

Fix

Add a TOOL_PARAM_EXAMPLES map for read/write/edit tools. When parameter validation fails, the error now includes a concrete example:

Missing required parameter: path alias. Supply correct parameters before retrying. Example: read({"path": "/absolute/path/to/file"})

This gives the model a template to follow on the next attempt, significantly improving self-correction rates for models with weak tool-call formatting.

One-file change, 14 lines added, 4 modified. No behavior change for well-formed tool calls.

Test plan

• Change is additive — parameterValidationError gains an optional second parameter with no default behavior change
• Existing tool param tests unaffected (validated locally)
• Manual verification: Kimi k2p5 self-corrects faster with example in error response

Fixes #56376 Related: #55005, #55039, #53747, #54366

Changed files

• src/agents/pi-tools.params.ts (modified, +14/-4)

PR #59440: fix: normalize Kimi anthropic tool payloads

• Repository: openclaw/openclaw
• Author: obviyus
• State: closed | merged: True
• Link: https://github.com/openclaw/openclaw/pull/59440

Description (problem / solution / changelog)

Summary

• mark Kimi and Kimi Coding as OpenAI-function-style Anthropic tool payload providers
• align the bundled plugin capability metadata with the fallback capability registry
• add regression coverage for Kimi payload normalization on the Anthropic Messages path

Fixes #59327 Fixes #54442 Fixes #58817 Fixes #53747 Fixes #56916

Changed files

• CHANGELOG.md (modified, +1/-0)
• extensions/kimi-coding/index.ts (modified, +2/-0)
• src/agents/pi-embedded-runner-extraparams.test.ts (modified, +71/-8)
• src/agents/provider-capabilities.test.ts (modified, +13/-6)
• src/agents/provider-capabilities.ts (modified, +2/-0)