mirror of
https://github.com/Kaelio/ktx.git
synced 2026-06-16 08:25:14 +02:00
494618ab14
* feat: add codex sdk runner foundation * feat: parse codex runtime events * feat: expose codex runtime mcp tools * feat: add codex llm runtime * feat: wire codex llm backend * test: avoid Array.fromAsync in codex runner test * docs: document codex llm backend * fix: tighten codex runtime config ownership * fix: use codex sdk env and thread options * fix: parse codex sdk event shapes * test: add codex backend live smoke * docs: clarify codex backend isolation * fix: drive codex loop metrics from mcp events * fix: enforce codex local step budget * docs: disclose codex isolation limits * fix: count all codex agent steps and stream step callbacks live The agent-loop step budget only counted completed mcp_tool_call items, so built-in command_execution steps (which the public Codex SDK/CLI surface can still expose) never decremented the budget, letting ingest/reconciliation run past stepBudget until Codex stopped on its own. onStepFinish was also replayed only after the whole stream drained, so live work_unit_step / reconciliation progress appeared stuck until the Codex process exited. collectEvents is now the single live step accumulator: it counts every completed agent-action item via a shared isCompletedAgentStep predicate (command_execution, mcp_tool_call, file_change, web_search), fires onStepFinish as each step completes, and enforces the budget on that broader count. A no-tool turn still counts as one step. toolFailures stays MCP-specific, since a non-zero command exit is normal agent exploration, not a loop failure. * test: align ingest llm-guard assertions with codex backend The skip-llm ingest guard message now lists codex as a valid backend and mentions a Claude Code/Codex session plus a codex setup hint, but this slow suite test still asserted the pre-codex wording. Update it to match the production message (already covered by the local-bundle-runtime unit test) and add the codex setup-line assertion. * fix: treat codex error:null tool calls as success The Codex SDK serializes error: null on successful mcp_tool_call items, so the failure check (item.error !== undefined) flagged every successful tool call as failed with the empty-payload default "Codex turn failed". This killed every ingest work unit under the codex backend before it could produce a patch. Key on status === 'failed' (authoritative, always set) and only treat a populated error object as a failure. Add a regression test built from a verbatim real-SDK event capture. * fix: default codex backend to gpt-5.5 and report real probe errors The previous default gpt-5.3-codex is an API-key-only model that the OpenAI API rejects under ChatGPT-account (subscription) auth, so codex status/setup failed with a misleading "authentication is not usable" message even though auth was fine. - Default codex model is now gpt-5.5 (works on both subscription and API-key auth); the curated setup picker offers gpt-5.5 / gpt-5.4 / gpt-5.4-mini and keeps free-form entry for account-specific ids (e.g. gpt-5.3-codex-spark). - runCodexAuthProbe now distinguishes "model not available" from an auth failure and surfaces the real API error: collectEvents retains stream events when the SDK throws on a non-zero exit, and the API error JSON envelope is unwrapped to its human-readable message. - The Codex isolation warning now renders inside the clack setup frame. - Docs updated to gpt-5.5 with a note that *-codex ids require API-key auth. * fix: require llm.models.default in status and match codex probe remediation Status reported a project ready when a non-none LLM backend was configured without llm.models.default, but the runtime (resolveModelSlots) hard-requires it, so ingest/scan/memory threw after `ktx status` said the project was usable. buildLlmStatus now fails for any non-none backend missing models.default and no longer invents a fallback model for claude-code/codex. Codex probe failures now carry a category-matched fix: a model-access failure steers the user at llm.models.default instead of the auth/install remediation. runCodexAuthProbe returns the fix and status consumes it; the message stays self-sufficient so setup output is unchanged. Docs: README now lists the codex backend and local Codex auth; ktx-setup.mdx states --llm-model only accepts codex/default or gpt-*/codex-* ids. Repaired four doctor fixtures that configured a backend without models.default (the now-correctly-blocked config) and added coverage for the new behavior.
90 lines
2.9 KiB
Text
90 lines
2.9 KiB
Text
---
|
|
title: LLM configuration
|
|
description: Configure ktx LLM providers, model roles, and prompt caching.
|
|
---
|
|
|
|
Configure text generation, structured extraction, and ingest or memory loops in
|
|
the top-level `llm` block.
|
|
|
|
## Backends
|
|
|
|
Set `llm.provider.backend` to one of these values:
|
|
|
|
- `anthropic`: Use the Anthropic API through `ANTHROPIC_API_KEY` or the
|
|
configured `api_key` reference.
|
|
- `vertex`: Use Vertex AI Anthropic models through Google Cloud credentials.
|
|
- `gateway`: Use AI Gateway-compatible Anthropic model ids.
|
|
- `claude-code`: Use your local Claude Code session through the Claude Agent
|
|
SDK. **ktx** strips provider-routing environment variables from child processes.
|
|
- `codex`: Use your local Codex authentication through the Codex SDK.
|
|
|
|
## Claude Code
|
|
|
|
Use aliases or full Claude model IDs in `llm.models`:
|
|
|
|
```yaml
|
|
llm:
|
|
provider:
|
|
backend: claude-code
|
|
models:
|
|
default: sonnet
|
|
triage: haiku
|
|
candidateExtraction: sonnet
|
|
curator: sonnet
|
|
reconcile: sonnet
|
|
repair: sonnet
|
|
```
|
|
|
|
During setup, choose the backend interactively or pass the model in automation:
|
|
|
|
```bash
|
|
ktx setup --llm-backend claude-code --llm-model opus --no-input
|
|
```
|
|
|
|
For Claude Code, `sonnet`, `opus`, and `haiku` map to **ktx** defaults. Full Claude
|
|
model IDs are also accepted.
|
|
|
|
`claude-code` exposes only **ktx** MCP tools for the current agent loop. SDK init
|
|
metadata may still list host slash commands, skills, and subagents; **ktx** does not
|
|
grant execution access to them.
|
|
|
|
## Codex backend
|
|
|
|
Use `codex` when you want **ktx** to run LLM-backed workflows through your
|
|
local Codex authentication instead of a direct provider API key.
|
|
|
|
```yaml
|
|
llm:
|
|
provider:
|
|
backend: codex
|
|
models:
|
|
default: gpt-5.5
|
|
```
|
|
|
|
Configure it non-interactively:
|
|
|
|
```bash
|
|
ktx setup --llm-backend codex --llm-model gpt-5.5 --no-input
|
|
```
|
|
|
|
This is separate from Codex agent-client setup. `ktx setup --agents --target
|
|
codex` installs instructions and MCP access for an end-user Codex session.
|
|
`ktx setup --llm-backend codex` makes **ktx** itself execute ingest, scan
|
|
enrichment, memory, and other LLM-backed work through Codex.
|
|
|
|
During runtime loops, **ktx** starts a temporary loopback MCP server for the
|
|
current run, exposes only the tools passed to that run, asks Codex to use a
|
|
read-only sandbox, sets `approval_policy=never`, auto-approves only those
|
|
run-scoped MCP tools, and disables Codex web search.
|
|
|
|
Codex backend isolation is currently limited by the public Codex SDK and CLI
|
|
surface. Codex may still load user Codex config and built-in command execution
|
|
or read-only file capabilities. Use `llm.provider.backend: claude-code` when
|
|
you need stricter Claude-Code-style runtime tool isolation, or remove host
|
|
Codex MCP and tool config before running untrusted prompts through the `codex`
|
|
backend.
|
|
|
|
## Prompt caching
|
|
|
|
`llm.promptCaching` has partial parity on `claude-code`. Status and doctor warn
|
|
when the Claude Agent SDK backend ignores configured cache fields.
|