ai: migrate LLM backbone from Kimi CLI to Codex CLI

Retires the Moonshot/Kimi subscription in favour of the already-paid ChatGPT
plan. Both CLI wrappers now run `codex exec`; the kimi-agent container is gone.

adolf-llm + hindsight-llm:
- runKimi -> runCodex (`codex exec --json --skip-git-repo-check`), resume via
  `codex exec resume <thread_id>`.
- MCP moves from a per-session .mcp.json (a workaround for Kimi having no
  --mcp-config-file flag) to a $CODEX_HOME/config.toml generated once at
  startup from shared-mcp.json. Field translation is load-bearing:
  bearerTokenEnvVar -> bearer_token_env_var, enabledTools -> enabled_tools.
- approval_policy="never" + sandbox_mode required, or unattended turns block
  on an approval prompt nobody can answer.

kimi-agent removed. It was the ONLY large-tier deployment behind LiteLLM, so
deleting it outright would have silently degraded every large-tier request to
the local 4B model via the existing fallbacks. tier-large, the auto_router
complex-reasoning route and their fallbacks now point at the codex-backed
adolf-llm wrapper (model_name: codex-agent).

Three environment blockers fixed along the way:
- OpenAI geo-blocks this host (403 unsupported_country_region_territory).
  Both containers now egress via the host xray proxy, with NO_PROXY keeping
  MCP and *.alogins.net traffic off the tunnel.
- node:22-slim ships no system CA store; the Rust codex binary validates TLS
  against it, so every HTTPS call failed with a generic transport error while
  Node's own fetch worked. ca-certificates added to both images.
- `codex exec resume` rejects -C/--cd (plain `codex exec` accepts it), which
  broke follow-up turns while first turns succeeded.

Known regression: Kimi's managed-usage API has no Codex equivalent, so the
/usage route returns 501 and there is no quota probe for the codex model.
The two quota plugins degrade quietly to no output.

Also: stop tracking cognee.env (live LLM + JWT secrets) and gitignore it.
The secrets remain in earlier history and should be rotated.

Verified live: plain turn, SSE streaming, session resume, MCP tool call,
bearer-token MCP call, and completions through both LiteLLM routes.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014Y5QPagv4iun1ghpwM96Ff
This commit is contained in:
2026-08-01 06:13:27 +00:00
parent a27bae828a
commit 9094d71e2f
66 changed files with 653 additions and 851 deletions

74
ai/gpu_preload_check.sh Executable file
View File

@@ -0,0 +1,74 @@
#!/usr/bin/env bash
# gpu_preload_check.sh — GPU residency guard (design DESIGN-a2a-agents.md sec 3b)
#
# Never-evict set on the 8GB GTX 1070: bge-m3 (embedder) + tei-reranker.
# Evicting either silently breaks Hindsight recall (the memory plugin's
# recall timeout just skips injection, no error surfaced) — the whole
# reason this guard exists.
#
# Usage: gpu_preload_check.sh <requested_mib> [gpu_index]
# requested_mib — VRAM footprint (MiB) of the model/process about to load
# gpu_index — nvidia-smi GPU index (default 0)
#
# Exit 0 — safe to proceed, never-evict set stays resident with headroom.
# Exit 1 — reject: loading this would eat into or evict the never-evict set.
# Exit 2 — reject: never-evict set isn't even currently resident (abort,
# something is already wrong — don't compound it by loading more).
#
# This is a guard for callers (workers/scripts) that are about to pull a
# model onto the shared GPU. It does NOT itself load or evict anything.
set -euo pipefail
REQUESTED_MIB="${1:?usage: gpu_preload_check.sh <requested_mib> [gpu_index]}"
GPU_INDEX="${2:-0}"
# tei-reranker measured footprint (2026-07-26, jina-reranker-v2-base-multilingual
# fp16 on CUDA torch): ~1690 MiB resident. bge-m3 measured ~882 MiB via ollama.
# Keep these as a documented floor, not just "whatever's currently resident" —
# a transient dip during another process's own load shouldn't false-negative us.
RERANKER_FLOOR_MIB=1690
BGE_M3_FLOOR_MIB=882
NEVER_EVICT_FLOOR_MIB=$((RERANKER_FLOOR_MIB + BGE_M3_FLOOR_MIB))
log() { echo "[gpu_preload_check] $*" >&2; }
# 1. Confirm the never-evict set is actually resident right now.
reranker_up=0
if curl -fsS -m 3 "http://localhost:8014/info" >/dev/null 2>&1; then
reranker_up=1
fi
bge_m3_up=0
if docker exec ollama ollama ps 2>/dev/null | grep -q '^bge-m3'; then
bge_m3_up=1
fi
if [[ "$reranker_up" -ne 1 || "$bge_m3_up" -ne 1 ]]; then
log "REJECT: never-evict set not fully resident (tei-reranker up=$reranker_up, bge-m3 up=$bge_m3_up)."
log "Something is already wrong — fix that before loading anything else onto the GPU."
exit 2
fi
# 2. Check free VRAM and whether the requested load would eat into the
# never-evict floor.
free_mib=$(nvidia-smi --id="$GPU_INDEX" --query-gpu=memory.free --format=csv,noheader,nounits | tr -d ' ')
if [[ -z "$free_mib" ]]; then
log "REJECT: could not read nvidia-smi free memory for GPU $GPU_INDEX."
exit 1
fi
remaining_after_load=$((free_mib - REQUESTED_MIB))
log "free=${free_mib}MiB requested=${REQUESTED_MIB}MiB never_evict_floor=${NEVER_EVICT_FLOOR_MIB}MiB remaining_after_load=${remaining_after_load}MiB"
if (( remaining_after_load < 0 )); then
log "REJECT: requested load (${REQUESTED_MIB}MiB) exceeds current free VRAM (${free_mib}MiB)."
log "The kernel driver would have to evict something to fit it — on this box that means"
log "risking the never-evict set (bge-m3 + tei-reranker). Refusing."
exit 1
fi
log "OK: load fits in free VRAM without necessitating eviction of the never-evict set."
exit 0