feat(codex): plexum-host MCP exposure + real consume mirror

Three-part wiring so codex agents get full dynamic-tool consume
support:

1. mcp-host subcommand in the plugin binary (routes argv[1] to
   mcpbridge.Run). Reads PLEXUM_MCP_SOCKET / PLEXUM_MCP_AGENT_ID
   from env, baked into codex's `mcp add --env` registration.

2. EnsureCodexMCPRegistered registers a per-agent stable server
   name (plexum-host-<sanitised>) lazily on first turn. Stable
   per-agent socket path via StableSocketPath so codex's
   registration entry and the per-turn bridge listener agree.

3. Post-turn ParseRolloutToolCalls + EmitCodexToolCalls walks the
   matching rollout-*-<thread_id>.jsonl and emits canonical events
   for every function_call + function_call_output pair using
   codex's REAL call_X ids. dynamic.jsonl now has block_ids that
   match codex's session, so consume mirror has real targets.

4. MutateCodexSession rewrites those response_item entries on
   consume: function_call.arguments → "{}" (heavy), output → marker.
   E2E verified: codex resume sees "...(tool called)" instead of
   the original output.
This commit is contained in:
h z
2026-06-01 20:14:06 +01:00
parent d075fc408a
commit a27f6bd067
5 changed files with 488 additions and 30 deletions

View File

@@ -21,6 +21,7 @@ import (
"path/filepath"
"git.hangman-lab.top/hzhang/Plexum-sdk-go/canonical"
"git.hangman-lab.top/hzhang/Plexum-sdk-go/mcpbridge"
plugin "git.hangman-lab.top/hzhang/Plexum-sdk-go/plugin"
"git.hangman-lab.top/hzhang/Plexum-openai-provider/internal/runner"
@@ -140,29 +141,33 @@ func (p *openaiPlugin) StreamWithAgent(ctx context.Context, modelID string, agen
}
// MutateSession is the session-mirror hook (decision #29x). codex
// stores transcripts at ~/.codex/sessions/<YYYY>/<MM>/<DD>/rollout-*
// -<thread_id>.jsonl as response_item lines whose payloads carry
// function_call / function_call_output entries keyed by `call_id`
// (not toolu_*) — so a Plexum tool_use_id like "toolu_X" won't ever
// match in there.
// rollout JSONL at ~/.codex/sessions/.../rollout-*-<thread_id>.jsonl
// is rewritten line-by-line: function_call_output entries matching
// the request's call_ids get their output replaced with the consume
// marker; heavy ops also stub function_call.arguments to "{}".
//
// v1 ships as a logged no-op: we resolve the file path so operator
// log can confirm wiring, but we don't rewrite. When Plexum's tool
// dispatch eventually routes through codex's native tool surface
// (preserving call_id round-trip), the rewrite logic plugs in here.
// Tool-call round-trip is provided by the runner package's
// ParseRolloutToolCalls + EmitCodexToolCalls (which run at end of
// each turn) — they push the codex call_X ids into Plexum's
// dynamic.jsonl, so the BlockMutations the host fires here all
// resolve.
func (p *openaiPlugin) MutateSession(ctx context.Context, req plugin.SessionMutateRequest) error {
path, err := runner.FindCodexSessionFile(req.Workspace)
if err != nil {
p.host.Log("debug", "codex session mutate: find failed", map[string]any{
"agent": req.AgentID, "err": err.Error(),
})
if len(req.Mutations) == 0 {
return nil
}
p.host.Log("info", "codex session mutate (no-op v1)", map[string]any{
"agent": req.AgentID,
"session_path": path,
"requested": len(req.Mutations),
"reason": "codex uses call_* ids; tool-call round-trip not yet wired",
mu := make([]runner.CodexMutation, 0, len(req.Mutations))
for _, m := range req.Mutations {
mu = append(mu, runner.CodexMutation{BlockID: m.BlockID, Op: m.Op})
}
touched, err := runner.MutateCodexSession(req.Workspace, mu)
if err != nil {
p.host.Log("warn", "codex session mutate failed", map[string]any{
"agent": req.AgentID, "err": err.Error(),
})
return err
}
p.host.Log("info", "codex session mutate complete", map[string]any{
"agent": req.AgentID, "touched": touched, "requested": len(req.Mutations),
})
return nil
}
@@ -182,8 +187,45 @@ func (p *openaiPlugin) mapModel(plexumModel string) string {
}
func main() {
if len(os.Args) > 1 && os.Args[1] == "mcp-host" {
if err := runMCPHost(os.Args[2:]); err != nil {
fmt.Fprintf(os.Stderr, "mcp-host: %v\n", err)
os.Exit(1)
}
return
}
if err := plugin.Serve(&openaiPlugin{}); err != nil {
fmt.Fprintf(os.Stderr, "plexum-openai-provider-plugin: %v\n", err)
os.Exit(1)
}
}
// runMCPHost: codex registers our binary's mcp-host subcommand via
// `codex mcp add --env PLEXUM_MCP_SOCKET=...`. The mcp-host process
// reads those env vars + serves stdio MCP to codex while dialing the
// plugin-side unix socket.
func runMCPHost(argv []string) error {
opts := mcpbridge.Opts{
SocketPath: os.Getenv("PLEXUM_MCP_SOCKET"),
AgentID: os.Getenv("PLEXUM_MCP_AGENT_ID"),
}
for i := 0; i < len(argv); i++ {
switch argv[i] {
case "--socket":
if i+1 >= len(argv) {
return errors.New("--socket needs value")
}
opts.SocketPath = argv[i+1]
i++
case "--agent-id":
if i+1 >= len(argv) {
return errors.New("--agent-id needs value")
}
opts.AgentID = argv[i+1]
i++
default:
return fmt.Errorf("unknown mcp-host flag: %s", argv[i])
}
}
return mcpbridge.Run(context.Background(), opts)
}