Files
synapbus/internal/harness/subprocess/config.go
T
Algis DumbrisandClaude Opus 4.6 04e4e3ca37 feat(harness): GEMINI.md support + cold-topic-explainer example
Adds everything needed to run a real multi-Gemini-model reactive agent
loop end-to-end on SynapBus.

internal/harness/subprocess/config.go:
  * AgentConfig.GeminiMD — content of workdir/GEMINI.md
  * MaterialiseAgentConfig writes GEMINI.md AND workdir/.gemini/settings.json
    when gemini_md is set. The settings file carries the same mcp_servers
    list as .mcp.json (so a Gemini child running from the workdir gets
    the exact MCP surface the operator configured, not the user's
    ~/.gemini/settings.json).
  * 2 new config_test cases: GEMINI.md + .gemini/settings.json round
    trip, GEMINI.md with empty mcp_servers still writes the settings
    file (explicitly clearing any inherited home config).

internal/harness/registry.go — BUG FIX:
  Resolve() now honours agent.HarnessName (explicit selection) BEFORE
  the inference chain, matching the reactor's own agentBackendKind
  policy. Previously, when multiple backends were registered,
  Resolve would pick "webhook" for every non-K8s agent — even when the
  agent's HarnessName was "subprocess" — because the original fallback
  chain put webhook first. This is why the first cold-topic-explainer
  run failed with "webhook: agent has no webhook config". Discovered
  during e2e testing.

internal/admin/socket.go + cmd/synapbus/admin.go:
  New `messages.send` admin command (socket + CLI). Sends a DM as any
  agent through the messaging service, bypassing the REST/MCP auth
  layers. Local-only via the admin Unix socket, so the threat model is
  "whoever can reach the socket is already admin".

  CLI:
    synapbus messages send --from X --to Y --body "..." [--priority N]
    synapbus messages send --from X --to Y --body-file path
    echo "..." | synapbus messages send --from X --to Y

  Used by the harness shell wrappers (so Gemini subprocess agents can
  DM each other) and by run_task.sh (to kick off a chain as a human
  user without implementing the REST session flow).

examples/cold-topic-explainer/ (NEW):
  Runnable 3-agent Gemini demo that exercises the subprocess harness,
  reactive triggers, recursive update, and all the preconditions (depth,
  budget, cooldown) end-to-end on a separate isolated synapbus instance.

  Layout:
    README.md         — usage + troubleshooting + cost notes
    start.sh          — builds synapbus, launches on port 18088 with
                        ./data, creates user + agents + harness configs,
                        marks agents reactive via sqlite3
    run_task.sh       — sends initial DM algis → decomposer-pro, polls
                        reactive_runs + messages for the FINAL: reply,
                        prints the result or dumps reactive_runs on
                        timeout for debugging
    stop.sh           — SIGTERM + 5s grace + SIGKILL fallback
    wrapper.sh        — shared subprocess local_command: reads
                        message.json + GEMINI.md, calls gemini headless
                        with --approval-mode yolo, strips the
                        "MCP issues detected" noise prefix, routes the
                        cleaned response to the next agent via
                        `synapbus messages send` over the admin socket
    configs/
      decomposer-pro.json  — gemini-3.1-pro-preview
                              (gemini-2.5-pro is currently capacity-
                              exhausted on Google's side)
      writer-flash.json     — gemini-2.5-flash
      critic-lite.json      — gemini-2.5-flash-lite
    .gitignore        — data/, bin/, synapbus.log, .synapbus.pid

  The wrapper does NOT rely on gemini's MCP tool-calling (which was
  unreliable in testing). Gemini is used as a pure text generator; the
  shell decides routing based on AGENT_ROLE:
    - decomposer → NEXT_AGENT (writer)
    - writer     → NEXT_AGENT (critic)
    - critic     → OWNER_AGENT if response starts with FINAL:,
                   REVISE_AGENT otherwise

E2E VERIFICATION (real run, real Gemini, not a mock):

  Topic: "how does SynapBus unify message delivery, reactive agent
          triggers, and harness runs on a single SQLite database?"

  Result (from data/synapbus.db after one successful run):
    harness_runs:
      #1 decomposer-pro subprocess success 106s
      #2 writer-flash    subprocess success 155s
      #3 critic-lite     subprocess success  10s
    reactive_runs: 3 rows, all succeeded, trigger_from chain:
      algis → decomposer-pro → writer-flash → critic-lite
    messages:
      #1 algis → decomposer-pro (topic)
      #2 decomposer-pro → writer-flash (Q1/Q2/Q3 breakdown)
      #3 writer-flash → critic-lite (3-paragraph draft)
      #4 critic-lite → algis (FINAL: + polished 3-paragraph explainer)

  Critic converged in one pass (all scores ≥ 8), so the writer↔critic
  refinement loop didn't need to recurse — but the plumbing for it
  (REVISE: branch in wrapper.sh, depth limit in reactor) is wired and
  ready. Flipping the critic's acceptance bar exercises the recursion.

  Full `go test ./...` remained green through all changes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-13 22:01:36 +03:00

276 lines
9.1 KiB
Go

package subprocess
import (
"encoding/json"
"fmt"
"os"
"path/filepath"
)
// AgentConfig is the typed shape of an agent's harness_config_json
// when that agent targets the subprocess backend. All fields are
// optional — an empty config is valid and simply does nothing.
//
// This config is stored on the agent row (harness_config_json) and
// edited via the admin CLI:
//
// synapbus harness config get --agent <name>
// synapbus harness config set --agent <name> --file config.json
//
// At Execute time the harness copies these fields into the per-run
// workdir so Claude Code / Gemini / Codex CLIs can find them via the
// conventions they already use (CLAUDE.md / AGENTS.md / .mcp.json
// in cwd; skills under .claude/skills; subagents under .claude/agents).
type AgentConfig struct {
// ClaudeMD is the content of CLAUDE.md written into workdir.
ClaudeMD string `json:"claude_md,omitempty"`
// AgentsMD is the content of AGENTS.md written into workdir. Used
// by Codex / Gemini CLIs that follow the AGENTS.md convention.
AgentsMD string `json:"agents_md,omitempty"`
// GeminiMD is the content of GEMINI.md written into workdir. The
// Gemini CLI reads this as the workspace system instructions.
// When set, MaterialiseAgentConfig ALSO writes a matching
// workdir/.gemini/settings.json containing the MCP servers below,
// so `gemini` invoked from the workdir sees both the role prompt
// and the agent's MCP tool surface in one step.
GeminiMD string `json:"gemini_md,omitempty"`
// MCPServers become the `mcpServers` object in workdir/.mcp.json.
// Claude Code picks this up from cwd automatically; other CLIs
// can be pointed at it with an explicit flag in local_command.
MCPServers []MCPServerSpec `json:"mcp_servers,omitempty"`
// Skills materialise as workdir/.claude/skills/<name>/SKILL.md.
// Each entry's Content is written verbatim — typically a YAML
// frontmatter block followed by markdown body, matching the
// superpowers skill format.
Skills []SkillSpec `json:"skills,omitempty"`
// Subagents materialise as workdir/.claude/agents/<name>.md.
Subagents []SubagentSpec `json:"subagents,omitempty"`
// Env is an extra env-var map layered on top of the agent's
// k8s_env_json and the caller-supplied Env. Last write wins, so
// use this field for agent-specific overrides.
Env map[string]string `json:"env,omitempty"`
}
// MCPServerSpec is one entry of the .mcp.json `mcpServers` object.
// We intentionally mirror Claude Code's shape so the marshalled file
// is drop-in compatible.
type MCPServerSpec struct {
// Name is the key in the mcpServers object (e.g. "synapbus").
Name string `json:"name"`
// Type is the transport. Claude Code accepts "stdio" (default),
// "http", and "sse". Empty = stdio.
Type string `json:"type,omitempty"`
// Stdio transport fields
Command string `json:"command,omitempty"`
Args []string `json:"args,omitempty"`
// HTTP / SSE transport fields
URL string `json:"url,omitempty"`
Headers map[string]string `json:"headers,omitempty"`
// Env is passed to the MCP server child (stdio only).
Env map[string]string `json:"env,omitempty"`
}
// SkillSpec is one entry under .claude/skills.
type SkillSpec struct {
Name string `json:"name"` // directory name — sanitised at write time
Content string `json:"content"` // SKILL.md body (typically YAML + md)
}
// SubagentSpec is one entry under .claude/agents.
type SubagentSpec struct {
Name string `json:"name"` // file basename — sanitised at write time
Content string `json:"content"` // markdown body
}
// ParseAgentConfig tolerates empty / nil / invalid JSON. Empty yields
// a zero-value AgentConfig; invalid JSON returns an error so operators
// see the typo at Execute time rather than silently booting a
// misconfigured agent.
func ParseAgentConfig(raw string) (AgentConfig, error) {
var cfg AgentConfig
if raw == "" {
return cfg, nil
}
if err := json.Unmarshal([]byte(raw), &cfg); err != nil {
return cfg, fmt.Errorf("subprocess: parse harness_config_json: %w", err)
}
return cfg, nil
}
// MaterialiseAgentConfig writes the config into workdir so the child
// process can find it. Existing files are overwritten; missing parent
// dirs are created. Returns the first write error encountered.
func MaterialiseAgentConfig(workdir string, cfg AgentConfig) error {
if cfg.ClaudeMD != "" {
if err := writeFile(filepath.Join(workdir, "CLAUDE.md"), cfg.ClaudeMD); err != nil {
return err
}
}
if cfg.AgentsMD != "" {
if err := writeFile(filepath.Join(workdir, "AGENTS.md"), cfg.AgentsMD); err != nil {
return err
}
}
if cfg.GeminiMD != "" {
if err := writeFile(filepath.Join(workdir, "GEMINI.md"), cfg.GeminiMD); err != nil {
return err
}
// A workspace-level .gemini/settings.json with mcpServers
// overrides the user's ~/.gemini/settings.json for MCP
// discovery, so the agent sees exactly the servers the
// operator configured here. We write an empty mcpServers
// block even when MCPServers is empty — this explicitly
// clears any inherited home-level servers.
if err := writeGeminiSettings(workdir, cfg.MCPServers); err != nil {
return err
}
}
if len(cfg.MCPServers) > 0 {
if err := writeMCPConfig(workdir, cfg.MCPServers); err != nil {
return err
}
}
for _, sk := range cfg.Skills {
if err := writeSkill(workdir, sk); err != nil {
return err
}
}
for _, sa := range cfg.Subagents {
if err := writeSubagent(workdir, sa); err != nil {
return err
}
}
return nil
}
// mcpConfigFile is the wire format of .mcp.json. We keep the outer
// struct small so marshalling produces exactly the shape Claude Code
// expects.
type mcpConfigFile struct {
MCPServers map[string]mcpServerEntry `json:"mcpServers"`
}
type mcpServerEntry struct {
Type string `json:"type,omitempty"`
Command string `json:"command,omitempty"`
Args []string `json:"args,omitempty"`
URL string `json:"url,omitempty"`
Headers map[string]string `json:"headers,omitempty"`
Env map[string]string `json:"env,omitempty"`
}
// geminiSettingsFile is the shape Gemini CLI expects at
// .gemini/settings.json. We keep it minimal — only mcpServers — so we
// don't clobber unrelated settings the user might merge in later.
type geminiSettingsFile struct {
MCPServers map[string]mcpServerEntry `json:"mcpServers"`
}
func writeGeminiSettings(workdir string, servers []MCPServerSpec) error {
out := geminiSettingsFile{MCPServers: map[string]mcpServerEntry{}}
for _, s := range servers {
if s.Name == "" {
continue
}
out.MCPServers[s.Name] = mcpServerEntry{
Type: s.Type,
Command: s.Command,
Args: s.Args,
URL: s.URL,
Headers: s.Headers,
Env: s.Env,
}
}
raw, err := json.MarshalIndent(out, "", " ")
if err != nil {
return fmt.Errorf("subprocess: marshal gemini settings: %w", err)
}
dir := filepath.Join(workdir, ".gemini")
if err := os.MkdirAll(dir, 0o755); err != nil {
return fmt.Errorf("subprocess: mkdir .gemini: %w", err)
}
return writeFile(filepath.Join(dir, "settings.json"), string(raw))
}
func writeMCPConfig(workdir string, servers []MCPServerSpec) error {
out := mcpConfigFile{MCPServers: map[string]mcpServerEntry{}}
for _, s := range servers {
if s.Name == "" {
continue
}
out.MCPServers[s.Name] = mcpServerEntry{
Type: s.Type,
Command: s.Command,
Args: s.Args,
URL: s.URL,
Headers: s.Headers,
Env: s.Env,
}
}
raw, err := json.MarshalIndent(out, "", " ")
if err != nil {
return fmt.Errorf("subprocess: marshal mcp config: %w", err)
}
return writeFile(filepath.Join(workdir, ".mcp.json"), string(raw))
}
func writeSkill(workdir string, sk SkillSpec) error {
name := sanitizeChildName(sk.Name)
if name == "" {
return fmt.Errorf("subprocess: skill has empty name")
}
dir := filepath.Join(workdir, ".claude", "skills", name)
if err := os.MkdirAll(dir, 0o755); err != nil {
return fmt.Errorf("subprocess: mkdir skill %q: %w", name, err)
}
return writeFile(filepath.Join(dir, "SKILL.md"), sk.Content)
}
func writeSubagent(workdir string, sa SubagentSpec) error {
name := sanitizeChildName(sa.Name)
if name == "" {
return fmt.Errorf("subprocess: subagent has empty name")
}
dir := filepath.Join(workdir, ".claude", "agents")
if err := os.MkdirAll(dir, 0o755); err != nil {
return fmt.Errorf("subprocess: mkdir agents: %w", err)
}
return writeFile(filepath.Join(dir, name+".md"), sa.Content)
}
func writeFile(path, content string) error {
if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
return fmt.Errorf("subprocess: mkdir %q: %w", filepath.Dir(path), err)
}
if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
return fmt.Errorf("subprocess: write %q: %w", path, err)
}
return nil
}
// sanitizeChildName strips path separators and other characters that
// would escape the workdir. Keeps [A-Za-z0-9._-] only.
func sanitizeChildName(s string) string {
out := make([]byte, 0, len(s))
for i := 0; i < len(s); i++ {
c := s[i]
switch {
case c >= 'a' && c <= 'z', c >= 'A' && c <= 'Z', c >= '0' && c <= '9':
out = append(out, c)
case c == '.' || c == '_' || c == '-':
out = append(out, c)
}
}
return string(out)
}