`agentsMd` value silently ignored by claude / gemini presets — Claude Code only auto-discovers `CLAUDE.md`

[marvin-69-jpg]

Summary

The chart's agentsMd value mounts the supplied content at {{ agent.home }}/AGENTS.md. This works for the default kiro preset (Kiro CLI follows the agents.md spec), but is silently ignored by the claude preset because Claude Code only auto-discovers CLAUDE.md, not AGENTS.md. The bot ends up with no project-level instructions and no warning is emitted anywhere.

Repro

helm install openab openab/openab --version 0.5.1 \
  --set agent.preset=claude \
  --set discord.botToken=... \
  --set-string discord.allowedChannels[0]=... \
  --set-file agentsMd=./AGENTS.md

Where ./AGENTS.md says something distinctive like:

# Identity
You are openab-bot. When asked "who are you", reply exactly with the
string "openab-bot identity confirmed".

Then in Discord:

@AgentBroker who are you?

Expected: bot replies with the distinctive string from AGENTS.md.
Actual: bot replies with the generic "I'm Claude, an AI assistant made by Anthropic..." — AGENTS.md is never loaded into context.

Evidence

AGENTS.md is correctly mounted in the pod:

$ kubectl exec deployment/openab -- ls -la /home/node/AGENTS.md
-rw-r--r-- 1 root node 4033 Apr  8 03:30 /home/node/AGENTS.md

But Claude Code's CLI source only references AGENTS.md in the context of /init-style bootstrapping (reading other AI tools' configs to generate a fresh CLAUDE.md):

$ grep -oE '.{60}AGENTS\.md.{60}' \
    /usr/local/lib/node_modules/@anthropic-ai/claude-code/cli.js
... existing CLAUDE.md, .claude/rules/, AGENTS.md, .cursor/rules ...
... reads existing AI coding tool configs (AGENTS.md, .cursor/rules ...)

Claude Code's project memory auto-discovery only loads CLAUDE.md. The Claude Agent SDK that claude-agent-acp wraps inherits this behavior.

claude-agent-acp itself sets systemPrompt = { type: "preset", preset: "claude_code" } and never injects the contents of AGENTS.md:

$ grep -nE 'systemPrompt' \
    /usr/local/lib/node_modules/@agentclientprotocol/claude-agent-acp/dist/acp-agent.js
957:    let systemPrompt = { type: "preset", preset: "claude_code" };
958:    if (params._meta?.systemPrompt) { ... }

So unless openab passes _meta.systemPrompt.append, no custom prompt reaches Claude Code at all.

Affected presets

Preset	Auto-discovery file	agentsMd works?
kiro (default)	AGENTS.md	✅
codex	AGENTS.md	✅ (presumably)
claude	CLAUDE.md	❌
gemini	GEMINI.md	❌ (presumably)

Workaround (in PVC)

Symlink works because the PVC persists:

kubectl exec deployment/openab -- ln -sf AGENTS.md /home/node/CLAUDE.md

The symlink survives restarts. But this defeats the chart's intent of treating agentsMd as a one-stop, declarative config — every operator will hit this and have to discover the workaround independently.

Suggested fix

Make the chart aware of which file the active preset wants. Something like: in templates/_helpers.tpl, define openab.agent.memoryFilename that resolves to CLAUDE.md for the claude preset, GEMINI.md for gemini, and AGENTS.md otherwise; then use it as the subPath mountPath in deployment.yaml.

Or: mount the same content at both AGENTS.md and CLAUDE.md (and GEMINI.md) so any preset is covered. Slight redundancy but zero preset-awareness needed in the helper.

Either fix is fully backwards compatible — existing users on the kiro preset are unaffected.