docs: correct stale agent-facing references in AGENTS.md and agents index

[mc856]

Contribution path

• Small, safe change that does not need a tracking issue

Problem

AGENTS.md and docs/capabilities/agents/index.md have drifted from the code. Following them
sends you to a source file, a doc path and a prompt section that don't exist, a pytest path
that fails, and a blueprint table that names the wrong model and says three agentic blueprints
have no MCP server when all three start one.

Solution

Audited the agent-related statements of both files against main; changed only what turns out
false, plus dead references hit on the way (pytest path, worker_manager line, two doc paths).
Style, omissions and the broader overview stay. Three edits are follow-through of a correction —
marked below.

AGENTS.md

Was	Evidence
GPT-4o in Quick Start + blueprint table; no MCP server on three agentic blueprints	no blueprint passes a model (default: mcp_client.py:62); all three compose McpServer.blueprint() (g1/blueprints/agentic/_agentic_skills.py, manipulators/xarm/blueprints/agentic.py)
Tree: agents/agent.py	removed in #1657; the LangGraph agent lives under agents/mcp/ — note moved there (follow-through)
Tree: worker_manager_*.py, "Python / docker worker pools"	the docker manager went with #2224 (issue #2038); left are worker_manager.py (a Protocol) and worker_manager_python.py, so the glob lost its underscore
None return → agent hears "It has started…"	wrong trigger — mcp_client.py:199 fires on non-text content; mcp_server.py str()-ifies anything lacking agent_encode(), so None arrives as "None"; step 4 updated to match (follow-through)
Step 5: add to the # AVAILABLE SKILLS section	only the G1 prompt has that section
"Supported param types" closed list	LangChain builds a pydantic schema from the signature (core/module.py:448); a shipped skill takes list[dict[str, float]] (gps_nav_skill.py:58)
docs/development/dimos_run.md (twice); tree: docs/agents/	now docs/usage/cli.md; renamed by #2316 — see below
uv run pytest dimos/core/test_blueprints.py	now under dimos/core/coordination/

docs/capabilities/agents/index.md

Was	Evidence
gpt-4o, "Best quality" (twice)	mcp_client.py:62; the quality claim went with it (follow-through)
WebInput: port 7779, "optional" Whisper STT	web_human_input.py binds 5555 and constructs WhisperNode() unconditionally; 7779 is the WebsocketVisModule default
skills "must return a string or image"	the contract is agent_encode() (SkillResult, Image); the rest is str()-ified — shipped skills return bool, dict, PoseStamped, generators
tag_location(name), map_query(query), PersonFollowSkill, GPSNavSkill	now location_name, query_sentence, PersonFollowSkillContainer, GpsNavSkillContainer

The two "Agent system documentation" labels. The label was wrong from day one: at
26e06b7951, docs/agents/index.md opened with "These docs are mostly for coding agents" —
the confusion #2316 names. #2316 renamed the directory and re-pointed the Further Reading line
itself (7fb373ba5), chasing the rename without revisiting the label; the tree line still
points at the removed docs/agents/. By #2316's own criterion both belong to
docs/capabilities/agents/ — never linked from AGENTS.md, while docs/coding-agents/ is
linked twice. I changed the target; if it was deliberate, I'll relabel instead.

Left alone. The architecture blurb/diagram in index.md never matched the old agent module
either — that's a rewrite, not a correction. A few falsifiable errors elsewhere in AGENTS.md
(--viewer values, DIMOS_ prefix, blueprint-run snippet, log path) — happy to send
separately. #3156 edits the Further Reading line between the two changed here; whichever lands
second needs a one-line rebase.

How to Test

Docs only. The only fenced blocks touched are in AGENTS.md, which run-doc-codeblocks
doesn't walk (it runs on docs/ and README.md); no fence in index.md changed. Locally:
pre-commit run --from-ref origin/main --to-ref HEAD passes all hooks, and the corrected test
path runs — uv run pytest dimos/core/coordination/test_blueprints.py -v, 23 passed.

AI assistance

Claude Code (Fable 5) and Codex (5.6-Sol) did the audit passes, drafted the corrections and
this text, and cross-reviewed each other's findings. I verified every changed line against
main and stand behind each one.

Checklist

• I have read and approved the CLA.

[greptile-apps]

Greptile has paused reviews on this repository — it used its 100 free open-source review credits for this billing period. Reviews resume automatically on August 3. To continue before then, an organization admin can keep reviews running past the free credits — those bill as normal usage.

[spomichter]

@greptile can you review accuracy

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

@@            Coverage Diff             @@
##             main    #3210      +/-   ##
==========================================
- Coverage   74.90%   74.62%   -0.28%     
==========================================
  Files        1114     1114              
  Lines      106086   106086              
  Branches     9665     9665              
==========================================
- Hits        79461    79170     -291     
- Misses      23816    24133     +317     
+ Partials     2809     2783      -26     

Flag	Coverage Δ
OS-ubuntu-24.04-arm	68.50% <ø> (-0.01%)	⬇️
OS-ubuntu-latest	70.52% <ø> (ø)
Py-3.10	70.51% <ø> (+<0.01%)	⬆️
Py-3.11	70.50% <ø> (-0.01%)	⬇️
Py-3.12	70.51% <ø> (+<0.01%)	⬆️
Py-3.13	70.51% <ø> (-0.01%)	⬇️
Py-3.14	70.51% <ø> (-0.01%)	⬇️
Py-3.14t	70.51% <ø> (-0.01%)	⬇️
SelfHosted-macOS	?

Flags with carried forward coverage won't be shown. Click here to find out more.
see 24 files with indirect coverage changes

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[mc856]

The codecov/project drop here isn't from this PR — .codecov.yml carries forward SelfHosted-macos, but ci.yml uploads the flag as SelfHosted-${{ matrix.os }} → SelfHosted-macOS, and codecov flag matching is case-sensitive. So macOS coverage never carries forward on fork PRs (Linux and Large do — exact-case names in the same block). The missing 291 of 106,086 lines are exactly the −0.28%. One-character fix in #3253.