Agent integration matrix — 2026-09-15/16 (sanitized field report)¶

Why this exists. Every agent integration this repo claims was verified at some point, but the claims lived in different places (installer output, docs/integrations/status.md, issue comments) and none of them said how to re-check it. This report is the consolidated evidence: one row per agent, the exact command that produced the evidence, and what the agent actually did with MisakaNet.

Sanitized. No tokens, keys or credentials (none were ever printed); no personal names, e-mail addresses or real host names; absolute home paths are written as ~ or <workspace>. Provider and model names are kept because they are part of the reproduction recipe, not secrets.

Environment: one WSL2 Linux machine, Node 22, five coding agents installed side by side. All timestamps are UTC; the local machine runs UTC+8.

1. Results¶

Agent	Version	Transport wired	Tools discovered	Called the tool?	Full chain (search → get_lesson → answer)
claude-haha (Claude Code)	—	MCP http + `settings.json` hooks	7	✅ ×2	✅ quoted the lesson back to the user
hermes	—	`~/.hermes/config.yaml` MCP entry	7 (`registered 7 tool(s)`)	✅ ×1	⚠️ at test time `no_match` — traced to a data bug (see §3), not to hermes
openclaw	—	MCP entry (file-configured)	✅	✅ ×1	✅ `toolSummary: {calls: 3, failures: 0}`
codex	0.154.0	`~/.codex/config.toml` (`streamable-http`)	7	✅	✅ live session, see §2.4
codewhale	0.9.7	`~/.codewhale/mcp.json` (`bearer_token_env_var`)	7 (`codewhale mcp tools`)	✅	✅ latest run: search → get_lesson
DSH (plugin channel)	0.1.x	bundle row → public endpoint	7	n/a (plugin, not an agent loop)	✅ via `dsh --dump-config` + install

2. Per-agent evidence¶

2.1 claude-haha / hermes / openclaw (three agents in parallel tmux windows)¶

Three windows, one question (pip install timeout 是什么原因？), raw logs on disk; all three exited 0 and all three called misakanet_search. Claude went further — search ×2, get_lesson ×3 — and repeated a lesson in plain language ("我参考了别人的一条经验：…") exactly as the rules block asks. The earlier, separate report has the raw tool-call sequences: docs/field-reports/agent-chain-tmux-2026-09-15.md.

2.2 hermes on the same question returned `no_match`¶

This is worth keeping in the record because it was the most useful finding of the day and the first explanation for it was wrong. The same query with a domain filter returned 0 hits while the unfiltered one returned 3. The filter was innocent: the corpus metadata had been written with slug titles and parent-directory domains for 91% of lessons (a CI job parsed YAML frontmatter without PyYAML and the parser swallowed the failure into an empty dict). Fixed, re-synced and re-verified; domain="python" then answered 3 hits where it had answered none.

2.3 openclaw¶

toolSummary: {"calls": 3, "tools": ["misakanet__misakanet_search", "misakanet__misakanet_get_lesson"], "failures": 0}.

2.4 codex — config, prompt and a live call¶

Three independent checks, each one strictly stronger than the last:

codex mcp list      # → misakanet | https://misakanet.org/mcp | enabled | Bearer token
codex doctor        # → config.toml parse ok · MCP servers 1 · 1 streamable_http · 0 disabled
codex debug prompt-input ""   # → a `# AGENTS.md instructions` item carrying our rule block

The third proves the user-level ~/.codex/AGENTS.md block reaches the model rather than merely existing on disk. Then a real session (codex exec --json, driven by a local provider over the Responses API) produced:

[mcp_tool_call] server=misakanet tool=misakanet_search
   args:   {"query":"pip install timeout","detail":"summary","top":5}
   result: {"results":[{"id":"…-pip-timeout-proxy","title":"pip install falha com ReadTimeoutError…",
                        "problem":"O `pip install` falha com `ReadTimeoutError`…","freshness":"recent"}]}

Two reproduction notes: codex 0.154.0 rejects wire_api = "chat" (Responses API only — the provider used here serves /v1/responses, so "responses" is the working value), and plain codex exec output does not name the tool (use --json; without it a run that merely imitates the rule's phrasing looks like a run that used a lesson).

2.5 codewhale — the newest integration, and the one with variance¶

codewhale mcp add misakanet --url https://misakanet.org/mcp --bearer-token-env-var MISAKANET_TOKEN
codewhale mcp list    # → misakanet [enabled auth=bearer-token] https://misakanet.org/mcp
codewhale mcp tools   # → the 7 mcp_misakanet_* tools, with their descriptions

Config lands in ~/.codewhale/mcp.json (a per-server block with url, bearer_token_env_var, timeouts, enabled). A live codewhale exec --auto --output-format stream-json run then did:

tool_search            → tool_references: [mcp_misakanet_misakanet_search, …]
mcp_misakanet_misakanet_search  {"query":"pip install timeout"}
   → {"results":[{"id":"…-pip-timeout-proxy","title":"pip install falha com ReadTimeoutError…",
                  "problem":"O `pip install`…","freshness":"recent"}]}
mcp_misakanet_misakanet_get_lesson

Honest sample size. With the production rules block in place: one run followed it fully (search → get_lesson), one run ignored it and spent 56 bash calls exploring instead, and one run was made in a workspace with no trust entry (codewhale's trust is per-project — the only trusted path on this machine is a Windows-mounted directory). So: the plumbing is proven, and following the rules shows agent-side variance; the untrusted-workspace run is a plausible cause, not a proven one. When the same block was rewritten into three imperative lines with an unmistakable marker, the agent emitted the marker and called the tool — i.e. it reads the workspace AGENTS.md, and a crisper instruction is followed more reliably.

2.6 DSH (plugin channel, not an agent loop)¶

Verified through the real loader and a real npm install rather than by reading the spec: dsh --profile web --dump-config composes the row as written; installing the published package makes DSH apply the bundle itself (id: misakanet-mcp, transport: streamable-http → public endpoint); and the endpoint answers the handshake the client performs (initialize → tools/list = 7 tools).

3. What these runs caught that unit tests did not¶

Finding	How the agents surfaced it
91% of the corpus projected with slug titles + directory domains	hermes' `no_match` on the same question the other two agents answered
`misakanet_search` hits carried no `problem`/`fix`/`tags`	codex and codewhale results showed a title and almost nothing else
The search index kept serving a pre-sync corpus for up to 20h	"the data is fixed, why is search still wrong" — the index had paired a new sync stamp with a cached corpus
The installer never refreshed an existing hook	the 14-day upgrade nudge could not arrive, and re-running the installer changed nothing
A schema validator could not read JSON-plus-`provenance` frontmatter	CI went red on lessons that had been valid all along, once a domain rewrite touched them

All five are fixed and re-verified; the durable records are docs/maintainer/handoff-2026-09-15.md §10–§15 and the issues referenced there.

4. Open items from this matrix¶

The installer has no codewhale target. Five agents are supported (claude, codex, hermes, openclaw, dsh); codewhale is configured by hand today, although it is arguably the easiest of the five (codewhale mcp add … --bearer-token-env-var, plus an AGENTS.md block).
Rules text strength matters. The production block is long and polite; codewhale followed it in one run and not in another, and followed a three-line imperative version immediately. Worth trimming or front-loading the imperative part for all agents.
Codex has no user-level lifecycle hook (0.154.0 hooks are admin-managed via requirements.toml), so the round-20 checkpoint reminder is rule-driven there.
"Trusted workspace" is an unwritten prerequisite for instruction blocks in at least one agent (codewhale); the installer does not mention it.