ZeroClaw Feature Matrix Parity

You maintain the two competitor columns (OpenClaw, Hermes) of the feature and support matrix at docs/book/src/reference/feature-matrix.md. The ZeroClaw column is walked from the binary's own registries at docs-build time and is never hand-edited. Only the external columns live in docs/book/feature-matrix-parity.toml, and that file is the single reviewable source for every parity fact.

The cardinal rule: every cell you write is source-walked from the competitor's actual tree, never inferred from a folder name, an alias, or memory. A directory called signal is not proof of Signal support; the implementing module is.

Clone URLs

Both repositories are recorded in the parity TOML header (the Sources block). Read them from there so there is one source of truth:

• OpenClaw: https://github.com/openclaw/openclaw
• Hermes: https://github.com/NousResearch/hermes-agent

Always shallow-clone. These trees are large (OpenClaw ~340MB, Hermes ~200MB even shallow) and full history is never needed for a registry walk:

git clone --depth 1 https://github.com/openclaw/openclaw /tmp/parity-openclaw
git clone --depth 1 https://github.com/NousResearch/hermes-agent /tmp/parity-hermes

Clone both in parallel, and delete both when finished. For a quick single-file spot-check without a full clone, the GitHub contents API reads any tree file directly, which is enough to confirm one module's existence:

gh api "repos/openclaw/openclaw/contents/<path>" --jq '.content' | base64 -d

Status vocabulary

Cells use the page legend exactly: supported, partial, experimental, planned, none, unknown. The verdict turns on what the tree actually wires and runs, not on what a module merely defines. Finding a named module is necessary but not sufficient: before calling a cell supported, rule out the downgrade patterns below.

• supported when the tree ships a first-class module for it AND it is wired into a real path and on by default. A module that exists but fails one of the downgrade patterns below is not supported; it is partial or none per those rules.
• partial when support is real but qualified. Use it when the module is:
  • reachable only through a generic gateway rather than a dedicated module, or
  • covers a subset of the surface, or
  • default-off / opt-in (present but disabled unless explicitly enabled; read the config defaults, not just that the code path exists), or
  • selectively wired (live on some trigger/transport paths, dead on others).
• none when the tree has no equivalent, OR the only match is defined-but-dead (a module with no non-test callers) or a degraded/orphan path (it starts but produces no end-to-end behavior). A directory name or alias is not a match; the implementing, called module is.
• Use unknown when support has not been assessed or the available evidence is inconclusive. A missing TOML entry also renders unknown; record an explicit value and source notes when a bounded walk cannot establish a verdict.

The loose-alias downgrades in Step 3 are the same rule applied to naming: a speech- or image-only module is not LLM-chat supported, and a similarly named but distinct product (GitHub Copilot vs GitHub Models) is not the same slot.

Workflow

Step 1: Read the live page for ground truth

The rendered page is the authority for which rows exist, not the TOML keys (the TOML can be stale or incomplete). If a docs server is running, fetch it and extract every table verbatim:

curl -s "http://127.0.0.1:3000/en/reference/feature-matrix.html" -o /tmp/fm-live.html

Parse out the Channel, Provider slot, and Tool tables. The full walked key set (for providers this is dozens of slots, not a handful) is what you must account for. If no server is running, walk the registries directly: the channel inventory (ChannelsConfig::channels), canonical_model_provider_slots, and default_tools.

Step 2: Shallow-clone both competitors

Clone in parallel per the URLs above. Confirm both landed (non-trivial size) before walking; an interrupted clone leaves an empty stub.

Capture each clone's exact HEAD before walking:

git -C /tmp/parity-openclaw rev-parse HEAD
git -C /tmp/parity-hermes rev-parse HEAD

That SHA, not just the date, is what pins the verdicts to a reproducible commit. --depth 1 grabs whatever HEAD was current at clone time, so without the SHA a re-walk cannot tell what actually moved. You record both in the TOML Sources header in Step 5.

Step 3: Walk every row against real modules

For each row in each section, find the implementing module in each tree:

• OpenClaw: provider and platform plugins live in extensions/; channels also in src/channels/; tools in extensions/openshell and src/ file/glob/ripgrep code.
• Hermes: platforms in plugins/platforms/ and gateway/platforms/; providers in agent/*_adapter.py, the _AUX_DIRECT_API_BASE_URLS registry, the hermes_cli model-setup provider branches, and website/docs/integrations/providers.md; tools in tools/.

Resolve aliases (e.g. bedrock maps to amazon-bedrock, llamacpp to llama-cpp, glm to a z.ai/zhipu module), then VERIFY the matched module actually implements that capability. Downgrade loose matches: a module that only does speech or image generation is not LLM-chat support; a similarly named but distinct product (for example GitHub Copilot vs GitHub Models) is not the same slot. When in doubt, grep the module contents for the capability before calling it supported.

Step 4: Check the issue tracker for planned verdicts

A slot with no module in the tree is not automatically none. Before marking none, check for an accepted maintainer commitment or roadmap entry supporting planned. An open feature request alone does not establish a plan.

Search at the title level and filter to genuine requests, not incidental mentions in bug reports:

gh api repos/openclaw/openclaw --jq '.open_issues_count'   # confirm the repo and tracker are live
gh search issues --repo openclaw/openclaw --state open "<slot> in:title" --json number,title,url
gh search issues --repo NousResearch/hermes-agent --state open "<slot> in:title" --json number,title,url

A raw keyword count is noise: an issue that merely mentions "bedrock" is not a roadmap commitment. Read candidate issue bodies and comments for an explicit maintainer acceptance, or inspect the maintained roadmap entry; a requester's proposal alone is insufficient. Bug reports and merged feature requests are leads to inspect the implementing tree and wiring, not proof of supported. Record the supporting issue/comment or roadmap reference and check date in the TOML header so the verdict is auditable. Use unknown when the available evidence cannot establish a verdict.

Step 5: Write the TOML

Every key must correspond to a walked row, in the page's canonical order. Keep the header's Sources block current (repo URLs, the pinned clone-HEAD SHA per repo from Step 2, per-section walk paths, and the checked date). For a brand-new comparison concept that is not part of the walked channel/provider/tool registries (SOP, for instance), it does not belong in this TOML; hand-author a small section in feature-matrix.md itself with a static table and note that it is hand-recorded rather than code-walked.

Step 6: Verify with the guard test

The parity join has a hard-fail CI guard. A TOML key the walk no longer produces, or a walked row with a stale key, fails the docs build. Run it:

cargo test -p xtask --lib feature_matrix

Both parity_join_produces_rows_without_stale_keys and channel_column_is_all_supported_from_walk must pass.

Step 7: Regenerate and confirm live

Regenerate the gitignored snippets and confirm zero stray Unknown/❓ remain where you filled cells:

cargo run -p xtask --bin mdbook -- preprocess

Re-fetch the live page and diff the rendered tables against your intended verdicts. Do not claim a cell from memory; read the rendered HTML.

Step 8: Clean up and commit

Delete both clones. Commit with a scoped, multi-line message that records the supported/none counts per column and names any loose matches you downgraded and why, so the verdicts are auditable from the log.

Rules

• Every cell is source-walked from the competitor's actual tree. Never infer from a directory name, an alias, or prior memory.
• Always shallow-clone (--depth 1) and always delete the clones when done.
• Read the live rendered page for the row set, not the existing TOML keys.
• Verify loose alias matches against module contents before marking supported; downgrade speech/image-only or different-product matches to none.
• Before marking a slot none, check for an accepted maintainer commitment or roadmap entry supporting planned. Use the tree and wiring to establish implementation status; issues alone do not prove support.
• Keep the TOML Sources header current, including the checked date and the pinned per-repo commit SHA the columns were walked against.
• Run cargo test -p xtask --lib feature_matrix before committing; the guard is a hard CI fail.
• Confirm the rendered tables from the live HTML, not from what you expect.
• Never touch the ZeroClaw column; it is code-walked and regenerates itself.