Agent capabilities

Read this before you conclude a dashboard panel is broken.

Supported agents do not record the same things. A panel that is empty for one agent and full for another is usually reporting that difference correctly, not failing.

Three agents write a local transcript, and whodunit reads it: Claude Code, Codex CLI and Antigravity CLI. Two more declare themselves in the commit message and write no transcript at all — GitHub Copilot and Cursor — and those are covered further down, along with what a declaration cannot tell you.

Every row below was verified by reading real transcripts on a real machine — 1,862 Claude Code transcripts, 125 Codex rollouts, 8 Antigravity databases — rather than from vendor documentation, which describes intent rather than what lands on disk.

What each agent reports

The three that write a transcript. Copilot and Cursor are declaration-only and appear in their own section.

	Claude Code	Codex CLI	Antigravity (agy)
Confidence ceiling	intersected	intersected	intersected
Which files were edited	✅	✅	✅
The exact lines produced	✅	✅	✅
Model	✅	✅	✅
Branch	✅	✅	❌ none recorded
MCP server	✅	✅	✅
Accept / reject outcome	✅	✅	✅
Human edited the output	✅	❌	❌
Input / output tokens	✅	✅	❌ none at all
Cache reads	✅	✅	❌
Cache writes	✅	❌	❌
Reasoning tokens	❌	✅	❌
Per-turn timing	❌	✅	❌
Reasoning effort	✅	✅	partial
Permission mode	✅	✅	partial
Context compaction	✅	✅	❌

What this means in practice

Antigravity contributes no cost data. Not "reports zero" — records none. Every token and timing column is NULL for an agy session, and panels built on them exclude it from the denominator rather than counting it as free. An agy session showing 0 tokens would make Antigravity look like the cheapest agent available, which is the opposite of a measurement.

Only Claude Code reports whether a human edited the output. This is the most interesting signal in the set — the difference between "the agent wrote this" and "the agent wrote this and it was kept" — and it exists for one agent out of three. The panel's denominator counts only calls that carried the signal, so the other two do not dilute the rate.

Only Claude Code reports changed_by. It is the one agent that says whether a human edited its output, so an observed commit from Codex or agy cannot say whether the text was revised by a person or replaced by a later turn. The key is absent there permanently, not pending.

Only Codex reports timing. Claude Code and Antigravity record none, so a latency panel is structurally empty for them. It says so rather than rendering zero, which would make them appear instantaneous.

Codex reports cache reads but never writes. That matters more than it looks: the cache write payback figure is uncomputable for Codex, and a missing denominator is not a payback of zero. See What the numbers mean.

Antigravity records no branch. Its store does contain branch-shaped strings, but they are a tracker's suggested branch name inside an MCP response — not the checked-out branch. Verified absent rather than merely unread, which is why the column is NULL instead of being filled from something that looks close enough.

Agents that declare themselves

Some agents write their own trailer into the commit and leave no local transcript. whodunit reads those, and grades them at the weakest rung the spec has.

Agent	Signal	Ceiling
GitHub Copilot	Agent-Logs-Url, Co-authored-by: Copilot	declared
Cursor	Made-with: Cursor	declared

The ceiling is the strongest state that agent can ever reach. A transcript agent can reach intersected; a declaring agent cannot, because there is no transcript to intersect against. The full ladder, and the four reasons a commit carries no state at all, are in What the numbers mean.

What a declared-only agent cannot report, and why an empty panel is not a zero:

	Copilot	Cursor
That an agent was involved	✅	✅
Which files were edited	❌	❌
The exact lines produced	❌	❌
ratio — the agent's share of the commit	❌	❌
Tokens, cost, session shape	❌	❌
Model	❌	❌

None of that is missing data waiting to be collected. A trailer states that an agent took part; it says nothing about which lines, so every panel built on line counts or token use is structurally empty for these agents rather than reporting zero. A ratio of 0.00 would assert the agent contributed nothing, which is the opposite of what a declaration means.

Why declared and not something stronger

A trailer is the author's own assertion about their own commit, verified by nothing. That is the definition of declared — "weakest: the author declared it, nothing verified it".

The distinction is not academic. In April 2026 VS Code began adding Co-authored-by: Copilot by default, and reverted it a month later after the line appeared on commits made with Copilot switched off. A self-applied label is evidence, and it is the least of it.

Some tools treat the same class of signal as high confidence. whodunit grades it lowest, and the method mix exists so a reader never has to take a declared figure for a measured one.

A transcript always wins. If an agent that writes a transcript also declares itself, the transcript's observed or intersected finding is what gets stamped — the trailer is a fallback, not a shortcut.

What is read

Only the commit message and its author fields. Not the diff, not the files, not anything the agent said. A declaration is a line of metadata, and the reader is never given anything else.

Agents not supported

Agent	Status	Why
Cursor	Partial	declared via its trailer; line-level attribution needs its local store — see below
Windsurf	Blocked	Chat history in an undocumented store with no compatibility contract
Antigravity IDE	Partial	Message bodies are encrypted; the CLI (agy) is supported
Gemini CLI	Blocked	Free tier removed 2026-06-18; no account to verify against

Cursor is a partial case rather than a blocked one. Its trailer is read, so its commits are attributed at declared. Reaching intersected would need its local session store, and a 2026-08-12 survey found that store readable rather than opaque — unencrypted session blobs, and a separate database in which Cursor computes its own commit-level AI percentage.

What blocks the step up is narrower than the store: no file-edit record was found in any of 111 readable session databases, so the shape of a Cursor edit is unverified. Building a parser against a format nobody has observed would be inventing one, which is the same reason Gemini CLI is parked.

The adapter interface is open for a contribution.

The measured detail

This page is the version you read before installing. The full inventory — every field, with occurrence counts and a privacy classification — is in docs/adapters/field-inventory.md and the per-agent store formats are in docs/adapters/agent-support.md.