What's captured — and what never is

Agent Karma’s entire value rests on trust, so it earns that trust the only honest way: by recording the shape of your work — never its content. It watches that you validated, not what you wrote. Everything below stays on your machine; there is no account, no telemetry, and a build-time guard that fails the release if a single network call sneaks into the bundle.

The one rule: metadata, never content

Agent Karma records	Agent Karma never touches
That a file changed — its name, extension, and whether it’s a test file	Your source code — file contents are never read
That a validation ran — its type (Test / Build / Lint / Type Check / Security) and result	The raw command string — it’s classified, then discarded
That a commit landed — its short SHA	The commit message, diff, or file list
Git diff counts — files changed, lines added/removed	The diff content itself
Your intent text — only if you leave capturePromptText on (default), so cards read back meaningfully	Terminal output — capture is hard-wired off, regardless of settings

If you turn capturePromptText off, your Karma Score and cards are still computed — the intent string is simply redacted.

Every event it records

A session is just an ordered list of small, typed events. Here is the complete set:

Event	What it carries	Captured how
session.started / session.ended	nothing — just the boundary	automatic
intent.captured	your intent text (only if capturePromptText is on)	you, at start
prompt.scored	a prompt-clarity score (0–100) + reasons	derived
dharma.generated	the Dharma Card (risk, context, expected validation)	derived at start
file.saved	fileName, extension, isTestFile (full path optional, off by default)	observed
validation.command	commandType, result, and source (observed | logged)	observed or logged
git.commit	short sha only	observed
git.diff.summary	filesChanged, linesAdded, linesDeleted	observed at end
karma.score.generated	the score + its rule breakdown	derived at end
phal.generated	the Phal Card (outcome readiness)	derived at end
outcome.reported	your optional reflection note	you, optional

Note what’s missing: there is no event that stores code, a command, a prompt response, or a diff. The taxonomy itself is the privacy guarantee.

How it observes — the collectors

Five small collectors watch your workflow and emit those events. None of them read content.

• Files — a save in the editor (or, optionally, an external write by an AI agent or CLI) emits one file.saved. The same file is de-duplicated per session, so it can’t inflate counts.
• Validations — with VS Code’s shell integration, Agent Karma sees that a command finished and its exit code. It classifies the command (is this a test? a build?) and keeps only the type and pass/fail — the command text is thrown away immediately.
• Commits — it watches git’s reflog (not a hook), because commits made by an AI agent or the CLI never fire VS Code’s APIs. It keeps the short SHA and nothing else. Commits are recorded for your timeline but do not affect your Karma Score.
• Ambient mode (optional) — instead of one session per task, it keeps one rolling session per day, capturing saves and validations hands-free.
• Claude Code usage (optional, off by default) — reads only the usage, model, and timestamp fields from Claude Code’s local logs. Never your prompts, replies, or code, and never over a network.

Observed vs. logged — and why it matters

Every validation is tagged with how it was learned:

• Observed — Agent Karma saw the command run and saw its exit code. This is the gold standard.
• Logged — you told it (“I ran the tests, they passed”) via the @agentkarma /verify chat command or the Log a Validation action. Essential for browser AI, copy-paste, or work outside the IDE.

Both count toward validation. But only an observed passing test earns the tests-passed bonus — you can’t award yourself a green check you didn’t actually run. See How the Karma Score works for the full rule set.

The guarantees, enforced

• No network, ever. A CI check (check:no-network) scans the shipped bundle for any HTTP API and fails the build if one appears.
• Local-only data. Sessions, events, and your Karma trend live in VS Code’s storage on your machine. Exports (JSON / Markdown) and the Karma Card are written wherever you choose.
• Your name stays yours. The name on a Karma Card is resolved locally — a setting, else your local git config user.name, else your OS user — and is never sent anywhere.

This is the difference between a coach and a dashboard: a dashboard measures you for someone else; Agent Karma measures your validation habit, for you, and keeps it. See the full Privacy & data contract.