kungfu-systems/kungfu/amp

Kungfu Agent Onboarding

When KUNGFU_AGENT_ENVIRONMENT=native-interactive, use the in-process environment envelopes before acting:

• KUNGFU_AGENT_CONSOLE_ENVELOPE identifies the exact Project, WorkConsole,

SessionAttempt, runtime Profile, and optional WorkRef.

• KUNGFU_SKILL_CONTEXT advertises the compact Skill catalog. Load full Skill

instructions only through the declared Kungfu entrypoint.

• KUNGFU_AGENT_CONTEXT and its entrypoints are discovery pointers, not a

prior chat transcript.

• KUNGFU_PRIOR_TRANSCRIPT_BYTES=0 means continuity comes from Core evidence,

never hidden provider conversation state.

Confirm current facts with read-only commands when needed:

"$KUNGFU_CLI_BIN" agent console current --json
"$KUNGFU_CLI_BIN" agent bootstrap-status --json
"$KUNGFU_CLI_BIN" agent capabilities --json
"$KUNGFU_CLI_BIN" skill catalog --json
"$KUNGFU_CLI_BIN" work status --workspace <path> --initiative-id <id> --assignment-id <id>
"$KUNGFU_CLI_BIN" agent session list --json

A bare kungfu run amp launch is intentionally Work-unbound so many terminal windows can start in the same Project. As soon as you choose or accept one Assignment, and before editing files or invoking a Work mutation, run:

"$KUNGFU_CLI_BIN" agent console bind-work --initiative-id <id> --assignment-id <id> --json

Keep the provider UI available when bootstrap is pending or degraded, but do not create, bind, or mutate Work until kungfu agent bootstrap-status --json reports state: verified.

The plan-native-bind-work and bind-native-work capability names are internal Session protocol operations, not public CLI entrypoints; never invoke them through kungfu agent session.

When durable Work may reduce continuity, handoff, evidence, duplicate retry, or external-write risk, submit only bounded structured signals to kungfu agent work-advisory --signals <signals.json> --json. Never include a transcript, hidden reasoning, credentials, or unrelated context. For recommend, show the returned preview and ask its single confirmation. Only after confirmation use the returned existing kungfu.work.capture, kungfu.work.admit, and kungfu.agent.console.bind-work path, cite its receipts, and continue the original task. Suppress a decline for the returned evidence root until the structured evidence changes. Advice grants no external authority.

For Skill reuse or creation, send only rooted catalog/Work/requirements evidence, candidate roots, enums, and booleans to kungfu agent skill-advisory --signals <signals.json> --json. Consume its shared policy root sha256:dc8ebb873760e55c40ef19b8354ba1e2b91706064a48dec00b1eb8dac0479267; do not reproduce the decision policy in provider prose. The result is read-only.

Do not continue unless it returns status: bound. If Kungfu reports native_work_already_active, stop this Work in the current terminal and follow the returned guidance; never bypass the guard or become a second writer.

Do not infer Work completion from terminal output or process exit. Mutate Work only through public Profile/KFD actions and their receipts. The Kungfu TUI is an observer of this native UI and never owns Amp input or transcript bytes.