sift-workflow

sift-workflow

This skill is the entry point — start here. It drives the whole loop for you; the /sift-harness:sift-* slash commands are manual controls for stepping a single phase by hand when you want to. For normal work, follow this skill end to end.

When a task deserves more than a quick edit — it's multi-step, it must be verified, or "done" needs to mean checked — drive it through the harness instead of declaring completion yourself. The harness turns the work into a packet and only lets it reach confirmed when independent witnesses pass.

Run these from the repo root (the plugin's bin/sift; in Claude Code the same verbs are the /sift-* commands).

The loop

1. Scaffold a packet. sift packet new <id> --profile toy (or software) writes a packet + an acceptance-test stub. Then sift packet validate <id> checks it. Fill in the goal, the scope.paths (the files this work may touch), and the acceptance test (what proves it done).

2. Plan. sift plan <id> moves it submitted → packeted and sets it as the active focus. From here the PreToolUse scope guard fences your edits to the packet's scope.paths — if you try to edit outside, it blocks you.

3. Do the work — in scope. Make the change and produce the declared artifact / make the acceptance test pass. Stay within scope.paths.

4. Execute. sift execute <id> runs the acceptance check (→ acceptance_met). If it fails, fix and re-run; don't proceed on red.

5. Review. sift review <id> runs the witnesses (W2 reproduce, W3 review, W4 wiring) and, only if they pass, writes a witness-bound confirmed event and clears the focus. Check sift state <id> → confirmed.

If you're resuming after a break, sift next <wave> tells you the next actionable packet (state is replayed from the log, so it survives a context compaction). sift doctor flags a confirmed packet whose review artifact went missing.

What's real (and what isn't)

• The witnesses, the witness-binding, the log, and the scope guard are the enforcement. A confirmed that no witness backed replays as corrupt.
• The scope guard is advisory — it protects only when the host runs the PreToolUse hook; it is not a kernel sandbox.
• Review backend & egress: W3 runs keyless (deterministic, no network) by default. Sending the diff to a network model backend requires explicit consent (SIFT_W3_EGRESS=allow); secrets are best-effort redacted first, and every send is recorded in the egress log. Best-effort, not a guarantee.

Honesty

This skill points you at the loop; the machinery enforces it, and where it enforces depends on the host:

• Under Claude Code: the Stop hook makes the turn un-endable while a packet is planned-but-not-confirmed, and the scope guard fences edits to the packet's scope.paths. That is runtime enforcement (host-honored, not a kernel sandbox).
• Everywhere (all hosts): the hash-chained trust-core is the hard backstop — a confirmed no witness backed, or a hand-edited log, replays as corrupt. Completion is witness-bound, not asserted.

On a host without hooks the live steering degrades to guidance, but the trust-core backstop never does.