Another take on
the agent harness.

1. Know / Tuft + Writfold

   Exact, cited evidence

   Tuft finds exact text in one pinned commit, and Writfold retains the matched bytes as citations bound by digest. A finding is not a proof, and a citation is not an approval.

   Runs (trial)
2. Conclude / Weft

   Check the plan

   Weft checks a plan's claims against declared assumptions: each is entailed, refuted or still open. Approval can require named checks to be entailed. Weft's results are recorded as hypothetical and carry no permission to change anything.

   Runs today
3. Change / Splicery + Writfold

   Edit only where owned

   Splicery expresses a byte-exact edit recipe and previews the candidate. Writfold applies it only in a Git worktree it created, under your signed approval and a current operator grant.

   Runs (trial profile)
4. Record / Writfold

   Evidence at every step

   Each step is a record in a validated ledger, so a run can be reviewed afterwards and an interrupted apply recovered. Reuse by the next session is planned.

   Runs (trial)

Runs today standalone or in the trialParts tested components exist, not yet connectedPlanned designed, not built

1. 01 / Writfold

   Retain exact inputs

   Keep the exact bytes of a file at a full commit id, bound as a citation by digest, so later edits cannot rewrite the evidence.

   Runs (trial)
2. 02 / Weft

   Assess the claims

   Check what the declared assumptions actually support: entailed, refuted or still open. A report citing evidence that was not retained is refused.

   Runs today
3. 03 / Splicery

   Prepare an edit

   Describe a byte-exact change and compute a candidate file. The original stays untouched.

   Runs today
4. 04 / Writfold

   Approve that edit

   You sign an approval bound to this exact prepared change, checked against a current operator grant and the context's policy.

   Runs (trial)
5. 05 / Writfold

   Apply it

   Writfold writes the approved change into its own Git worktree, after re-checking the file has not changed.

   Runs (trial)
6. 06 / Writfold

   Record the outcome

   A durable claim before the write and one recorded outcome after it. After an interrupted apply, recover decides the outcome from the file's exact bytes instead of applying the edit twice.

   Runs (trial)

# In a clone of https://github.com/writfold-dev/writfold
cargo build --release -p writfold

# Prepare, approve and apply one edit
export WRITFOLD_STATE=/tmp/writfold-trial
target/release/writfold trial init
target/release/writfold trial prepare --repo /path/to/repo --file src/lib.rs --recipe edit.spl
target/release/writfold trial approve <run>
target/release/writfold trial apply <run>
git -C "$WRITFOLD_STATE/worktrees/<run>" diff

# Find and cite exact evidence with Tuft
printf '@h := find text "trial_blob(" in "crates/" at "%s";\n@e := retain @h around lines 0;\n@v := project @e [path, line, binding, excerpt(max_chars: 72)] limit 6;\nemit @v budget_tokens 300;\n' "$(git rev-parse HEAD)" > q.tuft
target/release/writfold tuft run q.tuft --repo .

Your own checkout is never written: Writfold makes its own sparse worktree of the repository's HEAD, replays a Splicery recipe through the pinned kernel, and installs the result only after a signed approval and a current grant are verified. apply records a durable claim before it touches the file, and writfold trial recover <run> settles an interrupted apply. Splicery's examples/try/ has a recipe for a small Rust file; trial retain and approve --require CHECK are described in the trial profile. Weft and Splicery each have their own Try it: Weft, Splicery.

This is a cooperating single-user trial. It never writes your checkout, but nothing stops an agent's own shell from doing so, and it assumes nobody races it on its worktree. Search is exact text or a bounded regex; semantic and fuzzy search are deferred until the Omnigraph backend exists.

Found a problem or have an idea? Open an issue on Writfold, Weft or Splicery.