Work with no committed artefact at all (a feasibility answer, not a document) has no outcome that fits

[thedancingdeveloper]

Every outcome path in the harness assumes the deliverable is a diff. There
is no way to express an item whose product is knowledge — an investigation, a
feasibility answer, an ADR, a comparison of three approaches.

What happens today if you try

Write the item anyway and the agent does the work well. Then:

1. It leaves the worktree clean, because findings are not a code change.
2. session_executor sees a clean tree and records
   escalated / no_target — "the agent made no changes and left no reason"
   (An agent that correctly refuses an impossible item should be flagged as needing a person #174, shipped in An agent that says an item is impossible is asking for a person, not failing #175/An escalation lands in blocked, and does not read as a failed run #176).
3. A completed investigation is indistinguishable from an agent that did
   nothing.

If the agent writes its findings to a file to avoid that, it is now a diff, and
it meets the rest of the pipeline:

• the checks run against the worktree — cargo test has no opinion on an ADR;
• REVIEW_PROMPT asks the reviewer to judge "your diff against the item", so
  it grades a document as though it were an implementation;
• the definition of done is a pull request containing a change.

So the two available outcomes are "recorded as nothing" or "graded as code".

Why this matters beyond tidiness

It is the prerequisite for #181. Proposing a plan for an existing project
honestly requires investigating it first — reading the roadmap, checking what
is deployed, finding what the last release left open. If investigation cannot
be expressed as work, then plan generation is a single opaque model call rather
than something that can be split, checked, reviewed and retried like everything
else here.

It is also the shape of a large fraction of real engineering work. On rdpapp
the owner's own next package (P7.0) opens with "the first decision is the
attachment ADR: compare Node B's existing host-managed tailscaled, a managed
sidecar, and a tsnet/bridge approach". That is an item. The harness cannot
currently hold it.

What is actually needed

The interesting question is what "done" means when there is no diff, and it
should not be answered by loosening the gates:

• A declared deliverable. The item states what it must produce — a written
  answer, a recommendation with alternatives considered, a document at a named
  path. Checks that cannot apply are skipped because the item says so, not
  because they failed silently.
• A review rubric for findings, not diffs. Are the claims cited? Were the
  named alternatives actually examined? Is the recommendation supported by what
  was found, or asserted? A reviewer told to judge a document against the
  diff-rubric will reject it for being a document.
• An outcome that is not no_target. A finished investigation is
  completed, and the findings are its result, in the same way a diff is.

Deliberately not in scope

Letting an agent decide for itself that an item is "really" an investigation
and skip its checks. The item declares its kind; the agent does not choose it.
Otherwise the first hard test failure becomes an essay about why the test was
wrong.

Related

• The first run should be able to generate the plan, not only execute one #181 — the harness should accept an objective, not only a plan. That
  needs this to be honest about how it reached its proposal.
• An agent that correctly refuses an impossible item should be flagged as needing a person #174/An agent that says an item is impossible is asking for a person, not failing #175 — the clean-tree path that currently swallows this case. The
  fix there is right for what it covers; this is the case it cannot know about.

[thedancingdeveloper]

Narrowing this, after clarification. The motivating case is "review and generate a plan to upgrade to Node v22", where the deliverable is a PLAN.md — a file, therefore a diff, therefore something the existing pipeline can carry. That case is #181 and does not need this issue.

So the claim in the body above is too broad. Corrected:

• An investigation that produces a document already works in principle. The tree is not clean, there is something to review, and agent-harness plan --dry-run is a real check when the document is a plan.
• What remains genuinely unrepresented is work with no artefact at all — "is X feasible", answered in prose, with nothing to commit. That still lands as escalated / no_target, and is still worth solving.

The concerns about the review rubric stand for both: a reviewer told to judge a document against the diff rubric will reject it for being a document. But that is a rubric problem, not a missing-outcome problem, and it is smaller than this issue originally implied.

Keeping this open at lower priority, and removing it as a blocker on #181.