docs: one status document, and the honest headline at the top of it

[thedancingdeveloper]

What this is

docs/STATUS.md — the single document that says where this project actually
stands and what is left to do. There is no other current status document.

The headline is stated first and plainly: the harness has never delivered a
work item against a real workload. Four passes of the direct executor and one
standalone run of the agent loop against rdpapp produced no merged work, and no
other real workload has been attempted. Everything else in the document
qualifies that sentence.

Structure

1. Where it actually stands — the headline, what is nonetheless tested, and
   the The single-shot model call is the defect: every role that answers questions about a repository must be a loop #195 diagnosis (the single-shot model call is the defect; four format
   repairs on 2026-08-05 each improved the output and delivered nothing).
2. All pending work — every open issue, organised by what a reader can act
   on: what blocks everything (Build the agentic role runner, and put the implementer through it #215); the The single-shot model call is the defect: every role that answers questions about a repository must be a loop #195 programme (Build the agentic role runner, and put the implementer through it #215, Put the surveyor through the role runner: a plan should be written by something that read the repository #224, Put the assessor through the role runner: adopt asks a model to find evidence it cannot go and find #225,
   Put the reviewer through the role runner with a READ-ONLY environment, so it verifies instead of infers #226, Retire the planner and context pre-selection once the loop has delivered items #227); defects with no blocker (Two edit blocks naming one file by different path strings render its diff twice, and the second copy cannot apply #219, Two API routes compare a lease against the wall clock, not the queue's clock #220, Two holds opened by one attempt in the same tick raise a bare sqlite3.IntegrityError instead of a HoldError #221, _WIRE_ROLES permits a 'tool' message that _for_the_wire has already stripped the tool_call_id from #223, test_pausing_a_project_stops_claiming asserts completions stop within 100ms, which is a timing assumption about the host #207, A stored model answer is redacted, so it cannot be used to reproduce what the model actually said #209,
   Bug: silent-but-active CLI sessions are indistinguishable from hangs #103); blocked on a named decision or measurement (A generated phase heading is both a tracking umbrella and a claimable item, and it cannot be both #184, A correction learned on one item is paid for again on every item after it #189, An item the claim scan gives up on is left with no disposition, and empty means 'not finished with yet' #222 — D8);
   blocked on a real run (T19: 72-hour measurement run #33, T30: 48-hour soak #44, T37: 7-day unattended run #51, T43: A/B whether the reviewer seeing the plan changes its verdict #84). Each entry says what it is and
   why it is where it is.
3. rdpapp is the first application under test — TheDancingDeveloper-org/rdpapp,
   authoritative on Forgejo at repo.indexarr.net/indexarr/rdpapp with GitHub
   as the mirror. Chosen deliberately at the hard end: Rust, a 704 KB
   main.rs, a credential vault whose fixtures trip the redactor. The evidence
   package is summarised and linked, not duplicated, and its numbers are stated
   as rdpapp's.
4. How to run the harness against rdpapp — brought in from the unversioned
   ~/Working/Active/.harness-runs/rdpapp-m2/env.sh so it is versioned:
   verifiable preconditions (including the stale-base check that invalidated
   the abandoned 2026-08-04 attempt), the environment, the run command with
   every non-obvious flag explained, retrying without destroying last_error,
   monitoring with the A stored model answer is redacted, so it cannot be used to reproduce what the model actually said #209 redaction caveat, and cleanup naming the branches
   that belong to the abandoned attempt.
5. What is proven, observed and tested — a table using the README's three
   words exactly. The proven column is empty and says why.

What it deliberately does not do

• It does not describe how the harness works. docs/DESIGN.md owns that and is
  being written in parallel; STATUS.md links to it and says so. The link is
  currently dangling until that document lands.
• It does not preserve the structure of docs/FIT-FOR-PURPOSE-STATUS.md, which
  is frozen at 2026-08-04 and being deprecated separately. It is referenced as
  history only, never as current.
• It asserts no project's measurements as universal fact.

Only docs/STATUS.md is added. No other file is touched.

🤖 Generated with Claude Code