opentakeoff-mcp 0.9.21 → 0.9.26

package/README.md

@@ -108,14 +108,15 @@ includes document text, shape vertices, or result payload content.
 | `load_plan` | Open a plan PDF from disk. Default replaces the whole session; **`merge: true` ADDS the document to the working set** (#152) — plans + schedule + addenda as one takeoff, sheet graph spanning the whole set, marked set covering every worked sheet. Returns per-sheet dims, title-block `sheet_number`, and the detected drawn scale where present. |
 | `sheet_info` | One sheet's dims, vector segment count, scale status, detected suggestion, committed shape count. |
 | `set_scale` | Set a sheet's scale — exactly one of `label`, `upp`, `calibrate {p1, p2, feet}`, `use_detected`. |
-| `one_click` | One-Click Area at (x, y): flood fill bounded by the plan linework, traced, vertices snapped. On a SCANNED sheet (no usable linework) the flood falls back automatically to the rendered pixels — same engine as the canvas — with `raster_traced` disclosed on the reply and on the shape's origin (#154). Pass `condition` to commit; `role: "deduct"` subtracts. |
-| `detect_rooms` | Batch One-Click: reads every room-number label off the sheet's text layer and floods each — one call instead of `read_sheet_text` + reasoning + N `one_click` calls. Only cleanly-traced rooms come back; everything skipped is counted and reasoned in `withheld` (degenerate / duplicate / implausible / unresolved), never dropped silently. To commit: `assign_from_schedule: true` routes each room through its OWN room-finish schedule row and commits under the FLOOR finish that row states (rooms the schedule can't answer for return in `unresolved[]` with reasons and re-seedable coordinates); or pass `condition` to commit every room under one stated tag. |
+| `one_click` | One-Click Area at (x, y): the sealed flood engine bounded by the plan linework, traced, vertices snapped — the SAME feet-true arguments the canvas passes at a click (gap sealing up to a door width, door-swing wedge inclusion, the half-foot minimum-passage rule), so an MCP trace and a canvas click at one seed measure the same square footage (pinned against the bench corpus goldens in `test/parity.test.ts`). Every trace carries the engine's account of itself: `confidence` 0–1 with `confidence_factors` naming each deduction (`gap_sealed_px`, `door_wedges`, `min_pass_delta`, …) — a review prioritizer, never a verification; a low score is a `view_sheet {overlay: true}` audit prompt, not a fact. On a SCANNED sheet (no usable linework) the flood falls back automatically to the rendered pixels — same engine as the canvas — with `raster_traced` disclosed on the reply and on the shape's origin (#154). Pass `condition` to commit (the full account stamps `origin` centrally at the commit); `role: "deduct"` subtracts. |
+| `detect_rooms` | Batch One-Click: reads every room-number label off the sheet's text layer and floods each — one call instead of `read_sheet_text` + reasoning + N `one_click` calls, through the SAME sealed engine per room (confidence + the engine account ride each room and its committed origin). Only cleanly-traced rooms come back; everything skipped is counted and reasoned in `withheld` (degenerate / duplicate / implausible / unresolved), never dropped silently. To commit: `assign_from_schedule: true` routes each room through its OWN room-finish schedule row and commits under the FLOOR finish that row states (rooms the schedule can't answer for return in `unresolved[]` with reasons and re-seedable coordinates); or pass `condition` to commit every room under one stated tag. |
 | `measure_polygon` | Area + perimeter of a polygon you supply (min 3 verts). Requires scale. |
 | `measure_line` | Length of an open polyline (min 2 points). Requires scale. |
 | `derive_base` | **Base LF from committed rooms**: for every floor shape of a source condition, commits a linear base run tracing that room's boundary, quantified net of the door openings YOU state per room (`{shape_id, lf}` — your claim, recorded on `origin.derived`; the tool never guesses). All-or-nothing; one undo step. |
 | `measure_surface` | **Wall SF**: an open run traced along the wall, quantified as traced LF × the condition's height (the canvas's H knob — pass `height_ft` to set it, or set it once with `edit_condition`). Wall tile, wainscot, wall systems. Refuses without a height, minting nothing. |
 | `place_count` | **EA markers**: one point, one each — thresholds, stair nosings, floor boxes. No scale required (EA is scale-free). One shape per point; the whole call is one undo step. |
-| `symbol_sweep` | **Every instance of a repeated plan symbol, from ONE example**: marquee a tight `seed_rect` around a single drain/threshold/fixture symbol and the sheet's vector linework is searched deterministically for every other placement — translation plus 0/90/180/270 rotation and mirroring (both on by default). Score = length-weighted fraction of the seed's segments matched within `tolerance_px`; ≥ 0.92 is a match, the 0.75–0.92 band returns in `withheld` with reasons (never committed, never dropped silently), and the work cap is disclosed when it bites. `commit: true` + `condition` commits every match center as an EA count marker through the `place_count` path — one undo step, `origin.method "symbol_sweep"` with per-marker score and transform. No scale required. |
+| `symbol_sweep` | **Every instance of a repeated plan symbol, from ONE example**: marquee a tight `seed_rect` around a single drain/threshold/fixture symbol and the vector linework is searched deterministically for every other placement — translation plus 0/90/180/270 rotation and mirroring (both on by default). Score = length-weighted fraction of the seed's segments matched within `tolerance_px`; ≥ 0.92 is a match, the 0.75–0.92 band returns in `withheld` with reasons (never committed, never dropped silently), and the work cap is disclosed when it bites. **`scope: "set"` sweeps the whole working set, counting on PLAN-role sheets only** (the sheet graph decides; every excluded sheet disclosed in `skipped` with role and reason) — and the seed rect may sit on a detail or legend sheet, which then serves as the fingerprint SOURCE while staying excluded from counting: the estimator's "click the assembly in the detail, count it on the plans" gesture. Per-sheet results carry their own match/withheld lists, per-sheet cap accounting, and wall-clock `elapsed_ms`. `commit: true` + `condition` commits every match center as an EA count marker — the whole sweep (set-wide included) is one undo step, `origin.method "symbol_sweep"` with per-marker score, transform, and seed source (`origin.symbol.seed`). No scale required. |
+| `sweep_schedule_row` | **Take off a schedule row's mark from the row itself**: pass the row's key (e.g. `T1`) and the tool reads the row from the set's schedule tables (the row is the condition's cited source), anchors a fingerprint on the marker the tag is DRAWN as on a plan sheet (a deterministic pad ladder around the tag text; where the tag occurs more than once the fingerprint must recur at a second occurrence — `anchor.corroborated` — before it is trusted), and sweeps every plan-role sheet. **The count is geometry AND text agreeing**: drafting reuses one bubble shape across many marks, so a match counts only when the row's own tag sits within the marker footprint (its bbox rides the match as `tag_at` evidence); a match labeled with a sibling key is `excluded` and says whose it is, an unlabeled match is `withheld` as a question, a tag drawn with no matching marker is `text_only`. Refusal over guessing, each with the reason and the fix: no such row, an ambiguous key, a tag drawn on no plan sheet, no repeatable marker linework — a fingerprint is never guessed from text alone. `commit: true` commits the counted matches under the row's own key — one undo step, `origin.assignment {source: "schedule"}` plus the anchor and row citation on `origin.symbol.seed`. No scale required. |
 | `takeoff_summary` | Per-condition totals + grand totals, computed by the Report's rules. |
 | `export_takeoff` | The full `opentakeoff.takeoff_canvas.v1` payload — exactly what the app autosaves. Inline, and to disk with `path`. |
 | `delete_shape` | Remove a committed shape by id. |
@@ -128,8 +129,10 @@ includes document text, shape vertices, or result payload content.
 | `list_shapes` | The **mid-session inventory**: every committed shape's id, sheet, condition, role, quantities, review state, and assignment verdict (`schedule` \| `asserted` — where its finish tag came from) in one compact read — the ids `edit_shape`/`delete_shape` assume you have, without pulling the whole `export_takeoff` payload. Filters by sheet/condition narrow; empty is a result, not an error. |
 | `undo_last` | Step back over your own last `n` mutations, newest first. Exact inverses: a commit is removed, an edit restored verbatim, a delete re-inserted where it was, a materials edit's whole array restored, a condition edit's waste/multiplier pair restored. A whole `detect_rooms` sweep is **one** step. |
 | `annotate` | Place a note ABOUT the work — cloud/highlight (`rect`), text (`at`), callout (`at` + `target`), **arrow** (`from` + `to` — plank/seam direction), **bubble** (`at` + optional `r` — keynote circle, centered text), **dimension** (`from` + `to` — a dimension line with end ticks, labelled with the measured length at the sheet's scale; the one annotation the scale gate applies to — an unscaled sheet refuses like the measure tools). Attach to a condition and it wears that scope's colour on the canvas and in the marked set. No review gate: notes are not geometry. |
-| `list_annotations` | Every annotation with its condition RESOLVED to a finish tag, coordinates back in image px; filter by sheet/condition. `unattached` counts the link_annotation candidates. |
+| `list_annotations` | Every annotation with its condition RESOLVED to a finish tag, coordinates back in image px; filter by sheet/condition. `unattached` counts the link_annotation candidates. `verdicts[]` is the approval family's inventory — every mark with its actor stated, a condition filter reaching a verdict through its target shape. |
 | `link_annotation` | Attach an existing annotation to a condition (or detach with an empty tag) — the canvas's Attach/Detach control, reachable by an agent. |
+| `mark_verdict` | The **agent's pencil-signature** on work it checked — the agent half of the approval family (#176). Mints the graphite AGENT diamond, and structurally nothing else: the tool takes no actor input, so the estimator's APPROVED ring stays behind the canvas's human-only Approve tool. Target a committed `shape_id` (anchored on the shape — a room's centroid, a run's midpoint — with the id recorded as provenance) or a `sheet` + `at` point; optional short `text` rides the record. Touches no quantity; renders on the canvas and in the marked set, whose cover tallies the split (`Approval stamps: N estimator-approved · M agent-marked`); rides the annotations payload through `export_takeoff`/`import_takeoff` and the app's own saves. One mark per shape. |
+| `delete_verdict` | Lift an agent verdict mark by id. Agent marks only — the estimator's seal is human ink and is refused, the same line `edit_shape` holds on reviewed shapes. `undo_last` re-seats a lifted mark exactly where it was. |
 | `read_sheet_text` | Positioned page text (image px), optionally restricted to a region — title blocks, room labels, finish schedules. |
 | `find_text` | **Locate** a known string — the complement to `read_sheet_text` (which returns what a region *says*; this finds *where* a string sits). Case-insensitive substring match per pdf.js text run; each hit's center feeds straight into `one_click`'s seed. |
 | `sheet_graph` | The plan-set INDEX (#87): every sheet's role with evidence, the schedule tables found, every room tag with its stacked name, the detail callouts — how an agent decides WHAT to measure without a human enumerating rooms. |
@@ -152,7 +155,10 @@ Two rules hold the surface honest:
 - **Ink is not pencil.** A shape carrying `origin.reviewed === true` is work a
   human affirmed, and no agent verb touches it. This server has no review gate
   of its own, so the guard is inert here — it is the contract that makes the
-  surface safe to port to a host that *does* have one.
+  surface safe to port to a host that *does* have one. The approval family
+  holds the same line at the mark itself: `mark_verdict` can mint only the
+  AGENT diamond (there is no actor input to misuse), and `delete_verdict`
+  refuses the estimator's APPROVED seal outright.
 - **Self-revision is not correction.** `edit_shape` bumps `origin.agent_edits`
   and touches nothing in the human-correction vocabulary (`edited`, `edits`,
   `proposed_verts_norm`). Those fields mean *a human corrected the machine*;