Skip to content

PROGRESS.md

Latest commit

 

History

History
228 lines (195 loc) · 19.3 KB

PROGRESS.md

File metadata and controls

228 lines (195 loc) · 19.3 KB
Raw

PROGRESS — Current Development State

⚡ §RESUME_PROTOCOL — when an agent dies on a rate limit

A 429 terminates the agent but still fires a task-notification carrying the reset time. Parse it → record the agent + its last finding in prompts/AGENT_QUEUE.md §MID-FLIGHT → schedule a wakeup just past the reset → resume by SendMessage to the agent id (never re-dispatch fresh) → tell the user. Full text and the reasoning: prompts/AGENT_QUEUE.md §RESUME_PROTOCOL.

📋 The live worklist is prompts/AGENT_QUEUE.md — not this file

It carries §LIVE (which agent owns which files), the waves, the ⛔USER decisions, and the standing constraints. A session picking up work reads that; PROGRESS.md is state, not queue.

Current State — 2026-09-08 (session 2, take-over close) — ⚠ RESUME HERE: §36 W1–W6 ✅, W7 bakes gated; PR #1697 MERGED (Measure live), PR for feat/measure-indoor pending

Lane: prompts/MEP_CLASH_REVEAL_MOVIE.md — read §36 (worklist, all rows ✅ but W7), then §29.9, §37.4, §26.14, §33. bim-ootb: main carries PR #1697 (Measure checkbox + datum + slab/linear beats + opening gate, sw v1166, LIVE); branch feat/measure-indoor (= squashed main + W5 + W6, sw v1168) pushed, PR to open/merge next. Spec branch fable/meshdb-livewire.

DONE this session (each with its §-witness, logs /tmp/wt-storey-reveal/out/): W1 datum decide-once + §20.8 entry latch + drop ledger + cue span lock (witness_datum_stability.js 8/8 ×2) · W2 envelope = one connected structural component, measured, no code · W3 gate rebuilds the datum after Home · W4 §27 linear beat (witness_linear_beat.js 12/12 ×2; §14 across layers) · W5 §29 indoor beats — hall bounded at doors, stair going, door type, clear height (witness_indoor_beats.js 11/11 ×2) + the datum's Z anchor fix (Hospital storey rules were 9.2 m low) · W6 storey cards carry walkable m² + the datum's second life over the finished building (witness_storey_walkable_card.js 6/6 ×2, --life2 8/8 ×2).

⛔ NEXT — read prompts/MEP_CLASH_REVEAL_MOVIE.md §38 FIRST (user's two observations on the full film: plate-tint flicker + missing area info box at 9 s; the fly-out canvas unmeasured — wing spans, roof-edge-to-sill). Then: (1) PR #1699 MERGED (W5+W6 live, sw v1168) — nothing to merge. (2) W7 remaining bakes, user's go only: Hospital full 720p ✅ done (§37.5, film in Downloads); HHS full with --buildup --reveal --label; Hospital 1080p. (3) User rulings pending: Hospital opening datum now ends at 10.75 s (roof line) — keep or add "one storey below the roof"; HHS path skims the floor at 0.04 m and has a 7 s dive (path injection, §27g). (4) The 90–148 s pull-back is Measure-silent on Hospital, stated in §37.4.

Witness count: +5 files this session (slab, linear, indoor, datum stability, storey walkable card); gate witness 0 FAIL.

Current State — 2026-09-06 (session 3, final close) — ⚠ RESUME HERE: full bake confirmed, §PENDING.5 open, parallel session active

Full 195.8s film bake CONFIRMED (Hospital_FULL_allsystems_2026-09-06.mp4, ~/Downloads, 231 MB): 4,699/4,699 frames, 93.8 min wall clock, fileOk=true — every feature below shipped together in one real run. Two shorter verification bakes (fresh profile, 20-30s and 82-92s clips) ran first and matched prediction exactly: pre-topout elevation 50.0°→47.5° (original linear, unchanged), post-topout elevation 34.5°→32.0° (also original linear — revert holds throughout), plScale 0.8629→1.0000, poolSum 260.08→400.0 (genuinely double the old fixed-0.5 baseline, same fixture count). docs/BIMUserGuide.md updated: documented the (pre-existing, previously undocumented) Clash pairs checkbox + Silent-bake size selector, added --clash to the CLI table, added today's real full-bake timing.

New, untouched — §PENDING.5 in prompts/MEP_CLASH_REVEAL_MOVIE.md: user watched the full film, reports the HUD rotation left out Measure and Clash stats. Clash card exists in code and should have fired (--clash was on, 270 pairs built) — next session checks §CPE_BIG_STATS in this bake's own log to settle built-vs-shown before touching anything. No "Measure" HUD card exists at all — ask the user what they mean before building one.

A parallel Sonnet session is independently working §PENDING items 1-4 in /tmp/wt-clash-pending (items 2/3/4: clash-list depth display, cross-caller pair cache, tolerance-editor gear icon — coded, read-verified real by this session, not yet PR'd) and /tmp/wt-storey-reveal (§STOREY_HIGHLIGHT_REVEAL, status unknown to this session). Item 1 (bake-time estimates for 720p/1440p) was blocked on GPU availability while this session's full bake ran — GPU is free now. Next session: check that session's current status first, both worktrees are real and not this session's to touch.

Current State — 2026-09-06 (session 3, close) — ⚠ RESUME HERE: sun REVERTED (#1691), #1690 MERGED, clash lane CLOSED

bim-ootb main = 5daec9e0, sw v1155. Today's chain, all merged: #1686 label 3-row + clash card · #1688 mesh-true overlap depth · #1689 overlap-box marker + flat filter · #1691 revert of the sun snap (#1685, wrong starting spec — the linear 55°→6° crawl was always right) · #1690 post-topout fixtures 0.5 → 1.0 (the indoor-liveliness candidate, user: "use your own judgment" → pushed through). Witnesses on the merged tree: sun-linear 4/4, PL-topout 8/8. Worktrees pruned, nothing unpushed.

  • Clash-film lane CLOSED — further clash work (mesh-shape refinement, §TOUCH_BY_THICKNESS ⛔USER) goes to a separate session; do not re-open it from sun/indoor work.
  • Next bake (user's call, fresh --profile): 82–92 s on main ≥ 5daec9e0 shows everything together.
  • Pre-existing, untouched: tools.js (A._nightPLScale || 1) cannot express a 0 scale; narrowphase witness S7b/I1/I3 red on main.

Current State — 2026-09-06 (session 3, latest) — mid-session (superseded above)

User correction (relayed to and scolded at the other session): the sun-arc "topout snap" (#1685) was a wrong starting spec — the linear 55°→6° crawl was always correct. Reverted by hand on bim-ootb revert/sun-arc-topout-snap (PR opened with auto-merge; witness witness_sun_arc_linear.js 4/4: 34.47° at u=0.419, one-argument call site). TOPOUT_SNAP_EASE_U + _revealU kept (the Reveal round and #1690 use them). Spec §SUN_ARC_TOPOUT_SNAP carries a ⛔ REVERTED line — history, not live. Clash-film lane (#1686/#1688/#1689) is CLOSED — do not re-open or re-verify it from the sun/indoor task; hand any further clash work (mesh-shape refinement, §TOUCH_BY_THICKNESS) to a separate session. Indoor liveliness = the real ask. #1690 (§PL_TOPOUT_UNPIN, fixtures 0.5 → 1.0 post-topout, witnessed 8/8) is the built candidate and does not depend on the sun code; being synced onto the reverted main and pushed through as the cheapest path. The tools.js (A._nightPLScale || 1) finding is NOT relevant to liveliness (it only affects a 0 scale) — left alone.

Current State — 2026-09-06 (session 3, later) — earlier: 4 PRs merged, #1690 opened

User relayed a Sonnet review with advice on the three open items; taken as the go for 1 and 3, 2 built as a separate non-auto-merged PR (its "dim the fill" advice named no sourced value; the sourced lever is plScale 0.5 → 1.0). Detail: prompts/MEP_CLASH_REVEAL_MOVIE.md "The three follow-ons — MEASURED".

  • #1689 MERGED (sw v1153) — §CLASH_MARKER_OVERLAP_BOX: each marker is the oriented box of the real overlap solid, red/blue halves; markers W11 270/270 matched. §CLASH_FILM_FLAT_FILTER (Option 3): the film drops flat overlaps, 271 → 270, card says "1 flat touch dropped". Verdicts untouched.
  • #1690 OPEN, NOT auto-merged — ⛔USER take/leave (sw v1154) — §PL_TOPOUT_UNPIN: past topout the bake's fixtures ease 0.5 → 1.0 (nav Night Mode's tuned value) over the sun's own window; pre-topout byte-identical. Witness 8/8 on real staging (poolSum 200 → 400). If taken: merge, then a fresh-profile 82–92 s clip on main shows all four of today's features together.
  • Pre-existing finding: tools.js (A._nightPLScale || 1) cannot express a 0 scale (poolSum 400 at 0).

Current State — 2026-09-06 (session 3) — earlier today (#1685 verified, #1686 + #1688 merged, ⛔USER §TOUCH_BY_THICKNESS)

Done this session, all witnessed, detail in prompts/MEP_CLASH_REVEAL_MOVIE.md (three new dated sections at the end):

  • §SUN_ARC_TOPOUT_SNAP VERIFIED on a fresh Chrome profile (--profile /tmp/silent-bake-fresh-<epoch>) — elevation 14.5° at u=0.419, 6.0° from u=0.441, held to the clip end; last night's "still 34.5" was the stale-SW trap exactly as suspected. #1685 was already merged. Clip ~/Downloads/cll/Hospital_topout_snap_verify_clip82-92s_2026-09-06.mp4.
  • #1686 MERGED — §P2.4 label 3rd row [tol mm / clash mm] + §CLASH_HUD_CARD "271 mesh-true clashes flagged · 1,478 bbox candidates · 81.7% false". Labels witness 15 → 19 claims (P9a/P9b/H0/H1). sw v1151.
  • #1688 MERGED (273e1c59, both checks green) — §MESH_OVERLAP_DEPTH: the exact box of the overlap solid on every mesh-true pair (depthMeshM, overlapMaxM, overlapA/B, overlapCenter, overlapExact, overlapFlat); the label now reads the mesh figure — the shown pair is [50mm / 50mm], the OBB proxy said 344 mm. Hospital: SAT overstated by a median 2.78×, ≥3× on 3,185 of 6,749 pairs, worst 600×. Verdicts UNCHANGED. New I6 green on both buildings, red on main. sw v1152.

⛔USER §TOUCH_BY_THICKNESS (spec, end of the MEASURED section): make a flat overlap (< 1 mm thick) a touch instead of a clash? Fixes S7b (RED on main today) and drops 752 Terminal / 37 Hospital / 1 film pairs — but the witness oracle cannot judge multi-shell IFC meshes (3 formulations tried, each disagreed on ~200 pairs where the geometry shows real overlap). Options 1/2/3 listed there. Built, measured, reverted to annotation-only pending the call.

Pre-existing RED on main, now recorded (not from this session's code): narrowphase witness S7b (face-to-face cubes → CLASH), I3 n=223 Terminal / 326 Hospital, I1 n=2 Terminal — one hole: a coplanar-edge contact yields edge-length intersection segments, so "segment > 1 mm" ≠ interpenetration.

Still open from before: PL "real play" post-topout is a design lever (§BAKE_FILL_PIN pins ambient/hemi/ PL regardless of sun elevation; at 6° the sun 4.4 + hemi 0.617 still dwarf a 200-fixture pool ≈ 1.0 each) — ⛔USER, not more measurement. The real-overlap MARKER shape (oriented box from overlapCenter+overlapA) is a clash_film.js-only change now that the data exists — waits for a go since it changes the film's look.

Trap, still live: cli_silent_bake.js reuses /tmp/silent-bake-profile-<port> across runs and its sw= env line is read off the file on disk, not the active service worker — pass a fresh --profile on any verification bake.

Previous 2026-09-06 sessions — ARCHIVED into prompts/MEP_CLASH_REVEAL_MOVIE.md "Step 4 — CLOSED 2026-09-06"

GPU driver fixed (Secure Boot MOK); 5 PRs merged (#1676/#1678/#1679/#1684/#1683), revert tag pre-clash-pl-merge-2026-09-06 on 04e3dee2; #1681/#1682 merged earlier; eslint.globals.json fix.

Current State — 2026-09-04 (later)

ERP lane — §C2.3 DRAINED TO ZERO: all five ranked items closed, 4 PRs merged, sw v792

bim-ootb #1669 · #1672 · #1673 · #1675, every one CI-green. Three new witnesses, each RED on plain origin/main first: W-CALLOUT-CAMPAIGN 29/0 (was 9 PASS / 20 FAIL) · W-ADFORM-TRXMATERIAL 19/0 (was INCONCLUSIVE) · W-AGENT-ZIP-SYNC 11/0 (was 1 FAIL).

  • E-1, the callout campaign #1669 — live dispatch on the nine O2C/P2P tables 3 → 51 of 78 bindings. The 3 is not a typo: nothing had ever called AdCallout.installDefaultHandlers(), so six engine line callouts had never fired in a browser and §IC.3's "28 = 36%" counted a paper registry. Nine new atoms incl. all five CalloutPayment.*, with invoiceopen/invoicediscount/ paymenttermdiscount ported from the PL/pgSQL. → AGENT_QUEUE.md §CALLOUT-CAMPAIGN
  • Form #2 of 49 #1672 — AD_Form 103 VTrxMaterial, read face of the M_Transaction ledger. → §ADFORM-TRXMATERIAL
  • Four stale live witnesses #1673 — all GREEN, none a product defect; three named the wrong thing. Surfaced a CLASS: 12 more default to a /tmp/wt-* worktree that is GONE. → §STALE-WITNESSES §SW.2
  • The last two tracked binaries #1675 — preview_demo.db built from its own .dump; both *_agent.zip built by deploy-pages.yml. Found a live defect: odoo_agent.zip shipped without odoo_agent/extract_model.js. → §PREVIEW-DEMO-FROM-SQL · §AGENT-ZIPS-BUILT
  • §AZ.3 answered #1677 — idempiere_agent.zip nests. Not cosmetic: about_diy.js:199 had always printed cd idempiere_agent && … while the zip put the files at the ROOT, so the command the app prints could not run. New claim E1 checks the offer's cd <name> against the zip's top level.
  • 🔴 §PAGES-SERVES-THE-BRANCH #1680 — READ IT FIRST NEXT SESSION. Pages here is build_type: legacy, source {branch: main, path: /}: it serves the tracked files on main, and the artifact deploy-pages.yml uploads is published by nothing. Untracking the zips in #1675 404'd both live downloads; erp/version.json has been 404 since it was introduced; the minify step (claimed −36.9% gz) has never reached a user. Zips re-tracked, W-AGENT-ZIP-SYNC 15/0 with a new A0a (the zip is PRESENT). Verified by curl, not inferred: both 200, both extract to their own folder, diff -r clean. Standing rule: a deploy is not verified by a green workflow — fetch the URL. ⛔ Next (nothing on §C2.3 remains): the 12-witness stale-root class (§SW.2) · the 27 callout bindings still undispatched (§CC.7; four are navigate* atoms that should leave the denominator, not be ported) · 47 of 49 forms with no renderer · ⛔USER §PZ.3: switch Pages to build_type: workflow? (it would make version.json + minification live for the first time, and make the site stop updating whenever the workflow fails). Full restart brief: AGENT_QUEUE.md §RESTART.

Previous State — 2026-09-04 (earlier) — ARCHIVED to its spec section

ERP lane — §CLOSE.4's five open items closed, bim-ootb #1661-#1666, sw v789.

Six witnesses (W-KIND2-READBACK · W-DICT-LAZY 10/10 · W-INOUT-CALLOUT-ATOMS 9/9 · W-ADFORM-VMATCH 16/16 · §DIGEST 46 rows), five defects nobody asked for, three stale hygiene premises. Full detail: prompts/AGENT_QUEUE.md §ERP-SESSION-CLOSE-2. Its ⛔ list (§C2.3) is DRAINED — see the block above. Two of its numbers did not survive re-measurement: the "37-binding gap" was 50, and the "36% dispatch" was really 4% live.

Previous State — 2026-09-03

ERP lane, 2026-09-02/03 — 5 PRs, sw v777 (superseded above; full detail prompts/AGENT_QUEUE.md §ERP-SESSION-CLOSE)

The five document windows render their own AD tab (c_order 8→56 fields, c_payment 4→78, DisplayLogic 0→28/61, val-rule-filtered FK pickers); three shipped defects found by the new instruments; W-ERP-TWIN, W-MULTIHOST-SYNC and W-RELAY-AUTH became standing gates. Its ⛔ list is settled: W-SCALE-FORECAST's 0.8× batch was fixed at #1638 (3.53× median) and W-AD-DISPLAYLOGIC-LIVE was a stale witness, not a product failure. S-2 (no relay deployed) is still the real barrier to multi-writer, not S-1.

4D / viewer lane — 23 PRs merged 2026-09-02/03, release v1.58.0 — DONE, detail in the spec docs

§TM_REVEAL_TILED #1605 (dead air 44-71% → 0.0%) · §CACHE_PLAYED_LAYER #1607 · DAY-0 PASS=5 FAIL=8 → PASS=8 FAIL=5 #1551/#1615/#1625 (all 8 attributed; the "regression" premise retired) · §SUN_FILL_RATIO #1622 (1.0453 → 0.2388) · §MEP_COLOR_SURVIVES_PHOTOREAL #1621 (40,634 colourless MEP → 0) · §DUCT_SILHOUETTE #1631 (21.3 → 3.4 mm) · §CPE_AIM_DEPTH RETIRED #1619 (90.657° → 0.000°) · B-2 support consolidation #1627-#1630 (keep the AND) · 5D constants → JSON #1616 · nine vacuous § tags guarded #1608. → prompts/4D_GANTT_TM_REFACTOR.md · 4D_MODEL_INTEGRITY.md · PHOTOREAL_STILL_RENDER.md · CINEMA_DISCIPLINE_REVEAL.md · CINEMA_PATH_EDITOR.md ⚠ Clinic landing: local/live SYSNOVA/index.html still diverge 105 lines both ways — do NOT upload wholesale.

⛔ OPEN

  • Interior lighting / liveliness — running. "Lively" is variance, not brightness (user: "it should be lively, so far it has never been though lighting has been bright"). Measured on two real bakes bracketing #1622: CV 0.344 → 0.430, spread 42.9 → 59.1 — but the bright register drained (brightest fifth 140.7 → 79.2). Lead: the still's near-field boost never fires.
  • §Z.3 Clinic ground-slab appears late, then persists on scrub-back — untouched; two named hypotheses in LTU_TERMINAL_CLINIC_RENDER_CORRUPTION.md §Z.
  • A-28 §CPE_REVEAL_ARCH_HOLD overshoot (TOP of the queue, user-set 2026-09-03) — #1633 shipped, user sees the ARCH strip at the LAST stick of round 2, not the first. Suspect 1 (two reveal clocks) is DISPROVEN by code read: beats.flyback IS tF (effects.js:8656/:7624) and both callers pass the same film fraction (cinema_maxq.js:1511, cinema_path_editor.js:2532). Next: _flyBackPose direction, then whether cpeRevealApplyVisual's write survives the per-frame staging rebuild. Detail in AGENT_QUEUE.md A-28.
  • §FILM_UNSUPPORTED — not started; take the short --frames re-scope, not a full bake.
  • §LIGHT_SHAFT (D-5) — specced, queued behind the lighting lane (shares effects.js).
  • ⛔USER decisions (12 open) in AGENT_QUEUE.md: calibration lever (Hospital 318→940 d), LFS 8.53 GB pay-vs-rewrite, sub-element slab splitting, Terminal's 673 Ceiling Level NN elements.

Standing constraints added 2026-09-02/03

  • Bakes are a proven, expensive facility, NOT a measurement tool. Never launch a film to settle a number — ask first. Keep every § line on the bake path intact.
  • 4D generation is building-independent. No per-building constants or name-branching. Audited clean: every building name in the scheduling files sits in a comment; the only non-comment hits are the IFC classes IfcFlowTerminal/IfcAirTerminal/IfcFireSuppressionTerminal.
  • DBs stay on OCI, GH Pages serves the app. Settled: GitHub hard-rejects any file >100 MB; every major DB exceeds it.
  • A witness's own harness is a suspect before the product is. Four red live witnesses this session (§STALE-WITNESSES): every one was the instrument, and three printed an error naming a DOM node or a business gate while the PAGE was logging the real cause. Read the page log, not the harness verdict.
  • Beware checks that cannot fail. Five now: eslint | tail returning tail's exit code, a witness reading its own comment block, a 0-byte log from a buffered page.evaluate, a (none above = clean) line printed unconditionally — and A-28's 6/6-green witness that never judged a caller.

Older session log

Archived to prompts/archive/PROGRESS_sessions_archived_2026-09-01.md (2026-08-29 → 2026-08-14); the 2026-09-01/02 entries are superseded above. Per-lane detail lives in each lane's own prompts/*.md.