See interior dead air and colliding narration windows; validator hands over the summary - #15
Merged
Merged
Conversation
…, and hand the summary to the validator Sync of the narration-coverage fixes from Orkas release_1.6.5 (154c24196, 61cf39158, 8a94dc1c0, 66c1a720e), re-mapped onto the ovs toolchain: - assessVoiceoverCoverage now takes the per-line voiced spans instead of only min/max reach. coverageRatio only says how far the LAST line reached — a 60s cut with 29.7s of interior silence still scored 95% on it. The report adds voicedRatio (share of the clip that actually carries voice), interiorGaps (holes ≥0.5s, largest first) with a 'gapped' status past 2.5s, and raw-span overlap detection with an 'overlapped' status: a line whose audio runs past the next line's start mixes as two voices speaking at once. coverageForSegments already had the per-segment structure in hand and was collapsing it to min/max. - validateEdl rejects overlapping narration line windows (E_NARRATION_WINDOWS_OVERLAP) at free validation, before any synthesis is spent. The classic mistake is writing target_sec as the line's END time instead of its duration. - The three remaining terse enum rejections (reference media_type, reference role, edit_strategy decision signal) now name the legal values, like the intent/intent_basis messages already do. - `ovs plan validate` / MCP plan_validate attach the host-rendered plan summary to a passing result: it is the last thing the model reads before presenting the plan, and presenting a hand-written abstract instead of the real plan is how an unreviewed EDL got approved. Verified: build green; 244 unit tests pass and the full OVS_E2E=1 suite passes 254 (new coverage cases for gapped/overlapped/span-less, EDL window-overlap cases, and a built-CLI check that a valid plan prints its summary). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VL7mRpmEphLAbcH7YmABnV
…rate taste findings to advisory
Sync of the composition-QA fidelity fixes from Orkas release_1.6.5
(61cf39158, e6a71cc8f, ce7b2427c, 72e541a87, bafbdcda8, 4ef7be4ec),
re-mapped onto the ovs toolchain:
- htmlCopySearch: approved copy split across elements for per-word
reveals could never be found — a CJK line carries no whitespace, and
the raw-HTML haystack put tags between its characters. Script/style
bodies no longer count as visible copy. Applied to scene copy, the
cover headline, and cover signals.
- Accept the meaning, not one spelling: scenes writing the
background/midground/foreground fields the depth fixHint itself
prescribes now pass it; `motion` counts next to motion_verbs. Both
fixHints name the design contract's own scenes[] as the home so the
fix does not land in the manifest's strict canonical scenes[].
- DESIGN_CONTRACT_BUDGET_INCOMPLETE names each missing section's own
required fields — bare section names cost one structurally guaranteed
extra round (add shells, then get told the fields).
- designContractReadiness() runs the contract checks at PREPARE time,
when "before writing HTML" can still be followed, and rides the
prepare/reconcile result. Frame-evidence checks stay with inspect.
- Severity calibration: completeness stays blocking; grading authored
prose does not. GENERIC_AESTHETIC_THESIS, declared-signal count,
fidelity-contract completeness, rendered cover signals, and the cover
hero marker drop to advisory. A declared signal the frame renders as
readable copy counts however it is marked; a signal that only
restates the headline never counts.
- FROZEN_FRAME_RUN removed: identical sampled hashes on an
intentionally static composition are noise the preview already shows.
- runSourceAlignmentQa activates only on a real shotlist shape — a
stray {scenes:[...]} file under the retired name no longer wakes the
legacy alignment layer.
- BLANK_FRAME_MAX_CONTRAST exported so any capture-retry path re-shoots
exactly the frames QA would reject.
Verified: build green; full OVS_E2E=1 suite 266/267 (1 intentional
skip) with 12 new cases covering each behavior above.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VL7mRpmEphLAbcH7YmABnV
…eral second's replacement
Sync of the scene-anchor work from Orkas release_1.6.5 (77aaebf4c,
814e2c5a1, e78b20c0f context), re-mapped onto the ovs scaffold:
- buildCompositionScaffold defines S(id)/D(id) helpers that read each
section's data-start/data-duration, and positions its reveal tweens
from S(id) instead of a literal second. Scene windows move when
`ovs composition reconcile` rewrites the attributes (e.g. once
narration is measured, after the HTML is authored); an anchored tween
follows by construction, a literal keeps playing against the old
window and captures blank at that scene's QA frames.
- reconcileCompositionHtml keeps S()-anchored reveals anchored, and
keeps a legacy literal-positioned scaffold literal (rewriting it to
S() would reference an undefined helper).
- New AUTHORED_ABSOLUTE_TIMELINE_SECONDS advisory in contract QA: scans
tl.set/to/from/fromTo/add/addLabel/call positions (call's position is
its THIRD argument), skips S()/D() and relative string positions, and
hands back the exact replacement — `line N: tl.to(..., 6.5) ->
S("payoff") + 1.5` — instead of only the complaint. The checker knows
every window; the upstream incident left 46 stranded literals and 13
round trips of hand-retiming.
- stage-compose's Timing rule now prescribes S()/D() anchoring; the old
rule ("position every tween with an explicit time so it is
reproducible") is what authored the stranded literals.
Verified: build green; full OVS_E2E=1 suite 269/270 (1 intentional
skip), including a real hyperframes render of the S()-anchored
scaffold, a legacy-scaffold reconcile case, and detector unit cases.
Prompt audit: stage-compose "Timing" bullet replaced (explicit-time
rule -> scene-anchor rule with the retime rationale); no other prompt
clause added or removed; the new rule states what the scaffold now
actually emits, so prompt and code cannot disagree.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VL7mRpmEphLAbcH7YmABnV
… directly, and findings can be waived Sync of the gate-authority reversal from Orkas release_1.6.5 (d2c477f34, e6a71cc8f, 274678ccf, 33904ce59 context), re-mapped onto the ovs resolver and draft gate: - resolveGateTransition: an exhausted visual-QA cycle is now a USER FORK, not a silent wait. `present_findings_and_ask_user_direction` replaces `report_visual_qa_blocker`: show the current frames and remaining findings, offer another repair round or skipping the named check, and end the turn; the user's reply grants the next cycle, and the follow-up edit must be materially different — the failed strategies are recorded, and repeating one spends the new budget for nothing. A real revise decision still starts the fresh cycle directly, unchanged. - New `origin` input: who asked for the change decides whether to ask again. A gate_b_payload revision the CURRENT user turn names in the user's own words resolves to `apply_user_instruction_then_approve_plan` — the instruction is itself the authorization; asking the user to confirm a change they just dictated costs a full round trip and teaches them their instructions are not taken at face value. A model-initiated or mixed reply stays on the amendment path. Origin is the driving agent's honest self-report; nothing verifies it in OVS, and the skill says so. - QA waivers make the skip option honorable: `ovs draft --waive <codes>` (CLI + MCP) downgrades the named blocking findings to informational with a "[skipped by user decision]" suffix, persists them in qa/waivers.json so the user is never asked to skip the same check twice, and refuses evidence-integrity codes (missing frames/maps, parse failures) — those are repaired, not offered. - gate-control SKILL.md carries the matching contract: the exhausted fork with real choices, the origin classification step, the offer-the-waiver rule, and "repair passes that remain never stop for the user; an exhausted cycle does". Verified: build green; full OVS_E2E=1 suite 275/276 (1 intentional skip). New cases: exhausted fork (with and without an error code), user-dictated vs model/mixed amendment, waivable/refused codes, and a draft-gate integration proving a waiver moves past the blocking design finding, refuses the integrity code, persists, and never re-asks. Prompt audit: gate-control SKILL.md — replaced the "technical QA exhaustion must never create a recovery form" clause with the user-fork contract (behavioral reversal shipped in the resolver); added the origin classification step and the offer-the-waiver rule (both state what the resolver/draft now enforce); inverted the closing exhausted-cycle bullet. No clause duplicates another rule; each names the code path that backs it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VL7mRpmEphLAbcH7YmABnV
… the direction stop first Sync of the checkpoint-protocol craft from Orkas release_1.6.5 (c4ba98815 end-state, 707a8bf1f, 734cfc675, ed5d22684, 2182ad89a, 72e541a87, b0514f075, f0e0c348a), prompt-only, re-mapped onto the OSS orchestration/gate-control/video-router split: - orchestration: "Show the work; stop only five times" — the stops are a closed set and showing an artifact is not an ending in itself; the visual preview stops once per VISUAL IDENTITY (narration-only changes inherit the accepted frames); artifacts ride the message that ENDS the turn; an enumerated-option reply IS the decision; everything outside a tool call is user-facing copy (no finding codes or severity words; passing checks stay silent); repair every independent finding in one message and never re-read to confirm a successful write. - Gate A: the direction stop comes first, before any plan file exists; 2-3 genuinely different concepts; an exact brief still stops with ONE concept; propose instead of interrogating for creative preferences. - AUTO: an assembled production is ONE video — the stop count is fixed by the gate table, never per segment; an edit invalidates only its own segment. - gate-control: Gate B opens with the locked direction summary and restates settled facts instead of reopening them; a different-language reply is a revise instruction; amendment aftermath follows the visual identity; a recovery is executed, not narrated (the concrete mutation precedes the validator retry). - video-router: routing ends at the direction boundary — no plan file, script, narration copy, or art direction before the user picks. Verified: skills content tests pin each new clause; suite 266 pass. Prompt audit: every clause above is NEW behavioral guidance absent from the OSS pack (verified against current SKILL.md text); none duplicates gate-control's authorization rules — orchestration carries presentation /stop craft, gate-control carries authorization; the five-stops set matches the resolver's gate enum; no clause contradicts the PR5 user-fork/waiver contract, and the preview-identity rule is consistent with the PR5 amendment-aftermath resolver reasons. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VL7mRpmEphLAbcH7YmABnV
This was referenced Aug 11, 2026
Merged
Checkpoint craft: five stops, visual identity, and the direction stop first
Exhausted QA becomes a user fork; dictated changes apply directly; findings can be waived
Anchor tweens to scene windows; name every literal second's replacement
Repair QA self-contradictions, see split CJK copy, calibrate taste findings to advisory
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Batch 2/8 of the release_1.6.5 sync (narration coverage + plan validation). Stacked on #14 — merge that first, then re-target this to
main. Re-mapped by hand from Orkas commits154c24196,61cf39158,8a94dc1c0,66c1a720e.Real bugs this fixes in OSS today
coverageRatiois blind to interior silence — it is computed asclamp(voicedEnd / ref), i.e. how far the last line reached. The upstream incident: a 60s cut with 29.7s of interior dead air (seven holes up to 6.9s) scored 95%.coverageForSegmentsalready had the per-segment timings in hand and collapsed them to min/max. Now the per-line voiced spans flow through: the report addsvoicedRatio,interiorGaps(≥0.5s, largest first, capped),maxInteriorGapSec, and agappedstatus when a hole exceeds 2.5s.buildMixFilteradelays each line to itsstart_secand amixes — overlapping windows mix as double narration silently. Two layers now catch it:validateEdlrejects overlapping line windows (E_NARRATION_WINDOWS_OVERLAP) at free validation before any synthesis is spent (the classic mistake:target_secwritten as the line's END time instead of its duration), and the coverage report measures raw-span collisions (overlappedstatus) after the mix.ovs plan validateprinted{ok: true}and the model then wrote its own six-line abstract and asked for approval of an EDL the user never saw. A passing validate/plan_validatenow attaches the host-rendered summary — the last thing the model reads before presenting the plan is the real plan.media_type, reference role,edit_strategydecision signal) now name the legal values, matching the style theintent/intent_basismessages already had.Verification
Build green; 244 unit tests pass; full
OVS_E2E=1suite passes 254/254 including a built-CLI check that a valid plan prints itsTimeline:summary. New unit cases: gapped (reach 0.97 vs voiced 0.47), overlapped (1s collision), span-less backward compatibility, EDL window-overlap accept/reject/skip.🤖 Generated with Claude Code
https://claude.ai/code/session_01VL7mRpmEphLAbcH7YmABnV