docs(rag): close the corpus drift — findings/collision/fast-path/forge/env-ladder coverage (#596)

Seventeen-page sweep of the agent-facing KB against shipped behavior:
required covers_parent_criteria and per-AC criteria_verified reach the
QA/PM/task-tools pages (the QA docs also named non-callable pass_review/
fail_review — the MCP tools are pass/fail); collision_context lands in the
QA/gate/planning evidence docs; the possibilities matrix gets its own
architecture page + config entry; the auditor page gains its missing
waive_finding and playbook-curation verbs; git-pr-types.md is rewritten
off the long-dead is_root_pr model; PR/workspace/git-error pages stop
assuming GitHub (forge-agnostic + env-ladder semantics).

Co-authored-by: Renn F <rennf93@users.noreply.github.com>
This commit is contained in:
Renzo F
2026-07-19 19:00:48 +02:00
committed by GitHub
co-authored by Renn F
parent 700dbcd285
commit 862c0b27cd
17 changed files with 217 additions and 47 deletions
+1 -1
View File
@@ -26,7 +26,7 @@ All three open a normal, **assigned** UX/UI authoring task (balanced across the
## Artifact verification (request_render)
Authoring is gated on the RENDERED artifact, not just its source: the `request_render` do-tool (developer/QA) renders the caller's actual composition through the sidecar and extracts evenly spaced keyframe PNGs to a container-shared `.previews/` path, returning their absolute paths in the envelope's `evidence.frames`. The agent must Read every frame and verify each scene/feature from the brief appears fully and legibly — a 14-second cut that only ever shows its first scene is exactly what this catches. A developer renders their own working tree (worktree-aware, `head_sha`/`dirty` provenance stamped); QA renders a read-only `git archive` export of the assembled branch — never a working tree. A successful render stamps the task's `render_preview` marker, and `i_am_done` on a video-authoring task refuses without it (`Requirement.RENDER_VERIFIED`, mirrored in the possibilities-matrix fast path), so no video task can complete on a source-only self-review. QA's `claim_review` evidence carries a `video_context` block (composition id + the dev's stamped preview + an instruction to re-render the branch state) so the reviewer checks output, not source.
Authoring is gated on the RENDERED artifact, not just its source: the `request_render` do-tool (developer/QA) renders the caller's actual composition through the sidecar and extracts evenly spaced keyframe PNGs to a container-shared `.previews/` path, returning their absolute paths in the envelope's `evidence.frames`. The agent must Read every frame and verify each scene/feature from the brief appears fully and legibly — a 14-second cut that only ever shows its first scene is exactly what this catches. A developer renders their own working tree (worktree-aware, `head_sha`/`dirty` provenance stamped); QA renders a read-only `git archive` export of the assembled branch — never a working tree. A successful render stamps the task's `render_preview` marker, and `i_am_done` on a video-authoring task refuses without it (`Requirement.RENDER_VERIFIED`, mirrored in the possibilities-matrix fast path — see `docs/rag/architecture/possibilities-matrix.md`), so no video task can complete on a source-only self-review. QA's `claim_review` evidence carries a `video_context` block (composition id + the dev's stamped preview + an instruction to re-render the branch state) so the reviewer checks output, not source.
## Render loop and the sidecar