Star9
11
代码
代码
Issues
Pull Requests1
流水线
讨论
Wiki
分析
Star9
11
jokp_zjokp_zfeat(slidea): strengthen Phase 3 rules — defer export + edit in place
ca7d1b93创建于 7 天前历史提交
文件最后提交记录最后更新时间
assets
feat(slidea): bundle Noto Sans CJK SC as SVG-route font fallback Intranet hosts without system CJK fonts render Chinese as tofu in the cairosvg-rasterized PNG fallback, which also breaks VLM layout reflection. Bundle subsetted Noto Sans SC (Regular + Bold, ~22MB) under assets/fonts/ and inject a temporary fontconfig pointing at them via _bundled_fonts_context() in screenshot.py. The context is a no-op when fc-list :charset=4e2d already detects system CJK fonts, so user font preferences are respected. Include assets/ in the skill manifest so export_skill.py ships the fonts, and have install.py sanity-check the bundled fonts at the end of setup. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
30 天前
core
fix(slidea): staged execution output layout + session-id safety Six interrelated bugs uncovered when running --stages outline,render against a real Tavily-enabled environment. All changes orbit the same theme: ensure each run's artifacts land in exactly one directory, and make session-id a safe key for run_id recovery. 1. Staged render lost config — generate_pages_node did not accept or thread config, so prepare_generation_context_node inside the SVG subgraph could not see run_id and silently fell back to creating a new directory at the output root. Result: parse/research/thought/ outline landed in <run_id>/ but render landed in a fresh <new_ts>_<topic>/ (without the slides/ wrapper). Fixed by adding config parameter to generate_pages_node and passing config=config at the call site (run_ppt_pipeline.py:417). 2. Staged mode did not auto-recover run_id — only --resume went through _find_run_id_by_session. --stages <non-all> always generated a new run_id, so two staged invocations with the same --session-id created two directories. Fixed by adding a staged branch to the run_id resolver that prefers session-id recovery and only falls through to new_semantic_run_id when no prior run.json matches. 3. slides/svg/ nesting broke relative image paths — SVGs referenced images as images/xxx.png but lived under slides/svg/ while images were at slides/images/, so the relative path resolved to slides/svg/images/xxx.png (did not exist). On-disk SVGs were un- viewable in browsers; PPTX export only worked because the inliner had a .parent fallback. Root cause: _svg_page_filename returned svg/<name> (with prefix). Removed the prefix so SVGs sit directly under slides/, updated all readers (svg_dir fields in ppt.json, patch_render_missing.py, run_ppt_pipeline.py cached-svg lookup), test fixtures, and every doc that mentioned slides/svg/. 4. patch_render_missing save_dir fallback created orphan directory — when patch-render was called on a run that never rendered (no ppt.json), _resolve_save_dir fell back to <output_root>/<topic> with no timestamp prefix. Fixed to use <out_dir>/slides, matching the standard layout. 5. prepare_generation_context_node silently fell back when config was missing — this is why bug #1 above was able to lurk. Changed both SVG and HTML routes to raise RuntimeError when neither state.save_dir nor config-derived cache_dir is available, so future regressions surface immediately instead of contaminating the output root. 6. Default session-id "local" could collide across unrelated runs — staged-mode recovery would silently pick one of the matching runs (filesystem glob order, non-deterministic) and write into it, potentially overwriting an unrelated task's PPTX. Three-layer fix: - Auto-generate session-id (auto_<pid>_<ts>) when --session-id is omitted, so unrelated runs can never collide. - Detect multi-run collisions in _find_run_id_by_session: if multiple original runs match the same session-id, refuse to guess, log a WARNING listing the candidates, and return empty so the caller falls through to a fresh run_id. - Log INFO on successful recovery with the recovered run_id and a preview of the original request text, so users can verify the right run was picked. Also fixes: test stubs in test_cli_stage_smoke.py needed the new config kwarg on generate_pages_node, and needed explicit --session-id local to match existing fixtures once the default stopped being "local". Docs cleanup: --run-id is no longer advertised as a required flag for staged execution (auto-recovery handles it). staged-execution.md, quickstart.md, cli.md, svg-route.md examples all drop --run-id and use --session-id alone. caching-and-paths.md gains a Session-id Collision Detection section. Verified end-to-end with real Tavily keys: - Staged outline,render across two invocations with the same --session-id produces a single run directory (no split). - Deep-mode full pipeline with real images: 4 SVGs at slides/, 54 images at slides/images/, all 7 image hrefs resolve correctly, PPTX exported at 332 KB. - Collision detection unit-tested with two synthetic runs sharing a session-id: WARNING emitted, empty result, no contamination. - Auto session-id unit-tested: each invocation gets a unique auto_<pid>_<ts>, INFO log announces the generated value. - All 16 pipeline-contract + patch-render tests pass. cli_stage_smoke passes except 4 pre-existing langgraph-import errors unrelated to these changes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
13 天前
docs
fix(slidea): staged execution output layout + session-id safety Six interrelated bugs uncovered when running --stages outline,render against a real Tavily-enabled environment. All changes orbit the same theme: ensure each run's artifacts land in exactly one directory, and make session-id a safe key for run_id recovery. 1. Staged render lost config — generate_pages_node did not accept or thread config, so prepare_generation_context_node inside the SVG subgraph could not see run_id and silently fell back to creating a new directory at the output root. Result: parse/research/thought/ outline landed in <run_id>/ but render landed in a fresh <new_ts>_<topic>/ (without the slides/ wrapper). Fixed by adding config parameter to generate_pages_node and passing config=config at the call site (run_ppt_pipeline.py:417). 2. Staged mode did not auto-recover run_id — only --resume went through _find_run_id_by_session. --stages <non-all> always generated a new run_id, so two staged invocations with the same --session-id created two directories. Fixed by adding a staged branch to the run_id resolver that prefers session-id recovery and only falls through to new_semantic_run_id when no prior run.json matches. 3. slides/svg/ nesting broke relative image paths — SVGs referenced images as images/xxx.png but lived under slides/svg/ while images were at slides/images/, so the relative path resolved to slides/svg/images/xxx.png (did not exist). On-disk SVGs were un- viewable in browsers; PPTX export only worked because the inliner had a .parent fallback. Root cause: _svg_page_filename returned svg/<name> (with prefix). Removed the prefix so SVGs sit directly under slides/, updated all readers (svg_dir fields in ppt.json, patch_render_missing.py, run_ppt_pipeline.py cached-svg lookup), test fixtures, and every doc that mentioned slides/svg/. 4. patch_render_missing save_dir fallback created orphan directory — when patch-render was called on a run that never rendered (no ppt.json), _resolve_save_dir fell back to <output_root>/<topic> with no timestamp prefix. Fixed to use <out_dir>/slides, matching the standard layout. 5. prepare_generation_context_node silently fell back when config was missing — this is why bug #1 above was able to lurk. Changed both SVG and HTML routes to raise RuntimeError when neither state.save_dir nor config-derived cache_dir is available, so future regressions surface immediately instead of contaminating the output root. 6. Default session-id "local" could collide across unrelated runs — staged-mode recovery would silently pick one of the matching runs (filesystem glob order, non-deterministic) and write into it, potentially overwriting an unrelated task's PPTX. Three-layer fix: - Auto-generate session-id (auto_<pid>_<ts>) when --session-id is omitted, so unrelated runs can never collide. - Detect multi-run collisions in _find_run_id_by_session: if multiple original runs match the same session-id, refuse to guess, log a WARNING listing the candidates, and return empty so the caller falls through to a fresh run_id. - Log INFO on successful recovery with the recovered run_id and a preview of the original request text, so users can verify the right run was picked. Also fixes: test stubs in test_cli_stage_smoke.py needed the new config kwarg on generate_pages_node, and needed explicit --session-id local to match existing fixtures once the default stopped being "local". Docs cleanup: --run-id is no longer advertised as a required flag for staged execution (auto-recovery handles it). staged-execution.md, quickstart.md, cli.md, svg-route.md examples all drop --run-id and use --session-id alone. caching-and-paths.md gains a Session-id Collision Detection section. Verified end-to-end with real Tavily keys: - Staged outline,render across two invocations with the same --session-id produces a single run directory (no split). - Deep-mode full pipeline with real images: 4 SVGs at slides/, 54 images at slides/images/, all 7 image hrefs resolve correctly, PPTX exported at 332 KB. - Collision detection unit-tested with two synthetic runs sharing a session-id: WARNING emitted, empty result, no contamination. - Auto session-id unit-tested: each invocation gets a unique auto_<pid>_<ts>, INFO log announces the generated value. - All 16 pipeline-contract + patch-render tests pass. cli_stage_smoke passes except 4 pre-existing langgraph-import errors unrelated to these changes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
13 天前
scripts
fix(slidea): sanitize svg_to_pptx.py -n so Phase 3 re-export overwrites run_ppt_pipeline.py calls svgs_to_pptx(..., sanitize_filename(topic)) which converts spaces to underscores and strips illegal chars. The svg_to_pptx.py CLI, however, passed -n verbatim to svgs_to_pptx() without sanitizing. So when Phase 3 re-export ran with "-n <topic from ppt.json>", the new PPTX got a different filename (spaces preserved) than the original pipeline wrote (spaces underscored), and both files ended up coexisting in the run dir. ppt.json still pointed at the old file, so the user opened the unchanged PPTX and the edit appeared lost. Fix: apply sanitize_filename to args.filename in svg_to_pptx.py main(), exactly mirroring the pipeline. Idempotent for already-clean names, so no caller breaks. Doc updates in agent-edit.md Step 10: - Make explicit that <topic> is passed verbatim and the exporter sanitizes it (was wishful thinking before, now actually true) - Warn against pre-sanitizing -n (would double-transform) - Clarify the output path is <sanitized_topic>.pptx Also tightened the -n help text in svg_to_pptx.py to self-document the auto-sanitization. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
7 天前
skill
feat(slidea): strengthen Phase 3 rules — defer export + edit in place Two recurring failure modes in Phase 3 agent-led editing: 1. Auto-export. Despite the existing export-deferral rule, agents sometimes run svg_to_pptx.py right after an edit "to verify" or "to give the user something to look at", rewriting the PPTX and forcing the user to reopen the file. 2. Copy-then-edit. Agents "play safe" by copying output/<run_id>/ to a sibling directory (or a temp path, or a backup) and editing there. ppt.json still points at the original directory, so the user opens an unchanged PPTX and the edit appears lost. Same hazard on export: passing a different output_dir lands the new PPTX somewhere the user will never look. Strengthen the prompt so both failure modes are loud and repeated: agent-edit.md: - §0 restructured into "Critical Rules" with two named rules: Rule 1 (defer export) and Rule 2 (edit in place). Each has its own why and its own "when in doubt" guidance. - Rule 1 adds "When in doubt: do not export" and explicitly lists the bad reasons agents invent ("to verify", "to give the user something to look at", "because the edit is small"). - Rule 2 is new: never copy the run directory, never copy individual SVGs to backup paths, never create a working sub-directory, always pass original paths from ppt.json to svg_to_pptx.py. - Step 7 ends with an "edit in place" reminder cross-referencing Rule 2. - Step 9 heading becomes "Report the SVG change — STOP, do NOT export" with a callout enumerating the bad reasons. - Step 10 adds a pre-flight check: (a) did the user explicitly say to export? (b) are the paths from ppt.json? If either fails, stop. - §6 Common Pitfalls gets two new entries at the top: "Auto-exporting after every edit" and "Copying the run directory before editing". SKILL.md: - Phase 3 invariant now says edits happen "in place" and adds the no-copy / no-different-output-dir clause inline. Pointer to §0 updated to mention "two critical rules". No code changes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
7 天前
tests
fix(slidea): staged execution output layout + session-id safety Six interrelated bugs uncovered when running --stages outline,render against a real Tavily-enabled environment. All changes orbit the same theme: ensure each run's artifacts land in exactly one directory, and make session-id a safe key for run_id recovery. 1. Staged render lost config — generate_pages_node did not accept or thread config, so prepare_generation_context_node inside the SVG subgraph could not see run_id and silently fell back to creating a new directory at the output root. Result: parse/research/thought/ outline landed in <run_id>/ but render landed in a fresh <new_ts>_<topic>/ (without the slides/ wrapper). Fixed by adding config parameter to generate_pages_node and passing config=config at the call site (run_ppt_pipeline.py:417). 2. Staged mode did not auto-recover run_id — only --resume went through _find_run_id_by_session. --stages <non-all> always generated a new run_id, so two staged invocations with the same --session-id created two directories. Fixed by adding a staged branch to the run_id resolver that prefers session-id recovery and only falls through to new_semantic_run_id when no prior run.json matches. 3. slides/svg/ nesting broke relative image paths — SVGs referenced images as images/xxx.png but lived under slides/svg/ while images were at slides/images/, so the relative path resolved to slides/svg/images/xxx.png (did not exist). On-disk SVGs were un- viewable in browsers; PPTX export only worked because the inliner had a .parent fallback. Root cause: _svg_page_filename returned svg/<name> (with prefix). Removed the prefix so SVGs sit directly under slides/, updated all readers (svg_dir fields in ppt.json, patch_render_missing.py, run_ppt_pipeline.py cached-svg lookup), test fixtures, and every doc that mentioned slides/svg/. 4. patch_render_missing save_dir fallback created orphan directory — when patch-render was called on a run that never rendered (no ppt.json), _resolve_save_dir fell back to <output_root>/<topic> with no timestamp prefix. Fixed to use <out_dir>/slides, matching the standard layout. 5. prepare_generation_context_node silently fell back when config was missing — this is why bug #1 above was able to lurk. Changed both SVG and HTML routes to raise RuntimeError when neither state.save_dir nor config-derived cache_dir is available, so future regressions surface immediately instead of contaminating the output root. 6. Default session-id "local" could collide across unrelated runs — staged-mode recovery would silently pick one of the matching runs (filesystem glob order, non-deterministic) and write into it, potentially overwriting an unrelated task's PPTX. Three-layer fix: - Auto-generate session-id (auto_<pid>_<ts>) when --session-id is omitted, so unrelated runs can never collide. - Detect multi-run collisions in _find_run_id_by_session: if multiple original runs match the same session-id, refuse to guess, log a WARNING listing the candidates, and return empty so the caller falls through to a fresh run_id. - Log INFO on successful recovery with the recovered run_id and a preview of the original request text, so users can verify the right run was picked. Also fixes: test stubs in test_cli_stage_smoke.py needed the new config kwarg on generate_pages_node, and needed explicit --session-id local to match existing fixtures once the default stopped being "local". Docs cleanup: --run-id is no longer advertised as a required flag for staged execution (auto-recovery handles it). staged-execution.md, quickstart.md, cli.md, svg-route.md examples all drop --run-id and use --session-id alone. caching-and-paths.md gains a Session-id Collision Detection section. Verified end-to-end with real Tavily keys: - Staged outline,render across two invocations with the same --session-id produces a single run directory (no split). - Deep-mode full pipeline with real images: 4 SVGs at slides/, 54 images at slides/images/, all 7 image hrefs resolve correctly, PPTX exported at 332 KB. - Collision detection unit-tested with two synthetic runs sharing a session-id: WARNING emitted, empty result, no contamination. - Auto session-id unit-tested: each invocation gets a unique auto_<pid>_<ts>, INFO log announces the generated value. - All 16 pipeline-contract + patch-render tests pass. cli_stage_smoke passes except 4 pre-existing langgraph-import errors unrelated to these changes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
13 天前
.env.example
feat(slidea): configurable OUTPUT_DIR and source_dir in completion output Two related usability improvements: 1. OUTPUT_DIR env var (.env) redirects every slidea artifact to a configured directory. Empty (default) keeps the existing <slidea_install_dir>/output behavior; an absolute path replaces it verbatim; a relative path resolves against <slidea_install_dir>. Affects PPT runs, deep research workspace, sqlite checkpoints, and downloaded documents — everything that previously landed under <slidea_install_dir>/output. Implementation: settings.OUTPUT_DIR is resolved in core/utils/config.py right after settings load, producing the module-level output_files_dir that every existing importer already consumes. cache.run_dir drops its vestigial base_dir parameter and uses output_files_dir directly. 2. PPT generation completion now surfaces the editable SVG source directory alongside the PPTX path: >>> 生成PPT结束 >>> 生成文件：/.../<run_id>/<topic>.pptx >>> 源文件目录：/.../<run_id>/slides/svg So users can find the on-disk SVG source for further editing without having to read ppt.json. HTML route reports its save_dir as source_dir (HTML pages live directly there). Signed-off-by: Zhukunpeng <zhukunpeng8@huawei.com>
15 天前
.gitignore
use images of extracted documents Signed-off-by: SuBo <subo7@huawei.com>
2 个月前
README.md
fix(slidea): staged execution output layout + session-id safety Six interrelated bugs uncovered when running --stages outline,render against a real Tavily-enabled environment. All changes orbit the same theme: ensure each run's artifacts land in exactly one directory, and make session-id a safe key for run_id recovery. 1. Staged render lost config — generate_pages_node did not accept or thread config, so prepare_generation_context_node inside the SVG subgraph could not see run_id and silently fell back to creating a new directory at the output root. Result: parse/research/thought/ outline landed in <run_id>/ but render landed in a fresh <new_ts>_<topic>/ (without the slides/ wrapper). Fixed by adding config parameter to generate_pages_node and passing config=config at the call site (run_ppt_pipeline.py:417). 2. Staged mode did not auto-recover run_id — only --resume went through _find_run_id_by_session. --stages <non-all> always generated a new run_id, so two staged invocations with the same --session-id created two directories. Fixed by adding a staged branch to the run_id resolver that prefers session-id recovery and only falls through to new_semantic_run_id when no prior run.json matches. 3. slides/svg/ nesting broke relative image paths — SVGs referenced images as images/xxx.png but lived under slides/svg/ while images were at slides/images/, so the relative path resolved to slides/svg/images/xxx.png (did not exist). On-disk SVGs were un- viewable in browsers; PPTX export only worked because the inliner had a .parent fallback. Root cause: _svg_page_filename returned svg/<name> (with prefix). Removed the prefix so SVGs sit directly under slides/, updated all readers (svg_dir fields in ppt.json, patch_render_missing.py, run_ppt_pipeline.py cached-svg lookup), test fixtures, and every doc that mentioned slides/svg/. 4. patch_render_missing save_dir fallback created orphan directory — when patch-render was called on a run that never rendered (no ppt.json), _resolve_save_dir fell back to <output_root>/<topic> with no timestamp prefix. Fixed to use <out_dir>/slides, matching the standard layout. 5. prepare_generation_context_node silently fell back when config was missing — this is why bug #1 above was able to lurk. Changed both SVG and HTML routes to raise RuntimeError when neither state.save_dir nor config-derived cache_dir is available, so future regressions surface immediately instead of contaminating the output root. 6. Default session-id "local" could collide across unrelated runs — staged-mode recovery would silently pick one of the matching runs (filesystem glob order, non-deterministic) and write into it, potentially overwriting an unrelated task's PPTX. Three-layer fix: - Auto-generate session-id (auto_<pid>_<ts>) when --session-id is omitted, so unrelated runs can never collide. - Detect multi-run collisions in _find_run_id_by_session: if multiple original runs match the same session-id, refuse to guess, log a WARNING listing the candidates, and return empty so the caller falls through to a fresh run_id. - Log INFO on successful recovery with the recovered run_id and a preview of the original request text, so users can verify the right run was picked. Also fixes: test stubs in test_cli_stage_smoke.py needed the new config kwarg on generate_pages_node, and needed explicit --session-id local to match existing fixtures once the default stopped being "local". Docs cleanup: --run-id is no longer advertised as a required flag for staged execution (auto-recovery handles it). staged-execution.md, quickstart.md, cli.md, svg-route.md examples all drop --run-id and use --session-id alone. caching-and-paths.md gains a Session-id Collision Detection section. Verified end-to-end with real Tavily keys: - Staged outline,render across two invocations with the same --session-id produces a single run directory (no split). - Deep-mode full pipeline with real images: 4 SVGs at slides/, 54 images at slides/images/, all 7 image hrefs resolve correctly, PPTX exported at 332 KB. - Collision detection unit-tested with two synthetic runs sharing a session-id: WARNING emitted, empty result, no contamination. - Auto session-id unit-tested: each invocation gets a unique auto_<pid>_<ts>, INFO log announces the generated value. - All 16 pipeline-contract + patch-render tests pass. cli_stage_smoke passes except 4 pre-existing langgraph-import errors unrelated to these changes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
13 天前
README_CN.md
fix(slidea): staged execution output layout + session-id safety Six interrelated bugs uncovered when running --stages outline,render against a real Tavily-enabled environment. All changes orbit the same theme: ensure each run's artifacts land in exactly one directory, and make session-id a safe key for run_id recovery. 1. Staged render lost config — generate_pages_node did not accept or thread config, so prepare_generation_context_node inside the SVG subgraph could not see run_id and silently fell back to creating a new directory at the output root. Result: parse/research/thought/ outline landed in <run_id>/ but render landed in a fresh <new_ts>_<topic>/ (without the slides/ wrapper). Fixed by adding config parameter to generate_pages_node and passing config=config at the call site (run_ppt_pipeline.py:417). 2. Staged mode did not auto-recover run_id — only --resume went through _find_run_id_by_session. --stages <non-all> always generated a new run_id, so two staged invocations with the same --session-id created two directories. Fixed by adding a staged branch to the run_id resolver that prefers session-id recovery and only falls through to new_semantic_run_id when no prior run.json matches. 3. slides/svg/ nesting broke relative image paths — SVGs referenced images as images/xxx.png but lived under slides/svg/ while images were at slides/images/, so the relative path resolved to slides/svg/images/xxx.png (did not exist). On-disk SVGs were un- viewable in browsers; PPTX export only worked because the inliner had a .parent fallback. Root cause: _svg_page_filename returned svg/<name> (with prefix). Removed the prefix so SVGs sit directly under slides/, updated all readers (svg_dir fields in ppt.json, patch_render_missing.py, run_ppt_pipeline.py cached-svg lookup), test fixtures, and every doc that mentioned slides/svg/. 4. patch_render_missing save_dir fallback created orphan directory — when patch-render was called on a run that never rendered (no ppt.json), _resolve_save_dir fell back to <output_root>/<topic> with no timestamp prefix. Fixed to use <out_dir>/slides, matching the standard layout. 5. prepare_generation_context_node silently fell back when config was missing — this is why bug #1 above was able to lurk. Changed both SVG and HTML routes to raise RuntimeError when neither state.save_dir nor config-derived cache_dir is available, so future regressions surface immediately instead of contaminating the output root. 6. Default session-id "local" could collide across unrelated runs — staged-mode recovery would silently pick one of the matching runs (filesystem glob order, non-deterministic) and write into it, potentially overwriting an unrelated task's PPTX. Three-layer fix: - Auto-generate session-id (auto_<pid>_<ts>) when --session-id is omitted, so unrelated runs can never collide. - Detect multi-run collisions in _find_run_id_by_session: if multiple original runs match the same session-id, refuse to guess, log a WARNING listing the candidates, and return empty so the caller falls through to a fresh run_id. - Log INFO on successful recovery with the recovered run_id and a preview of the original request text, so users can verify the right run was picked. Also fixes: test stubs in test_cli_stage_smoke.py needed the new config kwarg on generate_pages_node, and needed explicit --session-id local to match existing fixtures once the default stopped being "local". Docs cleanup: --run-id is no longer advertised as a required flag for staged execution (auto-recovery handles it). staged-execution.md, quickstart.md, cli.md, svg-route.md examples all drop --run-id and use --session-id alone. caching-and-paths.md gains a Session-id Collision Detection section. Verified end-to-end with real Tavily keys: - Staged outline,render across two invocations with the same --session-id produces a single run directory (no split). - Deep-mode full pipeline with real images: 4 SVGs at slides/, 54 images at slides/images/, all 7 image hrefs resolve correctly, PPTX exported at 332 KB. - Collision detection unit-tested with two synthetic runs sharing a session-id: WARNING emitted, empty result, no contamination. - Auto session-id unit-tested: each invocation gets a unique auto_<pid>_<ts>, INFO log announces the generated value. - All 16 pipeline-contract + patch-render tests pass. cli_stage_smoke passes except 4 pre-existing langgraph-import errors unrelated to these changes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
13 天前
THIRD_PARTY_NOTICES
feat(slidea): add Phase 3 agent-led page editing workflow Introduces an agent-driven editing mode that modifies SVG pages directly under output/<run_id>/slides/svg/ and re-exports the PPTX, without re-running the generation pipeline. New references: - agent-edit.md: standard edit workflow (10-step process, required reading, template protection, QC self-check, PPTX re-export commands, ambiguity handling) - diagram-basics.md: shared SVG drawing foundation (layering order, 6 component patterns, spacing formulas, slidea SVG constraints, decision table for when to read each layout doc) - diagram-layouts/architecture.md, flowchart.md, sequence.md, structural.md: per-type layout algorithms and SVG skeletons SKILL.md updates: - Add Phase 3 section describing the responsibility chain (SKILL.md -> agent-edit.md -> diagram-basics.md -> diagram-layouts/<type>.md) - Extend frontmatter description to cover Phase 3 triggering (modify, redraw, diagram on existing slide) THIRD_PARTY_NOTICES: - Add baoyu-skills/baoyu-diagram attribution (MIT, Copyright 2026 Jim Liu) for the layout algorithm references adapted into diagram-basics.md and the four layout docs (color/font styling stripped, spacing rewritten as relative formulas) slidea pipeline code is unchanged; Phase 3 edits go through the existing scripts/svg_to_pptx.py exporter and the existing QC checker. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> Signed-off-by: Zhukunpeng <zhukunpeng8@huawei.com>
14 天前
requirements.txt
fix(slidea): remove unused package in requirements.txt
21 天前
README.md

Slidea

中文 | English

Demo

Slidea 是一个AI 驱动的 PPT 生成 skill，用于将高层次的PPT 生成需求转化为结构化研究材料、写作思路、幻灯片大纲，并最终生成可演示的 PPT 成稿。

样例一（英文） 样例二（英文）
SIGOPS Workshop（Kimi-k2.5）


查看 Prompt
Please make me an opening slides for the SIGOPS strategic workshop (https://ipads.se.sjtu.edu.cn/sigops-strategic/), emphasizing the long history of SIGOPS and the community and this is the 60th anniversary of SIGOPS (https://www.sigops.org/about/history/), review the 2015 SOSP History Day (https://sigops.org/s/conferences/sosp/2015/history/), the high-quality of the program, the two great keynote speeches, and the program for visionary talks, as well as the two great panels and workshop Session schedule. Please make around 8-10 slides.
 《呼啸山庄》介绍（Kimi-k2.5）


查看 Prompt
Help me create an English PowerPoint presentation to introduce the book Wuthering Heights
样例三（中文） 样例四（中文）
AI Agent 介绍（Gemini-3-pro）


查看 Prompt
帮我生成一个30页左右的PPT，内容是关于AI Agent，包括AI Agent基本原理，主要框架、面临的挑战、学术界进展，以及未来的机会点。
 幼儿园脱口秀（DeepSeek-V3.2）


查看 Prompt
请帮我生成一份5岁小朋友脱口秀的ppt，演讲题目是“假如我会魔法”

Slidea 能做什么

给定一个 PPT 生成请求，如“面向产品、技术与业务负责人，生成一份 10 页的 PPT，介绍 AI Agent”，Slidea 可以：

  • 将请求解析为结构化需求，
  • 从用户输入、URL 与可选搜索中收集资料，
  • 生成整份演示文稿的写作思路，
  • 将写作思路转化为幻灯片大纲，
  • 把每一页渲染为 SVG，并最终导出为可编辑的原生 PPTX。

这套系统面向 agent 驱动的使用场景，它支持分阶段执行、断开后续接执行等灵活的机制。之所以采用分阶段设计，主要有两个原因：

  • 研究、规划、大纲生成和渲染分离后，整体生成质量更稳定；
  • 中间产物可以被缓存、检查、编辑、恢复或重复利用。

快速开始：使用 Agent 安装 Skill（推荐）

Slidea 提供两个可安装到 agent 环境中的 skill：Slidea（PPT 生成）和 Deep Research（深度研究）。如果你的 agent 平台支持本地 skill，则可以安装其中一个或两个。安装完成之后，按当前二元路由配置 .env 即可开始使用。

两个 skill 均已适配 openEuler，Apple Silicon macOS，Windows WSL/PowerShell，以及部分其他 Linux 系统。可在主流 agent 环境中快捷安装并运行，如 OpenClaw、Codex、Claude Code 等。

Slidea Skill

Slidea 是 AI 驱动的 PPT 生成 skill。默认建议先配置 DEFAULT_LLM；如果你要开启 PREMIUM 模式，推荐保持 PREMIUM_LLM_MODEL=google/gemini-3.1-pro-preview 与 PREMIUM_LLM_API_BASE_URL=https://openrouter.ai/api/v1 不变，通常只补充 PREMIUM_LLM_API_KEY 即可。

安装 Slidea Skill

可将以下指令发送给你的 Agent：

请直接获取并遵循这里的说明安装 slidea skill：https://raw.gitcode.com/openeuler/capsule/raw/master/application/slidea/skill/slidea/INSTALL.md

安装并配置完 Slidea 模型资源后，重启 agent，使其重新加载已安装的 skill。随后通过你的 agent 环境支持的 skill 调用方式来触发 Slidea。

升级 Slidea Skill

要升级已有的 Slidea 安装，请将以下指令发送给你的 Agent：

请直接获取并遵循这里的说明更新 slidea skill：https://raw.gitcode.com/openeuler/capsule/raw/master/application/slidea/skill/slidea/UPDATE.md

使用 Slidea Skill

在 OpenClaw 这种环境中，你可能会这样调用：

使用 slidea skill 创建一份关于 AI Agent 的 PPT，10 页左右，面向产品、技术与业务负责人，不要深入洞察

在 claude code 这种支持 slash 风格 skill 命令的环境中，你可能会这样调用：

/slidea 创建一份关于 AI Agent 的 PPT，10 页左右，面向产品、技术与业务负责人，不要深入洞察

具体语法取决于宿主 agent，但预期体验是一致的：agent 加载 Slidea skill，在必要时补充缺失信息，并将幻灯片生成流水线执行到最终产物。

Deep Research Skill

Deep Research 是一个可独立安装和使用的深度研究 skill。它可以对给定主题进行多源调查，自动搜索网络、提取相关页面内容，并将发现综合为结构化的研究报告。

安装 Deep Research Skill

要安装 Deep Research skill，请将以下指令发送给你的 Agent：

请直接获取并遵循这里的说明安装 Deep Research skill：https://raw.gitcode.com/openeuler/capsule/raw/master/application/slidea/skill/deep_research/INSTALL.md

安装完成后，在已安装的 skill 目录中配置 .env，至少需要填写 DEFAULT_LLM_MODEL、DEFAULT_LLM_API_KEY 和 DEFAULT_LLM_API_BASE_URL。可选配置 TAVILY_API_KEYS 以启用网络搜索能力。

升级 Deep Research Skill

要升级已有的 Deep Research 安装，请将以下指令发送给你的 Agent：

请直接获取并遵循这里的说明更新 Deep Research skill：https://raw.gitcode.com/openeuler/capsule/raw/master/application/slidea/skill/deep_research/UPDATE.md

使用 Deep Research Skill

在 Agent 环境中，通过研究主题调用 Deep Research skill：

使用 deep_research skill 研究 <主题>

或使用斜杠命令风格：

/deep_research <研究主题>

该 skill 将自动搜索网络、提取并综合信息，最终生成一份结构化的 Markdown 格式研究报告。一次典型的研究运行需要 10-30 分钟。

支持平台

平台 架构 支持情况
Linux x86_64 / ARM64 支持openEuler
Linux x86_64 支持 Ubuntu/Debian
Windows x86_64 / ARM64 ✅
macOS Apple Silicon ✅

从源码使用

如果你是为了贡献 Slidea 本身，或需要在本地调试仓库代码，可以直接从源码使用 Slidea。

  1. 获取源码并进入目录：

    git clone https://gitcode.com/openeuler/capsule.git
cd capsule/application/slidea

  2. 使用脚本自动创建虚拟环境并安装相关依赖： 这一步会安装 SVG 渲染路线所需的 Python 依赖并准备 .env。

    python3 scripts/install/install.py

  3. 配置环境变量： 如果脚本还没有自动生成 .env，可以先执行：

    cp .env.example .env

    然后在 .env 中至少配置：

    • SLIDEA_MODE
    • DEFAULT_LLM_MODEL
    • DEFAULT_LLM_API_KEY
    • DEFAULT_LLM_API_BASE_URL 当前这些配置仅支持 OpenAI-compatible API。 最小可运行配置是 SLIDEA_MODE=ECONOMIC 加上三项 DEFAULT_LLM_*。 如果 Slidea 连接的是 model_service，并希望由 model_service 根据 AgentProfile 自动选择模型，可以设置 MODEL_INVOKE_HANDOVER=true。此时 Slidea 会忽略 SLIDEA_MODE、PREMIUM_LLM_* 和 DEFAULT_VLM_*，所有文本和视觉请求都发往 DEFAULT_LLM_API_BASE_URL，请求中的模型名留空，并携带 AgentProfile header；仍需要配置 DEFAULT_LLM_API_KEY 和 DEFAULT_LLM_API_BASE_URL，但不再要求 DEFAULT_LLM_MODEL。 Slidea 采用两层模型路由。DEFAULT_LLM 是必填项——绝大部分 pipeline 调用点（解析、深度研究、大纲默认分支、页面生成、HTML 视觉复核）都固定走它；没有它 pipeline 起不来。PREMIUM_LLM 是可选项——只有 SLIDEA_MODE=PREMIUM 时才会用在大纲主结构、SVG 页面生成这两个质量关键调用点；调用失败会自动回退到 DEFAULT_LLM。

    三种配置情况：

    • 只配 DEFAULT_LLM：pipeline 能完整跑完。即使 SLIDEA_MODE=PREMIUM，premium 路由的调用点也只会打一条警告并回退到 DEFAULT_LLM，效果等同 ECONOMIC。这是最小可用配置，常规使用足够了。
    • 只配 PREMIUM_LLM（DEFAULT_LLM 留空）：pipeline 在第一个 DEFAULT 路由调用点抛 Missing configuration for default_llm 后中断。不支持这种配置。
    • 两个都配 + SLIDEA_MODE=PREMIUM：premium 路由的调用点优先走 PREMIUM_LLM，调用失败时自动回退到 DEFAULT_LLM。

    如果你希望 premium 路由调用点优先使用高级模型，再额外补充 PREMIUM_LLM_API_KEY。默认 PREMIUM_LLM_MODEL=google/gemini-3.1-pro-preview 即可；GLM-5.2 也是 premium 槽位的推荐选项。

    推荐模型：

    • DEFAULT_LLM_MODEL：google/gemini-3.1-pro-preview、GLM-5.2 或 deepseek-v4-pro
    • PREMIUM_LLM_MODEL：google/gemini-3.1-pro-preview 或 GLM-5.2（默认为 gemini）
    • DEFAULT_VLM_MODEL：kimi-2.5 或 kimi-2.6

  4. 快速示例：

    .venv/bin/python scripts/run_ppt_pipeline.py \
  --text "生成一份 10 页的 PPT，介绍 AI Agent，不要深入洞察，面向技术人员介绍 Agent 技术发展趋势" \

    任务发起之后，大约 15 分钟左右，会生成 PPTX 文件，并在 Agent 上下文中告知文件路径。

  5. 详细示例 #1：

    .venv/bin/python scripts/run_ppt_pipeline.py \
  --text "生成一份 10 页的 PPT，介绍 AI Agent，不要深入洞察" \
  --session-id session_test \
  --run-id id_test

    其中，若不指定 session-id 和 run-id，系统将采用当前时间作为默认 id 值。上述例子，使用字符串 session_test 和 id_test 作为示例。

    恢复被中断的运行

    生成 PPT 的过程中会中断，并与用户交互。Slidea CLI支持恢复被中断的 PPT 生成任务。

    例如：当 scripts/run_ppt_pipeline.py 返回 stage: "input_required" 时，表示需要用户补充信息。这种情况下，需要使用相同的 run_id、session_id 和 --resume 再次调用 CLI。

    示例：

    .venv/bin/python scripts/run_ppt_pipeline.py \
--resume "面向技术人员介绍 Agent 技术发展趋势" \
--session-id session_test \
--run-id id_test

    其中 session_test 和 id_test 是之前已创建但被中断执行的 id 。

  6. 详细示例 #2：

    .venv/bin/python scripts/run_ppt_pipeline.py \
  --text "生成一份 10 页的 PPT，介绍 AI Agent，请深入洞察，面向技术人员介绍 Agent 技术发展趋势" \
  --session-id session_test \
  --run-id id_test

    slidea 自带深入洞察能力，如果 PPT 制作主题较为宽泛，可发起深入洞察（20-30 分钟），再基于洞察报告生成 PPTX 文件。

更多命令可以参考 docs/cli.md。

如果你不想使用步骤2中的安装脚本自动准备运行环境，也可以手动完成，步骤如下：

python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt

仓库结构

  • scripts/: 面向用户的 CLI 入口，包括 skill 导出、完整流水线、分阶段执行、补渲染以及嵌套的安装辅助脚本
  • skill/: 导出的 skill 包定义目录，包含 skill/slidea/（Slidea PPT 生成 skill）和 skill/deep_research/（Deep Research 深度研究 skill），各自包含 SKILL.md、INSTALL.md、UPDATE.md 及 skill 清单
  • core/: 主要的 LangGraph 应用，包括深度研究、PPT 生成以及共享核心工具
  • docs/: 面向公开仓库读者的文档，包括快速开始、CLI、架构和 app 说明
  • tests/: 针对可移植性、CLI 契约与运行时行为的回归测试

核心子系统

PPT Generator

core/ppt_generator/ 负责面向演示文稿的生成。

它会把源材料转化为：

  • 演示文稿的写作思路，
  • 幻灯片大纲，
  • 每页 SVG 渲染结果，
  • 以及最终可编辑的原生 PPTX 产物。

这个子系统被拆分出来，是为了把“如何思考这份 deck”与“如何把 deck 渲染出来”这两类问题分开处理。

使用示例：

.venv/bin/python scripts/run_ppt_pipeline.py \
  --text "生成一份 10 页的 PPT，介绍 AI Agent，不要深入洞察，面向技术人员介绍 Agent 技术发展趋势"

在 PPT 渲染过程中，会采用 fewshots 的模式，让模型输出的排版保持一致的风格。

当前，Slidea 内置了浅色通用，深色通用，红色政治，蓝色学术，幼小科普，中文艺术，英文艺术 7 种模板。Slidea默认会根据用户输入的主题自行选择最合适的模板渲染排版，但用户也可以直接在 PPT 生成任务的请求中，明确要求Slidea采用何种风格的模板。

Deep Research

core/deep_research/ 负责递归研究和长文综合。

它不负责幻灯片渲染，而是将一个宽泛的需求扩展成结构化研究过程，包括问题拆解、证据收集、缺口审视，以及生成可供演示流水线消费的研究产出。

当任务在进入幻灯片规划前需要先形成洞察时，应关注这一部分。

CLI 概览

Slidea 主要暴露三个脚本入口：

  • scripts/install/install.py: 初始化源码工作区或导出 skill 包的本地运行时依赖，skill 安装过程中会被调用
  • scripts/export_skill.py: 从源码树导出 skill 包，skill 安装过程中会被调用
  • scripts/run_ppt_pipeline.py: 主生成流水线，支持分阶段执行，PPT 生成任务发起时被调用
  • scripts/patch_render_missing.py: 对缺失页或指定页进行补渲染，PPT 内容生成不完整时被调用

完整参数说明和 JSON 返回契约请参考 CLI Reference。

输出与缓存

每次 PPT 生成任务的运行都由一个 run_id 标识，PPT 生成任务执行过程中的所有中间结果将缓存在 output/<run_id>/ 目录下。该目录默认位于 slidea skill 安装目录中；可在 .env 中设置 OUTPUT_DIR 将所有运行产物、缓存和中间结果重定向到其他目录。

缓存的常见文件包括：

  • run.json
  • references/
  • research/
  • thought/thought.md
  • outline/outline.json
  • ppt.json

最终渲染产物将保存在 ppt.json 记录的 slides_dir 路径下。SVG 渲染路线会产出可编辑的 slides/*.svg（磁盘上的唯一源文件，使用相对图片路径引用），以及一份位于 cache 根目录的原生可编辑 *.pptx。图片在 PPTX 导出时于临时目录中内联为 data: URI，因此磁盘上的 SVG 保持小巧可编辑。这种分离让系统可以在不重跑全流程的情况下，重新进入某一阶段或执行补渲染。

运行时降级行为

运行时是由配置驱动的。当可选服务缺失时，系统会降级，而不是整体失败：

  • 没有 Tavily 配置：跳过网页搜索
  • embedding 被禁用或未配置：跳过基于 embedding 的排序
  • 没有 VLM 配置：跳过基于 VLM 的图片评分与分发能力

文档导航

根据你的目标，可以从这里开始：

贡献

以下方向的贡献价值最高：

  • CLI 契约与运行时稳定性
  • research graph 质量
  • 大纲或渲染质量
  • 可移植性与环境处理
  • 对外公开文档

如果你修改了行为，请在同一个改动中同步更新 docs/ 下对应文档。

HTML 渲染路线（可选）

上文描述的 SVG 路线是默认路线，也是 skill/SKILL.md 暴露给宿主 agent 的唯一路线。Slidea 还保留了 HTML 渲染路线作为可选备用方案，适合需要 HTML/CSS 表现力、需要走 HTML→PDF→PPTX 转换流程的用户。该路线为按需启用，需要额外系统依赖。

HTML 路线会通过无头 Chromium 浏览器把每一页渲染为 HTML/CSS，合并每页 PDF 后再通过 LibreOffice 转换为 PPTX。它在代码库中完整保留（core/ppt_generator/thought_to_ppt/page_generators/ 与 core/ppt_generator/utils/browser.py），但默认不安装也不暴露。

何时使用 HTML 路线

仅在以下任一情况属实时才选择 HTML 路线：

  • 你确实需要 HTML/CSS 视觉模型（例如 SVG 无法表达的某些复杂 CSS 布局）；
  • 你希望最终产物中除 PPTX 外还包含 PDF 中间产物；
  • 你在调试旧版流水线。

其他场景下，请优先使用默认的 SVG 路线——它无需 Playwright 或 LibreOffice，可直接产出可编辑的原生 PPTX。

安装额外依赖

最简单的做法 —— 让安装脚本处理：

python3 scripts/install/install.py --with-html-route

各平台行为：

平台 LibreOffice 处理
Linux x86_64 下载 AppImage 到 libreoffice/。
Linux ARM64 不自动安装，打印发行版包管理器命令让你手动执行；RHEL 系还会准备 helper 脚本（见下）。
macOS（Apple Silicon） 下载 DMG，把 LibreOffice.app 拷到 libreoffice/，去除隔离属性。
Windows（x86_64 / ARM64） 运行对应 MSI 到 libreoffice/。

Playwright Chromium 在所有平台上都会装进 .venv。

手动兜底（脚本不适用时）：

.venv/bin/pip install playwright PyPDF2
.venv/bin/python -m playwright install chromium
# 然后用系统包管理器装 LibreOffice（≥ 25.2）

RHEL 系 Linux：额外的 helper 步骤

在 RHEL / CentOS / Fedora / openEuler / Rocky / Alma 上，Playwright 的 Chromium 需要一些安装脚本无法自动装的系统库。--with-html-route 的日志会打出一条精确的 bash "..." 命令，指向 scripts/install/extra_install_linux_rhel.sh —— 这条命令需要你手动执行（需要 sudo）。不跑这一步，HTML 路线只能产出 PDF，PPTX 导出会失败。

以 --render-mode html 运行

.venv/bin/python scripts/run_ppt_pipeline.py \
  --text "<请求内容>" \
  --render-mode html

HTML 路线的额外产出

HTML 路线会在渲染目录下额外写出：

  • 每页的 *.html 文件（一页一份），
  • 合并后的 *.pdf，
  • 通过 LibreOffice 把 PDF 转 PPTX 得到的最终 *.pptx。

HTML 运行的 ppt.json 会记录 render_mode: "html"，并相应记录 pdf_path 与 pptx_path。

HTML 路线专属的降级行为

启用 HTML 路线时，会多出一条降级规则：

  • 没有可用的 LibreOffice 转换：保留 HTML/PDF 输出，跳过 PPTX 转换。（默认 SVG 路线不依赖 LibreOffice，不受此规则影响。）

注意事项

  • --render-mode html 不会通过 skill/SKILL.md 暴露给宿主 agent。agent 使用 Slidea skill 时，始终走默认 SVG 路线，除非你显式把 HTML 路线命令交给它。
  • scripts/html_to_pptx.py（独立的 HTML→PPTX 转换器）以及 scripts/patch_render_missing.py（当 ppt.json 记录的是 render_mode: "html" 时）在补齐 HTML 路线依赖后仍然可用。
  • 如果你之前手动安装过 Playwright / PyPDF2 / LibreOffice，之后某次升级改动了 requirements.txt，.venv 中已装的包不会被自动卸载。HTML 路线通常仍可使用；若失效，按上述步骤重装即可。

第三方声明

本项目包含 PPT Master 的代码，采用 MIT 许可证，详见 THIRD_PARTY_NOTICES。