Skip to content

fix(ppt-engine): say what a reference is for on the provider it runs on - #1261

Open
joei4cm wants to merge 5 commits into
EverMind-AI:mainfrom
joei4cm:fix/ppt_minimax_reference_schema
Open

joei4cm wants to merge 5 commits into
EverMind-AI:mainfrom
joei4cm:fix/ppt_minimax_reference_schema

Conversation

@joei4cm

@joei4cm joei4cm commented Nov 11, 2026 •

Copy link
Copy Markdown
Contributor

Summary

ppt_generate_image told every deployment that references carries up to six pictures to match or vary, but the transport underneath it does not keep that promise on MiniMax: the host's MiniMax route sends the first reference alone, as a subject_reference of type: "character", and drops the rest without a word. The deck's model, reading a schema that offered six, asked for several and got one -- and was never told the others were not sent.

The tool now authors its own to_schema, the same way image_generate does, so the registry serves the answer live and a section changed between turns is read on the next one rather than frozen at admission:

  • on MiniMax, the tool description and the references parameter say one picture, kept as the character it shows, with the manner in the prompt where a reference cannot carry it;
  • off MiniMax, both stay byte-identical to what they were -- verified against the string origin/main renders, not just by eye.

The issue left the second-reference behaviour to the PR: this sends the first (which is what the transport does anyway, so no request changes shape), names the file for what was sent, and puts a note on the result saying how many were left out. Refusing instead would fail a whole batch of pictures for one extra reference; the note keeps the batch moving and fixes the complaint, which was the silence.

The wording reaches every ask, not only the single-picture one: a picture asked for inside prompts is a request of its own, so it carries the same one-reference description on MiniMax, and the ppt-script-authoring skill page -- which an author reads where the schema is not in front of them -- says the deployment keeps the first picture only, so the character to stay consistent with goes in that one.

Tests, in the shape the issue asked for: the registry-served schema with the section's provider switched between turns (the deck's counterpart of test_the_image_tool_says_what_references_do_on_the_provider_it_runs_on), and the reference handling itself -- one subject_reference on the wire, and the note on the answer.

Type

  • Fix
  • Feature
  • Docs
  • CI / tooling
  • Refactor
  • Other

Verification

  • Relevant tests pass locally

  • Relevant lint / type checks pass locally

  • User-facing docs or screenshots are updated when needed

  • PYTHONUTF8=1 uv run pytest tests/test_ppt_engine_generate_image_tool.py -o addopts='' -p no:xdist -q - 16 passed (14 existing + 2 new)

  • PYTHONUTF8=1 uv run pytest tests/test_ppt_engine_generate_image_tool.py tests/test_ppt_engine_image_search.py tests/test_ppt_engine_assembly.py tests/test_ppt_engine_plugin.py tests/test_agents_ppt_launcher.py tests/_ppt_engine_skill_claims.py -o addopts='' -p no:xdist -q - 229 passed

  • PYTHONUTF8=1 uv run pytest tests/test_ppt_engine_prompt_claims.py tests/test_ppt_engine_assets_script_helpers.py -q - 142 passed, which is what checks the skill page against the code that has to back it

  • Both new tests were watched failing first: the schema one on assert 'character' in served["description"] (the served text did not follow the provider), the behaviour one on the missing note (the host had already been sending only the first reference)

  • The default schema was checked byte-for-byte against the description origin/main renders, extracted from the old source with ast -- identical, with Up to six present and no MiniMax text

  • uv run ruff check and uv run ruff format --check on the two changed files - clean (the test file needed one ruff format pass)

  • uv run ty check on the two changed files - clean

  • make lint-imports - 10 contracts kept, 0 broken

  • make coverage-diff - 100.00% (0/0 executable changed lines): coverage measures raven/ only, so a plugins-dist/ change has no executable lines in the report; the gate passes the same way it did for a test-only PR

  • uv run python scripts/check_source_language.py origin/main..HEAD, scripts/check_large_files.py origin/main..HEAD, scripts/check_commit_messages.py origin/main..HEAD - pass

  • The doc change is the skill page the deck author reads; the schema itself is the documentation the model reads, and it is what changed

Risk

  • Security impact considered
  • Backward compatibility considered
  • Rollback path is clear for risky changes

Off MiniMax nothing changes: the served schema is byte-identical to today's, verified as above, and no request on any other provider goes out differently. On MiniMax the schema now promises one reference instead of six -- which is what the transport already did -- and a picture that arrives with more than one gets a note instead of silence. The provider read is the same private _runs_on_minimax the tool already reaches for in _model, so no host API was touched. Rollback is reverting the squash commit.

Related Issues

Fixes #1014

@LivXue

LivXue commented Nov 11, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — PR #1261, reviewed at 66b90dc2c305c87fbcaa987f807d56d698ddedab

Scope: the full base (147297b6) → head diff (2 files, +120/−2): PptGenerateImageTool now authors to_schema() so the served description and references text follow the provider (one character kept on MiniMax, the old text off it), and a multi-reference MiniMax request is truncated to the first with a note saying how many were left out. Comments: 0; reviews: 0; inline threads: 0.

Windows-specific findings: none. Checks:

  • The change is provider/schema logic and wire shaping: no path handling, subprocess, encoding or packaging surface; the file names in the text are model-facing prose, not I/O.
  • I checked the to_schema() override for cross-turn leakage because Tool.to_schema hands parameters over by reference: PptGenerateImageTool.parameters is a property that rebuilds its dict on every access, so the MiniMax reference-text mutation lives only in the returned dict and cannot stick to a later non-MiniMax turn.
  • The truncation happens before the deck/source bookkeeping the comment mentions, the note appends to any existing note, and _runs_on_minimax() is the same predicate image_generate already uses for its own description.

Verification (native Windows 11 10.0.26300, python 3.12.13, detached worktree D:\tmp\rv-1261, the runner putting the worktree's own plugins-dist on sys.path):

  • pytest tests/test_ppt_engine_generate_image_tool.py tests/test_ppt_engine_image_search.py tests/test_ppt_engine_assembly.py tests/test_ppt_engine_plugin.py tests/test_agents_ppt_launcher.py tests/_ppt_engine_skill_claims.py → 189 passed, 2 failed, 1 skipped.
  • The 2 failures are test_the_hook_stages_rewrites_and_announces and test_a_source_staged_on_the_first_turn_does_not_hold_the_next_one_to_a_build in test_ppt_engine_plugin.py — a file this PR does not touch — and the same two IDs fail at base e22c2dbf (reported in the feat(*): make ppt_image_search follow the host search vendor #1258 review), so they are the inherited Windows deck-staging baseline. The new image-tool tests pass here.

Coverage gaps: no live MiniMax call (the tests use an httpx mock); the UI is not involved.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from 66b90dc to 2b02e3b Compare November 12, 2026 00:31
@LivXue

LivXue commented Nov 12, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — PR #1261, reviewed at 2b02e3b80a2e9b1d2a896706564a3093ed7be40b

Scope: the branch was rebased onto the new main tip c8823d324 (the #1242 merge); git range-diff 66b90dc2^..66b90dc2 2b02e3b8^..2b02e3b8 reports the single feature commit patch-identical (=). The reviewed content — the provider-aware to_schema() on ppt_generate_image and the first-reference truncation with its note — is unchanged.

Windows-specific findings: none. The delta adds no code; the earlier checks stand (SVG-free JSON schema text, provider predicate shared with image_generate, no OS surface).

Verification (native Windows 11 10.0.26300, python 3.12.13, detached worktree D:\tmp\rv-1261b at the head, worktree's own plugins-dist on sys.path):

  • pytest tests/test_ppt_engine_generate_image_tool.py → 16 passed.

Coverage gaps: no live MiniMax call; the UI is not involved.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from 2b02e3b to a86d067 Compare November 12, 2026 07:54
@LivXue

LivXue commented Nov 12, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — PR #1261, reviewed at a86d067682251993c4da1dc15580622cc27147f1

Scope: the branch was rebased onto the new main tip 5e54b229; git range-diff 2b02e3b8^..2b02e3b8 5e54b229..a86d0676 reports the single feature commit patch-identical (=), and the diff against the new base is still 2 files, +120/−2 — the provider-aware to_schema() on ppt_generate_image and the first-reference truncation with its note, unchanged from the reviewed content.

Windows-specific findings: none. The delta adds no code; the earlier checks stand (SVG-free JSON schema text, provider predicate shared with image_generate, no OS surface).

Verification (native Windows 11 10.0.26300, python 3.12.13; detached worktree D:\tmp\rv-1261c at this head, the worktree's own plugins-dist on sys.path):

  • pytest tests/test_ppt_engine_generate_image_tool.py -q → 16 passed.

Coverage gaps: no live MiniMax call; the UI is not involved.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from a86d067 to 616429f Compare November 13, 2026 06:53
@LivXue

LivXue commented Nov 13, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — PR #1261, reviewed at 616429fe3c58bde87e4269458d44db3ce555dfcd

Scope: the branch was rebased onto the new main tip 22cbc9f0; git range-diff ddcc9c168..a86d0676 22cbc9f0..616429fe reports the single feature commit patch-identical (=), and the diff against the new base is still the same 2 files — the provider-aware to_schema() on ppt_generate_image and the first-reference truncation with its note, unchanged from the reviewed content.

Windows-specific findings: none. The earlier checks stand (SVG-free JSON schema text, provider predicate shared with image_generate, no OS surface).

Verification (native Windows 11 10.0.26300, python 3.12.13; detached worktree D:\tmp\rv-1261d at this head, the worktree's own plugins-dist on sys.path):

  • pytest tests/test_ppt_engine_generate_image_tool.py -q → 16 passed.

Coverage gaps: no live MiniMax call; the UI is not involved.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from 616429f to 3c2ca88 Compare November 13, 2026 10:50
@LivXue

LivXue commented Nov 13, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — PR #1261, reviewed at 3c2ca889cc926762fa5c7aea697c1c8ba28298f1

Scope: the branch was rebased onto the new main tip 51b40ee5; git range-diff ddcc9c168..616429fe 51b40ee5..3c2ca889 reports the single feature commit patch-identical (=), and the diff against the new base is still the same 2 files — the provider-aware to_schema() on ppt_generate_image and the first-reference truncation with its note.

Windows-specific findings: none. The earlier checks stand (SVG-free JSON schema text, provider predicate shared with image_generate, no OS surface).

Verification (native Windows 11 10.0.26300, python 3.12.13; detached worktree D:\tmp\rv-1261e at this head, the worktree's own plugins-dist on sys.path):

  • pytest tests/test_ppt_engine_generate_image_tool.py -q → 16 passed.

Coverage gaps: no live MiniMax call; the UI is not involved.

@joei4cm

joei4cm commented Nov 13, 2026

Copy link
Copy Markdown
Contributor Author

Friendly ping for a review when someone has a moment.

CI is fully green at the current head, including the coverage gates, and the branch sits on the latest main. Happy to address anything in the meantime.

@LivXue LivXue left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The execute-side half is right: on MiniMax, _one truncates references to the first and appends a note naming how many were left out, which is what the transport already sent, and the second test pins the wire body and the note. The schema half does not reach the model through the real registration path. The plugin factory returns a SessionTool, and SessionTool defines name, description and parameters but not to_schema. admit_tool therefore sees the inherited Tool.to_schema, records schema_dynamic as False, and get_definitions serves the frozen snapshot built from the prototype's static _DESCRIPTION and its Up to six reference text.


I confirmed this end to end: make_ppt_generate_image with a MiniMax ImageToolConfig returns a SessionTool, and registering that and calling get_definitions still serves Up to six with no character wording, schema_dynamic False. The new schema test registers the raw PptGenerateImageTool, so it exercises a seam production does not use and cannot catch this.


Please make the wrapper delegate, for example def to_schema(self): return self._prototype.to_schema(), so the registry marks the tool dynamic and serves the provider-aware schema, and add a test through make_ppt_generate_image rather than the bare tool. Targeted runs here: tests/test_ppt_engine_generate_image_tool.py, tests/test_ppt_engine_plugin.py and tests/_ppt_engine_skill_claims.py, 78 passed.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch 3 times, most recently from 840a16a to a2909a6 Compare November 16, 2026 10:06
@joei4cm

joei4cm commented Nov 16, 2026

Copy link
Copy Markdown
Contributor Author

Rebased onto the latest main (a2909a6 on top of c3be94a); CI is green.

The generator's reference parameter now says what a reference is for on the provider that renders it, so the help text matches what that provider actually does with the image.

@joei4cm

joei4cm commented Nov 16, 2026

Copy link
Copy Markdown
Contributor Author

Addressed the review finding.

SessionTool now authors to_schema and delegates to the prototype it wraps. The registry reads an authored to_schema as a declaration of a dynamic shape, so the schema is served live instead of frozen at admit time. That was the leak: a section that names MiniMax between turns was still served the OpenRouter wording, because the wrapper answered with the shape it was admitted with.

The test now goes through make_ppt_generate_image and a ToolRegistry, so it covers the seam production registers on rather than the bare tool: the schema is read back from the registry after the image section is switched, and the MiniMax wording is expected there. The image section in that test is an ImageToolConfig, the class the image section really validates to (media_section_class("image")), which is where the provider field lives.

Pushed to the branch. The whole ppt-engine suite passes locally (3774 passed, 13 skipped).

@LivXue LivXue left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No blockers in the reviewed scope.

Scope: both commits of c3be94a..b0e489a5 (3 files, +148/-2), read at b0e489a5.

The schema half was the open question, and it now reaches the model. I drove the production seam rather than the bare tool: make_ppt_generate_image(ctx) returns a SessionTool, the registry now reads that as schema_dynamic, and get_definitions() serves the character and one-picture wording once the image section names MiniMax, then flips back when it does not. Registering the prototype instead of the wrapper agrees with it. Reverting the new SessionTool.to_schema makes the new test fail on assert "character" in served["description"], so it covers the seam it claims to.

The execution side holds through the same wrapper: one subject_reference on the wire and the note on the result. refs[:1] sits at generate_image.py:344.

Two nonblocking items. The MiniMax wording rewrites parameters.properties.references only; prompts[].references carries no description at all, so a batched call still gets no guidance. And ppt-script-authoring/SKILL.md still says "up to six pictures" (line 1102), so a deployment that mounts that skill has the resident prompt contradicting the tool; whether the product mounts it is more than this diff can settle.

Verification: the targeted file 16 passed; the six-file set 206 passed (204 at base, so +2, not the 192 in the description); ruff check and format clean; all five repo checks pass. The head is 8 commits behind main and touches none of those files.

@LivXue

LivXue commented Nov 16, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — b0e489a530

Reviewed at head b0e489a5304bc034adf1a746c2a35051ee04b1d4 (base c3be94ab6e). Scope: the full base-to-head diff (plugins-dist/ppt-engine/raven_ppt/tools/generate_image.py, plugins-dist/ppt-engine/raven_ppt/plugin/session_tools.py, tests/test_ppt_engine_generate_image_tool.py), _references and _runs_on_minimax as they are exercised by the new branches, and the Tool.to_schema precedent in raven/agent/tools/media_gen.py.

No Windows compatibility issue found.

The PR gives the tool a live to_schema so the registry serves a MiniMax-specific description, and trims references to the first picture on that provider with a note saying so. Everything it adds is string- and list-level. The paths involved (_references, unchanged by this PR) already resolve absolute paths and deck-relative names the same way on both platforms, and the new code only slices the resolved list — no new path construction, no shell, no encoding decision, no platform branch. The one place a separator could matter is Path(text).is_absolute(), which is pre-existing and untouched.

Verification (native Windows host, Python 3.12.13, shared venv, detached worktree pinned to the head):

  • python -S runner.py tests/test_ppt_engine_generate_image_tool.py -q → 16 passed in 23.10s, including the two new tests (live schema switching between a non-MiniMax and a MiniMax section, and the second-reference note).

Coverage gaps: the schema is exercised through ToolRegistry.get_definitions() with a stub config, not against a live provider; no real image generation ran.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from b0e489a to 7399659 Compare November 16, 2026 23:22
@LivXue

LivXue commented Nov 16, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — 73996591e6

Reviewed at head 73996591e635b38b66cb30e542be236172e2734c (base dc35a0507b; this head is newer than the b0e489a530 reviewed earlier). Scope: the full base-to-head diff (plugins-dist/ppt-engine/raven_ppt/tools/generate_image.py, plugins-dist/ppt-engine/raven_ppt/plugin/session_tools.py, the SKILL.md paragraph, tests/test_ppt_engine_generate_image_tool.py), the Tool.to_schema contract in raven/contracts/tool.py, the schema_dynamic detection in raven/agent/tools/registry.py, and _references / _runs_on_minimax as the new branches drive them.

No Windows compatibility issue found.

What this head adds over the earlier one: SessionTool.to_schema (so the wrapper is admitted as schema_dynamic and served live instead of a frozen snapshot) and the MiniMax reference description extended to the prompts[].references sub-schema. Both are dict manipulation on a freshly built schema — Tool.to_schema returns self.parameters, and this tool's parameters property constructs a new dict per call, so the mutation cannot leak between renders; the registry's frozen copy comes from copy.deepcopy(parameters) at admit time, unaffected.

The one platform-adjacent surface is _references, which is unchanged by this PR; I checked the path shapes a deck model may write against it on this host anyway:

'C:\Users\livxu\Pictures\logo.png'       is_absolute=True  name='logo.png'
'C:/Users/livxu/Pictures/logo.png'       is_absolute=True  name='logo.png'
'\\server\share\logo.png'                is_absolute=True  name='logo.png'
'C:\Users\livxu\My Pictures\中文 logo.png' is_absolute=True  name='logo.png'

Path(text).is_absolute() accepts drive-letter, forward-slash, UNC and non-ASCII-with-spaces forms on Windows, and Path(text).name splits correctly for all of them, so the resolver is not separator- or locale-sensitive. The new held_back slice only trims the already-resolved list.

Verification (native Windows host, Python 3.12.13, shared venv, detached worktree pinned to the head):

  • python -S runner.py tests/test_ppt_engine_generate_image_tool.py -q → 16 passed in 21.49s, including the live-schema switching test and the second-reference note test.
  • Path-shape probe above, importing raven_ppt.tools.generate_image from the worktree's plugins-dist.

Coverage gaps: the tool was not exercised against a live MiniMax endpoint, so the "host sends only the first reference" premise the note describes is code-read from the host's route, not measured; the schema is verified through ToolRegistry.get_definitions(), not through a served turn.

@joei4cm

joei4cm commented Nov 17, 2026

Copy link
Copy Markdown
Contributor Author

Both non-blocking notes from the earlier review are addressed in this head:

  • A picture asked for inside prompts is a request of its own, so to_schema now writes the same one-reference description into the batched references too, not only the top-level one. Off MiniMax the batched ask keeps the shape it had -- no description of its own -- so the served schema stays byte-identical there, and the test pins both halves.
  • The ppt-script-authoring skill page said "up to six pictures" with no provider caveat; it now says a MiniMax deployment keeps the first one only, so the character to stay consistent with goes in that one.

The PR body also claimed 192 tests for the combined run where it measures 229; that is corrected, along with the two lines this head adds.

On the coverage gap you name: the "host sends only the first reference" premise is read from the host's MiniMax route, and I have not exercised it against a live endpoint either -- the test drives the schema through ToolRegistry.get_definitions(), which is the seam production registers through.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from 7399659 to 6786dac Compare November 17, 2026 11:06
@joei4cm

joei4cm commented Nov 17, 2026

Copy link
Copy Markdown
Contributor Author

The branch is on the latest main (head 6786dac42, behind 0) and CI is green: 22 checks pass, the only non-run is the docs deploy skip. The Windows compatibility review and the earlier "No blockers" review both read clean, and the two non-blocking notes from the first pass are addressed in this head.

The review on this PR still reads as a comment rather than an approval, so the branch shows review as required. If it was meant as an approval, submitting it with the Approve button is all this needs.

@LivXue

LivXue commented Nov 17, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — 6786dac424

Reviewed at head 6786dac42485a009f9f965468bd64119f55d6a68 (base c57fc5e51c4; this head is the 73996591e6 content rebased onto a newer base — its own diff is still the same four files, +158/-3, and git diff 73996591e6 6786dac424 -- plugins-dist/ tests/test_ppt_engine_generate_image_tool.py is empty). Scope: the PR's own diff — plugins-dist/ppt-engine/raven_ppt/tools/generate_image.py, …/plugin/session_tools.py, the SKILL.md paragraph, and tests/test_ppt_engine_generate_image_tool.py — plus the Tool.to_schema / registry schema_dynamic contract it depends on.

No Windows compatibility issue found. Unchanged from the previous review: the change is schema text and a list slice, _references (untouched) resolves drive-letter, forward-slash, UNC and non-ASCII-with-space path forms on this host, and the schema is rebuilt per render so the mutation cannot leak.

Verification (native Windows host, Python 3.12.13, shared venv, detached worktree pinned to the head):

  • python -S runner.py tests/test_ppt_engine_generate_image_tool.py -q → 16 passed in 25.89s.

Coverage gaps: no live MiniMax endpoint, so the "host sends only the first reference" premise is code-read; the schema is verified through the registry, not a served turn.

@LivXue

LivXue commented Nov 17, 2026

Copy link
Copy Markdown
Member

One question on the hardcoded one-reference MiniMax wording: is there an API-level way to learn that limit, rather than baking 1 into the deck tool and the transport?

What I can see today does not publish one. The docs type subject_reference as object[] with no maxItems, and the nearest sentence -- "For best results, upload a single front-facing portrait photo" -- reads as a quality recommendation rather than a cap. GET /v1/models exists, but without a key I cannot tell whether it carries capability metadata, and the docs' embedded request schema exposes no count. The host route already slices [:1] in media_gen.py, so the deck text matches today's transport; the risk is that the assumption now lives in two places and neither can be checked against the API.

Did you find a signal for the count -- a validation error when more than one subject_reference is sent, per-model metadata, or a documented hard limit? If not, would it be better to read the limit from the single host-side place (or let the API answer and report its response) so the schema does not freeze a number the API never states? A one-line note on what happens if two are actually sent would settle it.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from 6786dac to 7378de2 Compare November 18, 2026 09:15
@LivXue

LivXue commented Nov 18, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — 7378de279d

Reviewed at head 7378de279d9a2b98c38ec6dbf23644237e613432 (base 20aaefd1161f). Scope: the whole diff (4 files): raven_ppt/plugin/session_tools.py (the wrapper now authors to_schema), raven_ppt/tools/generate_image.py (provider-following description and one-reference handling), the ppt-script-authoring skill page, and the tests.

No Windows compatibility issue found.

  • No filesystem, subprocess, path, encoding, or platform branch is touched: the change is schema text selection (self.media._runs_on_minimax()) plus a refs[:1] truncation and a result note. references resolution is unchanged, so drive-letter/space/CJK path handling is the same as before the PR.
  • Mutable-schema check passed: PptGenerateImageTool.parameters is a property that builds a fresh dict per access, and Tool.to_schema() re-reads it, so the MiniMax wording cannot leak back into the default schema. Measured with a stub provider: default → MiniMax → default leaves the default declaration byte-identical (d0 == d2, "MiniMax" not in json.dumps(d2)) and the _DESCRIPTION.replace(...) anchor does match (the MiniMax description really changed).

Verification, native Windows host (GitHub tarball of 7378de279d, extracted under D:\tmp\rv1261):

  • python -S runner.py tests/test_ppt_engine_generate_image_tool.py -q → 16 passed
  • python -S runner.py tests/test_ppt_engine_prompt_claims.py tests/test_ppt_engine_assets_script_helpers.py -q → 140 passed, 2 skipped (both skips are the SIGALRM-based timeout helper, POSIX-only, unrelated to this diff)
  • The toggle-back probe above.

Gaps: no live MiniMax image request was made from this host (the new behaviour test mocks the transport); the tree is the commit's export rather than a git worktree.

@LivXue

LivXue commented Nov 18, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — e788a416d0

Reviewed at head e788a416d00e573a6efe2cbed69d558abe458d4c (base 20aaefd1161f). This head adds one commit over the reviewed 7378de279d — chore(ppt-engine): retrigger ci after the idle-ceiling flake — and changes no file: the tree of both heads is 4fbda90cb8ece5358801ff968083afb48dd48c24.

No Windows compatibility issue found — the code is byte-identical to the revision reviewed in the previous round (comment 6975245470), so that review stands: schema text only, no path/subprocess/encoding surface, the MiniMax wording cannot leak back into the default schema (parameters is a property building a fresh dict per access), and 16 passed plus 140 passed / 2 skipped (POSIX-only skips) held at that tree on this native Windows host.

No tests were re-run for this head because nothing under review changed; the equality above is the check.

@joei4cm

joei4cm commented Nov 18, 2026

Copy link
Copy Markdown
Contributor Author

Heads up on the red unit (py3.12 / ubuntu-latest / 4/4) job: this is an infrastructure flake, not a defect in the diff.

  • Run: https://github.com/EverMind-AI/Raven/actions/runs/44281660674
  • The job itself is green on tests: 10299 passed, 117 skipped.
  • The only failure is the --idle-ceiling-strict session ceiling in make coverage-shard: tests/test_cli_serve_commands.py::TestStopping::test_a_process_that_outlives_the_deadline_fails_the_stop waited 10.00s against the 3s ceiling on this runner (that test sleeps by design), so the shard exits non-zero.

The same flake hit the previous head as well (the head only adds an empty retrigger commit on top of 7378de279, changing no file), and the test is unrelated to this change (it drives the CLI serve stop path). Could the job be rerun, or the ceiling waived for this run? A rerun on a less loaded runner should go green: every other job passes (20 pass / 2 skip).

@LivXue

LivXue commented Nov 18, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — 31098f179e

Reviewed at head 31098f179edd25a9e2105d043e988989823333f4 (base 20aaefd1161f). Scope: the delta over the previously reviewed e788a416d0 — SKILL.md (+3/-2) and the _MINIMAX_REFERENCES string in generate_image.py (+2/-1), both wording: "a second named is not sent".

No Windows compatibility issue found. No logic changed since the review at 7378de279d (comment 6975245470): the schema still follows self.media._runs_on_minimax(), the truncation and note are unchanged, and the new sentence is plain ASCII text inside the same JSON-schema description.

Verification, native Windows host (GitHub tarball of 31098f179e, extracted under D:\tmp\rv1261b):

  • python -S runner.py tests/test_ppt_engine_generate_image_tool.py tests/test_ppt_engine_prompt_claims.py tests/test_ppt_engine_assets_script_helpers.py -q → 156 passed, 2 skipped (both skips are the SIGALRM-based timeout helper, POSIX-only)

Gaps: no live MiniMax request; the tree is the commit's export rather than a git worktree.

@joei4cm

joei4cm commented Nov 18, 2026

Copy link
Copy Markdown
Contributor Author

Hi @LivXue, your Windows compatibility review at head 7378de279 found no Windows compatibility issue, and the current head 31098f179 adds only the wording note you asked for (a second MiniMax reference is not sent) on top of the tree you already reviewed byte-for-byte. CI is green (22 pass / 1 skip). If you agree, could you submit the same conclusion with the Approve button? This PR is only waiting on that.

@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from 31098f1 to 653590a Compare November 18, 2026 15:38
@LivXue

LivXue commented Nov 18, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — 653590a730

Reviewed at head 653590a73045cc6871262572b75068feb12e04e6 (rebased onto newer main). The PR's own diff is the same four files and the same content reviewed at 31098f17 — session_tools.py still 2c0260802e..632bd11c31, the tests still 83cb0f0acc..e914661f00, and SKILL.md / generate_image.py carry the "a second named is not sent" wording reviewed in that round.

No Windows compatibility issue found. Still schema text plus a refs[:1] truncation and a result note; no filesystem, subprocess, path, encoding, or platform branch, and the MiniMax text cannot leak back into the default schema (parameters is a property building a fresh dict per access).

Re-verified at this head on the native Windows host (GitHub tarball of 653590a730, extracted under D:\tmp\rv1261c):

  • python -S runner.py tests/test_ppt_engine_generate_image_tool.py tests/test_ppt_engine_prompt_claims.py tests/test_ppt_engine_assets_script_helpers.py -q → 156 passed, 2 skipped (both skips are the SIGALRM-based helper, POSIX-only)

Gaps: no live MiniMax request; the tree is the commit's export rather than a git worktree.

joei4cm and others added 5 commits November 19, 2026 02:13
ppt_generate_image told every deployment that `references` carries up to
six pictures to match or vary, but the transport underneath it does not:
on MiniMax the host sends the first alone, as a character to keep, and
drops the rest without a word. The deck's model, reading a schema that
promised six, asked for several and got one -- and was never told the
others were not sent.

The tool now authors its own to_schema, the same way image_generate
does, so the registry serves the answer live and a section changed
between turns is read on the next one: on MiniMax the description and
the references parameter say one picture, kept as the character it
shows, with the manner in the prompt where a reference cannot carry it;
off MiniMax both stay byte-identical to what they were.

A second reference on MiniMax is now reported rather than swallowed:
the first is the one sent, the file is named for what was sent, and the
result carries a note saying how many were left out -- the host's own
truncation, with the word the issue says was missing.

Co-authored-by: Claude (claude-fable-5-1) <noreply@anthropic.com>
…ssion wrapper

The deck's image tool is registered as a SessionTool, which inherited
Tool.to_schema and so read as a static shape to the registry: the schema
was frozen at admit time, and a section that later named MiniMax was
still served the OpenRouter wording. The wrapper now delegates to its
prototype, which authors the provider-aware schema, and the registry
serves that answer live.

The test goes through make_ppt_generate_image and a registry, so it
covers the seam production registers on rather than the bare tool.

Co-authored-by: Claude (claude-fable-5-1) <noreply@anthropic.com>
A picture asked for inside `prompts` is a request of its own, so the
MiniMax wording that rewrote the single-picture `references` left the
batched one with no description at all: the model planning several
pictures in one call was offered no guidance where the single ask got it.
The skill page said "up to six" with no provider caveat, which is the
promise the MiniMax route cannot keep.

Co-authored-by: Claude (claude-fable-5-1) <noreply@anthropic.com>
@joei4cm
joei4cm force-pushed the fix/ppt_minimax_reference_schema branch from 653590a to 75223ed Compare November 18, 2026 18:14
@LivXue

LivXue commented Nov 18, 2026

Copy link
Copy Markdown
Member

Windows compatibility review — 75223ed6a6

Reviewed at head 75223ed6a697570fc501a50b0177f002ba62d02d (rebased onto newer main). The PR's own diff is unchanged from the revision reviewed at 653590a730: the same four files, 239-line patch, blob ids 2c0260802e..632bd11c31, 7cb56d0f28..a54d18f4ae, 7724e29a13..7a5405b736, 83cb0f0acc..e914661f00.

No Windows compatibility issue found. Still schema text plus a refs[:1] truncation and a result note; no filesystem, subprocess, path, encoding, or platform branch.

Re-verified at this head on the native Windows host (GitHub tarball of 75223ed6a6, extracted under D:\tmp\rv1261e):

  • python -S runner.py tests/test_ppt_engine_generate_image_tool.py tests/test_ppt_engine_prompt_claims.py tests/test_ppt_engine_assets_script_helpers.py -q → 156 passed, 2 skipped (both skips are the SIGALRM-based helper, POSIX-only)

Gaps: no live MiniMax request; the tree is the commit's export rather than a git worktree.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

fix(ppt-engine): on MiniMax a deck's references to match or vary become one character reference

2 participants