[t3dotgg]

Agents retype the whole page every time they call html_preview or html_render, because both tools take only inline html. Pages often already exist as files on disk. Discussion #23303 measured about 30K output tokens and four minutes per copy for a 40K-character page. Agents then skip the re-preview to save time.

Now both tools also take path to an absolute .html file, and a call passes exactly one of the two. The agent writes the page once, edits the file, and passes the path.

• File read. The server reads the file only for full-access threads. A supervised thread gets a clear error and passes html, the same as before. The read uses the same guards as other disk reads: no network paths, the real path must end in .html/.htm, server-internal paths are refused, the file is opened without following links, and it must have one hard link, be at most 25 MiB, and start like an HTML document. The page is saved as a snapshot at publish time.
• Render recovery. ACP providers that drop tool results recover path renders by path, in publish order. A call that sends both sources published nothing, so it recovers nothing.
• Shorter text. Tool descriptions and agent instructions are shorter, so the agent context budget goes down.
• Eval. The mcp-tool-use eval has a new revise-page-resends-html case. The grader reads the file behind path and fails page checks when it cannot replay the file exactly.

This takes over #32817 by @apostolos-geyer, whose commit is kept with its authorship. The review fixes are in separate commits.

Proof: vp run check:changed passes, and focused tests fail without each fix. A real-model eval (GPT-6.1 Sol, 6 visual cases × 2 samples) passes 12/12, and Sol chose path in both revise samples. A live headless-Chrome preview of an .html file passed by path rendered correctly.

Council review

Rounds: 9

Delegate	R1	R2	R3	R4	R5	R6	R7	R8	R9
Claude Opus 5.5	8/10	9/10	9/10	8.5/10	9/10	9/10	9.5/10	9.5/10	9.5/10
Muse Wonder Full	8/10	9/10	9/10	9/10	8.5/10	9/10	9.5/10	9/10	9/10
Codex GPT-6.1 Sol	9/10	9/10	9/10	9/10	9.5/10	9.5/10	9.5/10	9.5/10	9.5/10

Reviewed with sol-loop: 2 passes with GPT-6.1 Sol. Sol verdict: approved

Created with Claude Opus 5.5 in Claude Code (via T3 Code).

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Approved at c183eab

Macroscope's review found this PR approvable — HTML preview and render can now consume an HTML file path instead of requiring agents to repeat its contents. Full-access gating and constrained file reads limit the new exposure while preserving the existing rendering flow.

^{You can add or adjust custom eligibility rules. Learn more.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	21.5 KiB	21.5 KiB	0 B (0.0%)	24.4 KiB	✅
Codex	Live turn messages	1	1	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.9 KiB	21.9 KiB	0 B (0.0%)	24.4 KiB	✅
Claude	Live turn messages	1	1	0 (0.0%)	8	✅

Baseline: 20fcc6c · PR result: c183eab · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.