[Exploitacious]

Summary

Ports five features already in daily use in a private harness into the template, so a copy running in T3 Code works the same way out of the box. One PR, five parts:

1. SKILLS/t3-html-render (new). In a T3 Code thread it builds one self-contained interactive HTML page for data, a report, findings, a comparison or a plan, checks it with html_preview (728 px dark, 390 px light, zero console errors), and shows it inline with html_render. Ships assets/starter.html, a worked assets/example-report.html on synthetic data (a fictional ExampleOrg), and references/palette.md, whose series colors are validated for colorblind safety on T3's light and dark surfaces.
2. Render by default in T3 (CONTEXT/how-i-work.md, Reporting to the Operator; core.md Writing; BOOTSTRAP Round E). Replaces "offer a rendered page and wait for a yes" for T3 threads. Hosts without inline pages keep the offer.
3. Slash-only harness-update and remote-session (disable-model-invocation: true). Both run only when the Operator asks, so their descriptions leave the skill listing. SKILLS/README.md, README.md and AGENTS.md now say to type /harness-update or /remote-session, and how to undo it in a copy.
4. SKILLS/.trash/ gitignored, with a verify-ops-selftest.sh case that fails without the line.
5. Doctrine pack (generic rulings only):
   • at most two or three questions in a decision report, in prose; the procedure inline when the next move is the Operator's (Reporting);
   • a deploy is done when it is verified live, third-party updates included (Quality standards);
   • lanes and resources used freely with a clear purpose, with the existing spend gate and cost rule unchanged (P11, core Intake);
   • lanes never run session-close, because a lane inherits the foreman's session id (core Rhythm);
   • the T3 schedule_task wake rule: timers under about 115 minutes, structured schedule shapes, UTC nextRunAt, a one-off task deleting itself (Hosted inside T3 Code);
   • model-roles.md: Opus 5.5's price and API behavior, Sonnet 5.5's catalog figures and overturn condition, Haiku 5.5's 100K price step and how the Stage-1 settings hold that ceiling (and what would break it), Fable's cost in figures, and the T3 Code picker settings (customModels, hiddenModels).

Measurements

Skill listing (name plus description characters, the budget Claude Code caps at about 30,000), over the template's own SKILLS/:

	model-invocable entries	characters
before	17	3,575
after	16	3,366

harness-update and remote-session leave the listing; t3-html-render joins it.

Checks run locally

• Every suite in the CI loop (CONTRIBUTING.md step 2), sandboxed as gate.yml runs them, with the current Stage-1 settings bound for verify-ops-selftest.sh: 57 suites plus the two PowerShell suites, failed suites: 0.
• verify-ops-selftest.sh: the new .trash case failed before the .gitignore line and passes after it.
• bash -n on the edited shell script.
• The example page rendered headless at 728 and 390 px with zero console messages (a control page with a deliberate error was captured, so the check can see one).
• .claude-config/bin/secrets-scan.sh over every changed file: 0 hits added.
• Private-term scan over the changed files, every commit message, and this title and body: 0 blocking hits.