feat(positioning): runtime prompts carry the Trinity one-liner (abilityai/trinity-enterprise#801)

[vybe]

Fixes abilityai/trinity-enterprise#801 — the runtime-prompt half of the ADR-0011 positioning epic (abilityai/trinity-enterprise#796). Independent branch off dev — the four #796 children touch disjoint files and merge in any order (operator, 2026-10-05). Draft until the chain is in.

What

• src/backend/services/platform_prompt_service.py — the PLATFORM_INSTRUCTIONS header line every agent receives on every turn now reads "This agent runs on Trinity, the operating system for AI-native companies." (one neutral sentence, not the two-line pitch; about the same token count as before). Rebuilt on every invocation, so every agent on an instance carries it from its next turn after deploy — no restart, no fleet pass.
• config/trinity-meta-prompt/prompt.md — the sentinel copy's line 3 reworded to the same sentence (ruling: reword, not delete).
• config/agent-templates/trinity-system/CLAUDE.md — the system-agent seed now says "for this Trinity instance." and carries no positioning claim.
• tests/unit/test_ent801_positioning_prompts.py (new) — builds the live header through get_platform_system_prompt, reads the meta-prompt and the seed from disk, pins the three retired phrases absent (case-insensitive, matching the issue's grep -i) and the new sentence present in the two live surfaces.

Two commits: f8632f33 (the three edits + the test) and b6f0d83a (review fix I1, the case-insensitive guard). The issue's acceptance grep returns nothing on the branch.

Rulings carried (orchestrator, on the operator's behalf — plan file)

• Choice 1 (meta-prompt line 3, "delete or reword"): reword to the row-1 sentence. test_1402_prompt_contract.py pins six operator-communication phrases and one absence, nothing about line 3, so it stays untouched and the file remains a faithful copy of the live header.
• Choice 2 (a guard test): yes — the issue's own acceptance grep made permanent; scoped to the three absences and the new sentence, nothing about sovereign beyond absence, kept out of test_1402.
• The nine same-phrase hits under docs/ are not this issue's rows: six belong to abilityai/trinity-enterprise#800, three are dated records (docs/research, docs/security-reports, docs/archive).

Review + security

/review (claude-fable-5-1, report-only): MERGEABLE — 0 critical, 1 informational (I1: the absence guard compared case-sensitively while the acceptance grep is -i), fixed in b6f0d83a. Exactly four files in the range, every edit the plan's After text; the system-agent line ends "for this Trinity instance." with no claim; test_1402_prompt_contract.py untouched; the new test not being listed in tests/registry.json rated not a finding (CI collects by glob; 635 of 968 unit files are unlisted); .claude, src/backend/enterprise and the root CLAUDE.md untouched; the line-53 hunk sits outside the ask section that the chain-3242 drafts edit. /cso --diff: 0 findings.

Tests

• Targeted, foreground: 191 passed over the 7-file prompt set (test_ent801_positioning_prompts, test_1402_prompt_contract, test_ent243_prompt_tier, test_platform_prompt_unit, test_platform_prompt_runtime, test_1816_system_agent_adoption, test_1816_system_agent_convergence) at f8632f33; 23 passed over test_ent801 + test_1402 after the fix.
• Not run here: the full unit island (CI).

Before merge

• Sibling PRs sharing platform_prompt_service.py and prompt.md — feat(operator-queue): an agent sees its pending asks and can replace one (#3247) #3256, feat(operator-queue): atomic asks — authoring rules in the ask_operator contract, option and title limits (#3243) #3253, feat(operator-queue): approvals get a "Something else" answer (#3242) #3250 (chain-3242, stacked) and feat(workspace): Discuss and Dismiss on asks (abilityai/trinity-enterprise#747, #748) #3181 (ent#747/748): git merge-tree of each head against this branch is clean, and this branch adds no conflict beyond what each already has against dev (none, 2026-10-05 23:20 UTC). Merge in any order.
• Merge-tree against live dev @ 493d2a6 clean.
• Row 4 (operator / cloud-function owner): read the ask_trinity cloud function's own prompt (ask_trinity_client.ts:35, outside this repo) and confirm it carries no tagline; the corpus half is fixed by docs(positioning): user-docs 'What is Trinity' leads with the operating-system line (abilityai/trinity-enterprise#798) #3257 (ent#798).
• Row 5 (operator, instance admin): remove the stray line "Rabbit is a yellow monkey." from the trinity_prompt setting on trinity.ability.ai, unless it is a deliberate canary.
• Live-instance DoD: after deploy, an agent's turn prompt contains "operating system for AI-native companies" and not "Deep Agent Orchestration". Existing instances keep the old system-agent seed line until the system agent is re-created (the seed is copied once at creation).

Handoffs

• abilityai/trinity-enterprise#800 — the docs/memory / CLAUDE.md hits (branch feature/800-positioning-engineer-context).
• abilityai/trinity-enterprise#798 / PR docs(positioning): user-docs 'What is Trinity' leads with the operating-system line (abilityai/trinity-enterprise#798) #3257 — the docs-source half that feeds ask_trinity.

🤖 Generated with Claude Code

[trinity-ability]

merge-train: batch validated on train/20261006-1038 (#3270)