# Pi PR-rule delivery: real runtime proof

Two unmodified T3 checkouts, each running `T3CODE_DEV_SEED=0 vp run dev` on Linux, fresh checkout-local databases, Node 26.13.1, Pi 1.5.2, anthropic/claude-haiku-5-5. Web clients paired in the shared preview browser (1280x800). No tests or fake model responses are used as evidence.

Base: d6c1d17793948057d3bd40ac2534bae1527c6824 (merge-base with origin/main).
Head: 6141bdbd46c6a9efee2a5b52c3415132bc94a2f4.

## Steps on each checkout

1. Install dependencies with `vp i`; start a fresh dev server with `T3CODE_DEV_SEED=0 vp run dev`; pair the web client.
2. Settings > Providers > Add provider > Pi > Next > Next > Add instance. Pi reports v1.5.2, Authenticated.
3. Create one empty synthetic Git repository in the evidence directory's `project` folder, make an empty initial commit and set its origin to `https://github.com/example/pi-linking-evidence.git`. Nothing is fetched or pushed to that example remote. Add it through `node apps/server/src/bin.ts project add <project> --title 'Pi prompt evidence' --base-dir .t3`.
4. Set Pi launch arguments to `--extension <evidence-directory>/capture-final.ts`. The T3 adapter appends its own extension after this one. This observer does not add PR guidance or alter model requests.
5. Run `python3 drive.py base <base-checkout>` or `python3 drive.py head <head-checkout>`. It invokes T3's actual `thread start --provider pi --model anthropic/claude-haiku-5-5 --wait --json --base-dir .t3` CLI through the running server, once for each prompt path.
6. The sections run observes without overriding. For the fallback run, the observer's earlier before_agent_start handler returns a marked `systemPrompt`. Pi assigns the returned value to `forceSystemPrompt` before the observer's next handler and T3's handler. This is intentionally NOT `--system-prompt`, which only changes the preamble in Pi 1.5.2.
7. Capture `before_t3_handler`, then the actual `before_provider_request` payload's model, system, user messages and tool names, the real provider HTTP status and `agent_end`. Export readable transcripts and JSON with `python3 export.py`.

## Observed

Both base paths: orchestration is delivered but no pull_request_linking block, link_pull_request, watch_pull_request or unwatch_pull_request instructions.
Both head paths: the shared PR block is delivered, including created/taken-over/pushed PRs, every stack layer, watch-before-handoff conditions and unwatch-before-handoff.

The fallback captures have forceSystemPromptBeforeT3=true, start with EVIDENCE_FORCE_OVERRIDE, and have NO t3_code_orchestration wrapper anywhere in the actual request. T3 appends raw guidance. The sections captures have forceSystemPromptBeforeT3=false and have the wrapper produced by Pi's section renderer. These are distinct runtime branches.

All four final turns received HTTP 200, completed and replied `ok`. The transcripts include actual CLI completion results as corroboration. The JSON includes the pre-T3 prompt so the raw append can be independently compared with the final request.

## Scope and privacy

This proves instruction delivery, not live model compliance with linking/watching an actual PR. No forge mutation was attempted. The before-provider hook observes the actual payload immediately before its HTTP request and does not replace it. It captures selected request fields, not headers or tool argument schemas. Paths are replaced with <HOME> and <EVIDENCE>; exported assistant thinking blocks are omitted. No production database was read or seeded. No source changes were made.
