docs(research): re-measure pi/codex planner streaming — aic not the regressor

[CaicoLeung]

Why

Reported symptom: "CLI agents like Codex and PI have broken streaming output during the Planner phase." Suspects included the v0.5.1–0.5.3 reasoning-feed refactors (#130, #134, #135).

What this does

Re-ran the streaming attribution research from docs/research-cli-agent-streaming.md with today's CLI versions and a deterministic harness, and records the results there (+18 lines; no code change).

Findings

1. aic's relay/render path is green. A pty harness drove real aic 0.5.3 (HEAD build and installed Homebrew binary) against a scratch repo with a fake pi emitting 20 tagged thinking_delta events at a fixed 100 ms cadence. Both binaries repainted the reasoning window per delta — 20/20 markers at exact 100 ms spacing, 1.97 s spread. A render regression would show all markers in one late burst.
2. codex-cli 0.147 is silent by design. Raw probe: thread.started + turn.started instantly, then zero stdout until item.completed. No reasoning events exist in exec --json; the planner window correctly shows only the loading frame.
3. pi 0.84.2 regressed upstream since the 2026-08 measurement:
   • ~6–8 s of complete stdout silence precedes the first delta on every run (server-side thinking latency).
   • Short thinking phases no longer stream — the whole thinking lands as one ~10 ms burst of deltas (also pre-filled in message_start) right before the text phase, so the window jumps in one chunky dump, exactly like Claude Code's documented batching.
   • Long thinking phases still stream live (618 deltas over 13.3 s on a thinking-heavy prompt). The old "pi is always smooth" claim was prompt-dependent.

Not done (deliberate)

Flipping pi's Encoding::streams_reasoning_live() to false would trade the chunky dump for a plain loading frame, but would also hide pi's still-working live streaming on thinking-heavy runs — left as-is.