fix(coding-agents): survey prompt says Glob, not Read, for the directory layout

[MasterST1337]

The cold-repo survey spawns a headless claude -p --model haiku --allowedTools Read Glob Grep session. SURVEY_PROMPT asks the model to understand "the directory layout" but never says which tool shows it, and haiku reaches for Read(<repoRoot>) — a bare directory path, which errors EISDIR before the survey has read anything.

Evidence

Measured on one machine's OTel export, 2026-08-23..25: 10 distinct survey sessions failed this way on their first tool call. The transcripts all open the same way:

"I'll survey this repository's architecture to build Hindsight memory. Let me start by sampling the structure."
Read({"file_path": "/Users/…/<repo>"}) → EISDIR

Affected repos: kids-yt-factory, hindsight, sitebrain.eu, PlanView, Meridian, plus four separate git worktrees. That is every repo the survey ran cold against — and because a fresh worktree is a fresh bank, it re-fires per worktree rather than once per project.

The change

One sentence added to SURVEY_PROMPT naming Glob for the directory layout and forbidding Read on a directory. Glob was already in --allowedTools; the model was simply never told to prefer it here.

What this is not

A prompt sentence cannot guarantee a model's tool choice. The structurally robust fix is at the tool-wiring layer — reject Read on a directory client-side, or give the survey an ls-shaped tool — which this PR does not attempt. Happy to follow up there if you would rather fix it that way.

No test pins prompt content in this package (survey.test.ts asserts against the SURVEY_PROMPT constant, not its text), so no test changed. scripts/hooks/lint.sh clean; package suite 655 passing.

[strix-security]

Strix Security Review

No security issues found.

_{Updated for 7ab1283.}

---

Reviewed by Strix
_{Re-run review · Configure security review settings}