fix(shared): the session-ledger recipe writes /tmp quote files that trip the approvals reviewer; derive run-ids in a helper

[thewrz]

This was written agentically; verify its assertions:

North star

Six times on 2026-09-16 a Codex approvals-reviewer session was spun up to judge a session-ledger.sh append, costing 290k tokens and 5-9 s each, all verdict allow. Every one was triggered by the recipe at parallel-issues/SKILL.md:117-119: quote_file=$(mktemp) lands in /tmp, outside the workspace roots, so the harness asks. The helper already accepts --quote TEXT inline (.shared/scripts/session-ledger.sh:28-31). Separately the root hand-hashed three different strings to derive run-ids and produced two ids for the same PR (review-pr-7dcf51e0… and review-pr-f7aae90d… for #511), so the ledger holds three receipts for one grant.

North star rules this change must obey

• Runbook first, hooks second. The skill preamble (env contract, helper output, composed prompt) hands the agent a complete runbook before the first mutating call. Hooks only nudge when the agent drifts; a hook or guard that manufactures tool calls is itself a defect.
• No new refusal that a retry can trip. The remedy for every failure path is a helper that does the right thing or an exact line to paste, never "run it again". A stricter kit makes the agent try harder and burn turns getting blocked.
• Diff shape: negative in prose (agentkit/skills/**/*.md), positive in scripts, hooks, or tests. If the prose line count goes up, the PR needs a stated reason.
• Show the number it buys. Every acceptance criterion below is measurable on the next parallel-issues run from the rollout files.

Problem

A recipe that writes outside the workspace is a recipe that costs an approval; a run-id computed by the model is a run-id computed differently every time. Both are helper jobs.

Change

Prose (cut):

• parallel-issues/SKILL.md:117-119, review-remote-pr/SKILL.md and pr-to-green/SKILL.md equivalents: replace the mktemp recipe with the one-line session-ledger.sh append … --quote "$QUOTE"; delete the run-id derivation formulas (printf 'scope=…;flags=…' | sha256sum | cut -c1-32).

Scripts (add):

• .shared/scripts/session-ledger.sh: new run-id --procedure-set NAME --scope CSV [--flags CSV] --repo SLUG --base BRANCH printing the canonical id; append accepts --quote-stdin for multi-line quotes so no temp file is ever needed; if a temp file is still required, create it under $repo_root/.agent/cache.
• review-remote-pr/scripts/run-dir.sh: derive from session-ledger.sh run-id.
• tests/test-session-ledger-run-id.sh: same inputs → same id; scope order-insensitive; a second append with identical decision/scope/quote/run-id is a no-op that prints the existing record.

Acceptance criteria

• Next run: approvals-reviewer evaluations attributable to ledger appends = 0 (was 6).
• grep -rn 'mktemp' agentkit/skills/**/*.md returns 0 lines.
• One run-id per (procedure, scope); duplicate receipts = 0.

🤖 Co-authored by Claude Fable 5.1.