[don-petry]

Summary

The triage tier has been silently broken since the cascade landed (fbec114): the prompt expected $PR_METADATA / $PR_DIFF to be interpolated into the prompt body, but claude --print does no shell-variable substitution, so Haiku always saw the raw $PR_METADATA text and reported the data as missing. A silent-escalate fallback in review-one-pr.sh synthesized a fake escalate=MEDIUM verdict, which kicked every PR up to the deep-review tier — burning Sonnet/Opus tokens on every single PR while the workflow looked healthy.

The session circuit breaker / triage hard-fail in commit 4d2aee3 (on claude/practical-cannon-f1a4d4) correctly removed that fallback and finally surfaced the underlying bug, with Tier 1 returning conversational text on combsmith/ContentTwin#100 instead of JSON.

This PR fixes three things:

• No env-var substitution: scripts/review-one-pr.sh now builds a self-contained triage prompt at /tmp/cascade/triage-prompt.md by concatenating the static template with the actual PR context (URL, SHA, metadata JSON, truncated diff, prior review body). The model gets the data inline, not via env-var interpolation that never happened.
• --permission-mode plan was wrong: under --print with no human, plan mode makes the model propose a plan and ask for approval — that's exactly the "Before I proceed with the plan, I need to understand your intent…" output we saw. Removed; explicit --disallowed-tools covers every standard tool so the model has no choice but to answer.
• prs.txt context contamination: with Read/Glob available in plan mode, Haiku wandered the cwd, found the 31-line candidate pool, and assumed the task was about all of them. Disabling tools eliminates the contamination at the source.

Belt-and-braces: a sed step strips ```json ... ``` fences before jq empty, since Haiku still tends to fence its JSON despite the instruction.

Verification

• Reproduced both symptoms locally: empty env vars produced "PR metadata not provided", and adding prs.txt to cwd produced the "31 PRs" output verbatim.
• Ran the patched scripts/review-one-pr.sh against PR fix(pr-review): correct bot username to donpetry-bot (no hyphen) #100 in dry_run=true mode. Triage now returns escalate=false risk=LOW and the cascade proceeds to the single-reviewer Opus confirmation as designed. The Opus body even self-diagnosed: "The recurring re-reviews are caused by a triage-output-invalid infrastructure failure in the triage tier" — exactly the bug this PR fixes.

Out of scope

Per the bug report, this PR does not touch the session circuit breaker / per-tier timeouts / triage hard-fail logic that just landed in 4d2aee3. That code is correct — it caught this bug. This is a fix on main and merges independently of the practical-cannon branch.

Test plan

• claude --print --model claude-haiku-4-5-20251001 --disallowed-tools "..." < /tmp/cascade/triage-prompt.md returns valid JSON
• Markdown-fence-stripped JSON parses with jq empty
• Full scripts/review-one-pr.sh dry run on fix: update code-quality ruleset to full required check list ContentTwin#100 completes through Tier 1 → single confirmation → would-post
• Workflow dispatch on real PR after merge (delegated to maintainer)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes Tier 1 triage reliability by ensuring the model receives the actual PR context (metadata/diff/prior review) inline in the prompt, and by preventing tool-driven “wandering”/context contamination that broke the JSON contract.

Changes:

• Build a self-contained triage prompt file by concatenating prompts/triage.md with inlined PR context in review-one-pr.sh.
• Update run_triage to remove --permission-mode plan and deny tools explicitly for the Claude engine.
• Harden triage JSON parsing by stripping Markdown code fences before jq validation and update the triage prompt wording to reflect the new context source.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
scripts/review-one-pr.sh	Generates /tmp/cascade/triage-prompt.md with inlined PR context; strips JSON fences before validation.
scripts/engine.sh	Adjusts Claude triage invocation to remove plan mode and deny tools.
prompts/triage.md	Updates prompt instructions to reference inlined context and reinforce “raw JSON only” output.

Comments suppressed due to low confidence (1)

scripts/engine.sh:85

• run_triage (copilot engine) passes the entire prompt via a command-line argument: -p "$(cat "$prompt_file")". With PR metadata + up to 3000 diff lines inlined, this can exceed ARG_MAX ("Argument list too long") on large/minified diffs, causing triage to silently fail and default-escalate. Prefer feeding the prompt via stdin or a file-based flag if the copilot CLI supports it, or otherwise enforce a byte-size cap when building the prompt file.

    copilot)
      copilot \
        -p "$(cat "$prompt_file")" \
        --model "$ENGINE_TRIAGE_MODEL" \
        -s --no-ask-user
      ;;

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

TRIAGE_PROMPT_FILE is a fixed, predictable path in /tmp. That can cause collisions across concurrent runs (multiple PRs / parallel jobs) and is vulnerable to symlink clobbering on shared hosts. Consider using mktemp (e.g., under /tmp/cascade) and ensure the file is removed on exit (trap) after triage completes.

Suggested change

	TRIAGE_PROMPT_FILE="/tmp/cascade/triage-prompt.md"
	mkdir -p /tmp/cascade
	TRIAGE_PROMPT_FILE="$(mktemp /tmp/cascade/triage-prompt.XXXXXX.md)"
	cleanup_triage_prompt_file() {
	rm -f "$TRIAGE_PROMPT_FILE"
	}
	trap cleanup_triage_prompt_file EXIT

[Copilot]

The fence-stripping sed only removes lines that are exactly or with no trailing whitespace/CR. If the model outputs fences like "```json " (trailing space) or uses uppercase/language tags with punctuation, jq validation will still fail and force escalation. Consider making the pattern more permissive (trim whitespace, accept optional language token) and/or normalizing CRLF first.

Suggested change

	TRIAGE_RESULT=$(printf '%s' "$TRIAGE_RESULT" | sed -E '/^```[a-zA-Z]*$/d; /^```$/d')
	# Normalize CRLF first, then drop fence lines with optional indentation,
	# optional language tags, and trailing whitespace.
	TRIAGE_RESULT=$(
	printf '%s' "$TRIAGE_RESULT" |
	tr -d '\r' |
	sed -E '/^[[:space:]]*```([^[:space:]]+)?[[:space:]]*$/d'
	)