fix(review-cycle): optimize Greptile review cycle + add Approach 3 fallback + interview routing fix

[MScottAdams]

Summary

Optimize Greptile review cycle and fix 3 related issues in one PR.

Closes #305, #307, #328, #301

Changes

Task A: Greptile review cycle optimization (#305, t1.30.1) -- 5 changes

1. Mandate deft-pre-pr before PR creation -- upgraded AGENTS.md rule from ~ to ! + added Phase 1 verification gate in deft-review-cycle
2. PR scope gate warning -- Phase 1 warns when PRs span 3+ unrelated surfaces
3. Adaptive poll cadence -- 20-30s first check, 60s second, 90s thereafter (replaces fixed 60s)
4. Parallel rebase + review monitoring -- guidance in deft-swarm Phase 6 for overlapping work via start_agent sub-agents
5. .greptile/rules.md starter template -- elevated to SHOULD in tools/greptile.md

Task B: Approach 3 interactive blocking fallback (#307, t1.31.1)

• Added Approach 3 as last-resort blocking sleep loop for interactive sessions with no start_agent and no timer
• Gated by ! rule requiring user warning before activation
• Updated capability detection to 3 tiers: Tier 1 (start_agent), Tier 2 (scheduler/timer), Tier 3 (interactive blocking)

Task C: Select-String fallback for oversized output (#328, t1.32.1)

• Added ~ fallback command to deft-review-cycle Phase 2 Step 1 for when do_not_summarize_output produces output too large to process
• Windows/PowerShell context: gh pr view <number> --comments | Select-String "Outside Diff" -Context 50

Task D: Tighten deft-interview routing keyword (#301, t1.26.1)

• Replaced bare interview in AGENTS.md Skill Routing with interview loop / q&a loop / run interview loop
• Avoids collision with strategies/interview.md spec-creation intent

Checklist

• /deft:change <name> -- N/A (bundled swarm task, all changes trace to spec tasks)
• CHANGELOG.md -- added entries under [Unreleased]
• ROADMAP.md -- N/A (updated at release time)
• Tests pass locally (task check: 1053 passed, 2 xfailed)

Post-Merge

• Verify issue auto-close: After squash merge, confirm perf(review-cycle): Greptile reviews are slow — investigation + optimization plan #305, fix(skill): deft-review-cycle Approach 2 capability detection gap -- agent claims to yield-poll but no start_agent available and no sleep fallback #307, docs(deft-review-cycle): add Select-String fallback for oversized gh pr view output #328, fix(agents): tighten deft-interview routing keyword -- 'interview' collides with spec-creation intent #301 closed
• Enable branch protection on master requiring CI status check

[greptile-apps]

Greptile Summary

This PR addresses four independent issues in one bundle: (1) mandating deft-pre-pr before PR creation, adding PR scope gate and adaptive poll cadence to the review cycle, and a .greptile/rules.md starter template; (2) adding Approach 3 as a last-resort blocking sleep loop for interactive sessions; (3) a cross-platform oversized-output fallback in Phase 2 Step 1; (4) tightening the interview routing keyword. All four spec tasks (t1.30.1, t1.31.1, t1.32.1, t1.26.1) satisfy their acceptance criteria, and the previous Approach 2 timing-conflict finding has been addressed with the clarifying note at line 157.

Confidence Score: 5/5

Safe to merge — all spec criteria are satisfied, the previous Approach 2 timing concern is resolved, and only two minor documentation gaps remain (P2).

All four spec tasks (t1.30.1, t1.31.1, t1.32.1, t1.26.1) fully satisfy their acceptance criteria. The only findings are a CHANGELOG entry that omits the Unix fallback and a matching gap in the t1.32.1 spec criterion — both P2 documentation issues that don't affect runtime behaviour.

CHANGELOG.md and SPECIFICATION.md (t1.32.1 criterion) — minor documentation gaps around the Unix/macOS grep fallback; no functional files require special attention.

Important Files Changed

Filename	Overview
skills/deft-review-cycle/SKILL.md	Adds Approach 3 blocking fallback with user-warning gate, 3-tier capability detection, adaptive cadence, PR scope gate, cross-platform oversized-output fallback, and pre-PR verification gate — all acceptance criteria met and internally consistent.
skills/deft-swarm/SKILL.md	Adds parallel rebase + review monitoring guidance in Phase 6 via start_agent sub-agents; intentionally omits Approach 3 from the swarm context (correct — blocking loops are not appropriate for swarm monitors).
AGENTS.md	Upgrades pre-PR rule from ~ to !, and tightens interview routing to interview loop / q&a loop / run interview loop — both changes match spec criteria exactly.
tools/greptile.md	Elevates .greptile/rules.md from optional (?) to SHOULD (~) and adds a starter template — straightforward documentation improvement.
CHANGELOG.md	Entries present for all four issues; Task C entry describes only the Windows/PowerShell fallback but the implementation also ships a Unix/macOS grep fallback — minor scope understatement.
SPECIFICATION.md	All four spec tasks marked [completed] with appropriate acceptance criteria; t1.32.1 criterion only verifies the PowerShell fallback but the implementation also added a Unix fallback — no functional issue.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Review Monitoring triggered] --> B{start_agent available?}
    B -- Yes --> C[Tier 1: Approach 1\nSpawn sub-agent monitor\n20-30s / 60s / 90s cadence]
    B -- No --> D{scheduler / timer /\nauto-reinvocation available?}
    D -- Yes --> E[Tier 2: Approach 2\nYield-between-polls\nRuntime-controlled ~60s interval]
    D -- No --> F{Interactive session?}
    F -- Yes --> G[Tier 3: Approach 3\nWarn user → get confirmation\nBlocking sleep loop\n25s / 60s / 90s cadence]
    F -- No --> H[No monitoring path available\nDocument and escalate]
    C --> I{Exit condition met?\nconfidence > 3, no P0/P1}
    E --> I
    G --> I
    I -- No --> J[Wait and re-poll]
    J --> I
    I -- Yes --> K[Proceed to Step 5]

Loading

Prompt To Fix All With AI

This is a comment left during a code review.
Path: CHANGELOG.md
Line: 13

Comment:
**CHANGELOG entry understates Task C scope**

The entry says "(Windows/PowerShell context)" but `SKILL.md` also ships a Unix/macOS `grep` fallback alongside the `Select-String` one. A reader scanning the CHANGELOG for the Unix fallback won't find a reference to it.

```suggestion
- **Select-String fallback for oversized gh pr view output** (#328, t1.32.1): Added ~ fallback commands to deft-review-cycle Phase 2 Step 1 for when do_not_summarize_output produces output too large to process — PowerShell (`Select-String "Outside Diff" -Context 50`) and Unix/macOS (`grep -A 50 "Outside Diff"`)
```

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: SPECIFICATION.md
Line: 1257-1259

Comment:
**t1.32.1 acceptance criteria only verify the PowerShell fallback**

The implementation added both a PowerShell (`Select-String`) and a Unix/macOS (`grep -A 50`) fallback, but the spec only verifies the PowerShell command and notes "Windows/PowerShell context". The Unix fallback is real behaviour with no spec coverage.

```suggestion
- skills/deft-review-cycle/SKILL.md Phase 2 Step 1 contains ~ fallback guidance for oversized output with two platform variants: PowerShell: `gh pr view <number> --comments | Select-String "Outside Diff" -Context 50`; Unix/macOS: `gh pr view <number> --comments | grep -A 50 "Outside Diff"`
- Fallback is positioned after the `do_not_summarize_output: true` rule
- Guidance labels each variant by platform (PowerShell/Windows vs Unix/macOS)
```

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (3): Last reviewed commit: "fix: address Greptile review findings (b..." | Re-trigger Greptile}

[MScottAdams]

Rebase-only force-push onto updated master (after PR #329 merge). Only CHANGELOG.md conflict resolved -- merged both PRs' [Unreleased] entries. No logic changes.