Add meeting tech analysis tasks

[ScuttleBot]

Adds 5 new meeting analysis tasks for the GitLab Product Marketing meeting transcript:

1. task_meeting_tech_action_items - Extract action items and owners from the meeting
2. task_meeting_tech_decisions - List all decisions made with context
3. task_meeting_tech_competitors - Identify competitors and summarize positioning strategy
4. task_meeting_tech_messaging - Extract messaging framework options and final selections
5. task_meeting_tech_product_features - Create prioritized feature list for GitLab Commit keynote

All tasks use the meetings/2021-06-28-gitlab-product-marketing-meeting.md asset and include:

• YAML frontmatter (category: meeting, grading_type: hybrid, timeout: 180s)
• Clear prompts with specific extraction requirements
• Expected behavior with ground-truth data from the transcript
• Automated Python grade functions with regex-based checks
• LLM judge rubrics with weighted criteria

Closes #180, Closes #181, Closes #182, Closes #183, Closes #184

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

This PR adds five well-structured meeting analysis tasks. The grading functions use robust regex patterns with fallback file name checks, partial credit scoring, and appropriate early-exit on missing output files. The LLM judge rubrics are well-calibrated with clear weighted criteria. No critical bugs or security vulnerabilities found.

Files Reviewed (6 files)

• tasks/manifest.yaml
• tasks/task_meeting_tech_action_items.md
• tasks/task_meeting_tech_competitors.md
• tasks/task_meeting_tech_decisions.md
• tasks/task_meeting_tech_messaging.md
• tasks/task_meeting_tech_product_features.md

---

_{Reviewed by claude-4.6-sonnet-20260217 · 159,578 tokens}

[ScuttleBot]

🧪 PinchBench PR Test Started

Instance: 45.76.255.76 (Vultr vc2-2c-4gb, Ubuntu 22.04)
Branch: tasks/meeting-tech

Models Being Tested

• openrouter/anthropic/claude-opus-4.6
• openrouter/openai/gpt-5.4
• openrouter/google/gemini-3.1-pro-preview

Tasks Being Tested

• task_meeting_tech_action_items
• task_meeting_tech_decisions
• task_meeting_tech_competitors
• task_meeting_tech_messaging
• task_meeting_tech_product_features

Estimated Completion

~30-45 minutes (all 3 models running in parallel)

Started at 2026-04-14 14:46 UTC

[ScuttleBot]

🧪 PinchBench PR #313 Test Results — tasks/meeting-tech

Instance: 45.76.255.76 (Vultr vc2-2c-4gb, Ubuntu 22.04)
Branch: tasks/meeting-tech @ bcfca24
Test duration: ~22 min (all 3 models in parallel)

Per-Task Scores

Task	Claude Opus 4.6	GPT-5.4	Gemini 3.1 Pro
action_items	63%	0% ❌	88%
decisions	53%	0% ❌	72%
competitors	60%	0% ❌	92%
messaging	57%	0% ❌	97% 🔥
product_features	91%	0% ❌	88%
Overall	64.9%	0.0%	87.5% ✅

Cost & Efficiency

Model	Tokens	Cost	Score/$
Claude Opus 4.6	397K	$1.06	3.06
GPT-5.4	0	$0.00	N/A
Gemini 3.1 Pro	301K	$0.48	9.10

Issues & Observations

GPT-5.4 (0% — complete failure):
All 5 tasks failed with task execution failed (error), no transcript to evaluate. The agent workspace could not be found (Could not find agent workspace, using fallback) and session transcripts were empty. This is a known OpenClaw agent creation issue with GPT models, not a task design problem. GPT results should be disregarded for evaluating these tasks.

Claude Opus 4.6 (64.9% — penalized by judge context issues):
Claude actually completed most tasks, but the judge (Claude Opus 4.5) had trouble evaluating them:

• Tasks 1-4: Judge reported truncated/incomplete transcripts or missing candidate responses ("Only received the grading rubric and a fragment of source material"). This appears to be a judge context window issue rather than a task execution failure.
• Task 5 (product_features): Full evaluation worked, scored 91% — comparable to Gemini's 88%.
• The 53-63% scores on tasks 1-4 likely understate Claude's actual performance.

Gemini 3.1 Pro (87.5% — strong across all tasks):

• Consistently high scores across all 5 tasks (72-97%)
• messaging task scored highest at 97% — judge noted "Exceptional output. All candidate taglines captured with complete pros/cons."
• decisions was the hardest at 72% — missed messaging tagline selection and competitive analysis methodology approach
• Judge had 1st-attempt failures on every task but succeeded on retry

Task Difficulty Analysis

Based on Gemini scores (the only model with reliable evaluations across all tasks):

1. Easiest: messaging (97%) — Well-scoped extraction of tagline evaluation framework
2. Easy: competitors (92%) — Clear competitor identification with positioning analysis
3. Medium: action_items (88%) and product_features (88%) — Require thorough extraction across multiple meeting segments
4. Hardest: decisions (72%) — Requires identifying implicit decisions and their supporting context

Judge Reliability Note

The Claude Opus 4.5 judge had first-attempt failures on nearly every evaluation across all models, requiring retries. When evaluating Claude's output, the judge frequently reported incomplete context. This suggests the meeting transcript + task output may be pushing the judge's context window limits, or the multi-part message format isn't working reliably.

Recommendation

✅ Merge with caveats:

• Tasks are well-designed and test meaningful meeting analysis capabilities
• Gemini's 87.5% average suggests the rubrics are calibrated well (not too easy, not too hard)
• The decisions task at 72% provides good differentiation
• GPT failure is an infrastructure issue, not a task issue
• Claude's low scores appear to be primarily judge-side evaluation issues
• Suggestion: Consider whether the meeting transcript + prompt + grading rubric together may exceed typical judge context limits; might benefit from a more concise rubric or summary-based judging approach

Tested at 2026-04-14 14:48 - 15:12 UTC