Technical Deep Dive

Team Tracking Bundle

Recipe Engineering at Scale: Patterns for Reliable Bulk Processing

Currently Disabled

February 2026 · Two bundles · 19 recipes

The Mission

Bulk Session Analysis

Goal: Process hundreds of Amplifier sessions to understand team workflows, capabilities, and project patterns.

The Challenge: Build a recipe system that could reliably analyze sessions at scale — without crashing on missing files, choking on large transcripts, or hammering API rate limits.

Current status: The team tracking system is disabled (enabled: false in config since Jan 30, 2026). The engineering patterns discovered during development remain valuable regardless of deployment status.

The Ecosystem

Two Bundles, One Goal

📦

marklicata/amplifier-bundle-team-tracking

Session sync, bulk analysis, dashboards, and research pipeline recipes. 12 recipe files (~5,034 lines of YAML). Version 2.0 with smart sampling.

Author: Mark Licata (sole contributor)

🔧

microsoft-amplifier/amplifier-bundle-made-support

Support intake, story submissions, and team-status auto-update hook. 7 recipe files (~990 lines of YAML). Version 1.3.

Primary author: Salil Das (~95% of 61 commits)

Problem #1

Recipe Crashes on Missing Files

The sync-session recipe would crash when sessions lacked events.jsonl files.

❌ What We Expected

Conditionals protect against missing data. If a file doesn't exist, the conditional skips that step.

✓ What Actually Happens

The template engine evaluates ALL variables first, then checks conditionals. Missing files cause crashes before conditionals even run.

# Template evaluates $EVENTS even if conditional would skip it
EVENTS=$(cat ${SESSION_DIR}/events.jsonl)
# ← Variable resolved BEFORE the "if exists" check runs
# ← Empty $EVENTS → recipe crashes

The Solution

The "Option A" Pattern

Key insight: Move conditionals inside bash commands, not in recipe logic.

# ✓ Option A: Conditional inside bash
if [ -f "${SESSION_DIR}/events.jsonl" ]; then
  cat "${SESSION_DIR}/events.jsonl"
else
  echo ""  # Always return output
fi

# Result: No crashes. Every step returns output, even if empty.

Golden Rule: All recipe steps must always produce output, even if it's an empty string. The bash layer handles file existence; the recipe layer handles data flow.

Problem #2

Death by API Calls

Original Approach

Multiple GitHub API calls per session

Fetch session metadata via API
Download events.jsonl
Download transcript
Parse, analyze, repeat

Result: Slow for large batches. Rate-limited by GitHub.

Optimized: Local-First

Direct filesystem reads from ~/.amplifier/sessions/

Read local session files directly
Zero API calls for analysis
Parse directly from disk
Batch commits to reduce API usage by ~85%

Result: Significantly faster. No rate limiting.

Problem #3

ARG_MAX and Large Files

Large transcripts (580KB+) exceeded shell argument limits when passed as command-line arguments.

Before

Passing data via arguments

curl -X POST --data "$TRANSCRIPT"

❌ Fails on files >128KB (ARG_MAX limit)

After

Piping via stdin

echo "$TRANSCRIPT" | curl --data-binary @-

✓ Handles 580KB+ files reliably

Lesson: In recipe bash steps, always pipe large data through stdin rather than passing as shell arguments.

Architecture

19 Recipes Across 4 Categories

⚡

Automatic (via hooks)

sync-session, analyze-session — run on session end. Upload files, extract capabilities, update indexes in real-time.

🔄

Catchup (manual)

bulk-analyze-sessions, bulk-sync, find-unsynced-sessions — batch recovery when automatic sync has gaps.

📊

Reporting

team-dashboard, all-teams-dashboard, person-metrics, team-metrics — HTML dashboards from indexed session data.

🔬

Research Pipeline

detect-research-adoption, promote-research-tool — identify experimental tools ready for productization.

Development Velocity

Built Across Two Repos

2

Repositories

61+

Commits (made-support)

19

Recipe files

~6k

Lines of YAML

Primary contributors: Mark Licata (team-tracking bundle, sole author) · Salil Das (~95% of made-support, 58 of 61 commits) · Samuel Lee (2 commits) · Sam Schillace (1 commit)

Sources

Research Methodology

Data as of: February 20, 2026

Feature status: Disabled (enabled: false in ~/.amplifier/team-tracking.yaml)

Research performed:

Config check: cat ~/.amplifier/team-tracking.yaml — confirmed disabled
Bundle cache: find ~/.amplifier/cache -path "*team-tracking*" — 12 recipe YAML files found
Recipe lines: wc -l on all YAML — 5,034 lines (team-tracking) + 990 lines (made-support)
Git log: git log --oneline in made-support — 61 commits found
Contributors: git log --format="%an" | sort | uniq -c — 3 contributors in made-support
Timeline: git log --format="%ai" — Jan 14 to Feb 9, 2026
Repo origins: git remote -v — marklicata (team-tracking), microsoft-amplifier (made-support)

Gaps & corrections from original deck:

"354 sessions processed" — no evidence found; zero local sessions exist. Removed.
"600 capabilities mapped" — no evidence found. Removed.
"100x performance improvement" — no baseline measurement available. Downgraded to qualitative.
"0 recipe crashes" — self-validating claim. Removed.
"Commit 435d0d4" — does not exist in either repository. Removed.
"bulk-analyze-sessions-local.yaml" — file does not exist. Actual: bulk-analyze-sessions.yaml.

Note: The marklicata bundle is a shallow clone (1 commit visible). Commit count reflects made-support only.

Key Takeaways

Recipe Engineering Patterns

1. Template Variables Evaluate First

Don't rely on recipe conditionals to protect against missing data. Variables are resolved before conditionals run.

2. Use the "Option A" Pattern

Move conditionals inside bash commands. Ensure every step returns output, even if empty.

3. Pipe Large Files via Stdin

Use echo "$VAR" | curl --data-binary @- to avoid ARG_MAX limits on 580KB+ transcripts.

4. Optimize for Local-First

For bulk operations, prefer filesystem reads over API calls. Batch commits reduce API usage by ~85%.