What Did We Actually Ship?

Accountability from evidence, not reports

Team Pulse

A read-only lens over your team's outcomes that answers one question from real evidence

Team Pulse wraps a read-only HTTP surface over the team's structured vault of people, projects, initiatives, outcomes, tasks, and design docs. Grounded in microsoft/amplifier-bundle-team-pulse.

The question it exists to answer: what did we actually ship?

“a read-only HTTP surface over the team's structured vault of people, projects, initiatives, outcomes, tasks, and design docs”

Mined, not typed

Its knowledge isn't entered by hand — the corpus is mined automatically from real work

The bulk of live knowledge is a corpus of content collections mined for the team: conversation / decision wikis and code / repo wikis — pulled from the team's actual conversations and repos.

The evidence base already exists. So why do standups still run on memory?

Conversation / decision wikis
mined for the team from real discussions
Code / repo wikis
mined for the team from actual repositories

The gap

Reporting runs on self-report — people describe work instead of proving it

Team Pulse's answer to this is a design principle it enforces on every reply: retrieve, don't recall — never answer from pretraining. The fix is to pull the real evidence rather than trust the description of it.

If the answer must be pulled, not recalled — where does it get pulled from?

Anchored on outcomes

A structural graph rolls initiatives up to outcomes and down to shipped work

team_pulse_graph is the compact structural map: initiatives → outcomes → projects → tasks → members. Every answer anchors on an outcome, not a status update.

Anchored on outcomes — but what stops the answer from being made up?

1Initiativesroll up to…
2Outcomesroll down to…
3Projects → Tasks → Membersthe shipped work

One rule

One design rule governs every answer: retrieve, don't recall

The pull is grounded, not remembered. The rule is explicit: retrieve, don't recall — never fabricate a fact, a PR number, or a source.

Grounded evidence sets up the highest-value move: the JOIN.

Retrieve, don't recall
never answer from pretraining
Never fabricate
a fact, a PR number, or a source

The JOIN

Recover a decision in one corpus, confirm the PR that implemented it in another

The highest-value move: recover a decision in one sub-corpus, confirm its implementation (repo + PR# + author) in another, and cite both sides with each sub-corpus's freshness — as of last_updated.

A claim becomes proof: shipped work joined to the decision that ordered it.

# The JOIN
decision  ← conversation / decision corpus
PR        ← code / repo corpus (repo + PR# + author)
# cite both sides:
(source: corpus/<sub>/<page>.md,
         as of <last_updated>)

It refuses to editorialize

It reports the evidence and lets you judge — no scoring

Team Pulse deliberately will NOT score project health, risk, or velocity, recommend what someone should work on, invent fields the lens API doesn't return, or mutate anything — the lens API is read-only.

Accountability comes from the evidence, not the tool's opinion.

Will not score project health, risk, or velocity
Will not recommend what someone should work on
Will not invent fields the lens API doesn't return
Will not mutate anything — read-only

Built to survive scrutiny

Even reflections stay traceable back to real sessions

The single write path, team_pulse_submit_answer, records answers with session provenance — source_session_ids from Context Intelligence sessions — the sole mutation, so everything ties back to real, citable work.

The pattern generalizes: accountability comes from what shipped, not what was reported.

12

tools mounted — 1 configure + 11 data tools

1

write path: the sole mutation, session-mining provenance

Sources

Research Methodology

Primary source: microsoft/amplifier-bundle-team-pulse (created 2026-06-17, last pushed 2026-07-16). Public repo is a single squashed commit e89574d, "Initial public release," by Samuel Lee <samueljklee@gmail.com>, Mon Jul 13 23:50:40 2026 -0700. Project author in pyproject.toml: Microsoft Corporation.

Feature status: v1, read-mostly lookup. Bundle 0.4.1, tool module 0.2.1, team-pulse-lib 0.1.0.

Research performed:

gh repo view microsoft/amplifier-bundle-team-pulse --json name,description,createdAt,pushedAt,url
git rev-list --count HEAD and git log --format='%H%n%an <%ae>%n%ad%n%s' (count=1)
read_file README.md, modes/team-pulse.md, context/using-team-pulse.md, bundle.md (lens API, mined corpus, team_pulse_graph, the JOIN, retrieve-don't-recall, output contract)
read_file modules/tool-team-pulse/…/tool.py L835-847, L918 (12 tools = 1 + 11 data-tool classes)
grep -rn 'def test_' tests/ (74 module test fns / 8 files; 166 lib test fns / 19 files)

Gaps: Test counts are def test_ definitions via grep, not a pytest collection run. The lens API server itself is a separate deployment not in this repo; server-behavior claims are grounded in this bundle's client/docs. README/bundle prose saying "six"/"seven" tools and "20 unit tests" is stale relative to code — code-verified numbers are used here.

Primary contributor: Samuel Lee (sole commit, 100% of git history).