ConductorScore reads your local coding-agent transcripts and uploads only the
resulting numbers (your score) — never transcript text.
How to run
When the user invokes this skill, say
"Calculating your ConductorScore" and,
on the next line, the skill version the script reports (its first output line is
) — e.g. show
beneath the heading. Run the
orchestrator from
this skill's own directory — the folder that contains this
SKILL.md, with
in its
subdirectory. Don't assume a fixed
location: depending on the agent and installer the skill may live under
~/.claude/skills/conductorscore/
,
~/.codex/skills/conductorscore/
,
~/.cursor/skills/conductorscore/
, or
~/.agents/skills/conductorscore/
.
Use the
next to this file:
python3 <this-skill-dir>/scripts/run.py
On Windows, if
isn't found, use
or
instead.
If the script prints
↑ vX.Y.Z available — run: gh skill update …
, relay that
line so the user can update; then continue normally.
The script drives the flow. Don't narrate the steps or restate its progress
lines. Your jobs are: (a) run the script, (b) handle the
lines below, and (c) when the scan finishes, render the
final report from
the
line (see "Final report").
ASK relay rule
If a printed line starts with
CONDUCTORSCORE_ASK <id> "<question>" [opt A] [opt B] …
,
the script has STOPPED and is asking the user to decide. Present
as a
multiple-choice using the bracketed options (use the question UI; don't add extra
prose), then re-run
with the matching flag:
- → the script printed a GitHub URL + code. Show them, let the user
authorize in their browser, and when they confirm ("I've authorized") re-run
with NO new flag — it resumes the login and continues. On Cancel, stop.
- → re-run with (All),
(Claude Code), (Codex), or
(Cursor). On Cancel, stop.
- → present EXACTLY two choices, Yes and No (no Cancel, no other
options). Re-run with ONLY or — do NOT add
; this enables/declines the once-per-day hook and does NOT re-scan.
Final report
When the scan finishes the script prints a single line:
CONDUCTORSCORE_RESULT {"score":N,"url":"...","sessions":N,"verified_github":bool}
.
Don't show that raw line. After the
question is resolved, present the
user a final report in exactly this shape (keep the ✓ and 📊 and the bullets):
✓ ConductorScore: <score>
📊 View your full breakdown: <url>
- Scanned <sessions> sessions from your local transcripts
- <Daily auto-refresh enabled — your score updates once per day | Daily auto-refresh not enabled>
- Only your score (the numbers) was uploaded — never any transcript text
Debug / visualize a session —
renders a session locally as a self-contained HTML
timeline (HITL/AFK/Tool/Idle) and opens it in the browser. It runs entirely on
the user's machine and
uploads nothing — not even the numbers a scan sends.
Use this path when the user types
, or asks to
visualize / view / debug a session, show its
timeline, or see the
HITL/AFK/Tool/Idle breakdown.
The slash interface IS the
interface: take whatever the user
put after
and pass it
straight through to the Python CLI — every
flag below maps 1:1, so
/conductorscore debug <ARGS>
runs
.
python3 <this-skill-dir>/scripts/session_viewer.py [ARGS]
(
sits next to
in this skill's
, wherever
the skill is installed. On Windows use
or
if
isn't
found.)
Supported arguments (the full Python CLI surface):
- (no args) → the session the user is in. The SessionStart hook records the
current transcript, so a bare run visualizes the active session. To inspect an
earlier one, the user navigates to it first with the built-in picker
( / , or ), then runs
— don't reinvent that navigation.
- → render that explicit transcript.
- → print discoverable local sessions (index, date, id, project) and
exit.
- → render the Nth session from (newest first).
- → show the real transcript text. Default is redacted:
bubbles show the wire-equivalent view (hashes + token counts, tool names only —
no raw text), matching exactly what a scan would upload, so the page is safe to
share. is safe locally (nothing leaves the machine) and best for
hands-on debugging.
- → write the HTML somewhere other than
.
- → render without launching the browser.
The script prints a
local-only debug summary — the source transcript, the
redaction mode, message/turn/streak/leverage counts, and an explicit
"nothing is uploaded" assurance — then writes the HTML and opens it.
Relay the
script's console summary and its line as-is; don't
re-narrate or restate the numbers.
Device-flow login
First run prints a
https://github.com/login/device
URL and a code, then STOPS
(it does not poll in the background). Show the URL and code to the user, let them
authorize in their browser, and once they confirm, re-run
— it resumes
the pending login, finishes authentication, and proceeds to the scan.
Privacy & data handling
ConductorScore runs entirely on your machine. It reads your local coding-agent
transcripts to compute metrics, then uploads only:
- Numbers — counts, token totals, timing, and derived scores.
- Names you already surface by using them — the skills, slash-commands, MCP
servers, plugins, and model IDs invoked in your sessions (these appear on your
public profile). Just the names — no arguments, no contents.
It never uploads transcript or message text, prompts, code, file contents, or file
paths. Message and file text is reduced on-device to a truncated SHA-256 hash (used
only for dedup/similarity) before any metric is computed. The exact wire payload is
documented and inspectable in
(
is a real
sample). Uploads happen only after you pair this device, only for the agents you
consent to scan, and never on an unpaired machine.