kharmanskyi / os-done-or-not

os-done-or-not

One question, one screen: did the agent finish, and what actually happened,
in words a reader who does not code will understand. Nothing happened (pure
questions, no files touched) → one line saying so, no report.

Language

The user's language in this session, detected from the conversation. Translate
every template label. Code, files, commands stay English.

Step 1 - gather proof that takes seconds

Fast checks only; never re-run the test suite - use results this session
already produced. Not confirmable in seconds, or still running → **"not
checked"**, never "yes".

git status --porcelain            # uncommitted?
git log --oneline -10             # what landed
git log HEAD --not --remotes --oneline 2>/dev/null | head -20   # unpushed? (no remote: say so)
gh pr view --json state,mergeStateStatus && gh pr checks   # if a PR exists

Step 2 - name the outcome

Sessions end one of eight ways; pick the match before writing, or the
report says "fully done: yes" and "safe to close: no" in the same breath.

Outcomes 5 and 6 are where reports start lying; "the approach failed and was
rolled back" is a complete result.

Step 3 - write the report (translate the labels, keep the shape)

<Lead: 1–2 sentences. What changed for the product. Best result first.>

| | |
|---|---|
| ✅ | <done, and what proves it> |
| ⚠‹VARIATION SELECTOR U+FE0F› | <surprise or bad news> |
| ⏳ | <deferred, and until when> |

**Verdict**

| | |
|---|---|
| Fully done?              | Yes / No / Not checked |
| Anything needed from you?| No / <one concrete action> |
| New debt?                | No / <how many, where recorded> / Not checked |
| Safe to close?           | Yes / No - <reason in five words> |

Checkmark rows: two to five; drop an empty row, never a non-empty ⚠‹VARIATION SELECTOR U+FE0F›. Verdict
cells are one line each - detail lives in the checkmark rows, not the verdict.

Two conditional rows, only when the session makes them real - never "not
applicable": Easy to undo? (Yes - how / Hard - why) and **Security
touched?** only when yes - one line: exposed what, closed how. Deliberately no
"live for users?" row: unmerged it repeats the action line; merged but not
reaching users is a surprise - a ⚠‹VARIATION SELECTOR U+FE0F› row.

Step 4 - save it

Write both, creating folders as needed. The paths are the same on every tool.
If the file tool refuses them as outside the workspace, use the shell. Save
nowhere else, and never in the user's project. <project> is the name of the
folder the session started in, or of its repository's top folder when it is
inside one (the name only, not the path):

• ~/.claude/open-steps/reports/<project>/latest.md (overwritten)
• ~/.claude/open-steps/reports/<project>/history/<YYYY-MM-DD-HHMM>.md

Head the file with date, project, ticket. **Two parts; only part one goes in
the chat.** Part one - the report above, for the person. Part two - for the
next session, which reads it instead of re-exploring the repository; the only
home for engineering identifiers.

---
## Technical detail - for the next session, not the reader above

- Branch / worktree / PR and their state; commits made this session
- Files changed, with paths; test and check results as measured
- Commands worth re-running; the first thing to look at next

Terse, factual, a handover note; nothing changed → omit part two.

Step 5 - fold it into the map

The report describes one session. BIG-PICTURE.md holds the standing picture
of the product, and this is the moment it goes out of date. Use the
os-big-picture skill to update only the rows this session touched - a new
feature, a stage that moved, a deferred item joining the backlog.

Skip it in two cases only: outcome 8, or no BIG-PICTURE.md and the user has
never asked for one. A session that changed no row still runs it, because the
map's measured columns went stale while the session ran.

Jargon → plain words

Examples of the move - apply it in the user's language. A ticket or PR number
naming an action ("review PR #892") stays; elsewhere say what it changed.

Hard rules - these rules are the skill

1. One screen: about 10 lines, 15 the ceiling.
2. Lead with the outcome - best thing first, never the chronology.

3. Numbers a human can use - counts, money, time; hashes, branch names and
   build IDs stay out unless asked.

4. Translate every term; no plain equivalent → say what the user would see.
5. Every "yes" names its proof; no proof → "not checked".
6. Bad news gets its own ⚠‹VARIATION SELECTOR U+FE0F› row - never buried inside another line.

7. Exception, do not compress: an unhandled security risk or data loss is
   spelled out plainly - in the lead and its ⚠‹VARIATION SELECTOR U+FE0F› row, never by inflating a
   verdict cell. A handled risk is one line: exposed what, closed how.

Known gotchas

• "New debt? - No" gets written when there is no debt and when nobody looked;
  say "no" only after checking.

• Neither a green check nor a merge means users have it: landed-but-unreached
  gets a ⚠‹VARIATION SELECTOR U+FE0F› row naming what would ship it - the most common way a report ends
  up technically true and practically wrong.

• An approval can be revoked - a verdict is a snapshot; say so while a review
  is open.

• Do not grow the table: ten rows is a wall of text in a table costume.
• Matches none of the eight → say what happened; the list serves honesty.