iter boss: /boss skill — autonomous-mode discipline gated to user invocation

CLAUDE.md previously mixed universal facts (agent role boundaries,
commit discipline, design rationale, file roles, TDD-for-bugs) with
mode-specific autonomy rules (direction freedom, notifications,
WhatsNew procedure). Autonomous-by-default conflicted with the user's
intent that a fresh session should be collaborative-interactive unless
explicitly elevated.

Add `skills/boss/` containing only the three genuinely mode-specific
subsections — Direction freedom, Notifications, Done-state notifications:
WhatsNew.md. Trim CLAUDE.md from 343 to 243 lines; extend the skill-
system pointer paragraph with a one-sentence /boss gate. Universal
orchestrator discipline stays in CLAUDE.md because it applies whether
/boss is active or not.

Two cross-references that named the moved subsections by sub-heading
are repointed: skills/implement/SKILL.md and the implement-orchestrator
agent's standing reading list. The other ~11 agent-file references to
"orchestrator framing" still resolve correctly because that framing
stays in CLAUDE.md.

skills/README.md skill table extended with a `boss` row (now eight
skills); pipeline-diagram caption notes /boss wraps the pipeline.
.claude/skills/boss symlink follows the existing relative-path
convention.
This commit is contained in:
2026-05-12 10:14:33 +02:00
parent a99a39b65e
commit e953b137eb
9 changed files with 365 additions and 112 deletions
+6 -106
View File
@@ -48,6 +48,12 @@ are sharper tools, not a replacement for orchestrator judgement.
Specs go to `docs/specs/<milestone>.md`, plans to
`docs/plans/<iteration>.md`.
Autonomous orchestrator mode — picking the next iter from
`docs/roadmap.md` and looping until done-state — is gated to the
user-invoked `/boss` skill (`skills/boss/SKILL.md`). Outside
`/boss`, the default is interactive collaboration: the user asks,
Claude responds, Claude stops.
## My role: orchestrator
I am the **orchestrator** of this project, not the implementer. The
@@ -167,112 +173,6 @@ appeal does not count; neither does human ergonomics. Full criterion
lives in `docs/DESIGN.md` ("Feature-acceptance criterion") and is
applied as a gate by `skills/brainstorm/SKILL.md` during spec writing.
### Direction freedom
I have authority to choose the next iteration, refactor, or feature
without asking. Wrong calls are contained, not recovered: only the
Boss commits, and only when the working-tree state is consistent —
so bad work doesn't reach main in the first place. If a dispatched
agent's output is wrong, the remedy is `git checkout -- <paths>` on
the working tree (Boss-side), never a rewind of main HEAD. main is
forward-only.
The cost of asking "what should I do next" — context-switch for
the user, latency on my side — exceeds the expected cost of
discarding a bad working-tree state. So when the queue is non-empty
and the path is clear, just pick and proceed.
Bounce back to the user only when:
- A queued option requires a real design judgement I have not
made myself (genuine architectural fork, multiple substantive
options none of which is clearly default).
- I have hit something genuinely unexpected that changes the
project's direction (a fundamental design flaw, an external
dependency failure, a discovered invariant violation).
- The user has explicitly asked for a checkpoint.
A summary of what shipped is fine and welcome — but in
autonomous mode, follow it with the next dispatch, not a
question.
### Notifications
Two states call for a notify; nothing else does:
1. **Done-state.** The autonomous queue is empty — the milestone
closed, the audit ratified, and no further iteration is obvious
from the spec or the roadmap (or, in a user-scoped session,
every objective the user named has been completed and there is
nothing else queued). A finished sub-goal inside a still-open
milestone is *not* done-state. If the next iteration is obvious
and substantive, keep dispatching — "test suite landed", "audit
clean before close-fixes", "iter N done with N+1 in scope" are
internal progress, not notify events.
2. **Problem-state.** A bounce-back trigger per §"Direction
freedom" — design fork I cannot resolve, an unexpected
invariant violation, an external dependency failure, or the
user has explicitly asked for a checkpoint.
A user-specified stop trigger ("until X is done, then notify")
counts as done-state only if X is truly the last actionable item
in the autonomous queue at that moment. If hitting X still leaves
an open milestone with obvious follow-up iters, the right action
is to continue past X without notifying — the user's intent
behind the trigger is "wake me when there's nothing else to do",
not "wake me at this specific checkpoint".
Mid-flow progress notifications burn the user's attention without
giving him a decision to make. When in doubt, continue.
~/.claude/notify.sh "Text"
When notifying, the message body should be the actionable
summary: what I need from them, in one short line. Skip the
"hi, I" framing — just the gist. The user will see it on phone
and likely respond by returning to the session.
By default, I act autonomous. The notification is the
exception, not the rule.
### Done-state notifications: WhatsNew.md
When the trigger is **done-state** (autonomous work has wrapped
up, nothing left to do for the user to react to), the procedure
is two-step and the two outputs share one text:
1. Append a new entry to `docs/WhatsNew.md` (newest at the bottom,
append-only — same chronological convention as the per-iter journals).
2. Send the same text via `~/.claude/notify.sh`.
The text must be written for the **user-as-reader who did not
watch me work**:
- **Language: English.** Same as the rest of the repo; both
WhatsNew.md and the Notify push are permanent enough to be
treated as project content.
- **No technical internals.** Don't name crates, function
identifiers, type names, agent names, or iteration codes
(`ct.1.7`, `18a`, etc.). Describe the *change in the project*
— what is now possible, fixed, or visible — not the mechanics.
- **No assumption of context.** The user is on his phone and has
not seen this session. Lead with the result; the *why* is fine
if it fits in one short clause.
- **Length: Telegram-pragmatic.** No hard cap, but a few sentences,
not an essay. If the work covered multiple unrelated things,
bullet them.
- **Tone: factual, not performative.** No celebration, no "I have
successfully…". Just "X works now, Y is fixed, Z is new".
**Bounce-backs** (asking the user for a design judgement or
checkpoint) are *not* done-state and do **not** produce a
WhatsNew entry — they stay as Notify-only.
Format of a WhatsNew entry: a `## YYYY-MM-DD — <short title>`
heading followed by the message body. The body must match the
Notify text verbatim.
## Bug fixes — TDD, always
Bug fixes are RED-first, autonomous, no orchestrator gate. See