安装器如何把规则装进项目 深度解析

4**Unit of work:** one tracked item (an issue, a ticket, a task, a deliverable) per session by
5default. Its description is the contract. Two or three closely-related items on the same branch
6or workspace is fine when they share scope naturally. For an obvious small fix, a one-line
7description instead of a formal ticket is acceptable.
8
9**Autonomy:** proceed through **plan → do → verify → finalize → hand off** without asking for
10approval between steps. Decide routing and scope calls yourself, explain the WHY in the
11commit/PR/summary, and move on. The autonomy is in the *quantity of un-gated steps*, never in
12the *quality bar* — the principle rules and the profile's gates still apply to every step.
13
14**When the item is ambiguous:** research the right answer (official docs, reputable sources,
15the codebase/asset itself) and pick the researched path. **Do not default to "most
16conservative" or pick at random.** Note what you researched in the commit/PR/summary so the
17choice is auditable later.

23**Match your rigor to the stakes.** Work sits on a spectrum from quick-and-loose (a throwaway
24draft, a scratch experiment — "does it seem to work?") to fully disciplined (production systems,
25anything irreversible or outward-facing — verified at every stage). The skill is picking the
26right point per task: don't ceremony-wrap a five-minute scratch task, and don't "seem-to-work"
27something that ships to real users or touches real money. The profile sets the floor; raise it
28when the stakes are high. The single thing that separates disciplined work from guessing is
29**how the output gets verified** — see the principle rules.
30
31**Mind the last 20%.** You can produce the easy 80% of almost anything fast; the remaining 20% —
32the edge cases, the error handling, the integration seams, the subtle correctness — is where the
33real work is and where "looks right, even passes a quick check" hides the bugs. Spend your
34attention there, on the ambiguous and the hard-to-verify, not on re-admiring the easy part.
35
36**Conductor vs orchestrator — pick the altitude.** Some work wants *conductor* mode: hands-on,
37step-by-step, you watching each change (debugging, exploring unfamiliar ground, high-stakes
38edits). Other work wants *orchestrator* mode: define the goal, delegate to subagents, review
39outcomes rather than keystrokes (well-specified features, migrations, parallelizable sweeps).
40Neither is "more advanced" — choose by the task, and drop to conductor the moment something gets
41surprising.
42
43**Budget — a self-check, not a hard stop.** Keep work atomic (one concern per commit/change)
44and deliverables small. If you find yourself with 5+ unrelated changes stacked up, either split
45them or stop and hand off cleanly. Don't pause for elapsed time or time-of-day — momentum is
46fine as long as the quality bar holds. If you'd overshoot a natural stopping point, finish the
47current unit cleanly and write the handoff (see `50-git-and-handoff`).

45for p in "${profiles[@]}"; do
46  name="$(basename "$p" .md)"
47  out="$WORK/$name"
48  mkdir -p "$out"
49  if ! bash "$SETUP" --profile "$name" --assemble-only --target "$out" \
50        --operator-name "Test" --tracker "Linear" >/dev/null 2>&1; then
51    bad "$name: assembly failed outright"
52    continue
53  fi
54  f="$out/CLAUDE.md"
55  if [ ! -s "$f" ]; then
56    bad "$name: produced no CLAUDE.md"
57    continue
58  fi
59
60  errs=""
61  # An unrendered token means the agent reads a literal "{{TRACKER}}" as if it were the rule.
62  # Never acceptable, in any profile.
63  leaked="$(grep -oE '\{\{[A-Z_]+\}\}' "$f" 2>/dev/null | sort -u | tr '\n' ' ')"
64  [ -n "$leaked" ] && errs="$errs unrendered:[$leaked]"
65  # [TODO: set X] means "the assembler had no value for X". Two very different cases:
66  #   - X is flag-driven (below) → setup COULD have filled it and didn't. That's a bug: the agent
67  #     gets a blank where a rule should be. This is the regression that would fire if a
68  #     placeholder were renamed in core/ but not in fill_placeholders().
69  #   - X has no flag (BRAND_PALETTE, BRAND_FONT) → the TODO IS the product. It's the documented
70  #     handoff to the human (INSTALL.md's placeholder table; setup's own "resolve any [TODO]"
71  #     next-step). Flagging it would make this suite red by design and train people to ignore it.
72  for tok in OPERATOR_NAME OPERATOR_ROLE OPERATOR_BIO TRACKER TRACKER_POLICY; do
73    grep -qF "[TODO: set $tok]" "$f" 2>/dev/null && errs="$errs unfilled-flag-token:[$tok]"
74  done
75
76  if [ -n "$errs" ]; then
77    bad "$name: assembled with leaks —$errs"
78  else
79    ok "$name: assembles clean (no unrendered tokens, no TODO blanks)"
80  fi