UNRELEASED DEVELOPMENT SNAPSHOT · READ-ONLY EXCERPT

docs/the-cycle.md

This is the cited technical source at local revision 275a354, inspected on 22 September 2026. It differs from the public release. This reference publishes only the cited sections, not an installable release or the development branch.

Line numbers refer to the complete recorded file. Uncited lines are omitted. Full-file SHA-256: 8ea76909090be69064b08a34503f47299421a0c16fd2e24b5d10347b1bb8f5d3.

1# The cycle
2
3Lintel's main workflow. It turns a raw request into shipped, captured work through nine phases,
4each of which is its own skill, produces its own artifact, and declares what it costs you to skip.
5
6```
7SENSE → SCOPE → DEFINE → DISCOVER → PLAN → BUILD → REVIEW → SHIP → CAPTURE
8```
9
10Run the whole thing with `/li:cycle`, a subset with `--mode` or `--from`/`--to`, or any single phase
11on its own (`/li:define`, `/li:review`, …). This page explains what each phase actually does, where
12the gates are, and how the harness keeps your position when the session gets long.
13
14---
15
16## Why nine phases
17
18The phases are not ceremony. Each one exists because skipping it has a specific, observed failure
19mode — which is why every phase skill declares a machine-checkable `gap_if_skipped` field rather
20than a vague "this is important".
21
22The shape is deliberately front-loaded. Five of the nine phases happen before a line of code is
23written, because the expensive failures in agent-driven work are not bad code — they are correct
24code built against the wrong premise, discovered three hours later.
25
26---
27
28## Phase by phase
29
30### 1. SENSE — read the situation
31
32A silent, read-only diagnostic that runs before any decision. It detects operator intent, the active
33pack and its compliance mode, the active role, any cycle state a prior session left behind, and the
34current context budget. It ends with a mode recommendation.
35
36**Produces** a SENSE report and a *scale pre-read* — a first guess at how big this is, plus a
37`bimodal` flag when the request could plausibly be small or large.
38
39**Gate** none. SENSE is cheap and never blocks.
40
41**Skipping it costs** every downstream phase is mis-scoped, and a cycle left half-finished by the
42previous session gets silently resurrected or silently ignored.
43
44---
45
46### 2. SCOPE — size the request
47
48Light, skippable, and mostly silent. SCOPE turns a raw ask into a sized one. It fires **at most one**
49clarifying question, and only when SENSE flagged the request as genuinely bimodal — the discipline
50is that ambiguity gets resolved once, up front, not re-litigated in every later phase. It can also
51override a confidently-wrong route (a request that looks like a one-line deploy but is really a full
52cycle).
53
54**Produces** `scope.md` — resolved size, depth schema, chosen reading, and any route override. The
55depth schema is what PLAN later uses to pick its work-breakdown template.
56
57**Gate** at most one question, only when bimodal. Silent on clearly small work.
58
59**Skipping it costs** scale ambiguity is never resolved and PLAN has no depth schema to select
60against.
61
62---
63
64### 3. DEFINE — lock the design
65
66The forcing-question phase. DEFINE clarifies intent, locks premises, forces alternatives to be
67stated rather than assumed, and picks the wedge — the specific first cut of the problem that the
68rest of the cycle builds. It runs under the active role's lens.
69
70**Produces** an **approved** design document. Not a draft.
71
72**Gate** hard. The cycle does not proceed past DEFINE until you approve the design.
73
74**Skipping it costs** PLAN and BUILD consume ad-hoc prose with no locked premises, so nothing
75downstream can be checked against anything.