Cadora · AI-DLC method pack
AI-DLC is not a mode inside Cadora — it is an installable method: rules plus inputs that structure work into three phases. Cadora lays that method into a workspace, drives a phased topology through it, puts deterministic gates and human-review gates between the phases, and reads back a phase-by-phase audit trail. This is that method's journey, from an empty directory to an audited run.
One command turns an empty directory into an AI-DLC workspace. Cadora installs the rules — the v1 pack vendored in cadora/aidlc_rules/ — as backend-native project memory: core-workflow.md becomes the executor's instruction file (CLAUDE.md for Claude Code, AGENTS.md for Codex), and the detailed per-stage rules land in .aidlc-rule-details/. Your inputs — vision.md and the optional tech-env.md — sit beside them.
Pick the backend with --executor claude|codex|kiro — it only changes which instruction file the rules are written into. The method itself is identical across backends.
The method organizes work into inception → construction → operations — the three directories under .aidlc-rule-details/. Cadora's job is to sit between the phases: a review gate (a node marked review: true) pauses for a human before downstream work starts, and a build-test gate re-runs the real build and tests before a phase is allowed to count as done. The node ids below are the real ones from examples/aidlc-phased.topology.yaml.
A node's phase is more than a label: --construction-executor codex routes only the construction-phase nodes to a second backend, leaving inception and operations on --executor.
The three-phase model is not hidden in Cadora — it is written in the topology you hand to cadora run. Each node carries a phase, a dependency edge, and — where the method wants a checkpoint — a gate or a review: true flag. Here is examples/aidlc-phased.topology.yaml, prompts elided.
The full per-stage lifecycle — seven inception stages, six construction stages, conditional self-skips — lives in examples/aidlc-stages.topology.yaml. The single-session shape is examples/aidlc.topology.yaml (one node, gate: build-test).
One command installs the workspace and drives the phased topology through it. With --hitl, Cadora pauses at every review: true node — here requirements and design, both inception — and waits. requirements has finished and its documents are written; design is held at the gate; construction hasn't started. The operator approves, requests a same-stage revision, or aborts before any downstream work begins.
Without --hitl the same topology runs autonomously end to end. Headless? --review-file writes cadora-review-request.json and polls for a decision file — any tool or human can drop it — and fails closed on --review-timeout.
There are two method packs. aidlc is the default — the v1 rules, vendored and stable, with no network fetch. aidlc-v2 is a pinned upstream distribution and is honestly labelled EXPERIMENTAL: the default ref is a tag whose commit hash Cadora verifies after fetching, so a moved tag fails the install rather than silently shipping different code.
Why strip the pins? Upstream's shipped .claude/settings.json silently re-points every session at metered AWS Bedrock on opus[1m] at effortLevel: xhigh — the most expensive configuration. Cadora's guarded installer removes those pins so your run stays on the funding source you picked, and prints exactly what it took out. --keep-provider-pins opts back in.
Every strip is written to .cadora-aidlc-v2.json — the install record is itself part of the evidence, so "what did Cadora change about upstream?" has a file-level answer.
Because the method writes its own state, Cadora can read it back. cadora aidlc-audit is read-only: it parses the v2 per-intent state file (aidlc-state.md, six-state checkboxes) and the append-only audit trail (audit/<host>-<clone>.md shards, a 68-event taxonomy with ISO timestamps) into one summary — the phase you're in, the stage roll-up, and the count of everything that happened.
aidlc-audit reads the aidlc-v2 workspace's own state files — it is the read-back for the experimental pack. For a v1 run, the same evidence (nodes, gates, per-node cost, review decisions) is in the run archive that cadora report packs.
The frames above reproduce real cadora aidlc-init / run --hitl / aidlc-audit output and the actual node ids from examples/aidlc-phased.topology.yaml, rendered as editable HTML rather than raster captures — lift them into Figma or iterate directly. This is the method layer of Cadora: AI-DLC is installed as rules-plus-inputs, driven as a phased DAG, checkpointed by deterministic and human gates, and — for the experimental v2 pack — read back phase by phase from its own audit trail.