--- name: sdlc-align description: Periodic cumulative-drift audit — look at the last stretch of shipped goals as a whole and ask whether the work still matches the north-star's stated bets, or whether the strategy has been quietly rewritten by accumulation. Use when the user runs /sdlc-align, when /sdlc-status reports an alignment check is due, or after a north-star revision. allowed-tools: Bash(python3 *), Bash(bash *), Bash(git *), Bash(gh issue *) --- # sdlc-align The **window** check. Every other alignment gate in LoopSmith looks at one unit of work: `sdlc-plan-review` §4 holds a single plan to the north-star, and `sdlc-retro` asks what a single goal taught. Neither can see the shape of *twenty* goals, and that is where strategy actually drifts — no single decision violates the direction, but the trajectory has moved. **Read-only.** It produces a report and lets the user pick the reconciliation direction. It never edits the north-star, the goals, or any code. **Gate — a no-op without a north-star.** If `.sdlc/context/north-star.md` is absent, there is no stated direction to drift from: say so in one line and stop. (Drop-in projects skip this entirely, the same way `sdlc-plan-review` §4 does.) If it exists but states no priorities or bets, say that instead — the fix is `/sdlc-vision`, not a drift report. ## The window Default to everything since the previous report in `.sdlc/knowledge/align/`, or the last ~10 completed goals if there is none. State the window explicitly in the report; a finding is only as good as the span it was drawn from. Gather what actually shipped in it: - **Goals** — `.sdlc/goals/*.md` with `status: done` (local), or closed `sdlc:goal` issues (github). Include the parked ones too — `status: parked` locally, `sdlc:parked` on GitHub. Work that was started and abandoned still spent effort in a direction, and a *cluster* of parks on one theme is itself a drift signal: it usually means the backlog keeps reaching for something the strategy hasn't made room for. - **The work itself** — `git log` over the window: which areas of the tree absorbed the commits. - **The journey** — `.sdlc/journey/` notes or issue timelines, for the decisions behind the diffs. - **The stated direction** — the north-star's priorities, bets, and **non-goals**. **Ground it on evidence, not recall.** Before judging, run the read-only collector for a deterministic, secret-safe factual pack over the window — so the lenses reason from measured facts, not a re-derived impression of the git log: ```bash bash "${CLAUDE_SKILL_DIR}/scripts/alignment-collect.sh" --since-days ``` It emits `{"schema":"alignment-collect/v1", window, commits[], degraded[], dimensions:{d1..d7}}`: per-commit source/test/doc classification (d1/d3/d4), plan-adherence and churn hotspots, whether the project even documents how it verifies (d2, from `verify.command`), review-artifact presence (d5), decision-registry changes (d7), and **location-only** hard-stop flags — secrets/destructive-SQL/auth/contract touches, each `{commit,file,line,pattern_id}`, never the matched text (d6). It **renders no verdict** — that is your job below. Honour its `degraded[]` codes (`no_git`, `no_recognized_source`, `no_test_command`): a thin window is a thin finding, not a manufactured one. ## Lens A — cumulative direction Cluster the window's goals into themes and name the dominant two or three. Then compare against the north-star's stated priorities: - Does the dominant theme correspond to a stated bet? Good — say so and move on. - Is a stated priority getting **no** work at all across the whole window? That is drift by omission, and it is the quiet kind: nothing looks wrong at any point, and the priority simply never happens. - Is the dominant theme absent from the stated bets? Either a new bet emerged and the north-star has not caught up, or effort is misallocated. Both need the user; you don't get to decide which. Trace two or three concrete goals per theme. A theme you cannot name goals for is a hunch, not a finding — drop it. ## Lens B — implicit rewrite The higher-value lens. When a *cluster* of goals commits real effort to a direction the north-star never states, the strategy is being rewritten in the backlog rather than in the document — and nobody ever decided to do that. A single cluster is worth flagging. The tell is effort without a mandate: several goals, one coherent direction, no line in the north-star that asks for it. Distinguish it from ordinary maintenance — bug fixes and upkeep are not bets and do not constitute a rewrite. The recommendation is almost always the same: **make the implicit explicit.** Either the north-star gains the bet it has evidently been operating under, or the work stops. Leaving it unstated is the one option that guarantees the drift continues. ## What this check is not Do not re-do the per-unit gates. A single plan contradicting a rule is `sdlc-plan-review`'s FIX-FIRST, and one goal's lessons are `sdlc-retro`'s. If a finding here would have been caught by either, it belongs there — note that it slipped through (that is useful signal about the gates) and don't count it as cumulative drift. ## Output Exactly one shape — no neutral middle. Write it to `.sdlc/knowledge/align/.md` and include the `goals_reviewed:` count, which is how `/sdlc-status` knows when the next check is due. **The report is local in both backlog modes, and files nothing.** Drift is a question about the project's direction, not about any one issue, so there is no card to move and no issue to comment. If a finding deserves to become work, that is the user's call and a normal goal — say so in the report and let them file it. (Same restraint as `/sdlc-radar`: a scout that writes to a shared backlog unasked stops being useful the first time it's wrong.) **Drift found:** ```markdown # Alignment — : drift window: .. · goals_reviewed: ## TL;DR <2-3 sentences: the drift, its severity, the recommended direction.> ## DRIFT-1 — - **Lens**: A (cumulative direction) | B (implicit rewrite) - **Evidence**: <the goals + commits, named> - **North-star position**: <the quoted line, or "silent on this"> - **Divergence**: <what the work says vs what the document says> - **Reconciliation (user picks)**: - **A. Update the north-star** — <the specific line to add or change> - **B. Course-correct** — <what future goals should do instead> ## Reviewed, no drift <N goals across themes: <theme> (<count>), … — each tied to a stated bet.> ``` **Clean:** ```markdown # Alignment — <YYYY-MM-DD>: no drift window: <start>..<end> · goals_reviewed: <N> Dominant themes: <theme> → <the bet it serves>; … Stated priorities with no work this window: <none | the list> ``` Record the verdict too (optional, `telemetry.enabled`) — clean → `pass`, drift found → `warn`; align never blocks a merge (it's read-only/advisory, see above), so `block` is never used here. There is no single goal for a window check, so use the sentinel `(alignment)`, the same convention `(discovery-scan)` and `(decision-gate)` already use elsewhere in this stream: `python3 "${CLAUDE_SKILL_DIR}/../sdlc-loop/scripts/loop.py" emit .sdlc "(alignment)" gate --gate alignment --verdict pass|warn` ## Constraints - **Cap findings at four.** More than that and you are logging tactical defects, not strategic drift. Rank by effort committed to the divergence, not by how tidy the finding is. - **No fabrication.** Every finding names real goals and real commits. If you can't name them, it isn't a finding. - **Don't manufacture drift.** A clean window is the expected outcome most of the time. A thin "no drift" report is a real result — say it and stop. - **Never edit the north-star.** Propose both directions; the user chooses. This is the same propose-and-park contract `sdlc-retro` uses for standing docs. - **Short window, thin findings.** If the window holds only a goal or two, say the span was too short to read a trajectory and stop, rather than reading a trend into noise.