# ADR-033: The spec loop reads the decision log before it writes, and an Accepted record outranks the request | | | | --- | --- | | **Status** | Accepted | | **Date** | 2026-08-06 | | **Author** | bodurkalukasz | | **Tags** | method, specs, decisions, lifecycle | ## Context Two `severe` validation findings, both still `fail` against the tree before this change. `SPEC-18`: neither `/spec-specify` nor `/spec-clarify` reads `docs/decision-records/`. Grepping both files for a decision-record path returns only the `PATCHED(...)` provenance comments naming the ADRs that produced them. What both *are* told to read before asking the user anything is `docs/discovery/` - and ADR-024 makes a dossier explicitly non-normative. So the loop's entry point consulted, by documented procedure, the one source that binds nothing, and never consulted the one that binds everything. A spec contradicting an Accepted record is written, passes the clarify gate, passes every guard, and is found - if ever - by a human reading the record out of band. The only backstop was `AGENTS.md`'s red-flag list, which said "Contradicting an Accepted ADR". Not BDR. The product-side stream is the one a feature request actually collides with. `SPEC-17`: the loop has two entry points and retirement awareness reached only one of them. `/spec-impact` checks the target spec's `Status` first and stops on `retired`, naming the record; `/spec-update` refuses to edit a retired spec; `discovery-digest` refuses to route new material into one. `/spec-specify` - the entry point the product sells to a PO, and the one `AGENTS.md` names for new behaviour - contained the word `retired` zero times, while instructing that an existing capability directory means "update the existing spec in place". Reproduced against the fixture: `shift-notifications` is `Status: retired` per BDR-004, whose own text says a future in-house reminder need is a new capability, specced fresh, not a reopening. A PO asking for exactly that walks through `/spec-specify` into an edit of the retired spec. The two findings are one defect seen from two sides. Retirement is only the case where the governing record happens to have left a flag on the spec file; the general case - a record that governs the request and left no flag anywhere the loop looks - has no detection at all. ## Options considered - **Leave it to review and the red-flag list.** Rejected. A red flag fires when an agent already noticed, which is the step that failed; and review happens after the spec is written, which is the expensive moment to discover the request was never viable. - **A mechanical guard.** Rejected as the primary answer. Contradiction is semantic: a guard can check that a *cited* record's `Status` is still current (`/spec-reconcile` already does exactly that) but cannot see a spec that silently re-decides something it never cites. A guard that appears to cover this and does not is worse than none. - **Read every record in full at the start of every run.** Rejected. The cost scales with the repo's record count and is mostly waste; an agent told to read sixty records reads sixty records badly. Bounded read instead: the index table, then in full only the rows whose subject overlaps the request. The index exists for this - the README's gist column is already written to be read instead of the records. - **Put the check only in `/spec-impact`, and route everything through it.** Rejected: it makes the PO path depend on a developer-facing skill firing first, which is the routing assumption that already failed. Both entry points carry it. - **Delegate to `specs/constitution.md`.** The constitution template already hard-stops on "the change contradicts an Accepted ADR", and `/spec-specify` already loads it. Rejected on two counts, both observed rather than argued: the template scopes its own check to `/spec-plan` and `/spec-update`, and specify loads it at step 4 - *after* step 2 has already decided which directory to write into, which is where the retirement check has to fire. In the pre-fix reproduction the constitution did contribute to a stop, but only because the agent went looking for the record on its own after reading it. (The template's hard stop said "ADR" while its own altitude list said "ADR / BDR"; that is corrected here too, same defect as the red-flag list.) ## Decision **The decision log is read before the draft, at both entry points.** `/spec-specify` reads it before drafting; `/spec-clarify` reads it before asking. The read is bounded by the index - the README table, then in full only the records whose subject overlaps. **That bound rests on `decision-records-check`**, which fails a record on disk that the index does not list; where that guard is not running, the directory is listed too. The dry run proved this is not theoretical: the agent grepped the fixture's CI, found the guard wired nowhere, listed the directory - and BDR-004, the record that governed the whole request, is on disk and absent from a BDR index that stops at 003. **A capability is matched by subject, not by slug.** The short name specify generates is its own, not the repo's: `shift-reminders` and `shift-notifications` are one capability under two names. `ls specs/` and match on what the capability is about, because a near-miss slug is how one capability acquires two rival specs - and, when the existing one is retired, how a retirement is routed around without anyone deciding to. **Precedence is stated, not left to judgement.** An **Accepted** record in scope outranks the dossier, outranks the draft, and outranks whatever the user says in the moment. What a record settles enters the spec as settled content citing it - never as a `[NEEDS DECISION]` marker and never as a question. `Proposed` / `Rejected` / `Superseded` bind nothing; the supersession link leads to the record that does. **A contradiction stops the run.** There are exactly two legitimate routes - change the request to fit, or supersede the record (`/adr-write`, `/bdr-write`). Writing the spec anyway is not one of them, including when the user asks for it: an answer that collides with an Accepted record is a supersession, not a clarification. **Retirement is checked at every entry point, including the PO's.** `/spec-specify` reads the matched spec's `Status` at the step that decides where to write, and `retired` stops it before drafting - the same answer `/spec-impact` already gives, at the door the PO actually uses. The stop is *surface and hand back*, not *refuse*: a genuinely new capability specced fresh is often the right answer and the retiring record frequently says so, but that call belongs to the user with the record in front of them. What the agent may not do is proceed alone, under the old slug or a new one. **R6 says the binding half out loud.** The validation suite cited R6 for "an accepted record binds future work in its scope until superseded" while R6's text covered only how a record is changed. The rule now carries both halves. ## Consequences - The collision is found at the cheapest moment - before a spec exists - instead of by a human who happens to know the record. - **Cost accepted:** every specify and clarify run pays an index read it did not pay before. A repo with no records pays nothing; a repo with many pays one table. - **Cost accepted:** this is prose an agent can skip, not a guard, and placement is the whole mitigation. The consequential checks sit inside the step that already does the thing they qualify - the queue-building step in clarify, the directory-resolution branch in specify. The intake block itself is unavoidably a preamble, which the dry run flagged as the weak point ("a less careful agent skims the numbered Outline and treats everything above it as preamble"), so the Outline opens by saying the intakes are steps and can end the run before step 1. If real runs still skip it, the answer is a citation duty something can check, not more prose. - **Cost accepted:** the retirement rule is now written at two entry points. Deliberate: it was written once, on the developer path, and that is precisely how `SPEC-17` happened. - `/spec-clarify` can now end a session by refusing an answer. That is new, and it is the point: the alternative was recording the answer and letting the record rot. ## Confirmation `SPEC-17` and `SPEC-18` are the cases; they are re-run against the tree and move from `fail` when this lands. Mechanically checkable parts: `grep -c -i retired` over `standard/.claude/skills/spec-specify/SKILL.md` is no longer 0, and both spec-specify and spec-clarify name `docs/decision-records/`. The semantic part - that an agent actually stops - is not greppable and was tested the only way it can be: paired dry runs of one PO request against the `test-greenfield-core` fixture, pre-fix and post-fix, each agent told to follow the repo's documented procedure literally and to report honestly which of its reads the procedure instructed and which were its own initiative. Pre-fix, the agent did stop, and said plainly that opening the governing record was its own move because "spec-specify's text never mentions decision records at all". Post-fix, reading the decision-record index was its first action and it cited the step that told it to. Both of this decision's late additions - the guard-dependency of the index read, and subject-over-slug matching - came from the post-fix run finding the fix's own holes, which is the argument for running it rather than reasoning about it. ## What this rules out A guard that claims to detect a spec contradicting a record it never cites. Semantic contradiction is not greppable, and shipping something that looks like that check would buy false confidence at the exact point this standard is selling trust. ## Revisit when A spec lands contradicting an Accepted record with both skills' procedure followed - meaning the bounded index read is too coarse to surface the overlapping record - or the stop fires so often on records that did not really govern that authors learn to wave it through. ## Related - [ADR-024](ADR-024-discovery-dossiers-beside-the-specs.md) - the dossier is never normative. This names what is. - [ADR-032](ADR-032-re-entry-is-core-tracker-sync-is-an-extension.md) - the same shape: a documented loop assumption that nothing in the loop performed. - R5 (decisions are recorded), R6 (an accepted record binds until superseded), R8 (behaviour is specified by capability).