2026-08-15 — itd-84 (intent decomposition) ADOPTED at the MVP rung, per its own staged promotion ladder: the `decompose-before-filing` principle is filed, the hand-run four-piece protocol lives in the /abcd:intent surface page (capture and the planning interview both run it, loud-staged as not-yet-automated), a DECOMPOSITION rules domain injects the rule at recall time, and every hand-run is graded into `.abcd/development/research/notes/2026-08-15-decomposition-calibration.md` (the 2026-07-13 auto-merge SPLIT case seeds the corpus retrospectively). The deterministic Go pre-pass is admitted to the predictable-development plan's queue (front-door design open: `decompose` subverb vs a check in `intent ready`; collision with iss-210 noted). The capture-time agent rung stays DEFERRED until ~50 graded captures exist (itd-81 calibration); building or running it earlier is a recorded STOP. Rejected: overriding the bundled INTENTS rules domain to carry the new rule (a per-field override replaces the whole array and would withhold future bundled upgrades — iss-174's exact failure mode), hence the additive custom domain instead.

Frontmatter

.abcd/work/decisions/dcn-2608150000000002-itd-84-intent-decomposition-adopted-at.md
open on GitHub ↗ · commit history ↗

mentions4

• iss-174mergeDomain replaces a domain's recall and rules arrays wholesale, so a repo that overrides either array silently withholds every later security upgrade to the bundled defaults: internal/core/rules/rules.go copies the override array over the base one instead of unioning, with no schema bump and no notice, so a repo that pinned PII recall or PII rules before iss-156 keeps the old set and never sees the new network keywords or the never-commit-network-identifiers rule line. Per-field merge is the documented and wanted behaviour for state, but for security-bearing arrays the quiet outcome is a stale ruleset that looks current. Options to weigh: union rather than replace for recall/aliases (additive, no loss), a distinct replace-vs-extend marker in the override, or at minimum a loud diagnostic in abcd rules naming which bundled entries an override is withholding. First activated by the iss-156 PII upgrade, which is why it is captured now. Prior disposition to weigh when triaging: iss-66 (resolved, rules-loader trust boundary) document-accepted the adjacent risk that an override can weaken a default guardrail domain (its P15); this entry is the distinct case of an override silently withholding later upgrades rather than deliberately silencing a domain.
• iss-210A lone token that a user plausibly expects to be a subverb, but which is not one, is swallowed as free text and WRITES A RECORD on both mutating verb families — 'abcd intent status' files drafts/itd-N-status.md and 'abcd capture status' files issues/open/iss-N-status.md, each exit 0. Verified 2026-08-11 by building HEAD into a scratch repo and running both. This is a residual hole in the unrecognized-input-never-writes rule (iss-29) that iss-104's fix does not cover: suspectedTypoedSubcommand (internal/surface/cli/cli.go:2000) fires only on a NEAR-MISS of a real subverb, bestDist initialised to 3 so only edit distances 1 and 2 refuse. It is deliberately high-precision so a genuine multi-word prose title whose first word resembles a verb still files — the controls confirm that half works: 'capture resovle iss-1 note' and 'intent paln itd-1' both exit 2 with a did-you-mean and write nothing. But 'status' is edit distance >=4 from every real subverb in either family (intent: new/plan/ready/link/review/ingest; capture: list/resolve/wontfix), so the guard never engages and the word becomes the record title. The class is any plausible-but-nonexistent subverb far from every real one; 'abcd intent list' happens to be caught only by the accident that levenshtein(list,link)=2. NOT an intent-vs-capture asymmetry: the two behave identically, so a fix must land on the shared guard, not on the intent path alone. Fix direction: edit distance is the wrong discriminator for this class. A LONE single token with no other args is never a useful issue or intent title, so treat it as an unrecognized subcommand regardless of distance — refuse, exit 2, write nothing, name the real subverbs, and say how to force prose (quote it, or add words) so the terse-free-text escape hatch stays open. Detector: extend intent_surface_test.go and the capture surface test with a lone-token case per mutating verb family asserting exit 2 and zero records written, watched fail before the change; the existing near-miss cases must keep passing, and a multi-word prose title must still file. Impact on resolve is 'fix' at best and arguably 'breaking' — an unattended script or loop that shells 'abcd capture <one-word>' expecting a file would start failing — so it is NOT additive-class. Standing risk this closes: any autonomous loop that shells a bare word into either verb silently grows the ledger with junk records that a human must later triage.
• itd-81No Judge Ships Unmeasured, And A Clean Diff Is A Result
• itd-84An Intent Is Decomposed Before It Is Filed, And A Reversal Is Only Ever Flagged