Changelog

[0.2.4] - 2026-10-04 — SOW intake pilot

Candidate status: bounded intake corrections included; synthetic acceptance pending. This entry does not indicate npm publication or a completed four-app behavioral test.

• Ingest an SOW into stable question IDs, review source coverage, and track evidence-linked progress and amendments in a local Git repository.
• Carry SOW IDs through scoping, module briefs, workplans, and readouts.
• Bundle a portable skills entry point and native marketplace metadata. Live acceptance remains separate for each of the four target apps.
• Align selected-root isolation, staged handoffs/read-only resume, accepted amendments and gate invalidation, descriptive-question coverage, exact-file checkpoints, native update routing and runtime permission/completion rules. Source/package checks do not establish native behavior.

Notable changes to GDD, loosely following Keep a Changelog. Version corresponds to "version" in .claude-plugin/plugin.json.

[Unreleased]

[0.2.3] - 2026-10-04

Changed

• Discontinued Antigravity support. Retired installer flags fail clearly and leave existing installations untouched. Claude Code and Codex remain the CLI targets; Cowork and ChatGPT Work require separate native acceptance.
• Corrected Claude plugin installation scopes and removed unverified Cowork form-rendering claims.

Fixed

• Included the documented cloud bootstrap script and Claude SessionStart entry points in repository distribution.

[0.2.2] - 2026-08-20

Fixed

• Concurrent module dispatch could silently lose ledger rows (GDD-BUG-22). The workplan heuristics describe modules running "in parallel", meaning the work overlaps in the calendar — but this was read as authorisation to dispatch two module commands concurrently. LEDGER.md and SOURCES.md are shared append-only tables with no locking, no ownership window and no per-module staging, so concurrent appends interleave rows and collide on ids while leaving the file syntactically valid. The loss is silent: it surfaces only as a finding-id gap someone happens to notice, and the ledger is the spine that triangulation, the red team and the storyline all read. The planner heuristics now state the distinction, and every module's completion gate repeats it at the point of the write — research may overlap, ledger promotion is serial.

• Missing delegation degraded research invisibly (GDD-BUG-23). When no subagent tool is exposed to a module agent, the fresh-context research split collapses to a single inline pass at materially lower recall — and nothing fails, because every postcondition still closes on a well-formed ledger. The four delegating workflows now require a capability check before research begins; where delegation is unavailable the module must record it (blocked, or a delegation: unavailable note) and mark its recall degraded, rather than proceeding silently.

Changed

• Sources template v4 adds a Published column (GDD-BUG-25). The registry previously recorded only Accessed — when a source was retrieved. Any cutoff or vintage rule evaluated against that date is meaningless, because retrieval necessarily post-dates the analysis window. Published is now required on every row, with UNDATED as an explicit and countable category: a page whose publication date cannot be established cannot be shown to pre-date a cutoff, so it cannot carry a date-sensitive claim on its own.

[0.2.1] - 2026-07-22

Fixed

• Vestigial name: command frontmatter: removed the name: key from all 17 commands/*.md templates. Nothing downstream read it — the Claude Code (nested) layout copies command files through verbatim, so the field only landed in the installed copy and tripped Claude Code's autocomplete; the skills layout (Codex/Antigravity) already synthesizes its own gdd-<name> and ignored the source field. A pure subtraction, no logic change. /gdd:update users should re-pull. tests/check-references.mjs now fails if any shipped command reintroduces a name: key.

[0.2.0] - 2026-07-20

Feature-and-fix release. It completes the five-module CDD set (three new commands), adds the /gdd:update maintenance path plus a batch of installer hardening, and folds in the WS1-WS5 fixes from the v0.1.0 autonomous end-to-end test. Feature requests from that test are deferred. Ids like GDD-BUG-N and WS-N below are internal test-log references, kept as a traceable paper trail.

Changed

• Installer: gdd --version/-v; files whitelist narrowed so the npm tarball no longer ships publishing tooling; uninstall is contained to the config dir; install provenance recorded in the manifest; clearer messages on corrupt/partial installs and first-install collisions.

Fixed

• Engagement isolation (GDD-BUG-2, P1): the engagement root is <CWD>/.diligence — at exactly that path or not at all. New references/engagement-root.md contract, a preamble in all 16 workflows, an ISOLATION rule in all 10 agents, the root passed as an absolute path in every spawn prompt, provenance stamps (engagement_root + written_by) on STATE.md and the three report templates, and a D5 "foreign artifact" FAIL on stamp mismatch. Fleet- verified: all 16 commands refuse/route in an empty folder beside a full sibling engagement, zero sibling reads, sibling byte-identical.
• Subagent dispatch (GDD-BUG-1, P1): dispatch/return contracts in the agent specs (synchronous spawns only, postconditions verified on disk before returning) and completion gates in the five module workflows + triangulate + red-team — a command is not complete until its artifacts exist on disk; no "report back later" turns.
• Machine state integrity (GDD-BUG-3/4/6/11, P2/P3): hypothesis-tree registers every briefed module in state.json.modules; module workflows drive the pending→in-progress→done/blocked lifecycle; STATE.md's status column is constrained to the schema enum; gdd-scoper maps every interview boundary rule and test value into the machine lock; fx conversions must be registered in taxonomy_lock.currency.fx; D1/D3/D5 enforce all of the above.
• Workflow logic (GDD-BUG-5/8/9/14, P2/P3): D8 is N-A (not FAIL) before the first storyline, killing the built-in waiver deadlock; scope-deal runs a KQ feasibility check against the recorded constraints (unanswerable KQs are reframed or DEGRADED, never signed clean) with a KQ × instrument table in ENGAGEMENT.md; the storyliner may not mint unledgered numbers (needs-promotion round-trip through gdd-librarian) and must enumerate every CONTESTED row; artifacts are client-facing (no interview residue; amendment impact claims need a cited basis).
• Hygiene (GDD-BUG-7/10/12/13/15/16, P2/P3): the installer no longer silently clobbers locally modified files (sha256 manifest baseline, warn + skip, --force to overwrite) and the manifest lists itself; scope-deal's edit loop spec matches behavior (orchestrator applies section edits, scoper re-spawned for structural changes); codename discipline records its scope with an all-artifacts IDENTITY.md mode; one tier per SOURCES.md row with central banks pinned to tier 2; locators required (UNVERIFIED (snippet) for snippet-only access, counted by D4); rejected sources cited as (audit: S#) only; ledger table integrity; start's summary cap raised honestly to ≤8 sentences.

Added

• /gdd:update — update an install to the latest published version. Checks the installed version against npm, shows the changelog delta, confirms, then reinstalls in place. Locally modified files are backed up to gdd-patches/ and merged back; files a new version drops are pruned.

• Dogfood fold-backs from the kestrel-sound five-module run: D4 platform-metadata and citation-chain clauses, D6 staleness rule (waiver, not reclassification), D8 stale-trace fragility note, and a grown-ledger re-run rule in the check registry; unreachable screen-verdict state and citation-record promotion rule in risk-screens; enforced-against-whom test, proxy reference-class rule, and archived-pricing fallback in moat-evidence; platform-pool independence check in customer-evidence; filing-mirror recipe in research-recipes; market-risk-register anchor in the module-findings template; citation-chain and structural-defect-propagation red-team patterns; parallel-researcher source-registration protocol (researcher + analyst specs) and the report-filename write-guardrail fallback note.

• Three new module commands completing CDD's five-module set: /gdd:probe-customers (customer evidence up the evidence ladders — retention, satisfaction, KPCs), /gdd:assess-moat (moat claims tested mechanism by mechanism, with kill-tests run at build time), and /gdd:scan-risks (the standing risk screens — trip conditions, bounded deep dives, evidence-of-search for the rest). All three spawn gdd-analyst with gdd-researcher delegation, same pattern as the existing modules.

• Four new method references backing the modules above: customer-evidence.md (evidence ladders, KPC elicitation, review-mining discipline), moat-evidence.md (mechanism taxonomy, the one-release-cycle test, the circularity rule), risk-screens.md (the six standing screens, trip conditions, the evidence-of-search standard), and research-recipes.md (question-shape → recipe → tier index).

• Deal-objective quad in /gdd:scope-deal's Batch B interview: strategic intent, key concerns, value-creation levers, and deal breakers, captured verbatim into ENGAGEMENT.md; deal breakers seed /gdd:scan-risks's screens and the red team's mandatory attack list.

• gdd-researcher now matches each research question to an instrument in research-recipes.md before falling back to generic search, recording the recipe used in SOURCES.md's reliability notes.

• Red-team attack-pattern library extended with risk-screen and deal-breaker patterns: re-run an untripped screen's cheap test with hostile search terms; every deal breaker in the deal objective gets an explicit attack attempt.

[0.1.3] - 2026-07-09

Added

• CLI install support for Codex and Antigravity CLI (replaces the disabled gemini catalog stub — GDD targets Antigravity CLI as its Gemini-family runtime). Both runtimes previously had "enabled": false entries in runtime-catalog.json with no working converter.
  • Commands install as SKILL.md directories (skills/gdd-<name>/) for both runtimes, with each command's <execution_context>@...</execution_context> include resolved and inlined — neither runtime auto-inlines ${CLAUDE_PLUGIN_ROOT} references the way Claude Code does, so a straight copy would have installed a dangling reference.
  • Agents install as standalone TOML (agents/gdd-<name>.toml, developer_instructions inline) for Codex, matching its native custom-agent format, and as flat markdown with Claude tool names mapped to the Gemini/Antigravity vocabulary (Read→read_file, etc.) for Antigravity.
  • /gdd: prefixes in prose are rewritten to each runtime's actual invocation syntax ($gdd- / /gdd-) so cross-references between commands and agents read correctly post-install.
  • tests/smoke-install-codex.mjs and tests/smoke-install-antigravity.mjs added, wired into npm test.

[0.1.2] - 2026-07-09

Fixed

• AskUserQuestion interview steps that ran inside spawned subagents (gdd-scoper, gdd-storyliner) never rendered as native forms to the user — a subagent's interactive tool calls are headless, the form channel belongs only to the top-level command turn. Matches the pattern GSD already uses: orchestrators ask, agents write.
  • /gdd:scope-deal now runs the four-batch interview directly in the command; gdd-scoper is spawned afterward as a headless writer with the collected answers (tool grant narrowed accordingly).
  • /gdd:storyline now spawns gdd-storyliner twice: once to draft the governing thought, which the orchestrator confirms with the user directly, then again to build the full storyline from the approved thought (tool grant narrowed accordingly).
  • gdd-planner's AskUserQuestion grant was dead (no described use in either hypothesis-tree or workplan) and is removed; the branch/tree review in those workflows was already plain orchestrator conversation, not a form.

[0.1.1] - 2026-07-08

Fixed

• plugin.json now declares commands explicitly instead of relying on implicit discovery, and the five intake commands pair argument-hint with a named arguments: field — working theory for Cowork's fill-in-the-blank form rendering not applying to GDD's commands. Unconfirmed live; verify in Cowork.

[0.1.0] - 2026-07-08

Added

• Initial release: 13 commands, 10 agents, gdd-core (workflows, templates, references), installer, tests.