Skip to content

Commit 821681d

Browse files
committed
docs(state): record v3.4.5 (gsd-core #936 plan-phase convergence fix)
1 parent c8147b4 commit 821681d

1 file changed

Lines changed: 1 addition & 1 deletion

File tree

.planning/STATE.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -28,7 +28,7 @@ See: .planning/PROJECT.md (updated 2026-04-11)
2828
Milestone: v1.2 shipped (2026-04-24) — all 9 requirements satisfied, tagged, archived.
2929
Phase: none active. v1.2 phase artifacts moved to `.planning/milestones/v1.2-phases/`.
3030
Next action: `/gsd:new-milestone` to scope v1.3.
31-
Last activity: 2026-06-09, Committed v3.4.4 (local, awaiting release decision): added Claude Fable 5 (claude-fable-5) as a new top model tier ABOVE opus (1M ctx, ~2x Opus pricing $10/$50). New `fable` tier in model-catalog.json for every runtime (Anthropic-compat -> real model; non-Claude runtimes alias to their top model; Group-B null). Quality profile now uses fable for the 9 heaviest (routingTier:heavy) agents; standard/light + balanced/budget unchanged (cost only rises for quality users on heavy agents). Threaded fable through core.cjs VALID_TIERS + config-schema override regex + settings-advanced docs/table + sdk/src tier types (SDK rebuilt; dist reads catalog at runtime). tests/fable-tier.test.cjs 26 checks; suite 14/14; verified e2e (quality -> gsd-planner=claude-fable-5, executor=claude-opus-4-8). Commits 4a4b52a + 3fb0993. Prior: shipped v3.4.3 (local, awaiting release decision): adapted the highest-value deferred TIER-2 fix #730 (multi-milestone Phase Details scope) onto our older extractCurrentMilestone base -- a bespoke port (refactored the fence-aware end-scan into a computeSectionEnd closure + pull in the current milestone's versioned (Phase Details) section, boundary-aware to the version token). Two/three-milestone regression test added (confirmed it failed pre-fix); suite 12/12; no-op on single-milestone roadmaps incl. this project's own. Commits 82aa94d + e196a29. ALSO verified #768: Claude Code plugins CANNOT seed permissions (no plugin.json permissions field; bundled settings.json ignores them; fails --strict) -> #768 can only be a documented settings.json snippet, stays deferred pending that decision. See [[reference_cc_plugin_no_permissions]]. Prior: shipped v3.4.2 (hotfix follow-on): integrated both real fixes from gsd-core v1.4.2 -- #925 context-monitor echoes the invoking hook event (REQUIRED companion to 3.4.1's #770 SubagentStop wiring; without it CC rejects the monitor output on Stop/SubagentStop/PreCompact) + #921/#922 plan-phase attempt-based Agent gate (refines #913). New regression test (11/11 suite). Commits 2f9883b + d3972c1. Checking the next upstream version caught that 3.4.1's SubagentStop was latently broken without #925. Prior: shipped v3.4.1 (version scheme re-based: upstream renamed to @opengsd/gsd-core + restarted at 1.x, so plugin now tracks the gsd-core line with a +2 major offset, 1.4.1 -> 3.4.1; +2 keeps it above the prior 2.45.x lineage). Bundles the ad-hoc auto-memory feature AND 7 gsd-core v1.4.x cherry-picks: TIER-1 (#905 STATE preservation, #904 branch normalization, #771 14 agent colors, #25 verifier Step 7b, #913 plan-phase role-collapse) + TIER-2 (#892 checklist-roadmap validate, #770 SubagentStop hook). DEFERRED with reasons: #730 multi-milestone extractCurrentMilestone (97-line rewrite on patched divergent core.cjs, wants own task+fixture), #768 CC permissions prepopulation (no confirmed plugin-manifest mechanism). 3.4.x = selective cherry-pick of the gsd-core 1.4.x line, NOT 1.4.1 parity (multi-runtime/infra excluded by design). Commits 88d1288/8429b8c (auto-memory) + 32b1b31/8221979 (cherry-picks) + fe7c7f2 (release/re-version). 10/10 tests, jargon baseline regen'd (+1 legit STATE.md mention). README Versioning rewritten. NOTE the ad-hoc auto-capture context below:
31+
Last activity: 2026-06-09, Committed v3.4.5 (local, awaiting release): upstream check found gsd-core v1.4.3 (new stable). Cherry-picked #936 (the plan-phase fix the user asked about): our plan-review-convergence.md wrapped gsd-plan-phase in Agent() at BOTH sites (initial + replan); since gsd-plan-phase is a spawner and a depth-1 Agent on CC has no Agent tool, the replan loop silently produced no revised plan when reviewers flagged HIGHs. Fixed both to inline Skill() at depth 0; gsd-review stays Agent-wrapped (Bash leaf). New structural-guard test (5 checks). Other v1.4.3 fixes skipped: #934 (orphan gsd-pristine/get-shit-done snapshots from upstream rename, N/A to flat layout), #935 (changeset-CLI release infra). Commits a5bef1a + c8147b4. We are now current with gsd-core v1.4.3. Prior: shipped v3.4.4 (Claude Fable 5 top tier) (local, awaiting release decision): added Claude Fable 5 (claude-fable-5) as a new top model tier ABOVE opus (1M ctx, ~2x Opus pricing $10/$50). New `fable` tier in model-catalog.json for every runtime (Anthropic-compat -> real model; non-Claude runtimes alias to their top model; Group-B null). Quality profile now uses fable for the 9 heaviest (routingTier:heavy) agents; standard/light + balanced/budget unchanged (cost only rises for quality users on heavy agents). Threaded fable through core.cjs VALID_TIERS + config-schema override regex + settings-advanced docs/table + sdk/src tier types (SDK rebuilt; dist reads catalog at runtime). tests/fable-tier.test.cjs 26 checks; suite 14/14; verified e2e (quality -> gsd-planner=claude-fable-5, executor=claude-opus-4-8). Commits 4a4b52a + 3fb0993. Prior: shipped v3.4.3 (local, awaiting release decision): adapted the highest-value deferred TIER-2 fix #730 (multi-milestone Phase Details scope) onto our older extractCurrentMilestone base -- a bespoke port (refactored the fence-aware end-scan into a computeSectionEnd closure + pull in the current milestone's versioned (Phase Details) section, boundary-aware to the version token). Two/three-milestone regression test added (confirmed it failed pre-fix); suite 12/12; no-op on single-milestone roadmaps incl. this project's own. Commits 82aa94d + e196a29. ALSO verified #768: Claude Code plugins CANNOT seed permissions (no plugin.json permissions field; bundled settings.json ignores them; fails --strict) -> #768 can only be a documented settings.json snippet, stays deferred pending that decision. See [[reference_cc_plugin_no_permissions]]. Prior: shipped v3.4.2 (hotfix follow-on): integrated both real fixes from gsd-core v1.4.2 -- #925 context-monitor echoes the invoking hook event (REQUIRED companion to 3.4.1's #770 SubagentStop wiring; without it CC rejects the monitor output on Stop/SubagentStop/PreCompact) + #921/#922 plan-phase attempt-based Agent gate (refines #913). New regression test (11/11 suite). Commits 2f9883b + d3972c1. Checking the next upstream version caught that 3.4.1's SubagentStop was latently broken without #925. Prior: shipped v3.4.1 (version scheme re-based: upstream renamed to @opengsd/gsd-core + restarted at 1.x, so plugin now tracks the gsd-core line with a +2 major offset, 1.4.1 -> 3.4.1; +2 keeps it above the prior 2.45.x lineage). Bundles the ad-hoc auto-memory feature AND 7 gsd-core v1.4.x cherry-picks: TIER-1 (#905 STATE preservation, #904 branch normalization, #771 14 agent colors, #25 verifier Step 7b, #913 plan-phase role-collapse) + TIER-2 (#892 checklist-roadmap validate, #770 SubagentStop hook). DEFERRED with reasons: #730 multi-milestone extractCurrentMilestone (97-line rewrite on patched divergent core.cjs, wants own task+fixture), #768 CC permissions prepopulation (no confirmed plugin-manifest mechanism). 3.4.x = selective cherry-pick of the gsd-core 1.4.x line, NOT 1.4.1 parity (multi-runtime/infra excluded by design). Commits 88d1288/8429b8c (auto-memory) + 32b1b31/8221979 (cherry-picks) + fe7c7f2 (release/re-version). 10/10 tests, jargon baseline regen'd (+1 legit STATE.md mention). README Versioning rewritten. NOTE the ad-hoc auto-capture context below:
3232
Earlier this session: ad-hoc durable-decision auto-capture. /gsd:quick|debug|fast now save durable decisions to auto-memory at close-out (gated on `workflow.auto_memory_capture`, default on) so context persists between milestones without manual "remember". Discovered + fixed that gsd-plugin NEVER actually auto-captured before: the phase-memory writer was orphaned (never wired into a workflow) AND doubly broken (called undefined `memory.writePhaseMemory`; read phase from args[0]=command name). New `gsd-tools write-decision-memory` + memory.cjs helpers + references/auto-memory-capture.md + 10/10 tests. Commits 88d1288 (feat) + 8429b8c (release). NOTE: a cherry-pick survey (260609-kb0, 5 TIER-1 wins from v1.3.1->v1.4.1) is PARKED unapplied in its task dir. Prior: 2026-06-08 quick task 260608-vk9 retargeted the scheduled upstream-release notification (`bin/check-gsd-release.sh`) from the locked gsd-build/get-shit-done to open-gsd/gsd-core (the repo we now track); synced live cron copy + reseeded version baseline to 1.4.0 (no spurious email). Prior: 260607-w0s v2.45.10 selective cherry-pick of 5 TIER-1 bug fixes from upstream v1.3.1 (writer-agent Edit grants #582/#571, websearch fetch timeout #387/#308, decision-gate jq #283/#275, AskUserQuestion 4-option cap #243/#17). Chose cherry-pick over full sync: upstream restructured to build-at-publish + retired SDK + renamed to @opengsd/gsd-core, making a full v1.3.1 sync a ~350-file migration. gsd-core native plugin manifest (#766) is next-only; our manifest stays authoritative. 9/9 tests green, commits 7116f83/68a4851/dd24beb/1ab88fb. Prior: 2026-05-29 v2.45.2 landed @dboeckenhoff PR #12; v2.45.1 Opus 4.8 bump; v2.45.0 state-handler fix #9. 2026-05-22 verified upstream change set; case-insensitive patch retires on next sync.
3333

3434
```

0 commit comments

Comments
 (0)