paqad-ai 1.87.0 → 1.89.0

package/CHANGELOG.md

@@ -1,5 +1,48 @@
 # paqad-ai
 
+## 1.89.0
+
+### Minor Changes
+
+- e6c3737: feat(#567): run each feature-development stage in a fresh subagent, with the evidence bundle as shared memory
+
+  Stage isolation is core-engine behavior for Claude Code and Codex — there is no config knob.
+  On the graduated and full lanes, each mandatory feature-development stage runs in its own host
+  subagent seeded with the paqad contract, the previous pillar artifact, and its stage
+  instructions; the main chat stays a lean orchestrator that routes, narrates, asks the
+  decision-pause questions, and speaks the receipt. Nothing is compacted because nothing
+  accumulates — later stages no longer re-carry earlier stages' history, so input tokens drop on
+  every graduated and full-lane change. The fast lane and non-subagent hosts run the stages in a
+  single context, unchanged.
+
+  Install and update always render six stage agents per host at user scope (never into a project
+  directory), a record-only `SubagentStop` hook appends a per-bundle `context-efficiency.jsonl`
+  stream, and the in-flight fork guard refuses to mint a third bundle on one branch. Every stage
+  records under the orchestrator's session id, so the bundle keeps one identity — and each stage
+  subagent proves its own cold framework load through an `agent_id`-keyed entry sentinel (a
+  subagent shares its parent's `session_id`, so the sentinel keys on the distinct `agent_id`).
+  paqad's Node code still calls no LLM; the host's subagent runs the stage.
+
+## 1.88.0
+
+### Minor Changes
+
+- 1952ac7: feat(codex-cli): wire paqad's full hook set so feature-development stages are enforced on Codex exactly as on Claude Code (#566)
+
+  Codex CLI moves from the `live-completion-only` tier to `live-pre-and-completion`. One
+  native-hook renderer now drives every hook-capable host from a single ordered chain, so
+  Codex gets the entry-gate block, the plan-and-spec-before-code block, the decision-pause
+  block, the rules-loaded block, the capability kernel, and the spoken completion verdict —
+  the same guarantees as Claude Code, wired into `.codex/hooks.json` across all four
+  lifecycle events (SessionStart, UserPromptSubmit, PreToolUse `^apply_patch$`, Stop).
+
+  The same hook scripts run on both hosts; the host is one argv, never a fork. A single
+  `edit-targets` extractor parses the edited paths from a Codex `apply_patch` (or a Claude
+  `file_path`), so the gates and the stage writer are host-aware without duplicated parsing.
+  `paqad-ai doctor` gains a `Codex hooks wired` check, and onboarding with `codex-cli` prints
+  the one-time `/hooks` trust step. Gemini stays record-only and every advisory host is
+  unchanged; the Claude and Gemini generated configs are byte-identical to before.
+
 ## 1.87.0
 
 ### Minor Changes