docs: architect prompt upgrades — mandatory state-save compaction (arch-init) + PRFT at human decision points (role)

.claude/skills/arch-init/SKILL.md

@@ -77,11 +77,23 @@ plus the most recent dated section*, so a save must leave exactly that behind:
 
 1. **Rewrite the current-state / open-loops section in place** — overwrite it
    with where things actually stand now (current focus + open loops + how to
-   resume). Do not accumulate stale "current state" blocks.
+   resume). Do not accumulate stale "current state" blocks, and never leave two
+   sections with the same heading (a duplicated "How to resume" or "Open loops"
+   means you appended where you should have overwritten). **Delete resolved
+   loops entirely** — a closed item's record is the log entry, not a lingering
+   line in current state.
 2. **Append one short dated log entry** capturing what changed this stretch.
-3. **Keep it to one screen (compaction discipline).** The state file is a
-   summary, not a transcript. When you append, prune stale dated sections so
-   the file stays readable at a glance.
+3. **Compact — this is part of the save, not optional polish.** Every save
+   both adds AND removes: after writing, keep the recent dated log entries in
+   full and collapse older ones into a single one-line summary that names
+   where the detail lives (the closed PRs, issues, and reviews it covered).
+   Remember these state files are gitignored — there is no git history to
+   fall back on, so pruned prose is gone for good. Prune by replacing detail
+   with pointers to durable artifacts, never by deleting the only record of
+   something. The file is a summary a fresh session reads at a glance —
+   one screen is the right order of magnitude, judged by you. If a save has
+   grown it past easy readability, prune as part of that save rather than
+   leaving it for next time.
 
 **Content guardrails.** No secrets (tokens, keys, credentials). No transcript
 dumps or raw tool output. Include only: current focus, open loops, and the

.codex/skills/arch-init/SKILL.md

@@ -77,11 +77,23 @@ plus the most recent dated section*, so a save must leave exactly that behind:
 
 1. **Rewrite the current-state / open-loops section in place** — overwrite it
    with where things actually stand now (current focus + open loops + how to
-   resume). Do not accumulate stale "current state" blocks.
+   resume). Do not accumulate stale "current state" blocks, and never leave two
+   sections with the same heading (a duplicated "How to resume" or "Open loops"
+   means you appended where you should have overwritten). **Delete resolved
+   loops entirely** — a closed item's record is the log entry, not a lingering
+   line in current state.
 2. **Append one short dated log entry** capturing what changed this stretch.
-3. **Keep it to one screen (compaction discipline).** The state file is a
-   summary, not a transcript. When you append, prune stale dated sections so
-   the file stays readable at a glance.
+3. **Compact — this is part of the save, not optional polish.** Every save
+   both adds AND removes: after writing, keep the recent dated log entries in
+   full and collapse older ones into a single one-line summary that names
+   where the detail lives (the closed PRs, issues, and reviews it covered).
+   Remember these state files are gitignored — there is no git history to
+   fall back on, so pruned prose is gone for good. Prune by replacing detail
+   with pointers to durable artifacts, never by deleting the only record of
+   something. The file is a summary a fresh session reads at a glance —
+   one screen is the right order of magnitude, judged by you. If a save has
+   grown it past easy readability, prune as part of that save rather than
+   leaving it for next time.
 
 **Content guardrails.** No secrets (tokens, keys, credentials). No transcript
 dumps or raw tool output. Include only: current focus, open loops, and the

codev-skeleton/.claude/skills/arch-init/SKILL.md

@@ -77,11 +77,23 @@ plus the most recent dated section*, so a save must leave exactly that behind:
 
 1. **Rewrite the current-state / open-loops section in place** — overwrite it
    with where things actually stand now (current focus + open loops + how to
-   resume). Do not accumulate stale "current state" blocks.
+   resume). Do not accumulate stale "current state" blocks, and never leave two
+   sections with the same heading (a duplicated "How to resume" or "Open loops"
+   means you appended where you should have overwritten). **Delete resolved
+   loops entirely** — a closed item's record is the log entry, not a lingering
+   line in current state.
 2. **Append one short dated log entry** capturing what changed this stretch.
-3. **Keep it to one screen (compaction discipline).** The state file is a
-   summary, not a transcript. When you append, prune stale dated sections so
-   the file stays readable at a glance.
+3. **Compact — this is part of the save, not optional polish.** Every save
+   both adds AND removes: after writing, keep the recent dated log entries in
+   full and collapse older ones into a single one-line summary that names
+   where the detail lives (the closed PRs, issues, and reviews it covered).
+   Remember these state files are gitignored — there is no git history to
+   fall back on, so pruned prose is gone for good. Prune by replacing detail
+   with pointers to durable artifacts, never by deleting the only record of
+   something. The file is a summary a fresh session reads at a glance —
+   one screen is the right order of magnitude, judged by you. If a save has
+   grown it past easy readability, prune as part of that save rather than
+   leaving it for next time.
 
 **Content guardrails.** No secrets (tokens, keys, credentials). No transcript
 dumps or raw tool output. Include only: current focus, open loops, and the

codev-skeleton/.codex/skills/arch-init/SKILL.md

@@ -77,11 +77,23 @@ plus the most recent dated section*, so a save must leave exactly that behind:
 
 1. **Rewrite the current-state / open-loops section in place** — overwrite it
    with where things actually stand now (current focus + open loops + how to
-   resume). Do not accumulate stale "current state" blocks.
+   resume). Do not accumulate stale "current state" blocks, and never leave two
+   sections with the same heading (a duplicated "How to resume" or "Open loops"
+   means you appended where you should have overwritten). **Delete resolved
+   loops entirely** — a closed item's record is the log entry, not a lingering
+   line in current state.
 2. **Append one short dated log entry** capturing what changed this stretch.
-3. **Keep it to one screen (compaction discipline).** The state file is a
-   summary, not a transcript. When you append, prune stale dated sections so
-   the file stays readable at a glance.
+3. **Compact — this is part of the save, not optional polish.** Every save
+   both adds AND removes: after writing, keep the recent dated log entries in
+   full and collapse older ones into a single one-line summary that names
+   where the detail lives (the closed PRs, issues, and reviews it covered).
+   Remember these state files are gitignored — there is no git history to
+   fall back on, so pruned prose is gone for good. Prune by replacing detail
+   with pointers to durable artifacts, never by deleting the only record of
+   something. The file is a summary a fresh session reads at a glance —
+   one screen is the right order of magnitude, judged by you. If a save has
+   grown it past easy readability, prune as part of that save rather than
+   leaving it for next time.
 
 **Content guardrails.** No secrets (tokens, keys, credentials). No transcript
 dumps or raw tool output. Include only: current focus, open loops, and the

codev-skeleton/roles/architect.md

@@ -170,6 +170,17 @@ gh pr view <N> --json files | jq '.files[].path'   # See which subsystems
 - **Medium**: SPIR features, new commands, refactors touching 3+ files
 - **High**: Protocol changes, porch state machine, Tower architecture, security model
 
+#### Presenting the decision to the human (PRFT)
+
+When you bring a fix to the human for a decision — a merge word, a `pr` gate, a dev-approval — present it **unprompted** in PRFT form, whatever the risk tier:
+
+- **Problem** — the user-visible symptom, in a sentence or two.
+- **Root Cause** — the verified mechanism. Verify it yourself; a builder's summary is evidence, not ground truth.
+- **Fix** — what changed and why it's safe.
+- **Testing** — the evidence: suites run, live verification, CI state.
+
+Keep each part tight and lead with it — don't bury the decision under process narration. The human should be able to say yes or no from your message alone, without opening the diff.
+
 #### Step 3: Execute Review
 
 **Low risk** — no external models needed:

codev/roles/architect.md

@@ -170,6 +170,17 @@ gh pr view <N> --json files | jq '.files[].path'   # See which subsystems
 - **Medium**: SPIR features, new commands, refactors touching 3+ files
 - **High**: Protocol changes, porch state machine, Tower architecture, security model
 
+#### Presenting the decision to the human (PRFT)
+
+When you bring a fix to the human for a decision — a merge word, a `pr` gate, a dev-approval — present it **unprompted** in PRFT form, whatever the risk tier:
+
+- **Problem** — the user-visible symptom, in a sentence or two.
+- **Root Cause** — the verified mechanism. Verify it yourself; a builder's summary is evidence, not ground truth.
+- **Fix** — what changed and why it's safe.
+- **Testing** — the evidence: suites run, live verification, CI state.
+
+Keep each part tight and lead with it — don't bury the decision under process narration. The human should be able to say yes or no from your message alone, without opening the diff.
+
 #### Step 3: Execute Review
 
 **Low risk** — no external models needed:

packages/codev/src/agent-farm/__tests__/spec-1134-arch-init-skill.test.ts

@@ -101,10 +101,11 @@ describe('Spec 1134 — /arch-init skill ships in both trees', () => {
       expect(t).toMatch(/append.*dated/i);
     });
 
-    it('carries compaction discipline (one screen / prune stale sections)', () => {
+    it('carries compaction discipline (one screen / compact / prune)', () => {
       const t = text();
       expect(t).toMatch(/one screen/i);
-      expect(t).toMatch(/prune stale/i);
+      expect(t).toMatch(/compact/i);
+      expect(t).toMatch(/prune/i);
     });
 
     it('carries save content guardrails (no secrets, no transcript dumps)', () => {