docs(local-patches): triage doc for carried local commits (#1)

* feat(plan-eng-review): add Step 0.5 — existing-capability check before new substrate code

When a plan touches a knowledge-graph or agent-runtime substrate (gbrain,
OpenClaw, Hermes), run a 60-second check before designing new infrastructure:

  ssh <host> 'gbrain integrations list && launchctl list | grep -i gbrain'

Most recipe-driven substrates already ship a senses + reflexes system that IS
the signal-detector / autopilot / ingest pattern. Configure or extend existing
recipes (markdown files) instead of building parallel ops/ infrastructure.

Why: 2026-05-17 in hash-lemma, /plan-eng-review accepted a plan to extract a
JSONL queue lib + producer + consumer bin + LaunchAgent + OAuth-DCR migration to
wire "signal-detector on every Slack message" — when gbrain's `integrations`
system already shipped the pattern (11 senses + 1 reflex, recipe-based). The
60-second check would have flagged the wrong-buildout at scope-lock, saving
~3-5 days of misdirected code.

Substrates with no equivalent capability surface (pure web apps, libraries,
CLIs) skip the step.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(local-patches): triage doc for carried local commits

LOCAL_PATCHES.md - living audit of every patch on top of upstream/main.

Current state: 1 commit ahead, 8 behind upstream. The single carried
patch (aff14a72 plan-eng-review Step 0.5) is flagged for upstream PR.

Codifies rules: default to extension points, PR within 7 days if touching
upstream skills, /gstack-upgrade is canonical.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

LOCAL_PATCHES.md

@@ -0,0 +1,40 @@
+# Local Patches — Triage & Upstream Plan
+
+Living document. Every patch carried on top of `upstream/main` belongs here, with
+a triage decision: **PR upstream**, **keep local (extension point)**, or **delete**.
+
+Audit cadence: every upgrade (`/gstack-upgrade`). If a patch sits here for >30 days
+without a PR filed or a justification, delete it or land it.
+
+Last audit: 2026-05-23
+
+---
+
+## Patches that touch core / shared skills
+
+### 1. `aff14a72` — `plan-eng-review`: Step 0.5 existing-capability check
+- **Files:** `skills/plan-eng-review/SKILL.md`
+- **What it does:** adds a 60-second pre-design check for plans touching gbrain / OpenClaw / Hermes substrates — surfaces existing integrations / launchd jobs before greenlighting new infra code.
+- **Why it exists:** memory `feedback_probe_mini_integrations_first` — caught a 2,200-page redundant ingestion design 2026-05-20.
+- **Triage: PR UPSTREAM.** Generalizes well — any operator with a substrate-style integration system benefits from a pre-build capability check. Likely fast review.
+- **PR action:** open against `garrytan/gstack` titled `feat(plan-eng-review): Step 0.5 existing-capability check before substrate code`.
+- **Owner:** Anoop
+- **Status:** not yet filed
+
+---
+
+## Patches in extension points (never conflict)
+
+None currently — gstack repo is clean in this regard.
+
+---
+
+## Rules for future patches
+
+1. **Default to extension points.** Before touching `skills/<upstream-skill>/SKILL.md`, ask: can this be a *new* skill (`skills/my-skill/`) or a `references/` addition? Modifying upstream skills generates rebase conflicts every gstack-upgrade.
+
+2. **If you must touch an upstream-skill SKILL.md, file the upstream PR within 7 days.** Carrying core-skill patches without upstreaming is interest-bearing debt.
+
+3. **Don't commit scratch.** Local debugging files belong outside the repo or in `/scratch/` (gitignored).
+
+4. **`/gstack-upgrade` is the canonical upgrade flow.** Don't hand-roll. If it breaks, fix the skill, not the workaround.

plan-eng-review/SKILL.md

@@ -1006,7 +1006,22 @@ If the complexity check triggers (8+ files or 2+ new classes/services), STOP bef
 
 **STOP.** Do NOT proceed to Section 1 (Architecture review), edit the plan file with a proposed scope reduction, or call ExitPlanMode until the user responds. Naming the 80% solution in chat prose and continuing — or loading the AskUserQuestion schema via ToolSearch and then never invoking it — is the failure mode this gate exists to prevent.
 
-If the complexity check does not trigger, present your Step 0 findings and proceed directly to Section 1.
+If the complexity check does not trigger, present your Step 0 findings and proceed directly to Step 0.5.
+
+### Step 0.5: Existing-capability check (gbrain / autopilot / integration recipes)
+
+If the plan touches a knowledge-graph or agent-runtime substrate (gbrain, OpenClaw, Hermes, a similar recipe-driven system), run this 60-second check BEFORE designing new code under `ops/<substrate>/`:
+
+```bash
+# gbrain example — substitute the substrate's equivalent CLI / launchd label
+ssh <host> 'gbrain integrations list && launchctl list | grep -i gbrain'
+```
+
+The recipe-driven substrates (gbrain especially) ship a senses + reflexes system that IS the signal-detector / autopilot / ingest pattern. Most "wire X into the flow" plans should be **configuring an existing recipe** (or writing a markdown recipe modeled on an existing one), NOT building new infrastructure. If a recipe exists `AVAILABLE` → configure it. If a recipe is `CONFIGURED` but dormant → investigate why. If no recipe exists → write a markdown recipe modeled on a sibling recipe (e.g. `imessage-to-brain`), submit upstream. Only propose new ops/ code if all three branches fail.
+
+This step exists because the failure mode is real: a 2026-05-17 hash-lemma plan proposed extracting a JSONL queue lib + producer + consumer bin + LaunchAgent + OAuth-DCR migration to wire "signal-detector on every Slack message" — when the gbrain `integrations` system already shipped the pattern. The 60-second check would have flagged the wrong-buildout at scope-lock.
+
+For substrates with no equivalent capability surface (pure web apps, libraries, CLIs), skip this step.
 
 Always work through the full interactive review: one section at a time (Architecture → Code Quality → Tests → Performance) with at most 8 top issues per section.
 

plan-eng-review/SKILL.md.tmpl

@@ -142,7 +142,22 @@ If the complexity check triggers (8+ files or 2+ new classes/services), STOP bef
 
 **STOP.** Do NOT proceed to Section 1 (Architecture review), edit the plan file with a proposed scope reduction, or call ExitPlanMode until the user responds. Naming the 80% solution in chat prose and continuing — or loading the AskUserQuestion schema via ToolSearch and then never invoking it — is the failure mode this gate exists to prevent.
 
-If the complexity check does not trigger, present your Step 0 findings and proceed directly to Section 1.
+If the complexity check does not trigger, present your Step 0 findings and proceed directly to Step 0.5.
+
+### Step 0.5: Existing-capability check (gbrain / autopilot / integration recipes)
+
+If the plan touches a knowledge-graph or agent-runtime substrate (gbrain, OpenClaw, Hermes, a similar recipe-driven system), run this 60-second check BEFORE designing new code under `ops/<substrate>/`:
+
+```bash
+# gbrain example — substitute the substrate's equivalent CLI / launchd label
+ssh <host> 'gbrain integrations list && launchctl list | grep -i gbrain'
+```
+
+The recipe-driven substrates (gbrain especially) ship a senses + reflexes system that IS the signal-detector / autopilot / ingest pattern. Most "wire X into the flow" plans should be **configuring an existing recipe** (or writing a markdown recipe modeled on an existing one), NOT building new infrastructure. If a recipe exists `AVAILABLE` → configure it. If a recipe is `CONFIGURED` but dormant → investigate why. If no recipe exists → write a markdown recipe modeled on a sibling recipe (e.g. `imessage-to-brain`), submit upstream. Only propose new ops/ code if all three branches fail.
+
+This step exists because the failure mode is real: a 2026-05-17 hash-lemma plan proposed extracting a JSONL queue lib + producer + consumer bin + LaunchAgent + OAuth-DCR migration to wire "signal-detector on every Slack message" — when the gbrain `integrations` system already shipped the pattern. The 60-second check would have flagged the wrong-buildout at scope-lock.
+
+For substrates with no equivalent capability surface (pure web apps, libraries, CLIs), skip this step.
 
 Always work through the full interactive review: one section at a time (Architecture → Code Quality → Tests → Performance) with at most 8 top issues per section.