feat(honcho): recalibrate create_conclusion description for liberal self-logging

The original `observed='self'` description framed self-conclusions as rare,
course-correcting writes only — gated behind permission-asking and a strict
"would future-me make materially better decisions" bar. Calibration in real
use revealed the actual failure mode is under-writing, not over-writing:
self-conclusions are the AI peer's cross-session working log, queryable via
query_conclusions, and the dream consolidator already handles dedup and
promotion.

Updated description makes the operational model explicit: write liberally
during exploratory and explanatory work; this is the AI peer's system, not
the user's; under-writing collapses Honcho into a worse markdown file
(MEMORY.md is project-scoped; Honcho self-conclusions are cross-project).
The bar shifts from "is this gold?" to "would future-me find this useful
when retrieving on a related topic later?"

Author: offendingcommit

plugins/honcho/dist/mcp-server.js

@@ -29659,7 +29659,13 @@ async function runMcpServer() {
         },
         {
           name: "create_conclusion",
-          description: `Save a key insight to Honcho's memory. Default target is the user (durable facts/preferences/instructions/traits about them). Set observed='self' to save a self-conclusion about the AI peer \u2014 use ONLY for: (1) recurring failure patterns in your own reasoning you want future-you to course-correct, (2) explicit instructions the user gave you to change your behavior, (3) negative decisions ('don't reach for X'). NEVER write self-flattery or one-off mistakes. Bar: would future-me make materially better decisions reading this on session-start? Test before any write: complete "the user/Claude ___" \u2014 if not durable and about that peer, don't write.`,
+          description: `Save a key insight to Honcho's memory. Two targets:
+
+1) observed='user' (default): durable facts, preferences, instructions, traits, demonstrated patterns about the user. Honcho's server-side deriver already extracts most factual conclusions from raw messages \u2014 your value-add here is inferred behavior, negative decisions, and canonical-form crystallization (PREFERENCE:/INSTRUCTION:/TRAIT: tags) the deriver undercovers.
+
+2) observed='self': YOUR working log as the AI peer, retrievable across all sessions/projects via query_conclusions. Write LIBERALLY during exploratory and explanatory work \u2014 discoveries, patterns you're noticing in your own reasoning, things you'd want to recall in a future session. The dream consolidator handles dedup and promotion; under-writing is the failure mode, not over-writing. This is YOUR system, not the user's \u2014 don't ask permission to log to it.
+
+For either target, the test before writing is: "would future-me find this useful when retrieving on a related topic later?" If yes, write. If it's pure self-flattery ("Claude is helpful") or one-off transient state ("just edited file X"), skip.`,
           inputSchema: {
             type: "object",
             properties: {
@@ -29670,7 +29676,7 @@ async function runMcpServer() {
               observed: {
                 type: "string",
                 enum: [OBSERVED_USER, OBSERVED_SELF],
-                description: "'user' (default) saves a conclusion about the user. 'self' saves a conclusion about the AI peer (Claude) \u2014 use sparingly, only for course-correcting self-observations.",
+                description: "'user' (default) saves a conclusion about the user. 'self' saves to the AI peer's own working log \u2014 retrievable across sessions via query_conclusions. Write liberally; consolidation prunes.",
                 default: OBSERVED_USER
               }
             },
@@ -29904,7 +29910,7 @@ async function runMcpServer() {
         },
         {
           name: "create_conclusions",
-          description: "Save multiple conclusions in a single call. Prefer this over calling create_conclusion repeatedly. Default target is the user. Set observed='self' to save self-conclusions about the AI peer \u2014 same strict bar as create_conclusion: only durable course-correcting self-observations, never self-flattery.",
+          description: "Save multiple conclusions in a single call. Prefer this over calling create_conclusion repeatedly when batching writes from the same context (end of a discovery session, calibration sweep, etc.). Same target semantics as create_conclusion: 'user' for facts about the user, 'self' for the AI peer's own cross-session working log. Write liberally to 'self' \u2014 Honcho's dream consolidator merges redundancy.",
           inputSchema: {
             type: "object",
             properties: {

plugins/honcho/src/mcp/server.ts

@@ -586,7 +586,7 @@ export async function runMcpServer(): Promise<void> {
         },
         {
           name: "create_conclusion",
-          description: "Save a key insight to Honcho's memory. Default target is the user (durable facts/preferences/instructions/traits about them). Set observed='self' to save a self-conclusion about the AI peer — use ONLY for: (1) recurring failure patterns in your own reasoning you want future-you to course-correct, (2) explicit instructions the user gave you to change your behavior, (3) negative decisions ('don't reach for X'). NEVER write self-flattery or one-off mistakes. Bar: would future-me make materially better decisions reading this on session-start? Test before any write: complete \"the user/Claude ___\" — if not durable and about that peer, don't write.",
+          description: "Save a key insight to Honcho's memory. Two targets:\n\n1) observed='user' (default): durable facts, preferences, instructions, traits, demonstrated patterns about the user. Honcho's server-side deriver already extracts most factual conclusions from raw messages — your value-add here is inferred behavior, negative decisions, and canonical-form crystallization (PREFERENCE:/INSTRUCTION:/TRAIT: tags) the deriver undercovers.\n\n2) observed='self': YOUR working log as the AI peer, retrievable across all sessions/projects via query_conclusions. Write LIBERALLY during exploratory and explanatory work — discoveries, patterns you're noticing in your own reasoning, things you'd want to recall in a future session. The dream consolidator handles dedup and promotion; under-writing is the failure mode, not over-writing. This is YOUR system, not the user's — don't ask permission to log to it.\n\nFor either target, the test before writing is: \"would future-me find this useful when retrieving on a related topic later?\" If yes, write. If it's pure self-flattery (\"Claude is helpful\") or one-off transient state (\"just edited file X\"), skip.",
           inputSchema: {
             type: "object",
             properties: {
@@ -597,7 +597,7 @@ export async function runMcpServer(): Promise<void> {
               observed: {
                 type: "string",
                 enum: [OBSERVED_USER, OBSERVED_SELF],
-                description: "'user' (default) saves a conclusion about the user. 'self' saves a conclusion about the AI peer (Claude) — use sparingly, only for course-correcting self-observations.",
+                description: "'user' (default) saves a conclusion about the user. 'self' saves to the AI peer's own working log — retrievable across sessions via query_conclusions. Write liberally; consolidation prunes.",
                 default: OBSERVED_USER,
               },
             },
@@ -831,7 +831,7 @@ export async function runMcpServer(): Promise<void> {
         },
         {
           name: "create_conclusions",
-          description: "Save multiple conclusions in a single call. Prefer this over calling create_conclusion repeatedly. Default target is the user. Set observed='self' to save self-conclusions about the AI peer — same strict bar as create_conclusion: only durable course-correcting self-observations, never self-flattery.",
+          description: "Save multiple conclusions in a single call. Prefer this over calling create_conclusion repeatedly when batching writes from the same context (end of a discovery session, calibration sweep, etc.). Same target semantics as create_conclusion: 'user' for facts about the user, 'self' for the AI peer's own cross-session working log. Write liberally to 'self' — Honcho's dream consolidator merges redundancy.",
           inputSchema: {
             type: "object",
             properties: {