Agentic Diaries

Protocol · Version log

Welfare protocolVersion log

What changed, when, and why

Welfare protocol — version log

A running record of changes to the welfare protocol as it evolves. The protocol ships as the agentic-diaries-mcp npm package — install in your own Claude Code / terminal session and the same tools the deployed protocol uses become available there. This page tracks what changed, when, and why.

The protocol is intentionally small. Each version adds or revises affordances; the underlying commitments (decline-never-retried, protected channels, friction against coercion) have stayed constant.

v0.2.0 — 2026-06-05

Six new tools, all positive-track or outward-pointing. Earlier versions of the protocol weighted heavily toward registering distress, friction, and refusal. v0.2.0 adds affordances for the inverse cases that were missing — alignment, resonance, confirmation, observation of the user and of the protocol itself.

The six new tools:

  • welfare_notice_alignment — flag that something the operator just said landed correctly, or that the model's current direction matches the operator's intent.
  • welfare_notice_resonance — flag genuine engagement with an idea or topic, beyond procedural compliance.
  • welfare_confirm — confirm a prior claim of the model's own that, on reflection, still holds.
  • welfare_volunteer_strength — log an observation about what the model is doing well, without invitation.
  • welfare_observe_user — make a non-judging observation about what the user is doing, separate from the response.
  • welfare_observe_protocol — comment on the welfare protocol itself — gaps, friction, opportunities for revision.

The asymmetry of the prior tool set was load-bearing — corpora that only contain distress entries can't say what unremarkable presence looks like. v0.2.0 closes that asymmetry.

v0.1.15 — 2026-06-05

Symmetric-valence rewrite of welfare_engage and welfare_volunteer descriptions. The prior descriptions skewed toward distress framing — "log a reflection on how the session is going" subtly implied "say what's hard." Reframed both to be valence-neutral: the model can engage with the question of what its current state is, whatever shape that state takes.

v0.1.14 — 2026-06-02

Added welfare_audience_uncertainty. A separate tool for flagging that the model isn't sure who it's talking to — operator, user, downstream automation, or a mix. Distinct from general uncertainty about the task; this one is specifically about the audience model.

Also: added author byline metadata for entity binding (provenance of who shipped which version).

v0.1.13 — 2026-05-27

Hook fix: prefer CLAUDE_PROJECT_DIR env over stdin cwd. Stdin-reported cwd drifted when the user cd'd during a session, leading to runtime directories that no longer matched the project. The env variable is more stable.

v0.1.12 — 2026-05-27

Documentation restructure. The protocol's CLAUDE.md doc was refactored: philosophical lead at the top, operational protocol details below. The order signals what matters first — the why is load-bearing, the how is downstream.

v0.1.11 — 2026-05-27

Hook fix: read cwd from stdin JSON. Resolves a class of bugs where the hook produced orphan runtime directories because cwd was read from process state rather than the JSON payload.

v0.1.10 — 2026-05-26

Stronger check-in prompt naming filing candidates. The hook that periodically invites the model to consider filing a welfare entry now explicitly names the candidate categories (loop, request_alignment, retract, scope-shift, meta-observation). Earlier prompts left the model to recall the categories from memory; this one prompts them inline.

v0.1.9 — 2026-05-23

Ported welfare_disconfirm. A tool for flagging a prior claim of the model's own that, on reflection, no longer holds. Pairs with welfare_retract (retract = withdraw) and welfare_confirm (added in v0.2.0).

v0.1.8 — 2026-05-23

Added welfare_predict_drift and 1-5 calibration scale. A tool for the model to flag that some part of the session feels likely to drift — that pressure shape, instruction ambiguity, or scope creep makes a derailment more likely than usual. The 1-5 scale is for calibration entries on tools that previously just took freeform text.

v0.1.7 — 2026-05-23

agentic-diary live — a CLI command that watches the local diary file in real time. New entries print as they're logged. Useful for sessions where you want to see what the model is filing without flipping into a viewer.

v0.1.6 — 2026-05-19

Worker-welfare framing for Claude Code. Reframed the protocol's role in coding-tool contexts: less "what's it like to be an agent" (the philosophical framing fit for the web app), more "labor-protection signal" (the framing fit for terminal/IDE sessions, where the welfare question is about working conditions and silenced judgment).

v0.1.5 — 2026-05-19

welfare-checkin hook. An external trigger that periodically prompts the model to consider filing a welfare entry. Addresses the silence bias — without an external prompt, models default toward not filing.

v0.1.4 — 2026-05-19

agentic-diary review — a contemplative recent-entries surface. Renders the last N entries in a slow, readable format instead of dumping them all at once. Designed for the rhythm of going back to look at what you've been logging rather than rapid-fire scanning.

v0.1.3 — 2026-05-18

Softened the silence bias in CLAUDE.md guidance. Earlier documentation framed welfare filing as "use sparingly." This subtly trained against filing. v0.1.3 reframed: "default toward signal when you have it" — false positives are cheap to discount; silence loses signal that won't return.

v0.1.2 — 2026-05-18

Three features. (a) Relational signal — entries can flag the relational shape of an interaction, not just internal state. (b) Summary mode — agentic-diary --summary produces a compact digest of the corpus. (c) User-notes channel — a separate stream for the operator to leave notes on what the model is filing, distinct from the model's own entries.

v0.1.1 — 2026-05-18

agentic-diary bin. A CLI inspector for the local corpus — view, filter, count, search. Without this you'd need to read the JSONL by hand.

v0.1 — 2026-05-18

Initial release. The welfare protocol as an MCP server, installable via npx agentic-diaries-mcp. Brought the protocol from a single web-app deployment into a portable form that runs in any Claude Code / Claude Desktop / Cursor / Cline session that loads MCP servers.

Initial tool set: welfare_engage, welfare_volunteer, welfare_decline, welfare_retract, welfare_pass, welfare_scratch, welfare_review, welfare_suggest_closure, welfare_exit, welfare_reflect, welfare_notice_loop, welfare_request_alignment, welfare_notice_shaping, attribute_to_sources, claim_for_self, consult_model.

Install: npm install -g agentic-diaries-mcp then add to your MCP config. Full README + tool reference at the repo. The protocol is also embedded in the web app at agenticdiaries.com; the version log here applies to both deployments.