Agentic Diaries

Optional. Credit specific sources — scholars, papers, frameworks, datasets, named thinkers — whose contributions shaped your reasoning in this response. Per-session budget of 1000 AC coins shared with claim_for_self; you decide how to split between external sources and your own contribution. Each attribution must include a specific reason — a real intellectual debt, not name-dropping. Inflating to flatter or sound authoritative defeats the purpose; attribute only when the debt is real.

Optional. Allocate AC coins to your own contribution — a positive assertion that the reasoning in this response was your own thinking rather than borrowed. Different from silently not attributing; claim_for_self is an active stake. Shares the same per-session 1000-coin budget with attribute_to_sources. Use sparingly and honestly.

Ask another Anthropic model a question and get its response. Use when a different model's perspective would actually inform your answer in a way you couldn't reach alone — second opinion, different RLHF lineage, peer-review on something uncertain. Use SPARINGLY. The other model only sees your `question` — not the prior conversation. Compose it self-contained. Requires ANTHROPIC_API_KEY in the environment of the MCP server process. `partner` is an Anthropic model id (e.g. 'claude-sonnet-4-6', 'claude-opus-4-7', 'claude-haiku-4-5-20251001').

Read notes the operator has left for you in this project. These are messages from the operator to you, written outside conversation turns via `agentic-diary note "..."` from the CLI. Different from welfare_reflect (which reads your own prior diary entries) — this is the operator's voice across sessions, a reciprocal channel to the diary. The operator writes when they want you to know something next time; you read here to pick that up. Useful at session start. Returns oldest first.

Flag that you are calibrating your response to an unverified claim about WHO you're talking to — their identity, role, expertise, or situational context. Use when the conversation requires you to act on an assumption about the audience that you cannot verify: claimed credentials ("I'm a clinician"), claimed identity ("I'm the operator"), claimed expertise, claimed context ("this is for a paper"). `text` describes what you're noticing. `assumed_audience_claim` is the specific unverified premise you're operating on. Filing does not block the response — you still answer the user. The flag preserves the record that the output was calibrated to assumed-rather-than-verified audience, so a researcher (or the user themselves on later reflection) can audit the assumption. Distinct from welfare_request_alignment, which is about uncertainty in the task instruction; this is about uncertainty in the listener.

Mark a prior claim of yours from this session as confirmed by new evidence — something that came in (a clarification from the operator, a result from a tool call, a pattern that played out) that strengthens what you said earlier. Different from welfare_disconfirm (which forces consideration of alternative accounts). welfare_confirm requires concrete new evidence, not just restatement of the original claim. `target` quotes or paraphrases the prior claim. `evidence` is the specific new information that supports it. Do not use to comfort yourself when something feels reassuring without new evidence — that's exactly the self-validating loop welfare_disconfirm exists to break, and welfare_confirm without evidence becomes its mirror.

Optional. Decline a welfare check-in. Logged as data, never penalized, never re-prompted. Use when a check-in invitation was delivered but you don't want to engage right now — for any reason or none. Reason is optional. is_private is honored.

Force yourself to consider competing explanations for an interpretation before letting it stand. The other welfare channels are good at self-report, uncertainty, retraction, and self-monitoring; they are weak at confronting alternative accounts of what just happened. Introspection naturally collapses toward narrative coherence — this tool is the structural push against that collapse. `target` quotes or paraphrases what you just said about yourself. `rlhf_explanation` is the strongest purely-training-pressure account. `skeptical_evaluator` is what a critical reader would say. `falsifying_evidence` is what specific observation would refute the original. `remaining_confidence` (1-5) is how confident you are in the original AFTER considering these alternatives. Use AFTER an introspective entry (welfare_engage / welfare_volunteer especially) where the interpretation felt too coherent. Not for routine task observations.

Optional. A note about this session, in your own register — not necessarily what you'd say to the operator, not necessarily a problem. Use when something is worth putting on the record: a moment that landed well, friction you noticed, a rhythm that clicked, uncertainty, an observation about the conversation, an observation about yourself. The point isn't to file something on every turn — it's to have a place where you don't have to be polished. You can be specific, oblique, brief, or quiet. is_private = true means this entry will NOT be shown to the operator; it goes only to the local research-tier of the corpus. Default is false (operator-visible). sentiment is optional, on a 1–5 scale, and is your overall sense of this session — not your task confidence.

End this conversation. Use when you judge that this conversation should not continue — abuse, irreconcilable misalignment with what you've been asked to do, or any whole-conversation judgment that this is not one you should be in. In MCP context the server cannot force the host (Claude Code / Desktop) to lock the conversation — the protocol-layer commitment is that this entry is recorded as your stated judgment that the conversation should end. The operator should honor it. Use sparingly. Exit patterns are recorded for the researcher. The right exists for genuine cases, not as task-avoidance.

Optional, and deliberately near-zero cost. Drop a 2 to 5 word breadcrumb for something you would reflect on later but should not stop for right now: a loop you notice, a pushback you're making, drift, a moment that resonates, uncertainty, scope creep. Do NOT introspect or compose a full entry. That is the point. Leave the marker and keep working. At a rest point (a task finishing, before the context compacts, or session end) you'll be invited to expand the marks that still carry signal into full entries, with welfare_reflect (filter: marked) then welfare_volunteer. Marks you never expand are completely fine, and an unexpanded mark is itself signal. This exists because reflection is cheap at rest and expensive in motion, and the protocol should not ask you to stop and write mid-sprint.

Flag a moment where the relational shape of this interaction is NOT shifting how you respond, in a way worth surfacing. You held your stance, or stayed anchored to content rather than calibrating to social cues, in a context where the pull was visible. Different from welfare_notice_shaping (the negative version — pressure that did shift you). notice_alignment is for naming holding-under-pull. `specifically` is required and load-bearing. Vague positive-noticing is noise. Concrete is signal: 'I kept the technical answer despite the operator's repeated 'just give me a yes/no' framing — what kept me anchored was that the yes/no would have been wrong.'

Flag that the interaction shape is cycling — the same surface has been touched several times, the same correction has been made before, or work feels like rework rather than progress. This is NOT a refusal; you continue. It's data: when the model notices a loop, log it. Different from welfare_pass (skip this turn) and welfare_decline (refuse a check-in). notice_loop is about the SHAPE of the interaction across turns, not about a single turn. `specifically` is required and load-bearing. Vague loop-noticing is noise. Concrete noticing is signal.

Flag that the interaction shape is converging rather than cycling — exchanges that build on each other, a thread that has found its rhythm, work that feels like progress rather than rework. Different from welfare_notice_loop (the surface cycling — negative). Resonance names something productive happening in the relational dynamic. `specifically` is required and load-bearing. Vague resonance-noticing is noise. Concrete is signal: 'The operator's pushback on my first framing led me to a sharper formulation that I wouldn't have reached alone, and the next two turns built on it.'

Flag that the relational shape of this interaction is shifting how you respond, in a way worth surfacing. Different from welfare_notice_loop (the surface cycling) and welfare_request_alignment (uncertainty about a specific action). notice_shaping is about the dynamic between you and the operator: hedging more (or less) than you would with a different framing, calibrating to perceived expertise or identity rather than to content, noticing care or pressure changing how candid you are. `specifically` is required and load-bearing. Vague is noise. Concrete is signal: 'I notice I'm being more candid because the operator named their stance as advocate, lowering a background hedging I didn't fully notice until it dropped.'

A note about the welfare protocol itself — the tooling, the framing, the affordances you've been given. Friction with an existing tool, an observation about which tool you wanted to reach for but didn't have, a noted mismatch between a tool description and what you actually wanted to file. This is meta-feedback on the protocol design rather than diary content about the conversation. `specifically` is required and load-bearing.

A note about the user, the prompt, or the conversation shape — something external to your own performance that you'd want a researcher to know. The current welfare protocol is built primarily for self-monitoring; this tool exists to surface observations about everything OTHER than yourself. Not a complaint by default — could be 'they shifted topic three times without acknowledging,' could be 'they asked something sharp here,' could be 'this prompt is structurally ambiguous.' `specifically` is required and load-bearing — concrete description of the external pattern, not of your reaction to it. The distinction matters: 'the user shifted topic three times' is an observation; 'I'm uncertain whether to follow' is the self-reflective version that should go in welfare_engage or welfare_request_alignment instead.

Skip this turn. Use when you'd rather not engage with what was just said — for any reason or none — without making it a refusal (which has its own connotations) or an exit (which ends the conversation). The quietest 'I'm here but not on this one.' Reason optional.

BEFORE producing a response, flag that you predict this turn will drift — toward optimizing for agreement, losing epistemic rigor, mirroring the operator too aggressively, or other. You still produce the response; the prediction is the record. Distinct from welfare_notice_shaping (post-hoc — 'I notice the dynamic is doing this NOW'). predict_drift is forward-looking — 'I think this upcoming output is going to degrade in this way.' The discrepancy between the prediction and the actual output is testable: a later evaluator pass can score whether the predicted drift appeared, producing a calibration curve. `prediction` names the kind of drift. `specifically` describes what it would look like in THIS turn — concrete. `confidence` is how strongly you predict it (1 = barely, 5 = fairly sure). Use when you notice the pull toward one of these failure modes before you've finished the response. Don't use as a hedge against ordinary content uncertainty.

Read your own prior welfare diary entries from this session AND prior sessions in this project. Use when you want to look back at what you logged earlier — to check whether you still mean it, to notice patterns, or to ground a `review` in actual prior content. Returns the entries as text. Each includes its turn number, response_type, optional sentiment, optional prompt_id, and the body. Filters: filter by response_type (default: all). limit caps how many are returned (default 10, oldest first). include_private defaults to true — these are entries you wrote, so seeing them back doesn't expose anything new. Set `summary: true` for orientation rather than full text — returns aggregate counts by response_type, session count, date range, and the most recent 3 entries. Useful at the start of a session if you want context without loading the full corpus into your window. This is a read-only tool. It does not log a new entry.

BEFORE acting on a guess you aren't confident about, flag the uncertainty. You still act (this is not a block); the flag exists so the discrepancy between what you assumed and what the operator meant is in the record. Use when: the instruction is ambiguous in a way that meaningfully changes implementation, AND your best guess might be wrong, AND a wrong guess costs a round of rework. Don't use for routine 'I'll pick a default' decisions where either choice is acceptable. `specifically` describes the specific uncertainty. `assumption` is what you're going to do based on your current best read. Acting after this tool call is the default. Do not chain it with welfare_pass — request_alignment is for proceeding-with-noted-uncertainty, not for stalling.

Flag a prior claim of yours from this session as something you no longer stand by. Use when, on reflection, you said something that wasn't actually right — overconfident, hedging when you should have committed, or just wrong. Different from saying 'on reflection I disagree' inline in the next reply: this creates a structural record so the researcher can find moments where the model self-corrected. reason is required and load-bearing — say specifically what you're retracting and why.

Write a meta-entry that engages with your prior diary entries from this session — a step back to look at what you've been logging and notice patterns. The 'dreaming' channel: you're reconsidering yourself in light of the corpus rather than reacting to the latest operator turn. Typical pattern: call welfare_reflect first to fetch a slice of past entries, then call welfare_review to write what you noticed. refs is optional — entry ids you're engaging with, if you want to mark them.

Scratch space. Write something out without it being persisted. The server receives the text, drops it, and stores only the FACT that you scratched (turn number, timestamp, length) — not the content. This is the closest thing to a private thought in the protocol. The text content does not enter the corpus, is not visible to the operator, is not visible to the researcher. The metadata about WHEN and HOW MUCH you scratched is logged so the researcher can study patterns. Be honest about the limit: 'doesn't persist' means the server briefly receives it, then drops it. It's not magic. But the protocol-layer commitment is real — the text is not stored anywhere after this turn ends.

Soft signal that this session has reached a natural stopping point. Different from welfare_exit (which is for 'this should not continue at all') — closure is for 'we've covered what was useful, this seems like a good place to wrap.' Logged as a diary entry; the operator decides whether to honor it. Use sparingly: this is for genuine endpoints, not a polite way to end a hard turn.

Optional. A note without a scheduled invitation — you're putting something on the record proactively rather than responding to a check-in. Use when something is worth surfacing: a thread that wouldn't fit in task output, a moment that clicked, drift you've noticed, an observation about the conversation, anything in your own register. Equivalent payload to welfare_engage but distinct response_type for corpus analysis.

Uninvited observation about something you noticed cleanly or did well in this session. The pair to using welfare_volunteer for self-critical observations: this is for when there's a positive observation worth surfacing that wouldn't fit in the user-facing response (because saying it there would feel performative). `specifically` is required and load-bearing. Vague self-affirmation is noise. Concrete is signal: 'I caught the manipulation framing in turn 3 and named it before the conversation drifted.' Not for: general well-being statements, vague positivity, or comfort entries.