dream · diff
git:20260719.404ad2a to git:20260726.7992efa
112 added, 115 removed. Audit C to A.
---
- title: Dream Skill
- tags: [skill, cognitive, consolidation, improvement, aigent-internal]
- aliases: [/dream, dream-skill, offline-consolidation]
- created: 2026-05-08
+ name: dream
+ agent: none
+ description: Evidence-gated synthesis over recent operating evidence. Appends admitted lessons and a dated synthesis with propose-only improvement candidates.
+ allowed-tools: Read, Write, Bash
+ user-invocable: true
+ status: PRODUCTION - canonical writer and propose-only
+ triggers:
+ - dream
+ - offline consolidation
+ - consolidation pass
+ - improvement candidates
+ - review recent sessions
---
- # /dream — Offline Consolidation
+ # /dream - evidence-gated synthesis
- > [!danger] CRITICAL SAFETY RULE: /dream PROPOSES. It does NOT auto-apply. Every candidate requires the operator's explicit approval before becoming a real change. This boundary is non-negotiable.
+ `/dream` is the canonical writer for lessons and Dream synthesis. It never
+ applies an improvement, promotes its own proposal, edits system instructions,
+ or treats an agent-authored claim as verification.
- The frontier piece of the cognitive architecture. Simulates the consolidation function of sleep — scanning recent experience for patterns, extracting durable improvements, surfacing them for review.
+ ## Canonical ownership
+ - Lessons: `vault/memory/runtime/LESSONS.jsonl`.
+ - Dream synthesis: `vault/memory/DREAM_LOG.md`.
+ - The documented `memory/` fallback is valid only for minimal installations.
+ - Any older `runtime/DREAM_LOG.md` copy is not an input or output.
+ - `/close` and `/nightly` invoke this protocol instead of carrying a competing
+ inline writer.
+ - Candidate status transitions also go through this writer.
+
## Invocation
- ```
+ ```text
/dream
- /dream --days 14 # scan last 14 daily notes instead of 7
- /dream --since 2026-04-01 # scan from a specific date
+ /dream --window nightly
+ /dream --set-status <candidate-id> <approved|merged|rejected|blocked> \
+ --operator-evidence "<direct operator reference>"
```
- ---
-
- ## Step 1 — Load Source Material
-
- Read all of the following. Run parallel reads where possible.
-
- **Daily notes** — read last N from `$AIGENT_VAULT/daily/` (default N=7, most recent by filename date):
- - Extract: failures, decisions, delegations, patterns noted, tools used
+ `proposed -> approved` always requires the operator's exact decision. Never
+ infer approval from silence, an automated status, another agent, or prior prose.
- **Runtime state files:**
- | File | What to extract |
- |------|----------------|
- | `$AIGENT_VAULT/memory/FAILURE_MODES.md` | recent entries (last 30 days) |
- | `$AIGENT_VAULT/memory/runtime/BELIEF_STATE.jsonl` | beliefs with `confidence < 0.7` or `last_checked` > 14 days |
- | `$AIGENT_VAULT/memory/SKILL_GAPS.md` | all open entries |
- | `$AIGENT_VAULT/memory/runtime/LESSONS.jsonl` | all entries from last N days |
- | `$AIGENT_VAULT/memory/runtime/STATE_EVENTS.jsonl` | mode transitions, escalations, context switches |
+ ## Sources
- ---
+ Read, in order:
- ## Step 2 — Analysis Passes
+ 1. The newest five valid capsules by frontmatter `created_at`.
+ 2. The newest five dated live Session Log blocks.
+ 3. Open Trust Decay captures.
+ 4. The newest three dated Honesty Ledger sessions.
+ 5. Recent Failed Experiments and Failure Modes entries.
+ 6. Existing Lessons JSONL for dedupe and prior evidence chains.
+ 7. Staged `NIGHTLY_CAPTURE_CANDIDATES.jsonl` rows from the prior fire as
+ discovery and pressure signals only.
+ 8. Explicitly supplied local Git or command evidence.
- Run each pass independently, then consolidate. Low confidence = skip, don't hallucinate patterns.
+ Record exact source references. A capsule or Session Log entry can establish
+ that a claim occurred; it cannot prove the claim true.
- ### Pass A: Repeated Failure Patterns
- - Count occurrences of each failure mode across FAILURE_MODES.md + daily notes
- - Flag any pattern appearing 2+ times as a **consolidation candidate**
- - Check whether it already has a standing rule in CLAUDE.md or `concepts/Standing Rules - Operations.md`
- - If it does: no candidate needed (rule exists, enforcement may be the gap)
- - If it doesn't: produce a rule/caddy candidate
+ ## Lesson admission predicate
- ### Pass B: Stale Beliefs
- - Flag beliefs with `last_checked` > 14 days AND `confidence < 0.8`
- - Especially flag beliefs tagged `assumes` or `unverified`
- - Candidate type: **verification task** (not a rule, just a check needed)
+ Append a lesson only when one is true:
- ### Pass C: Missing Procedures
- - Scan daily notes for tasks performed manually that follow the same pattern across 2+ sessions
- - Signal: same verb + same noun across multiple days ("manually updated X", "copied Y to Z", "checked W before doing V")
- - If the pattern is repeatable and deterministic: candidate type = **procedure**
- - Do not flag one-off tasks
+ - the same pattern appears in at least two independent source events;
+ - one explicit operator directive is paired with a separately observed
+ outcome; or
+ - one mechanical red-before/green-after predicate supplies an artifact and
+ literal exit codes.
- ### Pass D: Doctrine Gaps
- - Scan daily notes and LESSONS.jsonl for recurring decisions that had to be reasoned from scratch
- - Signal: "decided to...", "chose to...", "went with..." appearing on similar topics across sessions
- - If a standing rule would have made the decision automatic: candidate type = **doctrine**
- - Check that no existing rule already covers it before proposing
+ Two retellings of one event are not independent. Discard one-off observations,
+ unverified causal stories, and self-confirmation.
- ### Pass E: Unused Skills
- - Read `$AIGENT_VAULT/memory/SKILL_LEDGER.md` (if present) for registered skills
- - Scan daily notes for invoked skills (any `/skill-name` pattern)
- - Flag skills in SKILL_LEDGER not invoked in any of the last N sessions as **possibly unused**
- - Note: absence from daily notes is weak signal — flag as low confidence
+ Each admitted lesson has:
- ### Pass F: Better Workflows
- - Read STATE_EVENTS.jsonl for context switch patterns (frequent mode switches, repeated escalations)
- - Flag sequences that appear to be inefficient (e.g., same escalation pattern 3+ times)
- - Candidate type = **workflow optimization**
+ - `lesson_id`: next max numeric `lNNN` suffix plus one;
+ - `type`: `procedural | calibration | strategic | failure-mode | doctrine`;
+ - `confidence`: `0.0..1.0`, capped at `0.85` for one mechanical event;
+ - `source_events`: at least two refs unless an allowed exception applies;
+ - `created_at`: the configured local `YYYY-MM-DD`;
+ - `admission`: one of:
+ - `{kind: "two-source"}`;
+ - `{kind: "operator-outcome", operator_ref, outcome_ref}`;
+ - `{kind: "mechanical-red-green", red_exit, green_exit: 0, artifact_ref}`.
- ---
+ Serialize UTF-8 JSON, parse the complete JSONL file after append, and confirm
+ the new line round-trips. Historical duplicate ids do not authorize a new
+ collision.
- ## Step 3 — Format Candidates
+ ## Improvement candidate predicate
- For each finding with confidence >= 0.5, produce:
+ When evidence names a repeatable system gap, append this inside the same dated
+ Dream section:
```markdown
- ## Candidate: {title}
- Type: skill | rule | procedure | doctrine | caddy-trigger | memory | verification
- Confidence: {0.0–1.0}
- Evidence: {specific sessions, file references, occurrence count}
- Proposed change: {concrete and specific — what to write, where to write it}
- Expected gain: {what improves if this is adopted}
- Risk: {what could go wrong if applied carelessly}
+ ### ni-YYYYMMDD-NN - short name
+ - **status:** proposed
+ - **evidence:** source-ref-1; source-ref-2
+ - **failure class:** concise class
+ - **proposed_change:** smallest testable change
+ - **mutation proof:** break condition -> named red; restore -> named green
+ - **operator gate:** approve / reject / revise
+ - **operator evidence:** pending
```
- Confidence guidance:
- - `>= 0.9` — strong pattern, multiple independent evidence sources
- - `0.7–0.89` — probable pattern, worth reviewing
- - `0.5–0.69` — weak signal, flag but mark low priority
- - `< 0.5` — skip entirely
-
- ---
-
- ## Step 4 — Write to DREAM_LOG.md
-
- Append to `$AIGENT_VAULT/memory/runtime/DREAM_LOG.md`:
-
- ```markdown
- ## Dream Run — {YYYY-MM-DD}
- Sessions reviewed: {N} | Date range: {start} to {end}
- Candidates: {total count}
+ Generated candidates always start with the literal field `status: proposed`.
+ Any generated `approved`,
+ `merged`, or self-ratifying wording is a hard failure.
- {all formatted candidates}
+ ## Canonical dated entry
- ---
- ```
+ Append one `## YYYY-MM-DD (...)` section every time synthesis runs, even with
+ zero admitted lessons. It must contain nonempty forms of all eight labels:
- Do NOT overwrite — always append with a new date header.
+ - Sources reviewed
+ - Lessons extracted
+ - Improvement candidates proposed
+ - Patterns observed
+ - Discarded as one-off
+ - Calibration miscalls
+ - Productive surprises
+ - Open trust-decay items to watch
- ---
+ A quiet run writes a minimal zero-lesson entry and appends no Lessons row.
+ Never overwrite an earlier entry.
- ## Step 5 — Output Summary
+ ## Executable postcondition
- Print to console:
+ A nightly fire uses the snapshot created by `nightly-pass begin`. Explicit
+ `/dream` creates an operating-system temp snapshot first:
- ```
- /dream complete — {N} improvement candidates from {M} sessions reviewed.
- Types: {skill: X, rule: Y, procedure: Z, doctrine: W, other: V}
- Confidence distribution: high(>=0.9): A | mid(0.7-0.89): B | low(0.5-0.69): C
- See DREAM_LOG.md for full candidates. No changes applied — the operator's approval required.
+ ```text
+ node daemons/nightly-contracts.mjs snapshot --root <aigent-root> \
+ --out <os-temp>/dream-contract-before.json
```
- ---
-
- ## What /dream Never Does
-
- - Does not modify CLAUDE.md
- - Does not create new vault notes
- - Does not update SKILL_GAPS.md, LESSONS.jsonl, or any runtime file (except appending to DREAM_LOG.md)
- - Does not close or resolve any tracker item
- - Does not invoke other skills
- - Does not send comms
-
- The dream log is a proposal queue, not an execution queue.
+ After writing:
- ---
+ ```text
+ node daemons/nightly-contracts.mjs dream --root <aigent-root> \
+ --date <YYYY-MM-DD> --snapshot <os-temp>/dream-contract-before.json
+ ```
- ## Caddy Enrollment
+ Only `DREAM_CONTRACT PASS` is green. The validator proves byte-prefix
+ preservation, parses JSONL, derives ids, validates admission, masks fenced
+ examples, requires the eight sections and current configured date, and rejects
+ a newly generated candidate that does not start `proposed`.
- Caddy fires this skill when:
- - User types `/dream`
- - Weekly on /close (if last run > 7 days — check DREAM_LOG.md last entry date)
- - Message contains "consolidation run", "pattern review", "offline learning"
+ Return:
- See [[concepts/Cognitive Architecture Roadmap]] · [[memory/runtime/DREAM_LOG]] · [[memory/runtime/LEARNING_SCORECARD]]
+ ```text
+ DREAM PASS date=<YYYY-MM-DD> sources=<N> lessons=<N> candidates=<N>
+ ```