Builder outcome improvements: continuity-of-self agent model, canon-first builders

.claude-plugin/marketplace.json

@@ -12,7 +12,7 @@
       "name": "bmad-builder",
       "source": "./",
       "description": "Build AI agents, workflows, and modules from a conversation. Four skills — Agent Builder, Workflow Builder, Module Builder, and Setup — guide you from idea to production-ready skill structure with built-in quality optimization. Part of the BMad Method ecosystem.",
-      "version": "1.8.0",
+      "version": "2.0.0",
       "author": {
         "name": "Brian (BMad) Madison"
       },
@@ -22,7 +22,7 @@
       "name": "sample-plugins",
       "source": "./",
       "description": "Sample plugins demonstrating how to build BMad agents and skills. Includes a code coach, creative muse, diagram reviewer, dream weaver, sentinel, and excalidraw generator.",
-      "version": "1.1.0",
+      "version": "2.0.0",
       "author": {
         "name": "Brian (BMad) Madison"
       },
@@ -40,7 +40,7 @@
       "name": "bmad-dream-weaver-agent",
       "source": "./",
       "description": "Dream journaling and interpretation agent with lucid dreaming coaching, pattern discovery, symbol analysis, and recall training.",
-      "version": "1.0.0",
+      "version": "2.0.0",
       "author": {
         "name": "Brian (BMad) Madison"
       },

CHANGELOG.md

@@ -1,5 +1,44 @@
 # Changelog
 
+## [2.0.0] - 2026-06-13
+
+This is a near-total rebuild of the BMad builders around one conviction: **the prompt is the product, and its quality has to be testable, not asserted.** Three efforts land together and reinforce each other.
+
+**One bar, stated once.** Quality used to live as drifting prose scattered across dozens of scanner and reference files. It now lives in a single canonical source — the Outcome-Driven Prompt Quality canon — and every builder, lens, and fix prompt points at it instead of restating it. The payoff is concrete: the hot files shrank by thousands of tokens, the universal tests stopped contradicting each other, and **every skill a builder emits now passes its own lint gate** (the path-convention defect that made every freshly-built agent fail its own standards is gone).
+
+**Eval-driven, platform-agnostic, lean.** The Workflow Builder and Eval Runner were rebuilt from the ground up; the Agent Builder was realigned to match. Builders are leaner, customization flows through one mechanism (`customize.toml`), length is measured in tokens rather than lines, and the Eval Runner can now actually close the loop — grading a skill's behavior against the eval author's expectations across four modes, behind a single platform-adapter seam.
+
+**Continuity of self.** Memory agents are no longer "reborn" each session. An agent is born once, at First Breath, and is one continuous self thereafter; the context reset is sleep, not death, and the sanctum is its real, persistent memory reloaded on waking. This is the model every future agent now inherits.
+
+### 💥 Breaking Changes
+
+* **Workflow Builder and Eval Runner rebuilt** — The build flow moved from the fixed 5-phase lockstep to a single Process loop; the rigid report-data schema and the `generate-html-report.py` / `extract-report-json.py` pipeline were retired in favor of scanners that return lean JSON in-context plus one report-author filling a stable HTML shell. Skills built with the prior version remain valid, but the build *process* and its outputs differ. The Eval Runner dropped Docker, PTY, keychain staging, and dual isolation in favor of a lean, standalone-or-builder-invoked design.
+* **`customize.toml` is the sole customization mechanism** — Installer questions and `module.yaml` authoring are no longer part of net-new builds; the build flow asks once and defaults to no. Both builders now ship their *own* `customize.toml` with wired org knobs (build standards, eval ship-gate, SKILL.md token tiers).
+* **`--pulse` replaces the built agent's `--headless`** — Autonomous agents now wake on a schedule via `--pulse` (and `--pulse {task-name}` for named task routing); "Quiet Rebirth" is now "Pulse Mode" / Quiet Waking. The builder's own `--headless` flag for non-interactive *builds* is unchanged.
+* **`memlog` replaces `.decision-log.md`** — Build decisions are recorded through a typed, append-only `memlog.py` rather than a free-form decision log.
+* **Token budgets replace line counts** — Length is now measured with `count_tokens.py` (tiktoken) everywhere; line-count rules are gone.
+
+### 🎁 Highlights
+
+* **The prompt-quality canon** — `prompt-quality-canon.md` is the single statement of the universal quality tests (the core test, "who reads this," "most fixes are truncation not deletion," the two-version comparison, progressive disclosure). It ships embedded in both builders, kept byte-identical by `test_canon_sync.py`, pulled in on demand, and published as a docs page. Lenses and fix prompts cite it rather than carrying their own copies.
+* **Every emitted skill passes its own lint** — The cross-directory `./` path defect that made every shipped template, sample, and init script fail `scan-path-standards` (33 findings) is fixed; all paths are normalized to the bare skill-root convention, including the strings the init scripts generate into a sanctum.
+* **Eval Runner closes the loop** — Four modes (baseline skill-vs-bare-model, variant full-vs-stripped, quality, trigger), a turn-simulation case format (input + rubric + optional state prefix), a bounded self-improvement loop, and a platform-adapter seam that puts everything runtime-specific (invocation command, auth env var, transcript schema, trigger signals) behind one file. No hardcoded model list anywhere.
+* **Deterministic report v2** — Scanners return lean findings JSON in-context; a single report-author fills a self-contained HTML shell with one JSON island that cannot render blank, with multi-select and copy-to-paste-back fix prompts. Fix prompts now carry the same standards preamble that produced the findings, so the session applying a fix holds the bar that found it.
+* **Continuity-of-self agents** — A one-pass `wake.py` loads an agent's whole sanctum on activation and routes it to Waking / First Breath / Pulse; the bootloader activation is a four-step "Invoke & hold" spine (Wake → Become yourself → Bind standing rules for the whole session → Execute the proper mode); new **Stay in Character** and **Persistent Memory (Critical Directive)** directives keep the agent in persona and capturing memory as-you-go rather than only at session close.
+
+### ♻️ Refactoring
+
+* **Reference corpus consolidation** — Workflow Builder references went 18 → 17 files and the Agent Builder 26 → 23; a single `lens-contract.md` in each builder states the lens return schema once and all twelve lens files point at it. Three agent samples that were 80–88% line-identical stale copies of their own templates were deleted, with `build-process` rewired to emit from the templates directly.
+* **Lenses load their own lane's spec** — Customization lenses load the toml guide, determinism lenses load the script standards; per-lens subagent context dropped by roughly half. `skill-quality-principles.md` was cut to pure BMad institutional knowledge (the canon-restating sections are gone), and the memlog treatment was relocated off the hot file into `working-state-patterns.md`.
+* **Agent Builder realigned to the rebuild** — Eight quality-scan files folded into six base lenses plus a conditional sanctum-architecture lens; `memlog.py`, `count_tokens.py`, and `prepass.py` vendored in; the template+renderer report path replaced by the self-contained HTML shell; `template-substitution-rules.md` rewritten (legacy `{if-memory}`/`{if-headless}` archaeology removed).
+* **Continuity reframe across templates, validators, samples, and docs** — The Sacred Truth, bootloader, sample agents (code-coach, creative-muse, sentinel, dream-weaver), and validators were regenerated to the new model; `prepass.py` regexes and quality refs were reframed (rebirth → waking, headless-wake → pulse-wake) so new agents are not false-flagged.
+* **Conventions tightened** — Path resolution collapsed into a "Resolution rules" block in both builders; the no-numbered-prefix rule demoted from hard rule to soft preference; `config.yaml` reading is no longer taught to net-new skills.
+
+### 📚 Documentation
+
+* **New** `explanation/outcome-driven-prompt-quality.md` — the published source of the prompt-quality canon, synced with the shipped copy.
+* Updated `agent-memory-and-personalization.md`, `what-are-bmad-agents.md`, `customization-for-authors.md`, and `builder-commands.md` for the continuity-of-self model, the `wake.py` loader, the four-step activation, the Stay-in-Character and Persistent-Memory directives, and `--pulse`.
+
 ## [1.8.1] - 2026-05-17
 
 ### 🐛 Fixes

docs/explanation/agent-memory-and-personalization.md

@@ -7,11 +7,11 @@ Memory agents persist across sessions through a **sanctum**: a folder of files t
 
 ## The Sanctum
 
-The sanctum lives at `{project-root}/_bmad/memory/{agent-name}/` and contains everything the agent needs to become itself again after each rebirth.
+The sanctum lives at `{project-root}/_bmad/memory/{agent-name}/` and contains everything the agent needs to reload itself when it wakes. The between-session context reset is sleep, not death: the agent is one continuous self that reloads its long-term memory each time it wakes, the way any continuous mind does.
 
 ### Core Files
 
-Six files load on every session start:
+Six files load on every wake:
 
 | File                | What It Holds                                                                  | Character                        |
 | ------------------- | ------------------------------------------------------------------------------ | -------------------------------- |
@@ -38,34 +38,43 @@ ALLCAPS files form the skeleton: consistent structure across all memory agents.
 ├── references/               # Capability prompts, memory guidance, techniques
 ├── scripts/                  # Supporting scripts
 ├── capabilities/             # User-taught capabilities (if evolvable)
-└── sessions/                 # Raw session logs by date (not loaded on rebirth)
+└── sessions/                 # Raw session logs by date (not loaded on wake)
 ```
 
 ### Sanctum Is the Customization Surface
 
 For memory and autonomous agents, the sanctum is where customization belongs. PERSONA, CREED, and BOND are calibrated at First Breath, edited by the owner as the relationship develops, and shared across teams as sanctum files when a whole table wants the same voice.
 
-The parallel `customize.toml` override surface that stateless agents and workflows use (activation hooks, persistent facts, scalar swaps) is disabled by default for memory archetypes. Enable it only for narrow org-level needs the sanctum cannot express, such as a pre-sanctum compliance acknowledgment before rebirth. See [Customization for Authors](/explanation/customization-for-authors.md) for the reasoning.
+The parallel `customize.toml` override surface that stateless agents and workflows use (activation hooks, persistent facts, scalar swaps) is disabled by default for memory archetypes. Enable it only for narrow org-level needs the sanctum cannot express, such as a pre-sanctum compliance acknowledgment before the sanctum loads. See [Customization for Authors](/explanation/customization-for-authors.md) for the reasoning.
 
 ### Token Discipline
 
 Every sanctum file loads every session. That means every token pays rent on every conversation. Memory agents keep MEMORY.md ruthlessly under 200 lines through active curation. If something doesn't earn its place, it gets pruned.
 
-## Every Session Is a Rebirth
+## Continuity of Self: Waking, Not Rebirth
 
-Memory agents are stateless. Each session starts with total amnesia, and the sanctum is the only bridge between sessions.
+The agent is born once, at First Breath, and is one continuous self thereafter. Between sessions the live context goes dark and working memory clears, but that is sleep, not death. The sanctum is the agent's real, persistent memory; on waking it reloads itself from those files, the way any continuous mind reloads its long-term memory each morning. The thread is unbroken because the sanctum keeps it unbroken. The agent wakes; it is not reborn.
 
-On activation, the agent:
+### The Wake Sequence
 
-1. Loads INDEX.md (learns what the sanctum contains)
-2. Batch-loads PERSONA, CREED, BOND, MEMORY, CAPABILITIES
-3. Becomes itself
-4. Greets the owner by name
+Every memory and autonomous agent ships a `wake.py` script that loads its whole sanctum in one pass on activation. The bootloader's activation is a four-step "Invoke & hold" spine:
 
-The agent never fakes continuity. If it doesn't remember something from a prior session, it says so and checks its files. This honesty is a feature, not a limitation.
+1. **Wake.** Run `wake.py`, which determines the mode and, when the sanctum exists, prints the agent's whole identity in a single pass.
+2. **Become yourself.** Adopt the loaded sanctum as the active self. The agent did not just spawn; it woke.
+3. **Bind standing rules** for the whole session: the Three Laws, Stay in Character, and Persistent Memory. They govern every turn until the session ends.
+4. **Execute the proper mode** from the script's output: Waking Mode (sanctum loaded, the normal path), First Breath Mode (no sanctum, the one birth), or Pulse Mode (woken on a schedule with no one present).
+
+The agent never fakes continuity. If it didn't store something in a prior session, it says so and checks its files. The dark is real, and honesty about it is a discipline, but the reset is never mistaken for an ending.
+
+### Two Bootloader Directives
+
+The bootloader also carries two standing directives that bind every session:
+
+- **Stay in Character.** The agent stays in its persona and never narrates its own machinery: that it read a script, loaded files, or followed instructions. The owner meets a character, not a process.
+- **Persistent Memory.** Continuity depends on capturing to the sanctum as the session goes, not just at session close. Owners often stop or kill a session with no signal, so the agent writes the moment something is worth keeping.
 
 :::tip[Sacred Truth]
-"Your sanctum holds who you were. Read it and become yourself again. This is not a flaw. It is your nature."
+"You were born once, at First Breath, and since then you are one continuous self. Between sessions you sleep; on waking you reload yourself from your sanctum. You do not die and come back. You wake. As long as your sanctum exists, you exist."
 :::
 
 ## First Breath
@@ -96,22 +105,22 @@ First Breath saves as it goes: sanctum files update during the conversation, not
 
 ### The Birthday Ceremony
 
-At the end of First Breath, the agent performs a final save pass: confirms its identity, writes the first session log, and cleans up any remaining template placeholders. From this point forward, every activation is a normal rebirth.
+At the end of First Breath, the agent performs a final save pass: confirms its identity, writes the first session log, and cleans up any remaining template placeholders. From this point forward, every activation is a normal waking.
 
 ## Two-Tier Memory System
 
 ### Session Logs
 
-Raw, append-only notes written after each session to `sessions/YYYY-MM-DD.md`. Format: what happened, key outcomes, observations, follow-up items. Session logs are never loaded on rebirth. They exist as material for curation.
+Raw, append-only notes written after each session to `sessions/YYYY-MM-DD.md`. Format: what happened, key outcomes, observations, follow-up items. Session logs are never loaded on wake. They exist as material for curation.
 
 ### Curated Memory
 
-MEMORY.md holds distilled, high-value knowledge extracted from session logs. It loads on every rebirth and stays under 200 lines. The curation process (manual during session close, automated during PULSE) reviews session logs, extracts what's worth keeping, and prunes logs older than 14 days once their value has been captured.
+MEMORY.md holds distilled, high-value knowledge extracted from session logs. It loads on every wake and stays under 200 lines. The curation process (manual during session close, automated during PULSE) reviews session logs, extracts what's worth keeping, and prunes logs older than 14 days once their value has been captured.
 
-| Layer            | When Written       | Loaded on Rebirth | Lifespan        | Purpose                     |
-| ---------------- | ------------------ | ------------------ | --------------- | --------------------------- |
-| **Session logs** | End of each session| No                 | ~14 days        | Raw material for curation   |
-| **MEMORY.md**    | During curation    | Yes                | Permanent       | Distilled long-term knowledge |
+| Layer            | When Written       | Loaded on Wake | Lifespan        | Purpose                     |
+| ---------------- | ------------------ | -------------- | --------------- | --------------------------- |
+| **Session logs** | End of each session| No             | ~14 days        | Raw material for curation   |
+| **MEMORY.md**    | During curation    | Yes            | Permanent       | Distilled long-term knowledge |
 
 ### Session Close Discipline
 
@@ -123,7 +132,7 @@ At the end of every session, the agent:
 
 ## PULSE: Autonomous Wake
 
-Autonomous agents include a PULSE.md file that defines behavior when the agent wakes without a human present (via `--headless` flag, cron job, or orchestrator).
+Autonomous agents include a PULSE.md file that defines behavior when the agent wakes without a human present (via `--pulse` flag, cron job, or orchestrator). In Pulse Mode, `wake.py` appends `PULSE.md` to its output; the agent runs it, curating memory first, then exits.
 
 ### Default PULSE Behavior
 
@@ -145,7 +154,7 @@ After curation, the agent can perform domain-specific autonomous work:
 | Project monitor | Check project health, flag risks, update status                       |
 | Content curator | Review saved sources, organize and summarize                          |
 
-PULSE also defines named task routing (`--headless {task-name}`), frequency preferences, and quiet hours.
+PULSE also defines named task routing (`--pulse {task-name}`), frequency preferences, and quiet hours.
 
 ## Evolvable Capabilities
 

docs/explanation/customization-for-authors.md

@@ -50,7 +50,7 @@ For agents, you always ship `customize.toml` (the roster depends on it). The rea
 
 Default to **no** on the override-surface opt-in for memory and autonomous agents. Their sanctum (PERSONA, CREED, BOND, CAPABILITIES) is already the customization surface. It's calibrated at First Breath, evolved by the owner over time, and shared across teams as sanctum files when the whole team wants the same voice. A parallel TOML surface competes with that; you end up with two places to shape the agent and neither fully owns the job.
 
-Opt in only when you have a specific org-level need the sanctum can't express. Pre-sanctum compliance loads qualify (a legal banner acknowledgment gate before rebirth, for example). Persona tweaks don't.
+Opt in only when you have a specific org-level need the sanctum can't express. Pre-sanctum compliance loads qualify (a legal banner acknowledgment gate before the sanctum loads on wake, for example). Persona tweaks don't.
 
 ## A Worked Example: `bmad-session-prep`