Builder outcome improvements: continuity-of-self agent model, canon-first builders#95
Merged
Merged
Conversation
…tic, lean workflow-builder: - New leanness/outcome scanner owns the minimal-baseline tests (core test, defend-against-its-own-absence, outcome-vs-prescription) that nothing applied before - Report pipeline collapsed: scanners return lean JSON in-context, one report-author fills a stable self-contained HTML shell with a single JSON island (cannot render blank; multi-select + copy-to-paste-back-prompt baked in). Retired generate-html-report.py, extract-report-json.py, the rigid report-data schema, report-quality-scan-creator.md - memlog replaces .decision-log.md (scripts/memlog.py, typed append-only) - customize.toml is the sole customizability mechanism; build flow asks (default no, headless no unless requested); no installer questions, no module.yaml authoring - tiktoken token counts replace line counts as the length metric (scripts/count_tokens.py) - Build flow rewritten from the 5-phase lockstep into the migration-guide Process loop - Enhancement scanner flipped to add-or-subtract; standard-fields numbered-steps fix - Script-opportunity seeking kept and strengthened (native python) eval-runner: - Rebuilt lean: dropped Docker, PTY, keychain staging, dual isolation - Standalone and builder-invoked; everything runtime-specific behind one platform-adapter seam (invocation command, auth env-var, transcript schema; skill_dir/load_signal for triggers) - Four modes: baseline (skill vs bare model), variant (full vs stripped), quality, trigger - Case = input + rubric + optional state_prefix (turn-simulation); bounded self-improvement loop - No hardcoded model list anywhere
Rename the Conventions heading to Resolution rules (binding, covers non-path variables) and collapse the redundant bare-path/skill-root bullets into one with an explicit not-the-project-directory guard.
Replace the pre-rebuild quality and infrastructure layer while keeping the agent-specific core (stateless/memory/autonomous gradient, sanctum, First Breath, persona-as-deliverable, script discipline). - Fold eight quality-scan files into six base lenses plus a conditional sanctum-architecture lens; each loads one shared agent-quality-principles.md and returns findings JSON in-context. - Add the gospel prompt-quality canon (references/prompt-quality-canon.md, pulled in on demand via a CREED standing order) so evolving agents stay lean instead of carrying a stale standards copy. - Vendor memlog.py and count_tokens.py from the workflow-builder; add prepass.py to gate the sanctum lens; switch length rules to tiktoken token budgets, never line counts. - Replace the template+renderer report path with a stable self-contained HTML shell plus a single JSON island and a report-author. - Rename the Conventions heading to Resolution rules and compress the path-resolution bullets to match the workflow-builder.
… patterns Customization - Ship the builder's own customize.toml with org knobs and wire each one: build_standards (enforced build criteria + Analyze conformance), evals_required (eval ship-gate), skill_md_token_desired/budget (tiered SKILL.md length), plus the standard activation/persistent_facts/on_complete. - Replace the re-typed customize merge-rule prose in activation with the one-line resolve_customization.py call. Reference sweep (bloat + reader-relative tenet) - skill-quality-principles.md gains a reader-is-an-LLM framing (cut what the reader knows; keep the why a goal needs to generalize) and the leanness/ enhancement lenses now cut both ways. - Cut duplication: standard-fields.md defers customize/path/description to their owners (-40%); script-opportunities defers authoring to script-standards; report-author and scan-orchestration de-duplicated. - scan-architecture rewritten to cite rather than restate. Working state - New working-state-patterns.md frames the choice (memlog vs structured working artifact vs both vs neither); relocate the full memlog treatment out of skill-quality-principles.md (~900 tokens off the hot file). - New producing-workflow-patterns.md (persona, intent modes, degradation). Conventions - Resolution rules heading and compressed path block. - No-numbered-prefix demoted from hard rule to soft preference everywhere. - config.yaml reading no longer taught to net-new skills. Handoff - Finalize now proactively offers full validation and opens the HTML report. - Add a 'harden the idea before you build it' beat to resist quiz-and-go.
- Replace the re-typed customize merge-rule prose in activation with the one-line resolve_customization.py call. - Sync the gospel canon shipped copy with the new 'Who reads this' section (write for the LLM reader; cut what it knows, keep the why a goal needs to generalize). - Make the leanness lens cut both ways: flag self-describing/negative-space/ wrong-altitude noise and a goal stripped of the why it needs (persona carve-out preserved).
Add a 'Most fixes are truncation, not deletion' section to the prompt-quality canon: keep the nudge, drop the over-explanation the reader infers, and cap the kept why at one clause rather than a justification essay. Propagate to the binary 'cut it' language so it changes behavior: the workflow-builder Core Test and both leanness lenses now recommend shrinking to the nudge and reserve outright removal for lines the model already does entirely.
Add docs/explanation/outcome-driven-prompt-quality.md, the published source of the Outcome-Driven Prompt Quality canon, synced with the shipped copy: the core test, Who reads this, Most fixes are truncation, the two-version comparison, progressive disclosure, and the habit.
…port fixes The canon (prompt-quality-canon.md) is now the sole statement of the universal quality tests; everything else points at it instead of restating it. Deduplication and re-routing: - skill-quality-principles.md cut to pure BMad institutional knowledge (3,943 -> 2,355 tokens); the canon-restating sections are gone - Both scan-leanness lenses cut to canon + lane mechanics only - Agent script-opportunities-reference rewritten from the retired-script graveyard to the live lean shape (2,720 -> 1,039 tokens) - Lenses load their lane's spec directly: customization -> the toml guide, determinism -> script-opportunities; per-lens subagent context drops by roughly half - Both build-process files open with a standalone canon mandate and no longer re-teach the small-version recipe, the two-version comparison, or the customization ask their spec files own Every path now carries the bar: - Agent edit-guidance loads canon + principles (it loaded no standard at all); numbered heading prefixes removed - Agent standard-fields drops the madlib Overview templates and the third statement of the sanctum-vs-override rationale Report fix prompts anchor to the standards: - findings.json gains a standards block (canon / principles / script-standards absolute paths, filled by the orchestrator) - Both report shells prepend it to every copied fix prompt, so the session applying a fix holds the same bar that produced the findings - Verified end-to-end in Chromium: theme-copy and findings-copy both open with the preamble in both shells Also lands the report v2 pipeline (deterministic render_report.py with synthesis layer, placeholder islands, run-folder + headless contracts), the eval-runner adapter seam with env isolation and trigger detection tests, builder ideation beats (propose-what-the-idea-implies, show-the-draft), agent-builder customize.toml (fixes activation step 1), process-template.py leftover-marker gate, the embedded canon copies guarded by test_canon_sync.py, and removal of the three superseded prepass scripts.
… their own lint
Corpus reorg (workflow 18 -> 17 files, agent 26 -> 23):
- lens-contract.md (both builders): single statement of the lens return
schema, in-context rule, id numbering, and don't-pad rule; all 12 lens
files point at it instead of carrying their own drifting schema copies
- report-author.md merged into its only loader (scan-orchestration /
quality-analysis) in both builders
- complex-workflow-patterns shrunk to its unique routing mechanics
(750 -> 455); deleted workflow's orphaned template-substitution-rules
- Deleted the three agent samples that were 80-88% line-identical
stale copies of their assets/ templates (sample-first-breath,
sample-memory-guidance, sample-init-sanctum.py); build-process now
wires emission to the templates, which were previously unloaded
Path-convention defect fixed (33 scanner findings -> 0):
- Every shipped template, sample, and init script used cross-directory
./ paths that scan-path-standards flags high, so every emitted agent
was born failing its own lint gate; all normalized to bare skill-root
paths, including the strings both init scripts generate into sanctums
- Two bare-_bmad path defects fixed; teaching examples rephrased so
they no longer literal-match the scanner
Agent reference staleness:
- template-substitution-rules rewritten (1151 -> 831): legacy
{if-memory}/{if-headless} archaeology deleted, Path References now
states the bare-path convention, process-template.py invocation up
front
- "Phase 3/5" references removed from mission-writing and first-breath
guidance (build-process is one loop, not phases)
- Customization-by-archetype ownership settled: principles owns the
contract, standard-fields owns the schema; agent-type-guidance keeps
only its discovery-time naming rule
agentskills.io adoptions (workflow-builder):
- build-process: "Ground it in real expertise" beat (mine runbooks,
incident reports, review comments, VCS history, done-by-hand
transcripts; generalize beyond the worked instance) and trace-reading
diagnostics at the eval beat
- skill-quality-principles: gotchas-stay-in-SKILL.md carve exception,
plan-validate-execute pattern, defaults-not-menus writing rule
SKILL.md token tiers raised: desired 1200 -> 2000, hard budget
2000 -> 3000 (still under the spec's 5k; budget is a drift guardrail,
the canon's tests remain the leanness bar).
Byte-identical copy of the canonical workflow-builder script that nothing referenced; transcript token accounting is inlined in run_evals.py. Eval-runner tests still pass.
Replace the "every session is a rebirth" framing with a continuous-self model across the agent-builder templates, validators, sample agents, and docs. An agent is born once at First Breath and is continuous thereafter; the context reset is sleep, not death, and the sanctum is its real persistent memory reloaded on waking. Templates (what future agents inherit): - SKILL-template-bootloader.md: continuity Sacred Truth, Stay in Character, Persistent Memory (Critical Directive), and a 4-step On Activation router (Wake -> Become yourself -> Bind standing rules every turn -> Execute the Proper Mode) that binds the standing rules for the whole session. - wake-template.py (new): one-pass sanctum loader; routes to Waking / Pulse / First Breath. Renames the runtime wake flag --headless -> --pulse. - CREED/PULSE/memory-guidance/first-breath templates updated to match. Validators & quality refs: widen allowed bootloader contents, drop the session-close-bloat flag, expect wake.py, and reframe prepass.py regexes (rebirth -> waking, headless-wake -> pulse-wake) so new agents aren't false-flagged. Samples (code-coach, creative-muse, sentinel, dream-weaver) and docs regenerated to the new model.
|
Important Review skippedToo many files! This PR contains 154 files, which is 4 over the limit of 150. To get a review, narrow the scope: ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (154)
You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
This pull request is abnormally large and would use a significant amount of tokens to review. If you still wish to review it, comment "augment review" and we will review it. |
Contributor
Author
|
augment review |
🤖 Augment PR SummarySummary: This PR reworks the BMad agent architecture around a "continuity-of-self" model — agents are born once at First Breath and wake thereafter, rather than being "reborn" each session. The sanctum becomes real persistent memory reloaded on waking. Changes:
Technical Notes: All sample agents (code-coach, creative-muse, sentinel, dream-weaver) regenerated. Multiple prepass scripts consolidated into one. Token counts via tiktoken (with chars/4 fallback) replace line counts everywhere. The memlog script ( 🤖 Was this summary useful? React with 👍 or 👎 |
![augmentcode[bot]](https://avatars.githubusercontent.com/in/1027498?s=60&v=4){kind=small}
| if isinstance(usage, dict): | ||
| found_usage = True | ||
| # result usage is authoritative; prefer it over the running sum | ||
| input_tokens = int(usage.get("input_tokens", input_tokens) or input_tokens) |
There was a problem hiding this comment.
The or input_tokens (and or output_tokens) pattern means an authoritative result event reporting 0 tokens is silently discarded in favor of the accumulated per-message sum, contradicting the comment on the line above. usage.get("input_tokens", input_tokens) already falls back to input_tokens when the key is absent, so the or only fires on falsy values — catching both None and the valid count 0.
Severity: medium
🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.
| 2. Check if `{project-root}/_bmad/config.yaml` exists — if a section matching the module's code is already present, inform the user this is an update (reconfiguration) | ||
|
|
||
| If the user provides arguments (e.g. `accept all defaults`, `--headless`, or inline values like `user name is BMad, I speak Swahili`), map any provided values to config keys, use defaults for the rest, and skip interactive prompting. Still display the full confirmation summary at the end. | ||
| If the user provides arguments (e.g. `accept all defaults`, `--pulse`, or inline values like `user name is BMad, I speak Swahili`), map any provided values to config keys, use defaults for the rest, and skip interactive prompting. Still display the full confirmation summary at the end. |
There was a problem hiding this comment.
This --headless → --pulse replacement appears semantically wrong: the context is about skipping interactive prompts during module configuration, not about the agent's autonomous Pulse Mode wake. The PR description explicitly states "The builder's own --headless flag for non-interactive builds is unchanged," and --pulse specifically means scheduled autonomous wake with no one present, which is not what module setup intends here.
Severity: medium
🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.
| # Persistent facts the agent keeps in mind for the whole session | ||
| # (org rules, domain constants, user preferences). Overrides append. | ||
| # These are static build-time config loaded on activation. They are not | ||
| # the sanctum: the sanctum is the agent's runtime memory across rebirths, |
There was a problem hiding this comment.
This newly added comment says "across rebirths" while the entire PR reworks the terminology from "rebirth" to "waking / continuity-of-self." Consider using "across wakings" to stay consistent with the new model.
Severity: low
🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.
| @@ -0,0 +1,78 @@ | |||
| # vendored from bmad-workflow-builder/scripts; canonical source there | |||
There was a problem hiding this comment.
The vendoring comment on line 1 pushes the shebang (#!/usr/bin/env python3) to line 2, where the OS cannot find it for direct execution (./scripts/count_tokens.py). The canonical copy at bmad-workflow-builder/scripts/count_tokens.py correctly has the shebang on line 1.
Severity: low
🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.
| ## On Quiet Waking | ||
|
|
||
| When invoked via `--headless` without a specific task, load `./references/memory-guidance.md` for memory discipline, then work through these in priority order. | ||
| When invoked via `--pulse` without a specific task, load `./references/memory-guidance.md` for memory discipline, then work through these in priority order. |
There was a problem hiding this comment.
This still uses the ./references/ prefix, while the other sample PULSE templates (code-coach, creative-muse) were updated in this PR to use the bare references/ path convention. The path-standards scanner would flag this.
Severity: low
Other Locations
samples/bmad-agent-code-coach/assets/PULSE-template.md:7
🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.
- dream-weaver/module-setup.md: revert --pulse -> --headless (this is the config non-interactive flag, not the agent's autonomous Pulse wake) - customize-template.toml: "across rebirths" -> "across wakings" for consistency with the continuity-of-self model - count_tokens.py: move shebang back to line 1 (vendoring comment had pushed it to line 2) - sentinel/PULSE-template.md: ./references/ -> bare references/ path; align memory-curation limit to token budget - eval-runner/adapter-claude-code.json: prettier formatting (fixes the failing lint check)
- CHANGELOG.md: add the 2.0.0 release section — the builder rebuild, the prompt-quality canon, and continuity-of-self, with breaking changes, highlights, refactoring, and docs - marketplace.json: bump all three plugin versions to 2.0.0 package.json version is left to the release process.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Outcome-driven improvements to the BMad builders. The latest and headline change reworks the agent birth model from rebirth to continuity-of-self; earlier commits on the branch rebuild the workflow-builder, align the agent-builder, and harden the prompt-quality canon across both.
Continuity-of-self (latest)
An agent is born once at First Breath and is continuous thereafter. The context reset is sleep, not death; the sanctum is its real persistent memory, reloaded on waking.
wake-template.py(new): one-pass sanctum loader routing to Waking / Pulse / First Breath. Runtime wake flag renamed--headless→--pulse(distinct from the builder's own--headlessbuild flag).wake.py, reframeprepass.pyregexes — so new agents aren't false-flagged.Earlier on the branch
Verification
wake.pycompile; sentinel test suite 10/10 green;prepass.pycompiles.