菜单
WHEN: You encounter an unfamiliar Forge term and need its canonical definition.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | forge-glossary |
| description | WHEN: You encounter an unfamiliar Forge term and need its canonical definition. |
| type | reference |
| version | 1.0.13 |
| preamble-tier | 1 |
| triggers | ["what does X mean","forge glossary","define forge term","forge terminology","product terminology","terminology.md","per-task terminology"] |
| allowed-tools | ["Bash","Read","AskUserQuestion"] |
Optional disambiguation via AskUserQuestion (allowed-tools) uses the same host mapping as all rigid skills — skills/using-forge/SKILL.md Blocking interactive prompts. See the term Blocking interactive prompt below.
Definition: The host-specific mechanism that gathers a human answer without advancing until answered — canonical skill tool AskUserQuestion (see allowed-tools); Cursor maps to AskQuestion; hosts without that tool use numbered choices in chat + stop. Full host table: using-forge Blocking interactive prompts. Applies to every Forge-supported IDE, not one vendor.
Cross-References: using-forge (Interactive human input, Stage-local questioning).
terminology.md) — not this glossaryDefinition: The per-task product term sheet at ~/forge/brain/prds/<task-id>/terminology.md (table of canonical names, disallowed variants, open_doubts in frontmatter). Authored in intake / aligned at council; used by planning, QA, and assertion text so human-facing copy matches contracts and the PRD lock.
Usage context: This forge-glossary skill documents Forge plugin and pipeline terms only. For branded or product vocabulary, read terminology.md and docs/terminology-review.md — not this file.
Cross-References: intake-interrogate, council-multi-repo-negotiate, docs/terminology-review.md, docs/templates/terminology.md.
Definition: In assistant dialogue (not static README/commands), name only the immediate next prerequisite, artifact, or skill — or a downstream step when the current question truly depends on it, or when the human asked “what comes next.” Do not preemptively enumerate full pipelines (council → tech plans → merge → …, or product-specific chains) during intake, council, planning, QA, or other upstream gates.
Usage Context: Reduces confusion when the user is mid-interrogation and the model front-loads later phases. Complements Stage-local questioning in using-forge.
What It's NOT: Not a ban on documentation — README.md, commands/forge.md, and brain templates may list full order. Not “never mention a later stage” — only don’t narrate it before the user is on that gate without cause.
Cross-References: using-forge (Horizon narration, Multi-question elicitation item 5). Canonical doc: docs/forge-one-step-horizon.md — also No defensive downstream-gate narration (repo-wide).
Definition: Assistant prose whose primary job is to explain why a later gate or artifact does not exist yet (semantic manifest isn’t ready yet, orphan automation, pasting a full State 4b chain, merge previews) while the human is still answering earlier phase questions.
Usage Context: Forbidden as default filler between sequential elicitation turns — any Forge phase (intake, qa-prd-analysis, council, planning, …). Allowed: static commands/ / README / SKILL.md tables; skip-ahead refusal (first missing prerequisite + next action); user explicitly asked why or full order.
Cross-References: docs/forge-one-step-horizon.md No defensive downstream-gate narration (repo-wide); using-forge Multi-question elicitation items 7–8.
Definition: One assistant message uses one blocking interactive affordance (e.g. AskQuestion for a single fork) while treating other needle-moving choices as prose “also answer…” without their own blocking turn — or pastes phase-specific waiver/roadmap copy from a later gate — or stacks two different primaries (e.g. Questions widget for task-id approval and a long markdown Q1 checklist in the same message). Intake example: task-id modal while Q9 design authority, net-new vs reuse, or Figma locks appear only in prose in the same turn.
Usage Context: Violates using-forge Multi-question elicitation item 6. Correct shape: one fork per turn (or a Confirm/Correct batch only where the active skill allows).
Cross-References: docs/forge-one-step-horizon.md No bundled unrelated decisions; commands/intake.md; intake-interrogate.
Definition: In live chat, when the next message’s job is to get one human answer, the assistant does not prefix with what commands/ or named skills do, which gates are open, or which later artifacts do not exist — unless the user asked for status. Does not paste defensive downstream-gate essays between normal Q→Q turns (forge-one-step-horizon No defensive downstream-gate narration). Does not suffix with not ready for … / needs … first unless that is the immediate blocker or the user asked. Reference material stays in commands/ and README.
Cross-References: using-forge Multi-question elicitation items 7–8; docs/forge-one-step-horizon.md Question-forward elicitation and No defensive downstream-gate narration (repo-wide).
Definition: The first non-skippable phase where a PRD is interrogated through 8 structured questions that lock scope, success criteria, constraints, and cross-service boundaries. Produces a locked PRD artifact.
Usage Context: Triggered at the start of every Forge run. The conductor invokes forge-intake-gate to execute intake. Output is immutable — no scope changes after intake without restarting the pipeline.
What It's NOT: Not a casual discussion or brainstorm. Not optional. Not a place to add features mid-interview. Not informal — every answer is logged in the brain with provenance (who answered, when, why that interpretation).
Cross-References: Enforced by forge-intake-gate (HARD-GATE). Output feeds into Council. Related to PRD lock.
Definition: Multi-surface contract negotiation where 4 domain surfaces (backend, web frontend, app frontend, infrastructure) reason about the locked PRD and agree on 5 contracts (REST APIs, event bus, cache, database schema, search). Produces the shared-dev-spec.
Usage Context: Runs after intake. Each surface reasons independently from its domain perspective, then surfaces negotiate compatibility. Results in a single canonical spec that all repos follow. Cannot proceed to build until council produces a frozen spec.
What It's NOT: Not a debate where one surface "wins." Not technical design — it's contract negotiation (interfaces, not implementation). Not optional; all 4 surfaces must participate. Not quick — requires disciplined negotiation across multiple services.
Cross-References: Enforced by forge-council-gate (HARD-GATE). Uses reasoning-as-backend, reasoning-as-web-frontend, reasoning-as-app-frontend, reasoning-as-infra. Output is the shared-dev-spec. Involves contract skills: contract-api-rest, contract-event-bus, contract-cache, contract-schema-db, contract-search.
Definition: Immutable lock on shared-dev-spec after council completes. Changes are not allowed without full re-negotiation through council. Signals transition from design to implementation.
Usage Context: Invoked via spec-freeze skill after council concludes. Once frozen, the spec becomes the single source of truth for all tasks in Tech Plan. Any surface discovering a conflict during build must escalate (not proceed).
What It's NOT: Not a soft freeze. Not a guideline. Not flexible to "quick fixes." Once frozen, the spec cannot be modified by individual repos or developers — it requires re-opening council.
Cross-References: Output of Council. Input to Tech Plan. Related to D24 (HARD-GATE discipline).
Definition: Per-project breakdown of the shared-dev-spec into bite-sized implementation tasks (2–5 minutes each), with exact code snippets and exact commands. One tech plan per repository.
Usage Context: Generated after spec freeze via tech-plan-write-per-project. Each task is atomic, measurable, and follows a standard format: description, code, commands, success criteria. Dev-implementer consumes these tasks sequentially in isolated worktrees.
What It's NOT: Not a high-level roadmap. Not aspirational. Not flexibility for developers to improvise. Each task has exact code and commands — deviation is a red flag requiring escalation.
Cross-References: Generated from shared-dev-spec. Consumed by Build. Related to D15 (TDD pressure scenarios).
Definition: TDD implementation phase where dev-implementer subagent executes each task from the tech plan in an isolated worktree, writing tests first, then code. Produces code ready for review.
Usage Context: Triggered after tech plan is complete. Dev-implementer is dispatched once per task. Each task runs in a fresh git worktree (D30) with no shared state. Reports status (DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, BLOCKED) at task completion.
What It's NOT: Not exploratory development. Not "let's see what works." Not a place to refactor specs. Not allowed to skip TDD — forge-tdd is a HARD-GATE.
Cross-References: Implements tasks from Tech Plan. Enforced by forge-tdd (HARD-GATE). Output goes to Review. Related to worktree-per-project-per-task (D30).
Definition: End-to-end product test that brings up the full stack (all services), executes evaluation scenarios (user journeys, cross-service flows), and verifies critical success criteria. Produces a verdict (GREEN, YELLOW, RED).
Usage Context: Runs after code is merged. Multiple eval drivers coordinate: API (HTTP), database (MySQL), cache (Redis), event bus (Kafka), search (Elasticsearch), web UI (Chrome DevTools Protocol), mobile (XCTest, ADB). All drivers report results to eval-judge which renders a verdict.
What It's NOT: Not unit testing — it's integration testing at scale. Not optional; all critical scenarios must pass. Not local — assumes full multi-service stack is running. Not quick fixes during eval — failures require escalation to self-heal loop.
Cross-References: Enforced by forge-eval-gate (HARD-GATE). Uses eval drivers: eval-driver-api-http, eval-driver-db-mysql, eval-driver-cache-redis, eval-driver-bus-kafka, eval-driver-search-es, eval-driver-web-cdp, eval-driver-ios-xctest, eval-driver-android-adb. Output feeds into Self-Heal (RED verdict) or Review (GREEN/YELLOW verdict).
Definition: The Forge machine-eval path: qa/semantic-automation.csv executed to produce qa/semantic-eval-manifest.json + qa/semantic-eval-run.log (CSV execution results) — NL-first steps + DependsOn DAG. Orchestrated by qa-semantic-csv-orchestrate / tools/run_semantic_csv_eval.py. eval-judge reads manifest + semantic-eval-run.log.
Usage Context: Gates verify_forge_task.py, review-readiness, prompt-submit-gates ([P4.0-SEMANTIC-EVAL] / [P4.4-EVAL-GREEN] path=semantic). Self-heal uses semantic-eval-run.log JSON lines as primary evidence on RED — not driver tables.
Cross-References: docs/semantic-eval-csv.md, docs/forge-task-verification.md, qa-semantic-csv-orchestrate, eval-judge § Semantic path.
Definition: Task-local CSV under ~/forge/brain/prds/<task-id>/qa/semantic-automation.csv defining semantic automation steps (Id, Surface, Intent, DependsOn, …) per docs/semantic-eval-csv.md. Validated by tools/verify/semantic_csv.py.
What It's NOT: Not manual-test-cases.csv (human TMS-style acceptance).
Definition: JSON under ~/forge/brain/prds/<task-id>/qa/semantic-eval-manifest.json recording schema_version, task_id, kind (e.g. semantic-csv-eval), outcome (pass | fail | yellow), recorded_at. Required for machine verification. semantic-eval-run.log holds per-step JSON lines for the run.
Definition: conductor.log marker logged after ~/forge/brain/prds/<task-id>/qa/manual-test-cases.csv has ≥1 approved row and Step 7 approval is granted in qa-manual-test-cases-from-prd. Format: [P4.0-QA-CSV] task_id=<id> rows=<n> approved=yes. When logging skipped=not_required, it is only valid on partial runs (/plan, /build, etc.) when forge_qa_csv_before_eval is false/unset in product.md. Never log skipped=not_required on a full /forge run.
Prerequisite for: [P4.0-SEMANTIC-EVAL] must come after this marker. conductor-orchestrate State 4b enforces the ordering.
Cross-References: qa-manual-test-cases-from-prd; docs/conductor-log-format.md; conductor-orchestrate State 4b step 0.
Definition: conductor.log marker logged after valid qa/semantic-eval-manifest.json exists on disk — State 4b machine-eval gate. qa/semantic-eval-run.log is the per-step CSV execution trace; commit it with the manifest whenever the runner produced it. Parsed by prompt-submit-gates.cjs (GATE_PATTERNS.SEMANTIC_EVAL).
Definition: conductor.log marker logged after TDD RED phase is confirmed per repo — failing tests have been written and watched fail before any production implementation code is committed. Format: [P4.0-TDD-RED] task_id=<id> repo=<role> test_files=<list> red_confirmed=yes. One marker per repo involved in the task.
Prerequisite for: [P4.1-DISPATCH] — no production feature code dispatched until RED is confirmed.
Written by: forge-tdd Output section (per the skill's Output instruction).
Cross-References: forge-tdd § Output; forge-verification § Phase Authority Check; conductor-orchestrate State 4b step 2.
Definition: conductor.log marker logged after design parity check completes per repo during Phase 4.2 Review. Format: [P4.2-DESIGN-PARITY] task_id=<id> repo=<role> reviewer=design-implementation-reviewer|figma-design-sync|skipped result=PASS|FAIL|SKIP. Only written when design_new_work: yes and the repo is web or app (no design_waiver: prd_only).
When result=SKIP: Harness not available — human sign-off required before advancing to Phase 4.3.
Cross-References: conductor-orchestrate § Phase 4.2; docs/conductor-log-format.md.
kind)Definition: Value for kind in semantic-eval-manifest.json when the manifest describes semantic CSV automation. verify_forge_task.py expects qa/semantic-automation.csv when this kind is set.
Definition: Automated fault-finding and repair loop triggered by a RED eval verdict. Sequences: locate fault → triage → fix → verify. Max 3 retries before escalation to human.
Usage Context: When eval returns RED, self-heal is invoked. self-heal-locate-fault identifies which service failed, self-heal-triage classifies the failure, self-heal-systematic-debug repairs, then eval re-runs. If 3 retries fail, escalates (BLOCKED).
What It's NOT: Not a blanket retry mechanism. Not allowed to modify the spec. Not permitted to work around failures — must find root cause. Not infinite retries — capped at 3 by self-heal-loop-cap.
Cross-References: Triggered by RED eval verdict. Uses self-heal-locate-fault, self-heal-triage, self-heal-systematic-debug, self-heal-loop-cap. Can escalate to Review after 3 retries fail.
Definition: Two-stage code quality gate. First stage: spec-reviewer verifies implementation matches shared-dev-spec line-by-line in actual code. Second stage: code-quality-reviewer checks 8-point quality framework, performance, security, and observability.
Usage Context: Runs after eval passes (GREEN or YELLOW). Both reviewers read actual code (D14: trust code), not reports. Produces APPROVED or CHANGE_REQUESTED. Required before PR merge.
What It's NOT: Not a formality. Not allowed to approve without reading code. Not a place to nitpick style — only substantial quality and spec compliance. Not optional; all PRs must pass both stages.
Cross-References: Triggered by passing eval. Enforced by forge-trust-code (HARD-GATE). Uses spec-reviewer and code-quality-reviewer subagents. Output feeds into PR Set merge coordination.
Definition: Post-merge retrospective where the dreamer subagent scores decisions, extracts patterns, and writes learnings to the brain. Captures what worked, what failed, and why.
Usage Context: Runs after all PRs in the PR Set are merged and feature is live. Dreamer reviews eval results, conflict resolutions, code review feedback, and produces structured learnings. These learnings inform future PRD interpretations and skill enhancements.
What It's NOT: Not a blame session. Not informal chat. Not optional — every shipped feature produces brain artifacts. Not skipped even for "small" features.
Cross-References: Triggered after PR Set merge. Uses dream-retrospect-post-pr skill. Outputs to brain via brain-write. Related to brain-recall for future pattern matching.
Definition: Coordinated set of pull requests across multiple repositories that must be merged in dependency order (services that others depend on merge first). Ensures cross-service compatibility during merge.
Usage Context: After review passes, all PRs are staged as a PR Set. pr-set-coordinate creates all PRs simultaneously, then pr-set-merge-order determines merge sequence. Merges proceed in order; downstream repos can only merge after dependencies are live.
What It's NOT: Not independent per-repo PRs. Not "merge whenever." Not allowed to reorder without service team sign-off. Not skipped for "simple" features.
Cross-References: Output of Review. Uses pr-set-coordinate and pr-set-merge-order. Related to council (which negotiates service boundaries).
Definition: The canonical contract document produced by Council and locked by spec-freeze. Lives at ~/forge/brain/prds/<task-id>/shared-dev-spec.md. Contains all 5 service contracts negotiated by the 4 surfaces: REST API (endpoints, payloads, status codes), event bus (topics, schemas), cache (keys, TTL, invalidation), database (tables, schema, migrations), search (document structure, analyzers). Immutable after [P2-SPEC-FROZEN] — changes require full SPEC-AMENDMENT Protocol (council re-vote + new [P2-SPEC-AMENDED] marker).
Usage Context: All downstream phases read from this file. tech-plan-write-per-project breaks it into per-repo tasks. spec-reviewer verifies implementation matches it line-by-line. forge-drift-check detects divergence. Do not modify post-freeze without re-opening council.
What It's NOT: Not a living document. Not per-repo. Not aspirational. Not a "first draft" — it is the signed contract.
Cross-References: Written by council-multi-repo-negotiate; locked by spec-freeze; read by tech-plan-write-per-project, spec-reviewer, forge-drift-check; amended via spec-freeze § SPEC-AMENDMENT Protocol.
Definition: Per-product workspace config file at ~/forge/brain/products/<slug>/product.md. Contains: product slug, repo paths (role → absolute path), start/health commands per service, deploy_doc for services with complex startup, and flags. The most important flag is forge_qa_csv_before_eval (boolean) — when true or when the entrypoint is /forge, qa-manual-test-cases-from-prd is mandatory before [P4.0-SEMANTIC-EVAL].
Key fields:
forge_qa_csv_before_eval: true|false — gates manual QA CSV requirementrepos: — role-to-path map (e.g., backend: /abs/path/to/backend)start: — how to start each service for evalhealth: — health check command per serviceCross-References: Read by conductor-orchestrate, eval-product-stack-up, deploy-driver-*; forge_qa_csv_before_eval enforced by conductor-orchestrate State 4b.
Definition: D30 discipline enforced by the worktree-per-project-per-task skill. Every dev-implementer task must execute in a fresh isolated git worktree — never on the main working tree or a shared branch. One worktree per repo per task. The branch follows naming convention task/<task-id>[-<repo-role>]. Worktrees are created before [P4.1-DISPATCH] and confirmed by forge-tdd Step 0.
Why isolated: Prevents state leakage between parallel tasks, ensures the RED test only sees the current task's code, and makes rollback clean (delete worktree, branch is gone).
Red flag: If git worktree list shows the current directory on main/master or HEAD without a task branch — STOP. Invoke worktree-per-project-per-task first.
Cross-References: Enforced by forge-tdd Step 0 HARD-GATE; conductor-orchestrate State 5 HARD-GATE; skill worktree-per-project-per-task.
Definition: The Forge brain is the append-only, git-backed evidence store at ~/forge/brain/. Two subtrees:
~/forge/brain/products/<slug>/ — persistent per-product config (product.md, codebase/, terminology.md, scan outputs)~/forge/brain/prds/<task-id>/ — per-task artifacts (all phases of a single pipeline run)Per-task layout (~/forge/brain/prds/<task-id>/):
| Path | Contents |
|---|---|
prd-locked.md | Intake output — immutable PRD with all 9 Q answers |
terminology.md | Canonical product term sheet |
shared-dev-spec.md | Council output — 5 contracts, frozen |
tech-plans/<repo>.md | Per-repo task breakdown |
qa/manual-test-cases.csv | Approved acceptance test cases |
qa/semantic-automation.csv | Machine-eval step DAG |
qa/semantic-eval-manifest.json | Eval run outcome |
qa/semantic-eval-run.log | Per-step JSON lines |
design/ | Figma MCP ingest or Lovable sync artifacts |
conductor.log | Append-only phase marker log |
decisions/ | Brain decisions (DREAM-, SPECCHG-) |
blockers/ | Escalation files when BLOCKED |
Cross-References: brain-read, brain-write, brain-recall, forge-brain-layout.
Definition: Immutable PRD artifact written to ~/forge/brain/prds/<task-id>/prd-locked.md at the end of intake. Contains all 9 intake-interrogate question answers as structured sections. Key frontmatter fields: product, goal, success_criteria, repos (role list), contracts (which of 5 apply), timeline, rollback_plan, metrics, design_new_work, design_intake_anchor. Once written and [P1-PRD-LOCKED] is logged, this file is read-only — reopen intake if it must change.
Cross-References: Written by intake-interrogate; read by all downstream skills; consumed by council-multi-repo-negotiate, tech-plan-write-per-project, qa-prd-analysis.
Definition: Per-task canonical term sheet at ~/forge/brain/prds/<task-id>/terminology.md. Table of canonical product names, disallowed variants, and open_doubts in frontmatter. Authored in intake, aligned at council. Consumed by QA authoring (for Intent/ExpectedHint wording), tech plans (UI copy), and assertion text so human-facing copy matches contracts.
What It's NOT: Not the forge-glossary (which covers Forge process terms). Not global — it is task-scoped. Not optional for UI-facing work.
Cross-References: intake-interrogate; council-multi-repo-negotiate; docs/terminology-review.md; forge-glossary § Product terminology.
Definition: A reusable, discipline-enforcing capability packaged as a SKILL.md file with YAML frontmatter (name, description, type, requires). Skills can be rigid (must follow exactly) or flexible (principles-based). Each skill has optional red flags (STOP conditions), anti-patterns (common excuses), and edge cases (unusual scenarios).
Usage Context: Invoked via the invoke command or through skill dependencies. A skill can require other skills, forming a dependency tree. Skills are discovered in skills/ directory (repo root) and symlinked from .claude/skills/. When you invoke a skill, the harness renders the markdown and passes context (PRD, spec, codebase) as needed.
What It's NOT: Not a tool (tools are CLI utilities). Not a hook (hooks are session-level plugins). Not an agent (agents are independent processes). Not a prompt — skills are reusable references that adapt to your context.
Cross-References: Related to superpowers (discipline-enforcing skills from Anthropic). Similar format to agent but lightweight. Enforced by forge-skill-anatomy (format checklist).
Definition: A skill that must be followed exactly as written, with no adaptation or shortcutting. Enforced by HARD-GATE markers and TDD pressure tests. Examples: forge-tdd, forge-intake-gate, forge-eval-gate.
Usage Context: When you invoke a rigid skill, you are agreeing to follow every step. Deviations are red flags. Rigid skills typically have anti-pattern preambles (D25) that rationalize common excuses and rebut them.
What It's NOT: Not flexible. Not subject to interpretation. Not "follow the spirit but skip steps." Not negotiable with schedule pressure.
Cross-References: Contrast with flexible-skill. Enforced by D24 (HARD-GATE tags on non-skippable steps). Related to D15 (TDD pressure).
Definition: A skill that establishes principles and constraints but allows adaptation to context. Example: contract negotiation skills use a framework but adapt to service-specific boundaries. Developers can skip optional sections if justified.
Usage Context: Invoke a flexible skill and apply its principles to your specific scenario. Document deviations. Flexible skills typically have edge cases (unusual conditions) that guide when to adapt vs. when to escalate.
What It's NOT: Not a free pass to ignore it. Not "do whatever you want." Not permission to skip mandatory sections. Flexible skills still enforce core discipline — just with more context-sensitivity.
Cross-References: Contrast with rigid-skill. Examples: reasoning skills (reasoning-as-backend, reasoning-as-web-frontend), contract skills.
Definition: An explicit STOP condition embedded in a skill that signals the skill cannot proceed as written. Red flags are safety valves — they prevent silent failures and force escalation. Each red flag is a single condition that, if true, halts execution.
Usage Context: When executing a skill, check red flags before each major step. If a red flag condition is met, stop and escalate (do not rationalize or work around). Examples: "if service is not responding, RED FLAG: BLOCKED"; "if more than 3 retries fail, RED FLAG: escalate."
What It's NOT: Not a warning. Not a hint to be careful. Not optional. Not a place for judgment — if condition is true, you must escalate.
Cross-References: Often paired with anti-patterns (common excuses NOT to apply red flags). Enforced by D24. Related to escalation (next step after RED FLAG).
Definition: A rationalization table (D25) that lists common excuses for skipping a discipline and rebuts each one. Embedded at the top of every discipline-enforcing skill. Example: forge-tdd lists "We're running late" → rebuttal: "TDD saves time because fewer bugs slip through."
Usage Context: Before skipping any required step, consult the anti-pattern preamble. If your excuse is listed, read the rebuttal. If your excuse is NOT listed, escalate (do not improvise). Anti-patterns are git-backed and immutable — they evolve through dreamer retrospectives but never disappear.
What It's NOT: Not permission to skip steps if your excuse isn't listed. Not a menu of excuses — it's a rebuttal table. Not formal policy; it's discipline enforcement via language.
Cross-References: Required by D25. Paired with red flags. Related to HARD-GATE skills. Examples in forge-tdd, forge-intake-gate, forge-eval-gate.
Definition: A non-negotiable process gate marked with a HARD-GATE label. Enforced by 5+ MUST bullets that cannot be skipped or rationalized away. Examples: intake (MUST satisfy mandatory prd-locked.md fields via intake-interrogate, confidence-first), council (MUST negotiate all contracts), eval (MUST pass critical scenarios), TDD (MUST write test first).
Usage Context: When you encounter a HARD-GATE, you have no choice but to execute it fully. It cannot be shortcut due to schedule pressure, complexity, or other factors. Red flags within HARD-GATE steps are not advisory — they are mandatory stop conditions.
What It's NOT: Not advisory. Not "try to do this if possible." Not flexible to context. Not negotiable with stakeholders or schedule.
Cross-References: Enforced by D24. Every HARD-GATE has associated skill. Examples: forge-intake-gate, forge-council-gate, forge-eval-gate, forge-tdd, forge-worktree-gate, forge-trust-code, forge-verification, forge-letter-spirit.
Definition: Collection of 10 discipline-enforcing skills from Anthropic (not Forge-specific) that cover planning, testing, debugging, code review, and development workflows. Superpowers are universal; Forge skills are product-specific. Superpowers include: writing-plans, brainstorming, executing-plans, dispatching-parallel-agents, test-driven-development, systematic-debugging, requesting-code-review, receiving-code-review, verification-before-completion, finishing-a-development-branch.
Usage Context: Use superpowers when facing any task that matches their domain. Example: before implementation, invoke superpowers:writing-plans. During test-driven development, invoke superpowers:test-driven-development. Superpowers often parallel Forge skills but provide deeper guidance.
What It's NOT: Not Forge-specific. Not baked into Forge pipeline — they are parallel resources you invoke as needed. Not required (though highly recommended). Not limited to Forge work — applicable to any Claude project.
Cross-References: Related to Forge skills but orthogonal. Invoked alongside rigid/flexible skills. Examples: superpowers:test-driven-development parallels forge-tdd, superpowers:writing-plans parallels tech-plan-write-per-project.
Definition: Sequential batch of skill enhancements organized in the Forge roadmap. Each phase adds anti-patterns, edge cases, and decision trees to existing skills. Phases: P0 (foundation, complete), P1 (critical eval drivers, complete), P2 (surface reasoning + brain + deployment, complete), P3 (remaining skills, in progress).
Usage Context: Phases guide skill maturity. P0 skills are foundational (always required). P1 skills enable multi-service eval. P2 skills add depth to reasoning and operations. P3 skills expand coverage to remaining domains. When invoking a skill, check which phase it's in — earlier phases are more stable.
What It's NOT: Not arbitrary groupings. Not a "nice to have" roadmap. Not flexible timelines — phases complete in order before moving forward. Not skippable — all phases are required for full Forge capability.
Cross-References: Related to batch (finer-grained grouping within a phase). Examples: P0 (intake, council, build), P1 (eval drivers: API, DB, cache, search, events), P2 (reasoning skills + brain skills + deploy drivers), P3 (remaining skills).
Definition: Finer-grained grouping of related skills within a phase. Example: P1 has 3 eval driver batches (HTTP/DB, cache/search, events/mobile) and 1 coordination batch (eval-judge, dream-resolve). P2 has 3 batches (reasoning, brain, deployment).
Usage Context: Batches allow parallel work within a phase. If phase P2 Batch 1 (reasoning skills) is complete, you can start using those skills while Batch 2 (brain skills) is still in development. Batches ship together but are tracked separately.
What It's NOT: Not independent from phases — batches are subdivisions of phases. Not arbitrary — batches group skills with strong dependencies.
Cross-References: Subdivides phase. Examples: P2 Batch 1 (reasoning-as-backend, reasoning-as-web-frontend, reasoning-as-app-frontend), P2 Batch 2 (brain-read, brain-write, brain-recall, brain-why, brain-forget), P2 Batch 3 (deploy-driver-pm2-ssh, deploy-driver-docker-compose, deploy-driver-local-process, deploy-driver-systemd).
Definition: Phase numbering for the Forge skill enhancement roadmap. P0: foundation skills (intake, council, spec-freeze, tech-plan, build, eval, review, dream). P1: critical eval drivers (API, DB, cache, search, events, web, mobile). P2: surface reasoning + brain operations + deployment drivers. P3: remaining skills (in progress).
Usage Context: Check phase number to understand skill maturity and feature completeness. P0 skills are required and stable. P1 skills enable multi-service product testing. P2 skills add reasoning depth and decision tracking. P3 skills expand to specialized domains.
What It's NOT: Not marketing phases. Not "versions." Not flexible — phase order is locked.
Cross-References: Each phase contains multiple batches. Progress tracked in memory file at /home/lordvoldemort/.claude/projects/-home-lordvoldemort-Videos-forge/memory/MEMORY.md.
Definition: ShopApp — a test e-commerce product used to pressure-test all Forge skills via realistic scenarios. Includes backend (Node.js), web frontend (React), mobile app (React Native or native), and infrastructure (Docker, Kubernetes). Lives in seed-product/ directory.
Usage Context: Every skill is validated against ShopApp before shipping. D15 requires all skills be TDD'd against seed product pressure scenarios. When developing a new skill, build a scenario on ShopApp first, then write the skill to handle it.
What It's NOT: Not the only product Forge can work on — ShopApp is the validation vehicle. Not a finished product — it's intentionally simple to isolate skill behavior. Not source of truth for Forge patterns; it's a test bed.
Cross-References: Related to D15 (TDD pressure scenarios). Used by skill tests. Part of forge-self-test.
Definition: A domain perspective in council and evaluation: backend (database, APIs, business logic), web frontend (browser UI, React), app frontend (mobile, native), infrastructure (deployment, operations). Each surface reasons about the PRD from its specialized viewpoint.
Usage Context: During council, 4 surfaces negotiate contracts from their perspectives. During eval, surface-specific eval drivers test the surface (web-cdp for web, xctest for iOS, adb for Android, API for backend, DB for schema). When reasoning about a PRD, switch surfaces to see blind spots in your design.
What It's NOT: Not vertical layers (frontend/backend split). Not silos — surfaces must negotiate with each other. Not optional — all 4 surfaces must participate in council.
Cross-References: Used by council via reasoning-as-backend, reasoning-as-web-frontend, reasoning-as-app-frontend, reasoning-as-infra. Tested by surface-specific eval drivers.
Definition: Non-negotiable practice embedded in Forge skills (TDD, HARD-GATE, two-stage review, isolation). Disciplines are enforced by anti-patterns (D25) and red flags (D24). Examples: test-driven-development (discipline), HARD-GATE enforcement (discipline), two-stage review (discipline), worktree isolation (discipline).
Usage Context: When a skill enforces discipline, follow it exactly. Disciplines prevent the most common sources of bugs: untested code, unreviewed changes, shared state, incomplete specifications. If schedule pressure tempts you to skip a discipline, consult the skill's anti-pattern preamble.
What It's NOT: Not bureaucracy. Not optional. Not "suggestions for quality." Not flexible to urgency or context.
Cross-References: Enforced by D24 (HARD-GATE tags) and D25 (anti-pattern preambles). Examples: forge-tdd, forge-trust-code, forge-worktree-gate, spec-freeze.
Definition: Explicit negotiated specification for the interface between services. Covers: REST APIs (methods, endpoints, payloads), event bus (topic names, message schema), cache (keys, TTL, invalidation), database (table names, schema, migrations), search (document structure, analyzers). Contracts are part of shared-dev-spec.
Usage Context: During council, services negotiate contracts. Each contract defines what data flows where, when, and in what format. Implementation must match contracts exactly (enforced by spec-reviewer). Changes to contracts require re-opening council.
What It's NOT: Not internal API design. Not documentation of what you built — it's the agreement before you build. Not flexible post-lock. Not micro-optimization territory.
Cross-References: Negotiated by council-multi-repo-negotiate. Skills: contract-api-rest, contract-event-bus, contract-cache, contract-schema-db, contract-search. Part of shared-dev-spec. Verified by spec-reviewer in review stage.
Definition: Implementation skill that "drives" a system by connecting to it, running operations, and verifying state. Two types: eval drivers (connect, run scenarios, verify results) and deploy drivers (start/stop services, check health). Examples: eval-driver-api-http (connect via HTTP, run API calls), deploy-driver-docker-compose (start containers, verify running).
Usage Context: Eval drivers are invoked during eval stage to test each service. Deploy drivers are invoked to bring up the stack for eval or production. Each driver exposes functions (connect, disconnect, run, verify) that skills use to automate integration tests.
What It's NOT: Not unit tests. Not mocks. Not local-only — drivers assume services are running. Not hardcoded to one service — drivers are reusable across products.
Cross-References: Eval drivers: eval-driver-api-http, eval-driver-db-mysql, eval-driver-cache-redis, eval-driver-bus-kafka, eval-driver-search-es, eval-driver-web-cdp, eval-driver-ios-xctest, eval-driver-android-adb. Deploy drivers: deploy-driver-pm2-ssh, deploy-driver-docker-compose, deploy-driver-local-process, deploy-driver-systemd. Stack-up: eval-product-stack-up. Semantic execution: qa-semantic-csv-orchestrate.
Definition: Signal that human judgment, context, or coordination is needed. Triggered by red flags, unrecovered failures, or scope ambiguity. Keywords: BLOCKED, NEEDS_CONTEXT, NEEDS_COORDINATION, NEEDS_INFRA_CHANGE. Escalation is not failure — it's the correct response when automation cannot proceed.
Usage Context: When a skill hits a red flag, escalate immediately (do not work around). When dev-implementer reports BLOCKED, escalate to conductor. When eval fails 3 times, escalate. When contracts conflict, escalate. Escalation triggers human review, context addition, or re-negotiation.
What It's NOT: Not a rare event. Not a sign of incompetence. Not shameful. Not the same as failure. Not permitting continued work around the issue.
Cross-References: Related to red flags, self-heal retry cap, dev-implementer status codes. Handled by conductor or human team.
Definition: Subagent that executes tech plan tasks sequentially. Writes tests first (TDD), implements code, and reports status (DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, BLOCKED). Each task runs in an isolated worktree with no shared state (D30).
Usage Context: Dispatched once per task from the tech plan. Receives task description, context (PRD, spec, codebase), and success criteria. Reports status at completion. If BLOCKED, escalates to conductor.
Cross-References: Related to worktree-per-project-per-task, forge-tdd. Part of Build stage.
Definition: Subagent that reads actual code line-by-line and verifies it matches shared-dev-spec exactly. Enforces D14: "trust code, not reports." Not allowed to skim or trust summaries. Runs a 9-phase verification scope: (1) API contract compliance, (2) DB schema compliance, (3) event bus contracts, (4) cache contracts, (5) search contracts, (6) cross-service integration wiring, (7) performance guard rails, (8) security requirements, (9) operational readiness (health endpoints, logging, metrics). Also checks for over-building (code not in spec) and under-building (spec requirements not implemented).
Usage Context: Invoked during Review stage. Must read full implementation (not diffs or summaries) and cross-reference against spec. Reports APPROVED or CHANGE_REQUESTED with findings per phase.
Cross-References: Enforced by forge-trust-code (HARD-GATE). Part of Review stage, first stage. Agent definition: agents/spec-reviewer.md.
Definition: Subagent that checks 8-point quality framework: (1) Naming Conventions & Clarity, (2) File Size & Organization, (3) Code Complexity & Readability, (4) Error Handling & Resilience, (5) Test Coverage & Quality, (6) Performance & Scalability, (7) Security Practices, (8) Observability & Debuggability. Plus Phase 4 cross-service quality checks (consistent error codes, cache key patterns, event schema field names, class/function naming conventions across services).
Usage Context: Invoked during Review stage after spec-reviewer approves. Must read code and apply all 8 checks. Reports APPROVED or CHANGE_REQUESTED with issues categorized as Critical, Important, or Minor.
Cross-References: Part of Review stage, second stage. Agent definition: agents/code-quality-reviewer.md.
Definition: Subagent that runs two functions: inline conflict resolution (when eval surfaces incompatibilities between services) and post-merge retrospective (scoring decisions, extracting patterns, writing to brain). Dual role during and after the pipeline.
Usage Context: Invoked during eval (if conflicts surface) and after PR set merge. Produces brain artifacts (decisions, learnings, patterns).
Cross-References: Uses dream-resolve-inline, dream-retrospect-post-pr. Related to brain skills.
| Status | Meaning |
|---|---|
| DONE | Task completed successfully, all success criteria met. Proceed to review. |
| DONE_WITH_CONCERNS | Task completed but code quality or correctness issues exist. Must be addressed before review. |
| NEEDS_CONTEXT | Missing information to complete task (spec ambiguity, missing API docs, unclear requirement). Provide context and re-dispatch. |
| BLOCKED | Cannot proceed. Escalate to conductor or human. (Example: required service not available, contract conflict, infrastructure unavailable.) |
Key locked decisions (only externally visible decisions are listed here; D1–D4, D6–D12, D16–D23, D26–D29 are internal implementation decisions recorded in brain/decisions/ and not surfaced in the glossary because they affect tooling internals, not skill-level behavior):
| Decision | Summary |
|---|---|
| D5 | No LangChain-style agent frameworks in Forge plugin code. Product eval may use CDP, Playwright, Puppeteer, Appium, XCTest, MCP on the host — ask the operator (browser MCP vs local CDP; Appium MCP vs ADB / XCTest for mobile) before locking the driver stack. |
| D13 | No runtime dependency on any external plugin at runtime |
| D14 | Trust code: spec-reviewer reads actual code, not reports or summaries |
| D15 | Skills are TDD'd against seed product pressure scenarios |
| D24 | HARD-GATE tags on every non-skippable step; red flags enforce them |
| D25 | Anti-Pattern preambles on every discipline-enforcing skill |
| D30 | Fresh worktree per project per task. No shared state. |
Definition: Integer (1–4) in a skill's YAML frontmatter that controls how much of the skill's context is inlined into session-start by session-start.cjs. Tier 1 = minimal (name + one-line description only); Tier 4 = full content inlined. Higher tiers burn more context budget; use sparingly for skills that must be available before any tool invocation.
How it works: On session start, session-start.cjs reads the active skill's preamble-tier from ~/.forge/.active-skill-tier. The cache file is per-active-skill (global path, tracks whichever skill is currently active). Format: line 1 = tier digit 1–4; line 2 = # sha256=<hex> (optional, new format). If line 2 is present and the SHA-256 of the current SKILL.md differs, the cache is invalidated and re-parsed. If line 2 is absent (older single-line format written by echo N > ~/.forge/.active-skill-tier), hash validation is skipped — the tier is used as-is. The tier determines how many sections are serialized into the IDE's system prompt injection.
Usage Context: Set preamble-tier: 1 for most skills (name + description sufficient). Set preamble-tier: 3 or preamble-tier: 4 only for foundational skills (using-forge, forge-glossary) that must deliver orientation context on every session. Never set preamble-tier: 4 on a skill that has 500+ lines — it will saturate context.
Cross-References: using-forge § ~/.forge/.active-skill-tier (cache format). session-start.cjs implements the read/write logic.
Definition: Boolean field in prd-locked.md (from intake-interrogate Q9) indicating whether the PRD requires net-new UI design artifacts. When design_new_work: yes, the pipeline MUST run State 4b-design (conductor-orchestrate) before P4.1 dispatch — materializing design to ~/forge/brain/prds/<task-id>/design/ (Figma MCP ingest, Lovable sync, or manual exports) and logging [DESIGN-INGEST] status=PASS to conductor.log. When design_new_work: no or design_waiver: prd_only, State 4b-design is skipped.
How to set: During intake-interrogate Q9, the agent asks whether new UI design artifacts are required. If yes, write design_new_work: yes in the YAML frontmatter of prd-locked.md. If no new design, write design_new_work: no (or design_waiver: prd_only with owner + risk justification).
Artifact Traceability: intake-interrogate Q9 → prd-locked.md design_new_work: yes → State 4b-design → ~/forge/brain/prds/<task-id>/design/MCP_INGEST.md (or LOVABLE_SYNC.md) → [DESIGN-INGEST] status=PASS in conductor.log → P4.1 dispatch unlocked for web/app repos.
Cross-References: conductor-orchestrate § State 4b-design; intake-interrogate Q9; docs/conductor-log-format.md [DESIGN-INGEST].
Definition: Eval failure classification for infrastructure failures that are not code bugs — ECONNREFUSED, Docker service down, MCP unavailable, device/emulator not running. Classified by self-heal-triage before consuming any retry budget. RED_INFRA bypasses the self-heal retry cap: infrastructure must be restored first, then eval re-runs from scratch (the failed attempt does not count against the 3-retry budget).
Usage Context: self-heal-triage checks for RED_INFRA symptoms before applying any other triage category. If RED_INFRA is detected: write BLOCKED escalation to ~/forge/brain/prds/<task-id>/blockers/, log [P4.4-RED-INFRA] to conductor.log, and stop. Do NOT log the attempt as a self-heal retry.
What It's NOT: Not a code bug. Not a test failure. Not a flaky test. Not a race condition. RED_INFRA is always an environment issue — fix the environment, not the code.
Cross-References: self-heal-triage § RED_INFRA Pre-Check; self-heal-loop-cap § RED_INFRA bypass rule; docs/conductor-log-format.md [P4.4-RED-INFRA].
Definition: Outcome value for a semantic eval step (in qa/semantic-eval-run.log) when the step could not be fully evaluated. Two forms: (1) an upstream dependency step passed but returned an empty result: {}, so downstream ${stepId.result.field} interpolation had no data; (2) external context required by the step (credentials, device, URL, test account) was not available at runtime. In both cases the step result is indeterminate — not a pass, not a product bug. eval-judge maps any CONTEXT_GAP entries to a YELLOW verdict if all non-skipped steps otherwise pass.
Usage Context: When eval-judge sees CONTEXT_GAP in semantic-eval-run.log, the verdict is YELLOW (not RED). The appropriate response is to provide the missing context (credentials, device, test account) and re-run — not to enter the self-heal loop.
What It's NOT: Not a test failure. Not RED_INFRA. Not a code bug. CONTEXT_GAP means "we don't know yet" — the step was not executed, not failed.
Cross-References: eval-judge § Semantic path verdict table; docs/semantic-eval-schema.md § outcome enum; qa-semantic-csv-orchestrate.
Definition: Outcome value for a semantic eval step when the step was skipped because a step it DependsOn returned a non-PASS result. Propagated automatically by the CSV eval runner when upstream steps fail. eval-judge maps runs where all non-PASS outcomes are BLOCKED_DEPENDENCY to a YELLOW verdict (dependency issue, not a code bug in the current step).
Usage Context: When debugging a RED eval run, distinguish BLOCKED_DEPENDENCY steps from genuine failures. A step marked BLOCKED_DEPENDENCY did not run — its result says nothing about the correctness of its own code. Fix the upstream failing step first, then re-run. When manifest.outcome = fail but all non-PASS steps are BLOCKED_DEPENDENCY, the eval-judge verdict is YELLOW (not RED) — the root failure is a dependency chain issue, not a code bug in the current implementation.
What It's NOT: Not the same as CONTEXT_GAP (empty result vs. upstream failure). Not a code bug in the blocked step. Not skippable — if the dependency step is genuinely broken, it must be fixed.
Cross-References: eval-judge § BLOCKED_DEPENDENCY verdict rule; docs/semantic-eval-schema.md § DependsOn propagation; qa-semantic-csv-orchestrate § dependency DAG.
| Verdict | Meaning | Next Step |
|---|---|---|
| GREEN | All critical scenarios passed. Ready to merge. | Proceed to Review stage. |
| YELLOW | All critical passed, some non-critical failed. Decide: fix or accept trade-off. | Review or return to Self-Heal. |
| RED | Critical scenario failed. Cannot merge. | Enter Self-Heal loop (max 3 retries). |
| NOT_EXECUTED | QA pipeline / orchestrator — no driver results (no stack, no env, agent session static check only). Not an eval-judge outcome; do not confuse with YELLOW. | Provide URL/device/credentials; re-run /qa-run. See qa-pipeline-orchestrate QA-P6, Edge case Static validation only. |
HARD-GATE Skills:
forge-intake-gate, forge-council-gate, forge-eval-gate, forge-tdd, forge-worktree-gate, forge-trust-code, forge-verification, forge-letter-spirit
Reasoning Skills (Council):
reasoning-as-backend, reasoning-as-web-frontend, reasoning-as-app-frontend, reasoning-as-infra
Contract Skills (Council):
contract-api-rest, contract-event-bus, contract-cache, contract-schema-db, contract-search
Eval Drivers:
eval-driver-api-http, eval-driver-db-mysql, eval-driver-cache-redis, eval-driver-bus-kafka, eval-driver-search-es, eval-driver-web-cdp, eval-driver-ios-xctest, eval-driver-android-adb
Deploy Drivers:
deploy-driver-pm2-ssh, deploy-driver-docker-compose, deploy-driver-local-process, deploy-driver-systemd
Brain Skills:
brain-read, brain-write, brain-recall, brain-why, brain-forget, brain-link
Self-Heal Skills:
self-heal-locate-fault, self-heal-triage, self-heal-systematic-debug, self-heal-loop-cap
Symptom: A user asks about "council" expecting a human committee or a governance process, not the Forge contract-negotiation phase between domain reasoning surfaces.
Do NOT: Explain the Forge term without acknowledging the ambiguity.
Action: When a term from the glossary collides with a common industry term (Council, Surface, Driver, Dream, Conductor), clarify the Forge-specific meaning upfront: "In Forge, 'Council' refers to the multi-surface contract negotiation phase — not a human committee. It's executed by 4 AI reasoning surfaces negotiating 5 service contracts."
Escalation: If the user's intent is genuinely unclear (asking about a human council process vs. the Forge phase), ask once: "Do you mean the Forge council phase, or are you asking about a governance process outside Forge?"
Symptom: A brain decision file says "negotiation round" where this glossary says "Council." A skill says "triage phase" where the glossary says "Self-Heal Triage." The user asks which is authoritative.
Do NOT: Guess or arbitrarily prefer the glossary over the source file.
Action: The source skill (self-heal-triage/SKILL.md, forge-council-gate/SKILL.md) is always authoritative for behavior. The glossary is authoritative for naming and definition. If they conflict on behavior, the skill wins. If they conflict on the name of a concept, the glossary wins. Document the discrepancy and suggest it be reconciled.
Escalation: NEEDS_CONTEXT if the phrasing difference implies a behavioral difference (e.g., glossary says "max 3 retries" but a skill says "max 5 retries" — that is a real conflict, not just a naming difference).
Symptom: A skill, agent, or brain file uses a term like "pressure scenario," "seeder," or "gate bypass" that does not appear in this glossary.
Do NOT: Invent a definition or silently treat it as undefined.
Action:
skills/ and agents/ with grep to find the canonical context in which it's usedEscalation: NEEDS_CONTEXT — request a definition from the skill author before proceeding with work that depends on this term.
Symptom: "We already know the contracts, can we skip council and go straight to implementation?"
Do NOT: Accept the skip request and proceed to tech plans without running forge-council-gate.
Action:
forge-council-gate with the existing contract proposals as input — surfaces still negotiate, verify compatibility, and sign off.[ABORT_TASK] logged to conductor.log by the human — which cancels the whole pipeline, not just council.Escalation: BLOCKED — council is non-negotiable per forge-council-gate Iron Law.
WHEN: You need to detect performance regressions before merge — TTFB, response time, bundle size. Run as part of forge-eval-gate or standalone before raising a PR.
WHEN: A decision is being superseded, deprecated, or has aged out. Archive it without deletion — marks as warm→cold→archived with full audit trail.
WHEN: You are writing a new decision, superseding an old one, or querying relationships across decisions. Create semantic edges between decisions and link concepts across products/projects/time.
WHEN: You are about to make a decision and need to check if prior art or past learnings exist. Recall decisions, patterns, and gotchas from the brain before proceeding.
WHEN: You need to trace the full provenance of a specific decision — who made it, when, why, and what alternatives were considered. Shows why, when, by whom, evidence, alternatives, outcome.
WHEN: You need to record a decision, lock a spec, log an eval run, or document learnings in the brain.