메뉴
WHEN: Shared-dev-spec is frozen and per-project tech plans must be written before dev-implementer dispatch. Output: 1 maximal plan per repo with Section 1b detail, 1b.2a exploration, and Section 2 tasks fully elaborated.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
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WHEN: A decision is being superseded, deprecated, or has aged out.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.
| name | tech-plan-write-per-project |
| description | WHEN: Shared-dev-spec is frozen and per-project tech plans must be written before dev-implementer dispatch. Output: 1 maximal plan per repo with Section 1b detail, 1b.2a exploration, and Section 2 tasks fully elaborated. |
| type | rigid |
| requires | ["brain-read"] |
| version | 1.0.3 |
| preamble-tier | 3 |
| triggers | ["write tech plan","create implementation plan per repo","plan for each project"] |
| allowed-tools | ["Bash","Write"] |
Judgment rounds (Section 0.1 / Section 0.2) run in chat — the plan file summarizes outcomes, not a substitute for dialogue. Follow skills/using-forge/SKILL.md Multi-question elicitation: transcript-visible questions, one coherent topic per message when multiple decisions remain, blocking interactive prompts / numbered options + stop for discrete forks (Cursor AskQuestion maps canonical AskUserQuestion). Do not paste the whole interactive workshop into Section 0 as if the human answered inside the markdown (Anti-Pattern rows below).
Cross-cutting assistant dialogue: docs/forge-one-step-horizon.md — using-forge Multi-question elicitation items 4–8.
| Rationalization | Why It Fails |
|---|---|
| "The implementer will figure out the details" | Vague tasks cause divergence. "Add the endpoint" is not a task — "Add POST /api/v1/orders to routes/orders.ts returning 201 with OrderResponse schema" is a task. |
| "I'll use pseudocode to keep the plan concise" | Pseudocode forces the implementer to make design decisions that should have been made in planning. Write complete code. |
| "The skill says brief / concise / compact, so I'll keep Section 1b and Section 2 short" | Misread. Those words apply to Section 0 topic labels, chat messages, and prescribed one-line N/A — not to Section 1b tables, fenced payloads, ### 1b.2a evidence, or Section 2. Thin Section 1b is incomplete, not disciplined. |
| "This task is too small to write out" | If it takes 2 minutes to execute, it takes 30 seconds to write. Small tasks that are written out get done correctly. Small tasks left vague get done wrong. |
| "I'll group related changes into one big task" | Tasks over 5 minutes need splitting. Big tasks hide complexity and make progress tracking impossible. |
| "The bash commands are obvious" | "Obviously" wrong commands waste a self-heal loop iteration. Write the exact command including flags, paths, and environment variables. |
| "I'll reference the spec instead of repeating details" | The implementer (dev-implementer subagent) works in an isolated worktree with only the plan. Self-contained tasks prevent NEEDS_CONTEXT status. |
| "I'll approximate inventory scale (~60+ files / many services) to sound thorough" | BLOCKED. Plans need what / where / how: repo-relative paths, entrypoints, how each row was verified (rg/Read/scan path), not headline counts. Enumerate services/files/slices the implementer must touch, or cite codebase/ + product paths per row. |
| "Source / export is huge — I'll skip deep reads or partial Section 1b" | BLOCKED. Line count and export volume are not discretionary relief. AGENTS.md Core rule 6 — batch reads, complete the skill’s required depth, or BLOCKED with evidence. |
| "I'll discover file paths by exploring the repo" | Duplicates work the scan already did and burns tokens. Default: read ~/forge/brain/products/<slug>/codebase/ first; put paths from index.md / modules/*.md / api-surface.md into tasks, then open sources when writing full file bodies. Exception: Section 1b.6 lists an UNKNOWN — you must deepen discovery (targeted rg/glob, read hub files, route tables, OpenAPI, client wrappers, test names) until resolved or BLOCKED — do not ship “mystery meat” tasks. |
| "Elaboration is optional — bite-sized tasks are enough" | Tasks without Section 1b.0, Section 1b.5 (synchronous API — REST / GraphQL / SOAP / … per lock) + #### 1b.5b when events/cache apply, Section 1b.1 / 1b.1a when persistence or search applies, Section 1b.6 unknown closure, and Section 1c review rounds hide integration risk. STOP. Elaboration is mandatory for E2E; micro-tasks execute the elaboration, they do not replace it. |
| "I've hit my question quota — ship the plan with lingering doubts" | There is no maximum question count during planning. Doubt left unasked becomes a gap in Section 2. STOP. Ask until confidence is high (see Section 0), then write the elaborative plan. |
| "Concise plan = professional" | Professional here means complete: the plan is the only input to sub-tasks. Concision that omits wiring, edge cases, or evidence is negligence. |
| "I'll cover PRD cases implicitly in tasks" | Every success criterion, edge case, and non-functional requirement from prd-locked.md + shared-dev-spec.md must appear in Section 1b.0 and map to Section 1b subsections or Section 2 tasks. Implicit coverage is invisible to review and ships gaps. |
| "Mechanics without intent — tables and tasks have no why or PRD trace" | BLOCKED. Every Section 1b.0 row needs a Why (rationale); every Section 2 task needs Traces to: + Rationale:; API / design / message rows must say which PRD or acceptance obligation they satisfy. Otherwise reviewers cannot tell purpose from shape. |
| "Schema / payload details can wait for implementation" | Forbidden when that subsection applies to this repo. Section 1b.1 / 1b.5 / 1b.1a must carry concrete persistence shapes (SQL DDL, Mongo validators/index specs, ClickHouse CREATE/mutation, … per contract), search/index definitions, or API request/response/error shapes (JSON, XML for SOAP, GraphQL operation + variables schema, … per lock) — or verbatim locked-contract excerpts — never TBD where the contract already decided them. Corollary: If this repo does not own persistence, search, or that API surface, a one-line N/A + spec citation (and sibling repo if needed) is the required elaboration — not empty tables and not invented schema. |
| "Every good plan has migrations and index mappings" | Stack bias. Many tasks are UI-only, docs-only, config-only, or consumer-only. Forcing relational-DDL + Elasticsearch-style sections when the product uses Mongo, ClickHouse, BigQuery, Typesense, etc. — or none in this repo — drives fake work or endless discovery. Derive applicability from the frozen spec + this repo’s role; elaborate maximally only on surfaces in play, in the contract’s schema language. |
| "Scan is optional if I know the repo" | The brain codebase/ is the default authority for where code lives. Skipping index.md, relevant modules/*.md, api-surface.md, and route/OpenAPI stubs before writing tasks is BLOCKED unless SCAN_INCOMPLETE / BLOCKED is explicitly recorded with owner. |
| "Touchpoint table is boilerplate — I'll mark N/A for everything" | ### 1b.2a is full exploration, not a formality. Every Y / PARTIAL needs paths + tools + behavior delta. Bulk N/A without opening the product repo is BLOCKED at self-review. |
| "I'll put the whole ‘interactive’ dialogue in Section 0 and ship" | Wrong split. The human answers in chat; Section 0 only summarizes decisions (short Question topic + USER: / TL: / verbatim-spec Answer). Nobody should answer planning questions inside the markdown file or paste walls of chat back into it — that defeats the point of an LLM-assisted interactive session. Rows with Frozen spec: + H for judgment without USER: are still SPEC_ROLEPLAY. |
| "I'll make the human review the plan file like a form" | BLOCKED UX. Rounds are in chat; the file is for implementers (outcomes + Section 1b + tasks). Do not require humans to “fill in” the plan as the primary Q&A surface. |
| "I'll wait for the user to say explore deeper / touchpoints / full plan" | BLOCKED. ### 1b.2a, Section 1b.6 deep discovery, and maximal Section 1b elaboration are default — run Read / rg / glob on the product repo + brain codebase/ until evidence exists or BLOCKED. Do not ask “should I continue exploring?” Judgment (ownership, product tradeoff, waiver) pauses only for Section 0.1 chat rounds — not for mechanical discovery. |
"REVIEW_PASS without pasting FORGE-GATE markers" | BLOCKED for any pipeline that runs verify_forge_task.py --strict-tech-plans. Self-review inventory + recross must sit in Section 1c with the two <!-- FORGE-GATE:… --> lines — see Section 1c item 2b. |
If you are thinking any of the above, you are about to violate this skill.
EVERY TASK IN A TECH PLAN IS SELF-CONTAINED, COMPLETE, AND EXECUTABLE IN ISOLATION. NO PLACEHOLDERS. NO PSEUDOCODE. NO "SEE SPEC" REFERENCES. THE PLAN IS THE ONLY THING THE DEV-IMPLEMENTER READS.
Default execution posture (MUST — no follow-up prompt required):
rg / Read on the product repo until ### 1b.2a has non-empty Evidence for every Y / PARTIAL row (or BLOCKED), Section 1b.6 has no lazy UNRESOLVED where tools could answer, and Section 1b.1 / Section 1b.5 / #### 1b.5b carry concrete shapes where applicable. Forbidden: shipping a thin outline and expecting the user to prompt again for “touchpoints” or “full exploration.”Normative claims (companion rule): Every interface claim in a task (path, field name, status code, topic name, column) must be copied from the frozen shared-dev-spec.md or the task-local inlined excerpt of contracts/* — not invented in the tech plan. If shared-dev-spec was thinner than reality, fix the spec (change request / re-council) — do not “paper over” in tasks. Program / rollout / sequencing lives in ~/forge/brain/prds/<task-id>/delivery-plan.md (non-frozen); tech plans may reference it by heading but must not rely on it for interface truth.
Product terminology (terminology.md): When ~/forge/brain/prds/<task-id>/terminology.md exists, treat it as the per-task product term sheet (distinct from the Forge plugin glossary — forge-glossary). Read it before writing user-facing strings, error messages, or marketing-adjacent labels in Section 1b / Section 2; align renames with open_doubts resolution and docs/terminology-review.md. If the file is absent and the PRD introduces named product concepts or branded labels, follow intake-interrogate to create it or record N/A with WAIVER in Section 0 when the task policy allows (see terminology-review slice table).
Optional PM traceability (inside each tech-plans/<repo>.md): You may group Section 2 tasks under IDs like REVERIF-<AREA>-<nn> with columns Est / Deps / Acceptance / Spec refs (link to shared-dev-spec heading or contracts/ heading). This does not replace one-file-per-repo or self-contained task bodies.
Cross-map between Anti-Pattern rows and Red Flags so you don't have to read both lists to find the relevant rule.
| Concern | Anti-Pattern row | Red Flag |
|---|---|---|
| Completeness scope (brief vs maximal) | "The skill says brief…" | "Plan has no Section 0 / 1b / 1c" |
| PRD traceability | "I'll cover PRD cases implicitly" | "Section 1b.0 missing, empty, or has PRD/spec rows without Section 2 task ids" |
| Intent / why columns | "Mechanics without intent…" | "Missing product intent trace" |
| Schema / wire shapes | "Schema / payload details can wait" | "Section 1b.1 / Section 1b.5 / Section 1b.1a use vague language" |
| Async contracts | (see Red Flags) | "Async contracts missing" |
| Touchpoint exploration | "Touchpoint table is boilerplate…" | "### 1b.2a missing, shallow, or misplaced" |
| Interactive vs form UX | "I'll put the whole dialogue in Section 0" / "I'll make the human review the plan file" | "Section 0 fake interactive" / "Section 0 cohorts via SPEC_INFERENCE" |
| Scan authority | "Scan is optional if I know the repo" | (see Overview — SCAN_INCOMPLETE failsafe) |
| Review gate | "REVIEW_PASS without FORGE-GATE markers" | "Tech plan status: REVIEW_PASS with no self-review round" |
| Phantom backend work | "Every good plan has migrations and index mappings" | "Section 2 tasks exist before applicable Section 0.2 work is done" |
Jump to: Anti-Pattern Preamble · Iron Law · Red Flags · Section 0 · Section 1b · Section 2 · Checklist
If you notice any of these, STOP and do not proceed:
codebase/ or repo paths and how each row was confirmed (rg/Read/scan artifact); implementers cannot act on scale alone.prd-source-confluence.md is treated as “too big” to read end-to-end or to mirror into Section 0c / 1b — STOP. AGENTS.md Core rule 6 — chunk reads, complete traceability work, or BLOCKED with a concrete tool limit — never self‑authorized truncation.### 1b.0b implementation at-a-glance (surface table + cross-repo deps + prose summary), Section 1b.1–1b.6 (each with either delta content or the skill’s explicit one-line N/A where not applicable), #### 1b.5b when round D applies (or its one-line N/A), ### 1b.2a touchpoint & boundary inventory (full exploration table + Exploration notes), or Section 1c. Micro-tasks without inventory and without cleared doubts hide gaps. STOP. Add Section 0 (incl. Section 0.2 rounds that apply) then Section 1b.0 → Section 1b.0b → … → ### 1b.2a → ### 1b.6 before Task 1.design_brain_paths / Lovable repo — Figma captured in Q9 is not decorative; it must drive the component/screen plan. STOP. Add the design→UI table and align tasks to nodes or brain paths.design_waiver: prd_only + scan reuse — Implementers cannot verify pixels or reuse. STOP.METHOD+path, GraphQL operation, SOAP action, …). STOP.Tech plan status: REVIEW_PASS with no tech-plan-self-review round logged in Section 1c revision log — Rubber-stamp. STOP.tech-plans/HUMAN_SIGNOFF.md + [TECH-PLAN-HUMAN] — Human feedback phase skipped. STOP.UNCONFIRMED rows) but Section 2 tasks are already written — Planning was short-circuited. STOP. Resolve or BLOCK before tasks.Frozen spec: paraphrase and Confidence H but no USER: / PO: / TL: prefix and no verbatim single-paragraph quote (Section 0.1 rule 3). That is SPEC_ROLEPLAY, not cleared doubt. STOP. Re-ask in chat or downgrade to M + SPEC_INFERENCE.SPEC_INFERENCE — Product segmentation, eligibility, or batch exclusions locked with SPEC_INFERENCE / H without USER: / PO: / TL: or brain touchpoints/COHORT-AND-ADJACENCY.md. STOP. REVIEW_PASS forbidden until fixed (Section 0.1 rule 6).contract-event-bus names topics / queues / exchanges / JMS destinations / partitions this repo produces or consumes, but #### 1b.5b is absent, N/A without citation, or prose-only (no fenced payload in the contract’s format — JSON, XML, Protobuf text, … — or verbatim contract). STOP. Section 1b.5 synchronous N/A does not remove the obligation to document messages/cache when in scope.Why: clause; or Section 1b.4 / Section 1b.5 / #### 1b.5b tables omit PRD / rationale where the skill prescribes them; or any Section 2 task omits Traces to: or Rationale: — STOP. Add intent lines until every change is tied to prd-locked.md / spec / contract obligation.### 1b.2a missing, shallow, or misplaced — No touchpoint inventory; table rows with empty Evidence / no repo paths for Y categories; Exploration notes missing or generic (“looked at repo”); or ### 1b.2a appears before Section 1b.5 / #### 1b.5b (cannot cite concrete ops/topics) — STOP. Run full exploration mode per ### 1b.2a.### 1b.2b missing when gate applies — Elaborative work (multi-file net-new, ≥3 integrations, or multiple PARTIAL touchpoints) but no first-session reconnaissance (git block, ≥5 minimum reads, ≥2 discovery commands) — STOP. Brain scan alone is not a work order.This skill converts a locked shared-dev-spec into bite-sized, executable technical implementation plans per project. Each task is 2-5 minutes of execution with exact file paths, complete code (no placeholders), and exact bash commands.
Primary audience: A human or agent who does not already know the product repos. The plan must stand alone: brain scan supplies where things live today; locked intake design (prd-locked.md and the Design source (from intake) section in shared-dev-spec) supplies what net-new UI should match (Figma nodes, design_brain_paths, Lovable GitHub tree, or an explicit waiver). Taking Figma in intake but omitting it from the tech plan is the same class of failure as skipping migrations — implementers will invent components and ship visual bugs.
Order of operations for paths: Before naming files in tasks, load ~/forge/brain/products/<slug>/codebase/ (at least index.md, SCAN.json, and the modules/*.md files that match the spec’s surfaces). Failsafe: if SCAN.json exists, run python3 tools/verify_scan_outputs.py <that/codebase> (up to 3 tries, 1s backoff). On persistent failure, prefix the plan with SCAN_INCOMPLETE and do not treat brain paths as authoritative until /scan passes verify — deepen with targeted rg/reads or BLOCKED per Section 1b.6. Derive exact repo-relative paths from verified brain material, then read the product repo only to pull current file contents for “complete code” blocks. If scan is missing or >7 days old, note it and align with product-context-load / user on /scan <slug> before finalizing paths.
Order of operations for UI: When this repo is web or app, read Design source (from intake) in the locked spec and any ~/forge/brain/prds/<task-id>/design/ ingest notes (MCP_INGEST.md, README.md, …). Complete Section 1b.4 before writing UI tasks so every screen/component change is tied to design anchors and/or scan-backed reuse paths — not to memory of a Figma URL from chat.
Order of operations for naming conventions: Before writing any task file paths, function names, class names, or variable names in Section 1b or Section 2, read ~/forge/brain/products/<slug>/codebase/code-style.md. All names, import shapes, async patterns, and error handling patterns in the plan must conform to this file. Fallback if absent (scan not yet run): open the 3 most recently modified source files in the product repo (git -C <repo> log --diff-filter=M --name-only -3 -- '*.ts' '*.py' '*.kt' | grep '\.') and infer conventions. Log [WARN] code-style.md absent — naming conventions inferred from recent files in Section 0 of the plan.
Elaboration bar (default = maximal where applicable): Tech plans are exhaustive by default for surfaces this repo owns, not minimal. Every PRD success path, edge case, failure mode, and non-functional requirement (latency, security, audit, rollback) that touches this repo must be visible in Section 1b.0 + Section 1b tables + ### 1b.2a touchpoints + Section 2 — if it is only in the planner’s head, the plan is incomplete. Prefer over-specifying (field names, types, indexes, status codes, idempotency keys) within the frozen contracts for those surfaces over leaving "reasonable defaults" to the implementer. Surfaces not owned here get explicit N/A, not filler. If the frozen spec is silent on a detail, record Section 1b.6 unknown or Section 0 question — do not silently invent.
Purpose: Sub-tasks inherit every gap you skip while “planning.” This section forces questions until doubt is low — no artificial cap on how many you ask.
USER: / role / spec / WAIVER), (c) Explicit one-line N/A for a Section 1b subsection this repo does not own. They do not mean the overall tech plan should be short.#### 1b.5b bodies, ### 1b.2a (including Exploration notes), Section 1b.3, Section 1b.6, and Section 2 must be as long as required for completeness — maximal where this repo owns the surface (Elaboration bar above). Forbidden: Trimming tables, skipping fenced examples, or thinning 1b.2a to “look professional.”delivery-plan.md, parity/ material, brain scan, another repo’s plan draft, or explicit BLOCKED with who must answer — all valid. Chat history alone is not durable; write outcomes into this plan or ~/forge/brain/prds/<task-id>/planning-doubts.md (optional file) and summarize in Section 1b.6 when they affect code paths.PO: / TL:). Then append rows to Section 0 that record what was decided — not a transcript of the whole thread. Hard ordering: Do not add Section 0 rows with answers until the corresponding topic has been resolved in this chat session. Do not add rows whose Answer pretends a human spoke when only the spec was read (see Red flags — Section 0 “fake interactive” / SPEC_ROLEPLAY). Question column: one short line naming the topic (e.g. “Who owns cron vs MQ for schedule?”) — not multi-paragraph copy-paste from chat or from the file as a fake “questionnaire.” Answer column: USER: / PO: / TL: + one-line outcome, or verbatim spec quote + path for non-judgment facts, BLOCKED, or WAIVER. Confidence H on judgment rows requires USER: / role prefix or verbatim spec — else M + SPEC_INFERENCE. Forbidden: Asking the human to “answer in the plan file” or requiring verbatim paste of chat into the repo — the value of the LLM is running the interactive round in the session; the file is the durable decision record for implementers.USER: / PO: / TL: (or verbatim spec quote that explicitly encodes the segment rule) — not SPEC_INFERENCE with Confidence H. REVIEW_PASS is forbidden until touchpoints/COHORT-AND-ADJACENCY.md is USER-backed or waived (docs/adjacency-and-cohorts.md). Optional: If discovery-adjacency.md lists hits for a shared entity this PRD touches but ### 1b.2a omits them → CHANGES.Forbidden: Dump a full Section 2 task list first, then add “follow-up questions” for persistence, search, or API contracts in an appendix. That inverts risk: implementers see tasks without locked contracts.
Required cadence with the human: Work in rounds — each round ends with explicit questions and a pause for answers before the next contract surface is finalized in Section 1b and Section 0.
Skip rounds that do not apply to this repo: If Section 1b.1 will be the one-line no persistence N/A, do not run round B as a “find migrations / collections / CH migrations” scavenger hunt — write the N/A ground (spec section / affected projects) and proceed. Same for C when Section 1b.1a is N/A, and A when Section 1b.5 is fully N/A (no synchronous API server or client work for REST/GraphQL/SOAP/etc.). When A is partial (e.g. consumer-only), run A only for client contract alignment — not for persistence another repo owns.
| Round | When to run | Cover | Before proceeding |
|---|---|---|---|
| A — Synchronous API surface | This repo serves or consumes a locked synchronous API for this task (REST over HTTP, GraphQL, SOAP / WSDL, gRPC, … — whatever shared-dev-spec + contract-api-* name) | REST: paths, methods, bodies, errors; GraphQL: operations, variables, errors; SOAP: operations, SOAPAction / QName, XML envelope samples or XSD refs; auth, versioning, idempotency | Section 1b.5 draft owner + consumer rows + fenced payloads in the contract’s wire format; user confirms or BLOCKED. If no synchronous API surface: skip — Section 1b.5 one-line N/A only. |
| B — Persistence / durable store | This repo owns schema or migration work for this task (any engine: relational SQL, MongoDB, ClickHouse, Dynamo, Redis persistence, … per contract-schema-db) | Collections/tables/partitions, indexes, TTL, nullability / sharding / backfill strategy as locked | Section 1b.0 rows for persistence + Section 1b.1 delta + verbatim or fenced schema in the contract’s native language per Section 1b.1 rules; user confirms or BLOCKED. If not owning persistence: skip — Section 1b.1 one-line N/A + sibling repo if applicable. |
| C — Search / ranked retrieval | This repo owns index, mapping, or ingest for this task (Elasticsearch, OpenSearch, Solr, Typesense, Meilisearch, vector DB index, … per contract-search) | Index or collection name, field definitions, analyzers / embedders, reindex or dual-write | Section 1b.0 rows for search + Section 1b.1a + fenced definition (often JSON; use XML/YAML if that is what the contract locks) or verbatim contract per Section 1b.1a rules; user confirms or BLOCKED. If not in scope: skip — Section 1b.1a one-line N/A with spec citation. |
| D — Cache / events | Contracts or spec assign cache/event work to this repo | Key patterns, TTL, destinations (Kafka topic, RabbitMQ queue+exchange+routing key, JMS queue, …), ordering, idempotency, payload keys (contract-cache / contract-event-bus) | #### 1b.5b (below) filled with tables + fenced payload (JSON / XML / …) or verbatim contract per rules or one-line N/A with citation; Section 0 ties open questions here; skip round D only when both contracts are out of scope for this repo. |
Chat style: Prefer short messages with numbered questions in chat (not a megabyte of Q&A in the thread). After chat resolutions, write Section 0 as compact outcome rows, then fully elaborate Section 1b (including Section 1b.0 synced to the latest PRD/spec rows) — Section 1b may be long; length is not a failure mode. Only then add/expand Section 2 tasks so they inherit the locked shapes.
tech-plans/<repo>.md or linked file)Include before ## Section 1b:
## Section 0: Planning doubt log
<!-- Optional one line: Planning rounds completed in chat (date) — rows below are outcomes for implementers, not the live Q&A transcript. -->
| Q# | Question (short topic — what was decided about) | Answer / resolution | Confidence (H/M/L) | Affects (Section 1b.x / Task ids) |
|----|---------------------------------------------------|----------------------|--------------------|----------------------------|
| Q1 | e.g. Schedule ownership (cron vs MQ) | USER: … | H | 1b.5b, T1–T4 |
No material doubts — ready to elaborate. (after confirming in chat per Section 0.1 rule 3).USER: / role / verbatim-spec discipline (Section 0.1), fail self-review. Do not require users or agents to paste entire chat logs into the repo.~/forge/brain/prds/<task-id>/shared-dev-spec.md (or the task’s frozen spec path from brain-read)prd-locked.md at ~/forge/brain/prds/<task-id>/prd-locked.md — success criteria, scope, design/Q10, and acceptance language (must be reflected in Section 1b.0)Read the spec file to understand:
Extract per-project work items by identifying:
Map to repositories (standard Forge topology):
shared-schemas/ — Shared TypeScript types, validation schemas, contractsbackend-api/ — Node/Express REST API, database migrations, business logicweb-dashboard/ — React SPA, UI components, state managementapp-mobile/ — React Native app, mobile UI, offline-first patternsAuthoring order in the saved tech-plans/<repo>.md file: (1) # title line, then Tech plan status: DRAFT (or REVIEW_CHANGES / REVIEW_PASS per Section 1c); (2) ## Section 0: Planning doubt log (outcome table after chat rounds; see Section 0, including Section 0.2 interactive rounds for applicable surfaces only); (3) this Section 1b starting with ### 1b.0, then ### 1b.0b, then ### 1b.1 and ### 1b.1a each as either the delta table or the skill’s one-line N/A, then ### 1b.2 (reuse vs net-new), ### 1b.2b (first-session reconnaissance or one-line N/A per gate), ### 1b.3 (spec trace bullets), ### 1b.4 as applicable, ### 1b.5 (synchronous API or one-line N/A), #### 1b.5b (brokers + cache or one-line N/A) when round A / D or contracts apply, then ### 1b.2a touchpoint & boundary inventory (full exploration mode — after wire maps so evidence cites real paths/operations), then ### 1b.6); (4) Section 1c body (revision log table + cross-repo notes); (5) Section 2 tasks.
Surface applicability (generic — no stack bias): Before deep-diving persistence, search, synchronous APIs, or message brokers, decide what this repo actually owns for this task using shared-dev-spec.md (affected projects / ownership), prd-locked.md, and ~/forge/brain/products/<slug>/codebase/ (e.g. SQL migrations, ORM models, Mongo migrations/validators, ClickHouse .sql / ALTER, Kafka/AMQP client usage, ingest workers, api-surface.md, GraphQL schema folders, WSDL paths, client modules). Maximal detail applies per applicable surface — not “always one RDBMS + one Lucene-derived search + REST-only.” Complete means: every in-scope case is in Section 1b.0, and every Section 1b subsection is either fully elaborated (tables + fenced blocks in the contract’s language — SQL, JSON, YAML, XML, SOAP, SDL, .proto, …) or an explicit one-line N/A + spec citation (and optional N/A (other repo: tech-plans/<file>.md) pointer). Forbidden: (a) leaving Section 1b.1 / 1b.1a / 1b.5 / #### 1b.5b blank without the prescribed N/A line; (b) inventing persistence schema, search definitions, operations, or destinations for a frontend-only (or otherwise non-owning) repo; (c) discovery loops (hunting migrations or index templates for engines this repo does not use) when the frozen spec + topology already show no ownership here — record N/A and move on.
Bite-sized tasks exist so a dev-implementer in isolation can execute without guessing. They do not replace Section 0 (cleared doubts + Section 0.2 user rounds where applicable), nor ### 1b.2b (ordered git + minimum reads + discovery commands when elaboration gate applies), nor ### 1b.2a touchpoint & boundary inventory (full exploration — every integration surface enumerated with evidence, not vibes), nor a short, explicit narrative of what changes in the world for this repo (data, search indices, reuse, design, synchronous API wiring (REST/GraphQL/SOAP/gRPC/…), broker destinations & payloads, cache, unknowns, review trail — omit subsections that are N/A, do not pad them). Subsection 1b.4 follows web/app rules; 1b.5 follows the locked API style when this repo serves or consumes that surface; #### 1b.5b follows contract-event-bus / contract-cache when this repo touches those surfaces; 1b.1a follows search when this repo owns index/mapping work; 1b.6 is always required (may be a single “no unknowns” line). Section 1c is always required. All of the above before Section 2.
Skipping them because “the tasks are obvious” or “only micro-steps matter” is BLOCKED — that is how schema drift, duplicate persistence shapes, wrong screens, wrong API wiring, and silent greenfield work slip through.
Purpose: Prove every in-scope requirement, success criterion, edge case, and non-functional constraint from prd-locked.md and shared-dev-spec.md is either implemented in this repo’s Section 2 or explicitly out of scope with a citation. Scan proves where work lands.
Build this table before ### 1b.1 (add rows until nothing material is missing):
| PRD / spec ref (id or heading) | Requirement or acceptance (one line) | Why (rationale) — what user/system outcome this row serves; what goes wrong if skipped (≠ restating the requirement cell) | Brain scan evidence (codebase/... path or SCAN.json note) | Owner Section 1b subsection(s) | Section 2 task id(s) or N/A (other repo) |
|---|---|---|---|---|---|
e.g. prd Q5, spec FR-2 | e.g. "User can revoke session" | e.g. "Stops stale tokens after compromise; audit trail for SOC2" | e.g. modules/web-auth-svc.md → src/auth/Revoke.ts | 1b.2, 1b.5 | T12–T18 |
Rules (MUST):
N/A + sibling repo file). No Section 2 task that does not tie back to a row or an explicit tech-debt / chore exception with human approval in Section 0.prd-locked.md / spec text.codebase/ paths (or UNKNOWN → Section 1b.6 until resolved). Guessing paths without scan is BLOCKED.web-dashboard must use N/A (other repo: …) in the task column and list only the Section 1b subsections this repo will touch (e.g. 1b.4, 1b.5 consumer rows). Do not duplicate backend Section 1b.1 / 1b.1a bodies here — point at the owning plan file instead.PRD trace gate (MUST — before the matrix is “done”): Read prd-locked.md end-to-end for this task. Every material bullet under scope, success criteria, acceptance / test language, NFRs, and edge / negative cases must land as a Section 1b.0 row (or a row pointing at shared-dev-spec.md / contract text that carries the same obligation), or as an explicit WAIVER row with citation. Silent omission of PRD language is BLOCKED at tech-plan-self-review — the matrix is not a summary of “main flows” only.
Purpose: One place—before dense Section 1b.1 / Section 1b.5 / #### 1b.5b—that states what changes, what this repo owns vs a sibling tech-plans/*.md, which locked contracts apply, and merge / land ordering across repos. Forbidden: Empty subsection, “see spec”, or only a link to shared-dev-spec.md without row-level ✓/N/A.
1 — Surface ownership (every row: ✓ = this repo elaborates in the cited subsection(s); N/A = one-line reason + spec section or sibling plan):
| Delta class | This repo | Detailed in subsection | If N/A: sibling plan / owner |
|---|---|---|---|
| Persistence / DDL / migrations | ✓ / N/A | Section 1b.1 | e.g. blue-ocean-server.md |
| Search / index / ingest | ✓ / N/A | Section 1b.1a | … |
| Synchronous API serve | ✓ / N/A | Section 1b.5 | … |
| Synchronous API consume | ✓ / N/A | Section 1b.5 | … |
| Async messages publish | ✓ / N/A | #### 1b.5b | … |
| Async messages consume | ✓ / N/A | #### 1b.5b | … |
| Cache (invalidate / read) | ✓ / N/A | #### 1b.5b | … |
| Cron / jobs / internal-only HTTP | ✓ / N/A | Section 1b.5 + Section 2 | … |
| Shared types / packages | ✓ / N/A | Section 1b.2 + tasks | … |
| Web / app UI | ✓ / N/A | Section 1b.4 | … |
2 — Cross-repo dependencies (add rows until none left implicit):
| Needs from (repo / plan file) | Interface (contract path, OpenAPI path/operation id, topic, table) | Blocks this plan? (Y / N / parallel OK) |
|---|---|---|
| … | … | … |
If none: one line — No cross-repo ordering risk — all interfaces frozen under task contracts/ + shared-dev-spec.md.
3 — Implementation summary (3–8 sentences, MUST): Prose for this repo only—runtime behavior change, main entrypoints (routes, workers, UI surfaces), data written/read, failure modes called out in PRD/spec, env / feature flags. Not marketing copy; enough that a cold implementer knows why Section 2 exists.
Store-agnostic: The frozen contract-schema-db names the engine(s) — relational (PostgreSQL, MySQL, …), document (MongoDB, …), columnar (ClickHouse, BigQuery table DDL, …), key-value, graph, etc. Match the contract’s vocabulary (table, collection, merge tree, materialized view, …); do not assume SQL unless the contract is SQL.
Ground this in the locked shared-dev-spec and any db-contract / contract-schema-db material — do not invent persistence objects here that the spec did not lock.
| Logical object (table / collection / topic / bucket / …) | Change type | Rationale (one line — product / PRD: cite prd-locked section/Q when not obvious from contract title alone) | Rollback or backward-compat note |
|---|---|---|---|
| (one row per schema-affecting change: CREATE, ALTER, DROP, new index, validator, TTL, partition, …) | use contract terminology | why this persistence exists for the user journey / compliance | e.g. nullable-first, dual-write, backfill, irreversible — link to contract if long |
No database or durable storage schema changes in this repo.Every later migration / schema-change task in this file MUST correspond to a row above (same logical object + change type). Orphans either way are a planning defect.
Schema detail (MUST — not optional): Applies only to rows in the delta table. For each such row, include either (a) the full proposed change in a fenced code block using the same language the frozen contract uses — e.g. SQL (CREATE TABLE / ALTER for relational or ClickHouse SQL dialect), JSON (Mongo validators, createIndexes, collection JSON Schema if locked in contract-schema-db), YAML/XML/DSL if the contract locks that — immediately under the table (or per-row sub-blocks), or (b) a verbatim paste from the locked contract-schema-db / shared-dev-spec excerpt with heading + file path. Spell out fields (columns, BSON paths, nested docs), types, nullability / required keys, defaults, indexes, sharding/partition/TTL, and referential rules (FKs, $lookup, weak refs) per what the store actually supports and the contract locks. Forbidden: "Add fields as needed", "migration per ORM", or deferring shape to Section 2 without an inlined or verbatim contract body here. When the section opens with the one-line N/A (no persistence in this repo), no schema block is required — do not fabricate persistence shapes to satisfy this skill.
When N/A: One line: 1b.1a not applicable — no search index ownership or contract-search surface in this repo for this task. (cite spec section if ambiguous.)
When the locked spec includes search / contract-search: Mirror the contract — do not invent fields in the plan that the spec did not lock.
| Index, alias, or collection (per contract-search) | Definition / mapping / schema change | Reindex, dual-write, refresh, or vector rebuild policy | Rollback / compat | PRD / rationale (one line) |
|---|---|---|---|---|
| (row per search surface touched) | e.g. new field spec, analyzer, vector params, Solr fieldType | e.g. reindex job, blue/green, incremental backfill | e.g. read old alias until drain | which search UX / reporting obligation this index serves |
Every later index / mapping / ingest task in Section 2 MUST correspond to a row above.
Index-definition detail (MUST — not optional): Applies only to rows in the table. For each such row, include either (a) the full proposed definition in a fenced block using the format the contract locks — often JSON (Elasticsearch, OpenSearch, Typesense, Meilisearch, Algolia index settings, many vector DB APIs), sometimes XML (Solr schema.xml snippets), YAML, or another DSL — immediately under the table (or per-row), or (b) a verbatim paste from the locked contract-search / shared-dev-spec with heading + file path. Cover field names, types, analyzers / tokenizers / normalizers (when applicable), vector / embedding dimensions and distance (when applicable), and dynamic / nested policies per engine. Forbidden: "mapping per team norm", "see dashboard X", or deferring shapes to Section 2 without an inlined or verbatim contract body here. When the one-line N/A applies, no definition block is required.
Summarize so reuse is not taken for granted from task ordering alone. Prefer evidence from ~/forge/brain/products/<slug>/codebase/ (index.md, modules/*.md, api-surface.md): reuse bullets should cite paths that appear in the scan when the repo was scanned. If the scan is missing or stale, state SCAN_MISSING_OR_STALE and trigger /scan <slug> (or equivalent) before claiming paths.
DISCOVERY_REQUIRED or HUMAN_CONFIRM — do not silently pick a module.HARD-GATE — Net-New File Declaration: Before any task can declare a file as net-new, the tech plan MUST show evidence from the brain scan or a grep command that no existing file serves this purpose.
Required evidence format:
Checked: brain/products/<slug>/codebase/modules/<role>-<module>.md — no UserService found
Checked: rg "class UserService" <repo> — 0 matches
Decision: net-new at src/services/UserService.ts
If evidence is missing → the task is BLOCKED until evidence is provided.
If a match IS found → the task must say "modify existing: <path>" not "create new".
Cross-task duplicate check (HARD-GATE during tech-plan-self-review): Before marking any file as net-new, scan all other ### Task N declarations in THIS plan for the same path. Two tasks declaring the same file net-new means both implementers will create it — last writer wins, first writer's work is silently discarded.
# Run during tech-plan-self-review to catch duplicate net-new declarations
grep -n "net-new\|Create:" tech-plans/<repo>.md | awk -F: '{print $3}' | sort | uniq -d
If duplicates are found: consolidate into one task (the one that owns the file's primary responsibility) and make the other task modify the existing file instead.
Why this exists: scan-codebase + codebase/index.md / modules/*.md answer where things probably live. They do not replace an implementer’s first hour in the product repo — the failure mode you feel as “scan/review don’t work” is often only brain stubs, no ordered assignment. This block is the handoff from navigation → execution.
HARD-GATE — fill when any of: net-new ≥2 files or packages in this repo for the task; ≥3 distinct integration touchpoints (HTTP, DB, bus, cron, third party); or ### 1b.2a marks PARTIAL for ≥2 categories. One-line N/A only when the repo change is trivial (single-file, single-call-site) and Section 1b.6 has no UNKNOWN for those paths.
git fetch + git branch -a (or equivalent) scoped to branches that might collide with this work; git log -1 --oneline on the branch the plan assumes. Use real remote/default names from discovery.md / context-loaded.md / product convention — not placeholders.validateSchema + auth pattern, controller index). For each path: one clause what to extract (e.g. “every exported symbol you will call”, “middleware order”). Forbidden: only [[wikilinks]] to brain codebase/ with no product-repo path.rg / grep (or fd) lines that locate extension points (e.g. banner system, layout shell, existing Modal import). Must run from repo root shown in the block.Self-review: If this subsection is missing when the gate applies, or paths are generic (controllers/foo.js without naming the real hub), → CHANGES — same bar as shallow ### 1b.2a.
Bullets mapping prd-locked.md bullets / Q ids, shared-dev-spec.md requirement headings, and contracts/* IDs to Section 2 task ids in this file.
Format (MUST — every bullet):
- **<source id>** (e.g. prd success #2, spec FR-3, contracts/api-rest POST /foo) → **Tasks T…** — **Why:** <one sentence — product or integrity reason this slice exists; what PRD pain is removed>
If a requirement has no task, STOP — fix coverage before dispatch. Forbidden: bullets that only list “FR-3 → T1–T4” with no Why: clause.
Purpose: Intake may lock figma_file_key + figma_root_node_ids, design_brain_paths, lovable_github_repo, or design_waiver: prd_only. Council may copy that into Design source (from intake) in shared-dev-spec.md. None of that helps E2E delivery if the tech plan never maps design to concrete components, routes, or files. This subsection is the handoff from design transport to implementation tasks.
When to fill the table vs one-line N/A
| Situation | Required content |
|---|---|
| This file is for a backend, worker, shared-schema, or other non-UI repo | One line: 1b.4 not applicable — not a web/app repo. |
design_ui_scope: not applicable in lock (no user-visible UI for this slice) | One line: 1b.4 not applicable — no user-visible UI in this repo for this task. |
design_new_work: yes OR implementable design paths/keys exist (Figma, design_brain_paths, Lovable repo) | Full table (below) + ensure each UI task later references a row id (e.g. D1) in its title or header |
design_new_work: no but UI still changes | Table or bullet list: which existing components (from scan) are extended; if any new file is unavoidable, say why PRD/engineering-only still requires it |
Design → implementation mapping table (add rows until every net-new or changed screen in scope is covered):
| Id | Design anchor (Figma node id / frame name / path under .../design/ / Lovable route) | UI deliverable (screen, component, layout region) | Existing code (path from brain scan) or NET_NEW | PRD section / Q / success criterion — rationale (what obligation this UI closes) | Design parity, tokens, a11y |
|---|---|---|---|---|---|
| D1 | (e.g. 123:456 or design/wireframes/checkout.png) | (e.g. Checkout summary card) | (e.g. src/features/cart/Summary.tsx or NET_NEW) | (e.g. prd success #2 — recruiter sees verification state before posting) | (states, contrast, keyboard) |
Rules
design_waiver: prd_only is locked, the table still lists screens/components and maps them to scan-backed files or NET_NEW, with one line citing the waiver for pixel latitude.D<n> or explicitly say waiver + PRD section so reviewers can trace intake → plan → code.Purpose: E2E delivery requires a written answer to: which component calls which operation, and where the handler / resolver / service implementation lives — using the same style the frozen contract uses (not “everything is REST”).
When N/A: This repo has no synchronous API server and no synchronous API client changes for this task — one line: 1b.5 not applicable — no synchronous API surface (REST/GraphQL/SOAP/gRPC/…) in this repo for this task. (cite spec section if needed.)
Contract styles (fill the tables that apply; omit styles the spec marks N/A):
REST (HTTP + JSON/XML as locked): When the repo implements or changes REST handlers:
| Endpoint (METHOD path) | Handler (repo-relative path : symbol or class#method) | Auth / versioning / idempotency | PRD / spec trace + rationale (one line — which acceptance or FR this operation fulfills; why it exists) | Consumers — component path or tech-plans/<other>.md Section 1b.5 row id |
When the repo consumes REST only (or add consumer rows in addition to the owner table above):
| Consumer (path : component or hook) | When it runs | METHOD + path | Client module (path) | Success + error handling (status / body shape) | PRD / rationale (one line) |
GraphQL: When the repo serves or consumes GraphQL per lock — add a final column PRD / rationale (one line) to each table you fill (same rule as REST).
| Operation (query / mutation / subscription + name) | Schema entry point or document id | Server handler or client call site (path : symbol) | Variables shape | Auth / errors (errors[] / extensions) | PRD / rationale (one line) |
SOAP / XML-RPC: When the repo serves or consumes SOAP per lock — append column PRD / rationale (one line) (same rule as REST).
| Service / port / binding (WSDL ref) | Operation (QName or SOAPAction) | Handler or client (path : symbol) | Request XML sample or XSD element ref | Response XML sample or fault shape | PRD / rationale (one line) |
gRPC / protobuf: When locked — append column PRD / rationale (one line).
| Package.Service/Method | .proto ref | Server impl / client stub (path) | Request / response message fields (or verbatim .proto excerpt) | PRD / rationale (one line) |
Cross-reference: If the owning plan is in another repo, cite tech-plans/<other>.md Section 1b.5 + stable operation key (METHOD+path, GraphQL operation name, SOAP QName, gRPC FullMethodName) so tech-plan-self-review Section 1d can diff them.
Wire shapes (MUST): For each new or materially changed operation in the tables above, add fenced examples in the contract’s format: json for typical REST/GraphQL variables and payloads, xml for SOAP bodies/faults, protobuf / text proto blocks if that is how the contract is written — or a verbatim OpenAPI / GraphQL SDL / WSDL / .proto excerpt with file path. Include errors, pagination, headers, or metadata when the spec defines them. Forbidden: "See WSDL / SDL / OpenAPI" without inlined shape, or TBD where the contract already decided.
contract-event-bus / contract-cache)Purpose: Services integrate through Kafka, RabbitMQ, ActiveMQ / JMS, Azure Service Bus, NATS, SQS, etc. — not only HTTP. Section 1b.5 synchronous N/A does not excuse missing #### 1b.5b when this repo publishes, consumes, or invalidates messages or cache entries named in the lock.
When N/A: One line: 1b.5b not applicable — no cache key or message-bus ownership in this repo for this task. (cite spec section / contract IDs.)
When applicable: Mirror contract-event-bus / contract-cache (or locked spec sections) — do not invent topic names or payload shapes.
| Kind | Destination id (Kafka topic[+partition strategy]; RabbitMQ exchange+queue+routing key; ActiveMQ/JMS queue/topic; SQS ARN/url; cache key pattern — verbatim from contract) | Producer (path : symbol) or EXTERNAL | Consumer or invalidator (path : symbol) or NONE | Ordering / TTL / idempotency / retry | PRD / rationale (one line — e.g. “prd section6 — notify recruiter of tier change”) |
|---|---|---|---|---|---|
| (row per changed contract item) | verbatim from contract | … | … | per lock | why this message/cache touch exists for the product |
Payload & key shape (MUST): For each new or materially changed message row, add a fenced block in the broker’s on-wire format as locked — commonly json (Kafka JSON, Rabbit JSON payloads), often xml (JMS/SOAP-wrapped bodies, legacy XML), or hex / text proto if binary contracts are documented that way — or verbatim paste from contract-event-bus. For cache rows, include value shape or invalidation rule as locked. Forbidden: Prose-only (“listens on queue X”) without inlined payload or verbatim contract when the contract defines it. If the PRD/spec never locked payloads, BLOCKED or SPEC_INFERENCE in Section 0 — do not silently invent in tasks.
Placement (read carefully): ### 1b.2a is written after ### 1b.5 and #### 1b.5b in the markdown file so rows cite real operations, topics, and handlers from those sections. Naming: 2a = integration / boundary exploration pass — not “optional appendix.”
Purpose: Full exploration — enumerate every class of external touchpoint this repo hits for this task, with scan-backed evidence and what changes. Rationale columns explain intent; this section proves you walked the walls (routes, auth, jobs, buses, third parties, flags, observability, …). Forbidden: empty table, “N/A all”, or rows with no codebase/ path / rg note / file:line.
Process (MUST — same planning session before REVIEW_PASS):
~/forge/brain/products/<slug>/codebase/ (index.md, modules/*.md, api-surface.md, route indexes, consumer registration files).rg / Read on the product repo (not only brain stubs) until you can mark Y (touched this task), N (out of scope + one-line why + prd or spec section), or PARTIAL (what remains → Section 1b.6 or Section 0).Inventory table (MUST — one row per category that is Y or PARTIAL; N requires one consolidated row per cluster or one row per N with citation):
| # | Touchpoint category | Y / N / PARTIAL | Evidence (repo paths + tool: e.g. rg pattern → src/...) | Current behavior (1 line) | What changes for this task (1–2 lines) | PRD / spec section | Section 1b subsection(s) + Section 2 task ids |
|---|---|---|---|---|---|---|---|
| 1 | HTTP ingress (router mount, gateway, version prefix) | Section 1b.5, … | |||||
| 2 | Authn / authz (session, JWT, API key, RBAC, moderator vs recruiter) | ||||||
| 3 | Middleware chain (CORS, body parser, rate limit, request id) | ||||||
| 4 | Feature flags / kill switches / env-driven behavior | ||||||
| 5 | Cron / scheduler / delayed jobs / internal-only HTTP | ||||||
| 6 | Outbound HTTP / RPC (other services, payment, KYC, webhooks you call) | ||||||
| 7 | Inbound webhooks (signatures, replay, idempotency stores) | ||||||
| 8 | Persistence (new migrations, triggers, read replicas, backfill jobs) | Section 1b.1, … | |||||
| 9 | Search / index / ingest workers | Section 1b.1a | |||||
| 10 | Message bus (publish, consume, outbox, ordering, DLQ) | #### 1b.5b, … | |||||
| 11 | Cache (read-through, invalidate-on-event, TTL races) | ||||||
| 12 | Email / SMS / push / in-app notification templates | ||||||
| 13 | File / object storage (S3, GCS, signed URLs, virus scan) | ||||||
| 14 | Observability (structured logs, metrics, traces, alerts touched) | ||||||
| 15 | Security (secrets manager, encryption at rest, CSRF, SSRF guards) | ||||||
| 16 | Compliance / audit (append-only logs, data retention, export) | ||||||
| 17 | i18n / l10n / copy tokens (if user-visible copy changes) | Section 1b.4 | |||||
| 18 | Mobile / web sibling consumers (deeplinks, App Links, shared cookies) | XALIGN / Section 1b.5 | |||||
| 19 | Shared libs / monorepo packages this repo imports | ||||||
| 20 | Build / deploy / runtime config (Docker, k8s, PM2, systemd — if this task changes them) |
Narrative (MUST): After the table, 3–10 bullets titled Exploration notes — surprises from rg/reads (e.g. “duplicate route registration in legacyRoutes.js”, “AMQP reconnect hides publish errors”). Each bullet ends with → task id or Section 1b.6 U#.
Minimum bar: If the task is non-trivial (≥3 Section 2 tasks or any Y on rows 1, 2, 5, 6, 8, 10), PARTIAL or empty Evidence on those rows is BLOCKED until resolved or BLOCKED with owner.
Purpose: Scans and specs leave gaps. Elaborative planning means closing gaps with repo evidence, not leaving them for implementers.
No material unknowns — scan + spec sufficient for every Section 2 path.):| U# | Unknown | Why it matters if wrong |
|---|
| … | Deep discovery performed (tools: e.g. rg, Read, hub file) | Resolution (RESOLVED: … | BLOCKED: …) |
Rules
/scan refresh — and Section 2 must not contain tasks that depend on that unknown until resolved.api-surface.md does not show registration, client wrapper, middleware chain, or feature flag gate: open those files in the product repo until the unknown is mapped or escalated.Tech plans are not one-shot documents. They go through review → revise → re-align until integration is legible on paper.
Status line — Placed immediately under the # title (see Section 1b authoring order). Values: DRAFT | REVIEW_CHANGES | REVIEW_PASS. Start at DRAFT. After tech-plan-self-review: all PASS → REVIEW_PASS; any CHANGES or BLOCKER → REVIEW_CHANGES, fix plans, set DRAFT, re-run review.
Revision log (append one row per meaningful edit):
| Rev | When (ISO8601 approx) | Who | Trigger (e.g. self-review Section 1b.5 FAIL, XALIGN drift) | What changed |
2b. FORGE machine gates (HARD-GATE when Tech plan status: REVIEW_PASS) — tech-plan-self-review Section 0c must be materialized in this file so CI can detect rubber-stamps (tools/verify_tech_plans.py / verify_forge_task.py --strict-tech-plans). Omit on DRAFT / REVIEW_CHANGES. Before flipping to REVIEW_PASS, append to Section 1c (above or below the revision log table, but inside Section 1c):
On its own line, immediately above the Section 0c requirement inventory table:
<!-- FORGE-GATE:SECTION-0C-INVENTORY:v1 -->
On its own line, immediately above the code recross-check evidence (paths + tools used, one bullet or row per checked path / sample rule from tech-plan-self-review Section 0c step 4):
<!-- FORGE-GATE:CODE-RECROSS:v1 -->
Why: Skills are prose; agents optimize for “sounds done.” Anchors force inventory + recross to exist in the same artifact the implementer reads — the failure mode you get without them is chat-only PASS.
First-class inputs (Section 0c inventory — HARD): If any of these exist under ~/forge/brain/prds/<task-id>/, the Section 0c requirement inventory (the markdown table immediately below <!-- FORGE-GATE:SECTION-0C-INVENTORY:v1 -->) must include ≥1 row that cites that source (source id / short text / trace columns — machine CI uses substring presence in the inventory block): prd-locked.md, prd-source-confluence.md or source-confluence.md (Confluence mirror / normative PRD body), shared-dev-spec.md, touchpoints/*.md (especially COHORT-AND-ADJACENCY, PRD-SIGNAL-REGISTRY), qa/manual-test-cases.csv. An inventory that only traces prd-locked.md while prd-source-confluence.md, touchpoints/*.md, or populated QA CSV exists is BLOCKED for REVIEW_PASS when CI enables verify_forge_task.py --strict-0c-inventory (see docs/forge-task-verification.md).
Minimum feedback loop
tech-plan-self-review on this file before treating the plan as final for dispatch.self-review round=<n> result=PASS|CHANGES|BLOCKED (short note).Cross-repo alignment (when this product task has ≥2 repos touching the same synchronous API or ≥2 repos with split producer/consumer message plans):
tech-plans/*.md are in DRAFT with Section 1b.5 filled (for the API styles in play), perform one cross-walk using the contract’s stable keys: e.g. consumer METHOD+path ↔ owner METHOD+path; GraphQL operation name ↔ server registration; SOAP QName / SOAPAction ↔ server operation; gRPC FullMethodName ↔ .proto method. Version/prefix/auth/metadata must match the lock.XALIGN PASS or XALIGN FAIL: <drift> and fix drift before setting REVIEW_PASS.Conductor logging (when orchestrated): emit [TECH-PLAN-REVIEW] and [TECH-PLAN-XALIGN] lines as specified in conductor-orchestrate State 4.
Human go-ahead (pipeline phase — after agent loops, before State 4b)
Automated tech-plan-self-review PASS + XALIGN PASS are not the same as stakeholder acceptance. Before semantic machine eval / RED (State 4b), a human must record one of:
~/forge/brain/prds/<task-id>/tech-plans/HUMAN_SIGNOFF.md from docs/tech-plan-human-signoff.template.md with status: approved (or changes_requested → revise plans → re-run steps 3–5 → new signoff), orstatus: waived with actor + reason (solo / unattended policy only).tech-plans/<repo>.md for traceability.[TECH-PLAN-HUMAN] task_id=<id> status=APPROVED|CHANGES_REQUESTED|WAIVED. No State 4b until APPROVED or documented WAIVED.Interconnection (no gaps): intake → parity (spec-freeze Step 0) → frozen shared-dev-spec → Section 0 doubt log → Section 1b (data, reuse, design, API map, unknowns) → Section 1c (agent review + XALIGN) → HUMAN_SIGNOFF.md → State 4b (QA CSV / semantic automation + manifest / RED) → implementation. Skipping a box forwards hidden gaps backward into cheaper phases — forbidden.
Immediately under each ## Task … title (before Files affected), include both blocks:
Traces to: — Bullet list of stable ids: prd-locked.md section / Q / numbered success or acceptance line; shared-dev-spec.md heading or FR id; contracts/<file>.md heading or operation id. Must tie to ≥1 Section 1b.0 row (or explicit chore exception approved in Section 0).Rationale: — 1–4 sentences: why this task exists (design intent, risk removed, user-visible outcome); what PRD or acceptance obligation it implements; what breaks in prod (or what review fails) if omitted. Forbidden: empty, or only repeating the task title.Each task must satisfy:
Traces to: + Rationale: present per Traceability & rationale aboveTest tasks must list every scenario by name, not just "write tests". Vague test tasks are rejected.
❌ Wrong:
- [ ] Write tests for UserService
✅ Required:
- [ ] Write failing tests — UserService test matrix (RED phase):
- `createUser_happyPath_returnsUser`
- `createUser_duplicateEmail_throwsConflictError`
- `createUser_dbDown_throwsServiceUnavailable`
- `createUser_emptyName_throwsValidationError`
- `createUser_integration_persistsToDb`
Each must be watched FAIL before proceeding to implementation tasks.
## ❌ Task: Add validation
- Files: backend-api/routes/auth.js
- Code: "Add validation logic here"
- Test: "Run npm test"
## ❌ Task: Create user model
- Files: backend-api/models/user.js
- Code:
```js
class User {
...
validateEmail() { /* validation */ }
}
### Correct Example
```markdown
## ✓ Task: Add email validation to User model
- Files: backend-api/models/user.js
- Code: (complete class with method)
```js
class User {
constructor(data) {
this.id = data.id;
this.email = data.email;
}
validateEmail() {
const emailRegex = /^[^\s@]+@[^\s@]+\.[^\s@]+$/;
if (!emailRegex.test(this.email)) {
throw new Error('Invalid email format');
}
return true;
}
}
module.exports = User;
npm test -- models/user.test.js (expect: email validation throws on invalid)
### Task Template
```markdown
## Task N: [Specific action] in [file/module]
**Traces to:**
- `prd-locked.md` — … (section / Q / success line id)
- `shared-dev-spec.md` and/or `contracts/…` — …
**Rationale:**
1–4 sentences: why this change; which PRD obligation it closes; what fails if skipped.
**Files affected:**
- /path/to/file1.js
- /path/to/file2.ts
- etc.
**Complete code for each file:**
(Show every line, no abbreviations)
**Exact bash command to test:**
```bash
npm test -- specific-test-file.test.js
Expected output: (What success looks like)
Git commit message:
feat: [action description]
---
## Section 3: Task Ordering
### Dependency-First Approach
Tasks must respect this order:
1. **Layer 0: Shared Schemas** (no dependencies)
- TypeScript types
- Validation schemas (zod/joi)
- Contract definitions
- Shared constants and enums
2. **Layer 1: Backend Infrastructure** (depends on schemas)
- Database migrations
- Models and DAOs
- Business logic services
- Error handling and middleware
3. **Layer 2: Backend APIs** (depends on models)
- REST endpoints
- Request/response handlers
- Authentication and authorization
4. **Layer 3: Web Frontend** (depends on API contracts)
- API client (generated or manual)
- State management (Redux, Zustand)
- Components and views
- Forms and validation
5. **Layer 4: Mobile App** (depends on API contracts)
- API client
- Navigation structure
- Screens and components
- Offline-first setup
### Within-Project Ordering
Order by dependency:
- Shared types → Constants → Validation → Models → Services → Routes → Middleware → Integration
### Dependency Markers
Mark explicit dependencies:
```markdown
## Task 3: Create user repository
**Depends on:** Task 1 (User schema), Task 2 (database connection)
## Task 5: Add POST /users endpoint