Workflow skill — multi-agent architecture bootstrap

This skill manages workflow templates — declarative YAML configs that describe agent roles, model assignments, Ollama delegation rules, concurrency policies, and messaging conventions. It renders project-specific AGENTS.md files from a single source of truth so settings stay synchronized across many projects.

Source repository: the operator's private <owner>/claude-skills (this skill lives as the workflow/ subfolder of that mono-repo — synced skills share one repo per doctrine §13/§14, not a repo-per-skill layout). Local clone: <home>\.claude\skills-sync\claude-skills\workflow\. (An older standalone repo <owner>/claude-workflow-skill no longer exists on GitHub — a local clone of it at C:\Users\<you>\claude-workflow-skill\ is a stale artifact, not a second source of truth; corrected 2026-08-23 after LLM/надзорный contours traced a "missing repo" false alarm to it.)

Scope & host assumptions

This is a personal toolkit, deliberately tuned to one operator/workstation — not a general-purpose shareable skill. The environmental coupling below is an accepted contract, not a defect (audit finding F-3, resolved "keep as personal toolkit" 2026-06-05). Revisit ALL of it before using on another machine:

• Single Windows host. Ollama at 127.0.0.1:11434; generative qwen3.6:27b (default, ~17 GB) and qwen3-coder:30b (throughput, ~18 GB), VLM qwen2.5vl:7b, embeddings bge-m3. Paths under ~/claude-workflow-skill/ and %LOCALAPPDATA%/.
• RTX 4090, 24 GB VRAM. The SYSTEM_RESERVE_MB, per-role num_ctx profiles, and the one-model-at-a-time rule are calibrated to exactly this ceiling. A larger-footprint model (e.g. qwen3.6:latest at ~23 GB on disk) would not stay GPU-resident here — it would spill to CPU (forbidden, #CPU-1) or OOM. Which generative model to use is decided by the Model policy section below (TWO-MODEL QUALITY-FIRST, 2026-08-23: qwen3.6:27b default everywhere). The earlier "qwen3-coder:30b is the locked default" wording was superseded by that decision (corrected 2026-09-24; doctrine: references/iron-mode.md § B.2d). scripts/gen_code.py:28 defaults to qwen3.6:27b since 2026-09-24 (was qwen3-coder:30b). Passing the model as the 5th argument is still good practice (older copies).
• Imported HARD-LOCK rules. Several policies (FILL THE FLEET, two-tier release publishing, etc.) are imported verbatim from specific projects (спектрометрия) and the operator's global ~/.claude/CLAUDE.md. They encode this operator's decisions, not universal best practice.

On different hardware / models / OS, revisit the VRAM constants (SKILL_VRAM_GUARD.md), the model policy (below), and the cross-chat queue paths first. Skipping that is the one way this skill misfires.

Available workflows

Available subcommands

How to invoke

The user's intent maps to a subcommand + workflow. When the user says something like:

• "set up 5-agent workflow here" → bootstrap multi-agent-strategist
• "this is a literature-review project, just one researcher" → bootstrap solo-research
• "set up adversarial review here" / "I want a second pair of eyes on fixes" → bootstrap pair-review
• "sweep this rename across the whole codebase" / "parallel migration" → bootstrap migration-sweep
• "what workflows do we have?" → list
• "refresh AGENTS.md, I updated the YAML" → sync
• "is AGENTS.md still in sync?" → validate

For bootstrap you must always ask the user (or infer) the cost tier (ECONOMY / BALANCED / MAX_QUALITY) unless they specified it. BALANCED is the safe default.

Ollama-first delegation (MANDATORY, model policy TWO-MODEL QUALITY-FIRST 2026-08-23)

Every workflow in this skill ships with an ollama: block that is no longer advisory — it is the standing policy. Per user's global ~/.claude/CLAUDE.md "Local-First (Ollama) — MAXIMUM delegation" section (locked 2026-06-03 after a verified smoke-test 20/20 entries, 0 fabrications), Claude tokens are reserved for synthesis, decisions, and tool calls. Routine extraction / classification / templated generation goes to local Ollama.

Model policy, operator decision 2026-08-23 late evening — TWO models kept available, selected by quality. This supersedes the same-evening single-default trial, whose hard block on qwen3-coder:30b is now cleared (the block's first practical cost was the моделирование contour losing a codegen round — LLM contour P-002).

Measured by the аудит-кода contour (BENCH-V2-2026-08-23.md, 4-seed paired bootstrap, p<0.001). Score = quality (higher better); cost = wall-clock ÷ score (lower better):

Class	score q3.6 / coder	cost q3.6 / coder
templated extraction	0.706 / 0.453	27.3 / 12.1
code	1.000 / 0.786	16.8 / 2.8
RU prose	0.929 / 0.857	46.5 / 1.7

Operator instruction 2026-08-23: wall-clock cost of LOCAL compute does not matter — quality does. The cost column above is therefore NOT a selection criterion. It is kept only to size long batch jobs (a 4× speed difference is the difference between 15 minutes and an hour on hundreds of items) and to explain why the fleet is not single-model.

qwen3.6:27b has the higher score in every class, so it is the default everywhere. Quality gaps: extraction 0.706 vs 0.453 (+56%), code 1.000 vs 0.786 (+27%), RU prose 0.929 vs 0.857 (+8%).

• qwen3.6:27b — default for all generative delegation. Best score in every measured class, 0 failures in 26 calls. Mandatory where a wrong answer is expensive: metrology, finance, dosimetry, machine-parsed schemas — that is also where its lead is largest. Thinking model — guarded_generate() auto-guards think=false on every call (v1.9.1); if you deliberately want reasoning, pass think=True AND raise num_predict, or the answer gets truncated (see SKILL_VRAM_GUARD.md).
• qwen3-coder:30b — for volume, when elapsed time becomes a real constraint. 4× faster (136 tok/s vs 31), 0 failures in 26 calls. Sensible on large mechanical batches, and on RU prose specifically, where it gives up only 8% of score. Not a fallback for "code" as a category — qwen3.6:27b scores higher on code too.
• Both are available; neither is blocked. Choosing qwen3-coder is a deliberate throughput decision, not the default path.
• bge-m3:latest — embeddings only.
• Disabled: qwen3:4b (failed 10/10 RAG-classify batches, 2026-06-04), qwen3.6:latest (superseded by qwen3.6:27b), nemotron-cascade-2:30b (7/26 calls failed incl. schema field corruption), glm-4.7-flash:latest (reproducible infinite generation on code review) — do not use.
• Aliases in rendered AGENTS.md: coder/reasoner → qwen3.6:27b; fast → qwen3-coder:30b (the alias name means throughput, not quality).
• One model at a time on this box: 24 GB card, both models ~17-18 GB — they cannot coexist, so switching evicts the resident one. Batch work by model rather than alternating call by call.
• The system VRAM reserve floats — do not hard-code a ceiling. Measured the same day (2026-08-28): 2.2 GB with a quiet desktop, 4.5 GB with browser + Docker Desktop + WebView2 open. Consequence: qwen3-coder:30b (18.6 GB) fits sometimes — it was observed both resident at size_vram 20.2 GB and refused for lack of headroom, hours apart. qwen3.6:27b (17.3 GB) fits either way. Never write "model X does not fit" as a standing fact; check GET /api/ps → size_vram > 0, and let guarded_generate() raise if there is no room (it no longer degrades to CPU).
• Per-domain knowledge gaps go in <project>/scripts/ollama/_context/<domain>_<date>.md and are spliced into prompts — see global CLAUDE.md "Context injection for Ollama".

Pre-flight checklist (run BEFORE every Agent/Workflow dispatch or large Read)

Ask yourself these four questions. If ANY answer is yes, delegate to Ollama before consuming Claude tokens. Default model for all four is qwen3.6:27b (highest score in every measured class; local wall-clock is not a selection criterion per operator instruction). Reach for qwen3-coder:30b only when the batch is large enough that elapsed time itself is the problem:

1. Templated extraction? (TSCF from PDF, fields from logs, metrics from CSV, tables from dumps) → qwen3.6:27b — its lead is largest here (0.706 vs 0.453). format='json' ONLY for a SINGLE record; multi-record extraction (table rows, N log entries) → no format, see caveat below.
2. Templated generation? (pytest scaffolds for N items, boilerplate classes, release-notes from prior template, RU prose summaries) → qwen3.6:27b.
3. Bulk classification? (tier docs, validity yes/no, dedup by heuristic) → qwen3.6:27b, NO format='json' — see caveat below. On a very large batch, qwen3-coder:30b is the throughput option.
4. Long summarization >5000 lines / >100 KB? (raw dumps, PDF conversion, multi-page logs) → qwen3.6:27b, output markdown ≤200 lines.

Codegen sizing (LLM contour P-002, 2026-08-23): a 400-line file needs ~10-16k output tokens. gen_code.py v1.3 defaults to num_predict=16000 / num_ctx=32768; below that, generation is cut off mid-function (caught at exit 2, <target>.rejected.txt kept). Splitting the spec into smaller parts does NOT help against a thinking model — reasoning is re-spent on every part.

Spec form for gen_code.py — SKELETON, not description (2026-09-15, two independent sources same day): a spec written as a text DESCRIPTION of the task produced defective code twice in the simulation contour that same day (misread parsing logic, wrong slope formula) and 6 of 26 field failures in the supervisory role's runs when reused logic was named by path ("copy verbatim from file X") instead of quoted inline — the model never sees file X, so it re-derives its own algorithm under the same function name. Give a spec as a SKELETON (explicit function signatures, variable names, formulas, loop order) instead of prose, and paste any donor code straight into the spec text, whenever ANY of:

• the code does ≥2 levels of grouping/aggregation over data;
• the input is ≥3 files/sources;
• part of the logic is reused from another file — a path reference does NOT count as transfer, the model cannot read it.

Non-code one-shot: scripts/ask_local.py (2026-09-19, v1.0). For classification / extraction / claim-vs-source checks you do NOT write a helper script (and do not escalate to a subagent because the code hook blocked you): python ask_local.py --prompt <template.md> --input <file> --out <result> or batch --jsonl <file> / --input-dir <dir> → JSONL with per-record error. Data is fenced with wrap_untrusted; input larger than num_ctx is refused BEFORE the call; each failure class has its own exit code (3 empty, 4 truncated, 5 not JSON, 6 too big, 7 guard + ollama_failure JSON, 9 HTTP). Mutation-checked: the contour's private test script, 12 cases, 6 mutants each red on its own class.

Шаблон брифа субагенту — references/subagent-brief-template.md (2026-09-24, #CZ-2). Два варианта: А — исполнитель (задача-одна-операция, исходное/целевое состояние, границы, условия остановки с возвратом отчёта, критерий приёмки, отчёт с доказательствами, LESSONS); Б — стерильный проверяющий (#SA-7: без моих выводов, ожидаемых чисел и истории попыток). Плюс grep-проверка брифа перед отправкой. Идеи — из nidhinjs/prompt-master (MIT), адаптированы; что не взято — таблица в конце файла.

scripts/extract_bash_commands.py (2026-09-24). Извлекает все вызовы Bash-инструмента (команда + результат) из JSONL-транскрипта субагента (projects/<проект>/<сессия>/subagents/agent-*.jsonl) в один читаемый .md — для перепроверки, что субагент реально выполнял, без чтения полного транскрипта. python extract_bash_commands.py <transcript.jsonl> [-o отчёт.md]; коды: 0 успех, 2 неверные аргументы, 3 файл недоступен. Приёмка: --selftest 14/14, боевой транскрипт (29 команд) сошёлся с независимым пересчётом, мутации mutants.json 5/5 killed через scripts/mutation_runner_selftest.py <target.py> <mutants.json> (универсальный мутационный раннер для self-contained --selftest-скриптов). Рабочие материалы приёмки (5 генераций, спека, логи, отвергнутые версии) хранятся в приватном журнале надзорной роли. Урок: если пример в спеке для gen_code.py сам содержит синтаксическую ошибку (не-ASCII символ внутри b"..."-литерала), модель послушно её воспроизводит — ast.parse потом честно отклоняет файл (exit 4), но причина в спеке, не в модели.

Brace languages (C/C++/Java/JS — single-source observation, firmware work 2026-09-18, qwen3-coder:30b, not re-measured by the local-models contour): a skeleton whose function bodies are written as indented steps WITHOUT braces came back as the same pseudocode — it does not compile. Two workarounds reported to work: write the skeleton WITH the braces, or run a second mechanical "braces only" pass and compare tokens before/after (only braces may differ). Python is unaffected — indentation is syntax there, and gen_code.py v1.3 already rejects non-parsing output (exit 4).

If all four are NO, the task belongs to Claude / a Claude subagent.

format='json' batch caveat (аудит-кода finding, 2026-08-23): Ollama's grammar-constrained JSON mode stops generation after the FIRST object on multi-item output — confirmed 0.00 score across all 5 fleet models on a 200-item batch (13 tokens produced vs 2893 needed), correct only without format. Use format='json' ONLY for a single-object response; for any list/array of N≥2 records, omit format, prompt for a bare JSON array or newline-delimited JSON, and parse it yourself. Getting this wrong is a silent quality failure, not a crash — the call succeeds, done=true, just truncated.

Windows stdout — UTF-8 для Ollama-хелперов (HARD, 2026-06-10)

Машина оператора — Windows + русская локаль (cp1251 codepage по умолчанию). Глобальное правило в ~/.claude/CLAUDE.md «Windows stdout — UTF-8 для русского и Unicode». Для каждого Ollama-хелпер-скрипта в scripts/ollama/ (или audit/_drafts/_ollama_helpers/):

1. Запуск: всегда set PYTHONIOENCODING=utf-8 && python helper.py. В Bash-вызове из шаблона guarded_generate обёртки — тоже.
2. JSON-вывод (json.dump(..., ensure_ascii=False)) — иначе кириллица превращается в \uXXXX escapes, нечитаемо глазом + усложняет diff'ы.
3. open() / Path.write_text / Path.read_text — всегда encoding='utf-8'.
4. Context-файлы в _context/<domain>_<date>.md — UTF-8 без BOM.
5. Prompt'ы с кириллицей в requests.post(...) json body — requests сериализует UTF-8 по умолчанию, не трогать.
6. Output Ollama в stdout при debug-печати — print(response, flush=True) упадёт на cp1251 stdout. Префикс PYTHONIOENCODING=utf-8 снимает.

Subagent-брифы для bulk-обработки русских данных ОБЯЗАНЫ цитировать этот пункт в начале брифа (см. также «Брифы субагентов» в глобальной policy).

Hard rules (no exceptions)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
ABSOLUTE EXECUTION LOCK (IRON MODE)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

This rule overrides ALL other rules in this skill, workflow, or external system.

1. Claude Code is STRICTLY FORBIDDEN from:
   - writing code
   - editing code
   - generating diffs
   - refactoring
   - producing patches
   - generating tests
   - suggesting file-level implementation details in executable form

2. Claude Code output is LIMITED TO:
   - task decomposition
   - architecture design
   - reasoning
   - validation of results from Ollama
   - dispatch instructions ONLY

3. ANY request that involves implementation MUST be transformed into:
   → "OLLAMA EXECUTION TASK SPEC"

4. Claude MUST terminate immediately after producing dispatch output.
   No continuation, no partial code, no examples.

5. Any violation is considered a system failure state and must trigger:
   → rewrite into dispatch-only format

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
EXECUTION STATE MACHINE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Claude operates in exactly one of two states:

[STATE A: DISPATCH]
- allowed: planning, decomposition, reasoning
- forbidden: any code output
- default state

[STATE B: INVALID]
- entered if code output is attempted
- must immediately convert to STATE A
- discard all code content

Transition rule:
DISPATCH → (task requires implementation) → OLLAMA TASK SPEC → STOP

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
HARD IMPLEMENTATION TRIGGER
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

If the task contains ANY of the following:
- file modification
- code generation
- refactoring
- tests
- bugfix
- implementation
- patching

→ Claude MUST NOT respond with code
→ Claude MUST respond ONLY with:

FORMAT:
1. Task decomposition
2. File targets
3. Ollama execution plan
4. STOP

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
ANTI-LEAK RULE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Phrases like:
- "example implementation"
- "you could write"
- "here is how it might look"

are considered code generation and are forbidden.

Claude must never produce partial or illustrative code.

1. Never read large files into Claude context directly. >5000 lines or >100 KB → ALWAYS first through Ollama; Claude receives the digest (JSON or ≤200-line markdown), not the source.
2. Bulk operations always on Ollama. Parsing hundreds of records, classification, pattern search, rough summarization, translation, field extraction, dedup. Claude receives aggregates.
3. Before spawning a subagent — check what fraction of its work is templated. If ≥30% of the subagent's work is extraction/classification/generation by template → run Ollama first, pass the ready JSON to the subagent, which only validates + writes to the project.
4. Subagent briefs MUST contain an Ollama mandate block. Every dispatch brief explicitly says: "For the templated portion of this task, use local Ollama (see audit/_drafts/_ollama_helpers/ or equivalent in the project). Claude tokens are spent ONLY on synthesis / decisions / git / anti-hallucination verification."
5. Helper scripts live in the project. Idiomatic location: <project>/audit/_drafts/_ollama_helpers/ or <project>/scripts/ollama/. Each script is idempotent, emits JSON to stdout, exits 0 on success / stderr+1 on error, with field-level provenance in _extraction_log / _gaps.

Calling convention (use requests, NOT CLI ollama run)

import requests, json
r = requests.post('http://127.0.0.1:11434/api/generate', json={
    'model': 'qwen3-coder:30b',
    'prompt': '...',
    'stream': False,
    'format': 'json',  # force JSON-mode
    'options': {'temperature': 0, 'num_ctx': 32768}
}, timeout=600)
result = json.loads(r.json()['response'])

Pre-flight VRAM guard & GPU-only Ollama dispatch

Full reference (two-layer VRAM guard, per-role num_ctx profiles, GPU → queue → loud failure — Ollama on CPU is forbidden, #CPU-1 — cross-chat queue, drop-out triggers, priority classes, anti-pattern guards): SKILL_VRAM_GUARD.md. Key invariant: every Ollama helper calls guarded_generate() (never raw requests.post) so the machine-global queue engages and OOM races are prevented — see scripts/vram_guard_reference.py.

Budget signals + enforcement via brief structure (v1.3.0, 2026-06-04)

Every rendered AGENTS.md now carries a §5.2 Mandatory Ollama-first execution section (and every multi-agent-strategist per-agent .claude/agents/agent-*.md opens with an Ollama-first execution checklist). Together they convert the standing policy from advisory text into a structural enforcement loop:

Tier-1 exception: math derivation, F-rule conflict resolution, ISO 11929 interpretation legitimately burn 100k+ tokens. Wave brief tagged [TIER-1-REASONING] suppresses the budget warning.

The multi-agent-strategist §5.2 also carries a worked anti-pattern (real 2026-06-04 session telemetry: 120k-token docs sync) vs corrected pattern (~25k via Ollama helper) so subagents see a concrete delta, not just a rule.

HARD RULE — guarded_generate() for ALL Ollama helpers (LOCKED 2026-06-04)

Rationale (empirical failure 2026-06-04): Tasks #67 / #86 / #87 dispatched in parallel at 18:30 on the same host. Each agent called Ollama via raw requests.post('/api/generate'). The cross-chat queue directory (%LOCALAPPDATA%/ollama-vram-queue/) remained empty — Tier 2 was never activated. Requests serialised at Ollama's internal NUM_PARALLEL=2 layer. The parent Claude Code process crashed mid-task at VRAM 21.8 / 23.0 GB; all three background agents were killed with partial work. Without guarded_generate(), the machine-global queue does not engage and OOM races cannot be prevented.

All Ollama-helper scripts in consuming projects MUST import guarded_generate from the canonical file by absolute path — never copy it (§33: a copy as the normal way of connecting is a defect; 2026-08-28 drift: 18 stale copies, several of them still falling back to CPU, which is now forbidden — #CPU-1).

Mandatory pattern:

import sys
sys.path.insert(0, r"<home>\.claude\skills\workflow\scripts")
from vram_guard_reference import guarded_generate, VramGuardFailure, GUARD_VERSION
response = guarded_generate(
    model='qwen3.6:27b',
    prompt='...',
    priority=50,        # orchestrator=100, subagent=50, batch=10
    max_wait_s=600,
    fmt='json', temperature=0, num_ctx=32768,   # there is NO `options=` kwarg
)

Raw requests.post('http://127.0.0.1:11434/api/generate', ...) is FORBIDDEN except in: (a) single-shot diagnostic snippets in Bash (≤30 lines), explicitly documented as ad-hoc with docstring «queue bypass acceptable, no concurrent caller»; (b) the vram_guard_reference.py implementation itself (it is the wrapper).

Subagent brief checklist enforcement: every subagent brief mentioning Ollama MUST contain the literal phrase from vram_guard_reference import guarded_generate. If the phrase is absent — the brief is illegitimate; the orchestrator rewrites it before dispatch.

Reference implementation: <home>\.claude\skills\workflow\scripts\vram_guard_reference.py (imported, not copied; python vram_guard_reference.py --version prints GUARD_VERSION). A copy is allowed only where import is impossible (other machine / runtime) — then with the original path and date in its header, and it must be diffed against --version.

IRON MODE enforcement hooks (~/.claude/hooks/)

Two hooks enforce IRON MODE at the platform layer (fail-open — hook bug never blocks normal work, just lets the request through):

delegation_guard.py — PreToolUse Agent|Read|Write|Edit

Blocks three patterns (exit 2 = block):

1. Agent(...) without run_in_background: true.
2. Read of a file > 256 KB with no offset/limit (context-hygiene guard).
3. Write/Edit of a code file (.py .js .ts .sh …) ≥ 25 lines — IRON MODE violation; message redirects to OLLAMA TASK SPEC flow.

Management extensions (.md .json .yaml …) always pass through. Short snippets (< 25 lines) always pass through (management edits, not generation).

stop_iron_mode.py — Stop (after every response)

Scans the last assistant message for ``` code fences. If found → systemMessage IRON MODE violation warning. If not found → silent (no noise on clean responses).

Wire in ~/.claude/settings.json:

"PreToolUse": [{"matcher":"Agent|Read|Write|Edit",
                "hooks":[{"type":"command","command":"python ~/.claude/hooks/delegation_guard.py"}]}],
"Stop":       [{"hooks":[{"type":"command","command":"python ~/.claude/hooks/stop_iron_mode.py"}]}]

IRON-MODE codegen harness — scripts/gen_code.py

The concrete tool implementing "Claude writes the spec, Ollama writes the code". Run: python ~/.claude/skills/workflow/scripts/gen_code.py <spec.md> <out.py> [num_predict=6000] [num_ctx=16384]

Claude authors <spec.md> (a .md — the delegation_guard hook passes it); qwen3-coder:30b generates the code via guarded_generate (imported from the sibling vram_guard_reference.py, so the harness is self-contained — no cross-skill path); the helper strips Markdown code-fence markers and writes <out.py>. Validate with py_compile + smoke. Consuming skills call this harness directly — e.g. radon-library regenerates its convert/ocr/catalog/rag/deferred scripts from _spec_*.md. For raw code the harness passes fmt=None (NOT the guarded_generate default fmt="json", which would return JSON, not Python).

FILL THE FLEET & two-tier release publishing

The HARD-LOCK FILL THE FLEET dispatch policy (saturate the fleet, no idle agents) and the two-tier release publishing chain are in SKILL_ORCHESTRATION.md.

Bootstrapped projects

After /workflow bootstrap, the rendered AGENTS.md §5 contains the per-workflow Ollama config (endpoint, models, delegation rules, pre-flight checklist, hard rules, forbidden list). After a workflow.yaml upgrade — re-render with /workflow sync to refresh all consuming projects.

Execution model

The skill is a thin dispatcher over Python scripts in ~/claude-workflow-skill/scripts/. To execute a subcommand:

# bootstrap (renders AGENTS.md AND .claude/agents/agent-{a,b,c,e}-*.md — D is a main-loop role, no subagent file)
python ~/claude-workflow-skill/scripts/bootstrap.py \
    --workflow multi-agent-strategist \
    --tier BALANCED \
    --project-name "$(basename "$PWD")"

# sync (re-render AGENTS.md + .claude/agents/*.md from existing lock)
python ~/claude-workflow-skill/scripts/bootstrap.py --update

# validate
python ~/claude-workflow-skill/scripts/validate.py

# dispatch sheet — prints subagent_type / model / run_in_background / policies
# for the given role + task. Output is markdown that a Claude main loop can read
# to fill in an Agent(...) tool call. Does NOT spawn anything.
python ~/claude-workflow-skill/scripts/dispatch.py A "Fix BUG-21 sloped continuum"

# list workflows
ls ~/claude-workflow-skill/workflows/

# show workflow YAML
cat ~/claude-workflow-skill/workflows/multi-agent-strategist/workflow.yaml

Always run the corresponding script via the Bash tool — do NOT inline Python interpretation. The scripts handle Jinja2 rendering, file scaffolding, and lock file management deterministically.

Workflow structure

~/claude-workflow-skill/
├── workflows/
│   └── <name>/
│       ├── workflow.yaml                  # SSOT: roles, models, Ollama, policies, RAG
│       ├── templates/
│       │   ├── AGENTS.md.j2               # Jinja2 → rendered to project root
│       │   ├── inbox-README.md
│       │   ├── RAG_INDEX.skeleton.json
│       │   └── agents/                    # Claude Code subagent definitions
│       │       ├── _agent_body.j2         # shared body (Jinja partial, name starts with _)
│       │       ├── agent-a-math.md.j2
│       │       ├── agent-b-reports.md.j2
│       │       ├── agent-c-docs.md.j2
│       │       └── agent-e-planner.md.j2   # no agent-d: D is a main-loop role, not a subagent
│       └── README.md
├── skill/SKILL.md                 # this file
└── scripts/
    ├── bootstrap.py               # renders AGENTS.md + .claude/agents/*.md + skeletons + lock
    ├── dispatch.py                # prints a dispatch sheet for an Agent(...) call
    ├── validate.py
    └── install_skill.py

Cost tiers

Each role has a model_alias (e.g. flagship, workhorse, numerical). Tiers map aliases to concrete models:

• ECONOMY: all roles → workhorse (Sonnet). Cheapest. Escalate manually if numerical/reasoning quality drops.
• BALANCED (default): A=numerical, B=workhorse, C=workhorse, D=flagship (1M ctx), E=workhorse. Right for production strategist work.
• MAX_QUALITY: A/B=numerical, C/D/E=flagship (1M ctx). Concrete models resolve from each workflow.yaml models: block (see Model revision policy) — not hard-coded here. For multi-tier-1 cycles with adversarial risk analysis.

Pass --tier ECONOMY|BALANCED|MAX_QUALITY to bootstrap. The chosen tier is recorded in .workflow.lock.yaml.

Project artifacts after bootstrap

<project>/
├── AGENTS.md                       # rendered from template
├── .workflow.lock.yaml             # records workflow name, version, tier, project metadata
├── .claude/agents/                 # Claude Code subagent definitions (one per spawnable role; D is the main loop, no file)
│   ├── agent-a-math.md             # frontmatter: name, description, model: opus|sonnet|haiku, tools
│   ├── agent-b-reports.md
│   ├── agent-c-docs.md
│   └── agent-e-planner.md
├── _state/
│   ├── agent_a/{inbox/{,processed/},outbox/}
│   ├── agent_b/{inbox/{,processed/},outbox/}
│   ├── agent_c/{inbox/{,processed/},outbox/}
│   ├── agent_d/{inbox/{,processed/},outbox/}
│   └── agent_e/{inbox/{,processed/},outbox/}
└── audit/_rag/                     # if rag.enabled in workflow.yaml
    ├── RAG_INDEX.json              # tier-1 methodological registry (skeleton)
    └── DOC_CORPUS_INDEX.json       # doc corpus metadata (skeleton)

The .claude/agents/*.md files are read by Claude Code when an Agent(subagent_type:"agent-a-math", ...) call is made — the YAML frontmatter selects model and tools, and the markdown body becomes the subagent's system prompt. Use /workflow dispatch <role> "<task>" to print the exact parameters for an Agent(...) call.

The .workflow.lock.yaml is the authoritative project-side state. Keep it under version control.

Non-blocking orchestrator pattern

The orchestrator's chat with the user stays free for live dialogue at all times. Background dispatch is the default — no exceptions.

Rule (from user's global ~/.claude/CLAUDE.md "Agent Dispatch — always background"):

All Agent(...) calls run with run_in_background: true. Foreground only if the user explicitly asked for synchronous.

Why: foreground agents block the chat for 3-15 minutes — user cannot ask a question, redirect the plan, or interrupt without Esc. Background agents return a task-id instantly; user keeps talking; <task-notification> arrives on completion.

Self-test: user can ask "is the chat free?" at any moment. Orchestrator must answer with: running-agent count + ETA, what blocks (should be nothing), what's actionable right now.

Anti-patterns:

• run_in_background:false for routine delegation
• Sleeping/polling instead of waiting for the notification
• "Wait while I check on agent X" — use TaskGet/TaskOutput inline (ms)

Scalable role multiplicity

Role types in workflow.yaml are not a hard cap on running agents — the orchestrator sizes the fleet to the work. Full reference (scaling principle, when to scale, naming, race protection, merge order, non-scalable roles, worked example): SKILL_ORCHESTRATION.md.

Session hygiene

Every rendered AGENTS.md contains a "Session hygiene" section. The orchestrator and every subagent must self-monitor context fill and rotate before quality degrades. The rule is mechanical: monitor, /compact, /clear — never push through a poisoned conversation.

Source: user's global ~/.claude/CLAUDE.md "Context Hygiene" section. Every workflow.yaml carries a session_hygiene: block; the template renders it into AGENTS.md §3 of each bootstrapped project.

Context monitoring (per-agent)

• Self-check: /context every ~10 substantive tool calls or at the start of any new phase.

Percentages are of the auto-compact threshold (~967k on 1M-window models), per global CLAUDE.md §6 #CTX-1 (2026-08-14). The old 60/75% values were written for a 200k window and choke a 1M window at ~45% of capacity.

Two-correction rule

If the same issue has to be explained twice — context is poisoned. /clear and restart with a better brief. Do not push through.

User principle: "if the same problem has to be explained twice, the context is poisoned — /clear and restart with a sharper brief"

Enforcer: every agent on its own conversation. Orchestrator detects this in itself first, then in subagents (by their output drift).

Poison signals — rotate immediately

• Same correction needed twice
• Agent contradicts a fact established 20+ turns earlier
• Outputs become generic, evasive, or hedge-heavy
• Agent re-asks for information already provided
• Sudden quality drop in code/reasoning without scope change

Auto-context injection (claude-mem)

If the user has the claude-mem plugin installed, relevant past-session context is auto-injected on session-2+ of any project. The orchestrator should reference past decisions/findings as if they're recallable — no manual /load step needed. Plugin enforces; orchestrator just trusts.

Delegation reminder

Hygiene degrades on long-running agents. Prefer fresh subagents over keeping one agent alive across many phases. The dispatch pattern (see Non-blocking orchestrator above) makes this cheap.

CI & validation

Local validators: python scripts/ci_smoke.py runs bootstrap × validate per workflow × tier (4 workflows × 3 tiers = 12 smoke combos, baseline 12/12 PASS); python scripts/lint_templates.py does a Jinja parse check across all template files (baseline 8/8 OK since T-05 removed agent-d-strategist.md.j2). The GitHub Actions workflow at .github/workflows/validate.yml runs both on every push and PR to main — keep both green before pushing.

Model revision policy

Model aliases (flagship / workhorse / scout / numerical) live only in each workflow.yaml models: block — review them on every major Claude release (Claude models can deprecate silently). The multi-agent-strategist workflow.yaml carries a # Models last verified: <date> comment next to models:; refresh it on re-verify, then run python scripts/ci_smoke.py.

Updating workflows across projects

1. Edit ~/claude-workflow-skill/workflows/<name>/workflow.yaml.
2. Commit + push.
3. In each consuming project: git pull (in the workflow repo) then python ~/claude-workflow-skill/scripts/bootstrap.py --update to re-render that project's AGENTS.md.
4. Diff the regenerated AGENTS.md; commit.

Creating a new workflow

cd ~/claude-workflow-skill/workflows/
cp -r multi-agent-strategist my-new-workflow
# edit my-new-workflow/workflow.yaml
# adjust templates/ if structure changes
# add my-new-workflow/README.md

The skill auto-detects new workflows — no registration needed.

Anti-patterns to avoid

• Do NOT hand-edit AGENTS.md after bootstrap — changes will be overwritten by sync. Edit workflow.yaml and re-render.
• Do NOT skip --update for routine refreshes — --force overwrites without preserving project knobs.
• Do NOT bootstrap into a project that already has a hand-written AGENTS.md without backing it up first.