bridgie 0.0.0-stage → 0.6.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Anoop Vijayan
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -1,3 +1,136 @@
-# Temporary Holding Version
+# Bridgie
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+A file-based operating system for running a company's work through AI agents:
+**decide → build → verify → accept → learn**. One tool, pointed at any folder,
+including one with no code in it yet.
+
+Humans own intent (the bet) and acceptance. Agents produce. The harness
+verifies. Only the runner writes state; every human choice happens on a
+rendered page and is written back to files you can diff.
+
+## Install
+
+Needs **git** and **Node 20+**. Works the same in PowerShell, cmd, Git Bash, macOS and Linux terminals.
+
+```
+npx bridgie@latest setup
+```
+
+One command, for you and your agents:
+
+1. puts `bridgie` on your PATH (`npm install -g bridgie`; npm writes the right launcher for your shell);
+2. installs the `/bridgie` agent skill for every agent it finds on the machine (Claude Code, Codex,
+   GitHub Copilot, Antigravity, Gemini CLI, Cursor and others) through the open
+   [`skills`](https://skills.sh) installer; pass `--agent claude-code --agent codex` to pick;
+3. writes `~/.bridgie/policy.yaml`: which model CLI each seat uses, from the ones it finds.
+
+Open a **new** terminal afterwards. To update the tool and the skill together, run the same command again;
+`bridgie doctor` warns when an installed skill is out of date. The skill alone, without the CLI:
+`npx skills add anoopvijayan-dev/bridgie` (the skill then runs Bridgie through `npx bridgie`).
+
+**Claude chat and Claude cloud sessions** read skills from your claude.ai account, not your disk: upload
+`skills/bridgie/SKILL.md` there once (and again after it changes).
+
+**Working on Bridgie itself:** clone this repo and run `node bin/bridgie.mjs setup` in it; `setup` then links
+your clone instead of installing from npm.
+
+## Start
+
+```bash
+bridgie new my-idea --lane product     # folder + git + Bridgie state, in one step
+cd my-idea
+bridgie                                # tells you exactly what to do next, at every stage
+```
+
+`bridgie doctor` checks tools and model adapters. `bridgie help` lists every command.
+
+## Finding and testing an idea (the decide loop)
+
+```bash
+mkdir ~/ideas && cd ~/ideas && git init
+bridgie init --lane product
+bridgie scout --from ~/Downloads/broker-emails        # ranks the problems people describe, by independent sources
+bridgie bet new packs --from broker-emails/O1 --risk money
+bridgie frame packs                                    # drafts who / problem / why now / today / outcome; assumptions → questions
+bridgie research packs                                 # competitors, prices, complaints, every fact linked (needs web access)
+bridgie grill packs --lenses                           # adversarial grill + a three-lens board (offer, moat, product) and a chair
+bridgie panel packs                                    # drafts customer.yaml and pitch.md; you correct them; run again for 20 simulated buyers
+bridgie numbers packs --template && $EDITOR .bridgie/bets/packs/numbers.yaml && bridgie numbers packs
+bridgie evidence packs                                 # interview guides; drop notes in .bridgie/bets/packs/evidence/ and run again
+bridgie open packs                                     # answer, accept the confidence proposal, sign
+```
+
+Every verdict is computed by the runner, never written by a model: the board's vote and average, the
+panel's buy rate and Van Westendorp price range, the unit-economics checks. Simulated buyers are an
+upper bound and never change a bet's confidence; only `bridgie evidence` (real interviews, data, tests)
+proposes a tier, and you accept it. A money bet cannot be signed without positive unit economics.
+
+## The loop
+
+```bash
+cd ~/some-project                     # a git repo; for a brand-new company, an empty folder works until Build
+bridgie init --lane engineering
+bridgie bet new net-collect --title "Net collection for failed payments" --risk money
+$EDITOR .bridgie/bets/net-collect/bet.md   # who, problem in their words, why now, outcome
+bridgie grill net-collect             # questions that need real-world evidence
+bridgie open net-collect              # answer, then sign / kill / need more evidence
+bridgie run net-collect               # shape → plan → build → verify → package; stops at each decision
+bridgie open net-collect              # approve the design, ack risk paths, unblock tasks, accept
+bridgie learn net-collect             # proposals for tests, gotchas and policy
+```
+
+What each step guarantees:
+
+| Step | What happens | Guards |
+|---|---|---|
+| `shape` | branch `bridgie/<slug>`, `requirements.md` (EARS) and `design.md` | signed bet, clean tree |
+| `plan` | `tasks.json`, every requirement mapped | approved design, no cycles; a requirement tagged `risk: money` raises the bet's risk; re-planning supersedes any acceptance package |
+| `build` | one task per run in `.bridgie/worktrees/<slug>/<task>` | human ack for risky bets; `.bridgie/` edits discarded; only declared inputs reach the model |
+| `verify` | the runner runs every required test tier, then a read-only reviewer reads the results in a detached worktree | changed risk paths raise the bet's risk (logged, never lowered) and need a fresh ack; a required tier with no command is not proven; every evidence row counts and the reviewer's verdict can only tighten the result; different vendor from the producer; 3 attempts then blocked |
+| `accept` | one-page acceptance package, bound to the branch commit and a digest of spec, tasks and evidence | every task passed, re-checked at the moment of approval; a stale package is refused and rebuilt by `bridgie accept`; approvers per the current risk (2 for money), each a different person |
+
+Verification commands live in `.bridgie/policy.yaml` (`verify.commands.<tier>`; see the comment `bridgie init`
+writes). Without them Bridgie uses the lane's commands for the stacks it detects; a tier the risk requires
+but nobody configured is never skipped silently.
+
+While an agent runs, Bridgie state is read-only (a project-wide seat lock): your decisions are refused with a
+message rather than rolled back, and anything the agent adds, changes or deletes under `.bridgie/` is restored.
+
+`bridgie open` prints a personal link with an access token; decisions are recorded under the identity of that
+link. `bridgie open --invite alice` prints a second link for a second approver. With `--host` beyond
+localhost the traffic is plain HTTP, so share each link only with its person.
+
+Bridgie never pushes and never merges into your base branch. Merge `bridgie/<slug>` yourself after acceptance.
+In an agent harness, say `/bridgie` (the installed skill drives the CLI and leaves decisions to you).
+
+## Layout
+
+| Path | What |
+|---|---|
+| `bin/bridgie` | the command |
+| `bin/bridgie.mjs`, `runner/cmd/` | the CLI: one Node module per command, no bash anywhere |
+| `runner/` | dependency-free Node modules (state, context, surface, guard, doctor) |
+| `seats/<loop>/<seat>/SKILL.md` | what each agent seat is for, its inputs, what it may write (decide: scout, framer, researcher, three lenses, chair, profiler, buyer, grill, evidence) |
+| `runner/panel.mjs`, `runner/numbers.mjs` | the computed parts of the decide loop: persona cards, tallies, price range; unit economics |
+| `lanes/<lane>/` | risk paths, attack lists, checklists, verifiers per domain |
+| `adapters/` | claude-code, copilot, codex, antigravity, fake: how a seat is launched on a model runtime |
+| `skills/bridgie/` | the skill agent harnesses use to drive bridgie |
+| `schemas/`, `templates/`, `policy/` | the contracts |
+| `ui/templates/` | decision surface pages |
+
+## Status
+
+v0.4.0 (M3 decide loop): scout, frame, research, lenses, panel, numbers, evidence; swarm seats and web
+capability in the runner; confidence decisions. Borrowed with thanks from
+[founder-skill](https://github.com/Jakeschincariol/founder-skill) (MIT): the three-lens board, the
+runner-dealt buyer panel, Van Westendorp pricing and computed unit economics. v0.3.0 (M2 hardening): enforcement gaps from an external review closed (acceptance binding, evidence
+reconciliation, effective risk and mandatory tiers, state guard and seat lock, authenticated decision
+server, reviewer works from runner-run test results). v0.2.0 (M1): the whole loop runs end to end on the
+fake adapter and is covered by tests for spec requirements R1–R14 and properties P1–P3. Real adapters are verified against stub CLIs with the
+flags of Claude Code 2.1, Codex 0.167 and Copilot CLI 1.0; Antigravity follows the v1 notes and its
+write mode is unverified (see `adapters/antigravity/gotchas.md`). Next: a first real run of the decide loop on live
+models, recovery for interrupted runs, seeded-bug evals for the reviewer, marketing and sales lanes.
+
+## Tests
+
+`npm test` runs the unit tests; `tests/run.sh` runs everything (unit, `bats tests/` for the CLI, doctor). CI runs both on Linux, macOS and Windows with Node 20 and 22, and checks that `bridgie` launches from PowerShell and cmd after `setup`. Tests ignore `~/.bridgie`.

package/VERSION

@@ -0,0 +1 @@
+0.6.0

package/adapters/README.md

@@ -0,0 +1,25 @@
+# Adapters
+Each adapter is `adapters/<name>/run.mjs`, run with the same `node` as Bridgie (run.mjs is plain Node, so adapters behave the same on Windows, macOS and Linux). A legacy `run.sh` is still launched with bash if no run.mjs exists.
+
+An adapter launches one seat invocation on one model runtime. Contract:
+
+    run.mjs --seat <seat dir> --prompt <prompt file> --out <output file> --cwd <dir> \
+           --mode readonly|write --model <name> [--effort <level>]
+    run.mjs --capabilities --seat <seat dir>   # prints {"vendor": "...", "reads_repo": true|false, "reads_web": true|false}
+    run.mjs --probe                            # prints OK when the CLI is installed (no model call)
+
+Rules:
+- Read the prompt from the file; never pass it inline (Windows argument limits).
+- The seat ends its reply with one fenced ```json block. The adapter saves the full reply
+  next to `--out` as `reply.txt` and extracts the last JSON block into `--out` with
+  `adapters/lib/extract-json.mjs`. If the model wrote `--out` itself, that file wins.
+- `readonly` mode must not let the model write in `--cwd`. The runner also checks state files
+  and the reviewer worktree afterwards and fails the run on any change.
+- `reads_web: false` makes the runner refuse seats whose manifest lists the `web` tool (research).
+- `vendor` drives cross-family checks; `reads_repo: false` makes the runner refuse to use the
+  adapter as a reviewer.
+- Exit 0 only when `--out` exists. Any other exit code with a reason on stderr.
+- Exit 6 when the CLI itself failed (signed out, over its usage limit, crashed) and left no JSON; put the CLI's last words on stderr. Exit 4 means no reply at all, exit 5 a reply with no valid JSON block.
+- Keep the traps you learn in `gotchas.md` next to the adapter.
+
+Adapters: `fake` (tests), `claude-code` (anthropic), `codex` (openai), `antigravity` (google).

package/adapters/antigravity/gotchas.md

@@ -0,0 +1,9 @@
+# antigravity adapter gotchas (carried from Bridgie v1, verified 2026-08-12)
+- The binary is the CLI `agy` (`%LOCALAPPDATA%\agy\bin\agy.exe`, often not on PATH in a spawned shell), not the IDE. The old `gemini` CLI's OAuth tier is retired.
+- `--print` takes the prompt as its VALUE: put `--print "<prompt>"` last, or the next flag becomes the prompt.
+- `--project` is load-bearing. Without it agy uses `default-cli-project`, clears the permission grants and every tool is denied. The adapter finds the project whose folderUri covers the working dir in `~/.gemini/config/projects/*.json`; override with `BRIDGIE_AGY_PROJECT`.
+- Headless auto-denies tools. Grant read tools in `~/.gemini/antigravity-cli/settings.json`: `permissions.allow: ["read_file(*)","list_directory(*)","glob(*)","grep(*)","search_file_content(*)"]`, and deny `write_file(*)`, `edit_file(*)`, `replace(*)`, `run_shell_command(*)`, `command(*)` for reviewer use. Never use `--dangerously-skip-permissions`.
+- `--sandbox` hangs on Windows. Do not use it.
+- Probe first: `agy --project <id> --mode plan --print "Reply with exactly: OK"`.
+- A QA seat that cannot read the repo invents findings (v1, 2026-08-08). Task worktrees live inside the repo (`.bridgie/worktrees/`) so the project's grants cover them.
+- UNVERIFIED: write mode. The adapter uses `--mode ${BRIDGIE_AGY_WRITE_MODE:-agent}` for implementer seats; confirm the mode name on first use and record it here.

package/adapters/antigravity/run.mjs

@@ -0,0 +1,49 @@
+// Antigravity CLI adapter (google). agy has no stdin mode: it is told to read the prompt file.
+import { existsSync, readdirSync, readFileSync } from 'node:fs';
+import { dirname, join, basename } from 'node:path';
+import { homedir } from 'node:os';
+import * as A from '../lib/common.mjs';
+
+const o = A.parse(process.argv.slice(2));
+function findAgy() {
+  if (process.env.BRIDGIE_AGY_BIN) return process.env.BRIDGIE_AGY_BIN;
+  const onPath = A.which('agy');
+  if (onPath) return onPath;
+  const la = process.env.LOCALAPPDATA;
+  if (la && existsSync(join(la, 'agy', 'bin', 'agy.exe'))) return join(la, 'agy', 'bin', 'agy.exe');
+  return null;
+}
+// findProject: Antigravity project id whose folderUri covers dir. Without it agy runs in
+// "default-cli-project" and wipes the permission grants (see gotchas.md).
+function findProject(dir) {
+  if (process.env.BRIDGIE_AGY_PROJECT) return process.env.BRIDGIE_AGY_PROJECT;
+  const want = dir.replace(/\\/g, '/').toLowerCase();
+  const d = join(homedir(), '.gemini', 'config', 'projects');
+  let best = null;
+  for (const f of existsSync(d) ? readdirSync(d) : []) {
+    try {
+      const j = JSON.parse(readFileSync(join(d, f), 'utf8'));
+      const uris = [j.folderUri, ...(j.folders || []).map(x => x.uri || x)].filter(Boolean);
+      for (const u of uris) {
+        const p = decodeURIComponent(String(u).replace(/^file:\/\//, '')).replace(/\\/g, '/').replace(/^\/([a-zA-Z]:)/, '$1').toLowerCase().replace(/\/$/, '');
+        if (want.startsWith(p) && (!best || p.length > best.len)) best = { id: j.id || basename(f, '.json'), len: p.length };
+      }
+    } catch { /* skip unreadable project files */ }
+  }
+  return best?.id || null;
+}
+const bin = findAgy();
+if (o.probe) A.probe(bin, 'agy (PATH or %LOCALAPPDATA%/agy/bin)');
+if (o.caps) A.caps({ vendor: 'google', reads_repo: true, reads_web: false });
+A.requireRunArgs(o);
+if (!bin) { process.stderr.write('agy not found\n'); process.exit(1); }
+const project = findProject(o.cwd);
+if (!project) { process.stderr.write(`no Antigravity project covers ${o.cwd}; open it in Antigravity once or set BRIDGIE_AGY_PROJECT\n`); process.exit(1); }
+const mode = o.mode === 'write' ? (process.env.BRIDGIE_AGY_WRITE_MODE || 'agent') : 'plan';
+const instruction = `Read ${o.prompt} in full. It contains your role, inputs and output contract. Do the work now in ${o.cwd}, then end your reply with the fenced json block it asks for.`;
+const args = ['--project', project, '--mode', mode, '--effort', o.effort || 'high', '--print-timeout', process.env.BRIDGIE_AGY_TIMEOUT || '15m', '--output-format', 'text'];
+if (o.model !== 'default') args.push('--model', o.model);
+args.push('--print', instruction);   // --print takes the prompt as its VALUE and must come last
+const reply = A.replyPath(o);
+const rc = A.run(bin, args, { cwd: o.cwd, stdout: reply, stderr: join(dirname(o.out), 'agy.log') });
+A.finish(o, reply, rc);

package/adapters/claude-code/gotchas.md

@@ -0,0 +1,5 @@
+# claude-code adapter gotchas
+- Flags verified against Claude Code 2.1.319: `-p`, `--output-format text`, `--permission-mode dontAsk|acceptEdits`, `--allowedTools`, `--disallowedTools`, `--add-dir`, `--model`, `--effort low|medium|high|xhigh|max`.
+- The prompt goes on stdin; never inline it (Windows command-line length limit).
+- `dontAsk` refuses any tool not in `--allowedTools` without prompting, which is what keeps a reviewer read-only in print mode.
+- Override the binary with `BRIDGIE_CLAUDE_BIN` (e.g. a Copilot-hosted wrapper).

package/adapters/claude-code/run.mjs

@@ -0,0 +1,26 @@
+// Claude Code adapter (anthropic). Prompt on stdin, reply on stdout, JSON extracted to --out.
+import { dirname } from 'node:path';
+import * as A from '../lib/common.mjs';
+
+const o = A.parse(process.argv.slice(2));
+const bin = A.which(process.env.BRIDGIE_CLAUDE_BIN || 'claude');
+if (o.probe) A.probe(bin, 'claude CLI');
+if (o.caps) A.caps({ vendor: 'anthropic', reads_repo: true, reads_web: true });
+A.requireRunArgs(o);
+if (!bin) { process.stderr.write('claude CLI not found on PATH\n'); process.exit(1); }
+const args = ['-p', '--output-format', 'text', '--add-dir', dirname(o.out)];
+if (o.model !== 'default') args.push('--model', o.model);
+if (o.effort) args.push('--effort', o.effort);
+if (o.mode === 'readonly') {
+  // Read and inspect only; everything not allowed is refused without prompting.
+  args.push('--permission-mode', 'dontAsk',
+    '--allowedTools', 'Read', 'Grep', 'Glob', 'LS', 'WebSearch', 'WebFetch', 'Bash(git diff:*)', 'Bash(git log:*)', 'Bash(git show:*)', 'Bash(git status:*)',
+    '--disallowedTools', 'Edit', 'Write', 'NotebookEdit');
+} else {
+  args.push('--permission-mode', 'acceptEdits',
+    '--allowedTools', 'Read', 'Grep', 'Glob', 'LS', 'Edit', 'Write', 'Bash',
+    '--disallowedTools', 'Bash(git push:*)', 'Bash(git remote:*)', 'Bash(git checkout:*)', 'Bash(git switch:*)');
+}
+const reply = A.replyPath(o);
+const rc = A.run(bin, args, { cwd: o.cwd, stdin: o.prompt, stdout: reply });
+A.finish(o, reply, rc);

package/adapters/codex/gotchas.md

@@ -0,0 +1,6 @@
+# codex adapter gotchas
+- Flags verified against codex-cli 0.167.0: `exec -s read-only|workspace-write -C <dir> --skip-git-repo-check -o <last message file>`, `-m`, `-c key=value`.
+- Pipe the prompt on stdin (v1: inlining a large diff broke on Windows).
+- ChatGPT-auth Codex rejects `-m` for models it does not offer: leave the model as `default` unless you know.
+- "requires a newer version of Codex" → `npm install -g @openai/codex@latest`.
+- v1 incident: a packet that swept in the run directory hit Codex's 1,048,576-character input limit; packets exclude `.bridgie/runs/**`.

package/adapters/codex/run.mjs

@@ -0,0 +1,18 @@
+// Codex CLI adapter (openai). Prompt on stdin; last message written with -o; JSON extracted to --out.
+import { dirname, join } from 'node:path';
+import * as A from '../lib/common.mjs';
+
+const o = A.parse(process.argv.slice(2));
+const bin = A.which(process.env.BRIDGIE_CODEX_BIN || 'codex');
+if (o.probe) A.probe(bin, 'codex CLI');
+if (o.caps) A.caps({ vendor: 'openai', reads_repo: true, reads_web: true });
+A.requireRunArgs(o);
+if (!bin) { process.stderr.write('codex CLI not found on PATH\n'); process.exit(1); }
+const reply = A.replyPath(o);
+const args = ['exec', '--search', '-s', o.mode === 'write' ? 'workspace-write' : 'read-only', '-C', o.cwd, '--skip-git-repo-check', '-o', reply];   // --search: web search tool on
+// ChatGPT-auth Codex rejects -m for models it does not offer: only pass it when one is configured.
+if (o.model !== 'default') args.push('-m', o.model);
+if (o.effort) args.push('-c', `model_reasoning_effort="${o.effort}"`);
+const log = join(dirname(o.out), 'codex.log');
+const rc = A.run(bin, args, { stdin: o.prompt, stdout: log, stderr: log + '.err' });
+A.finish(o, reply, rc);

package/adapters/copilot/gotchas.md

@@ -0,0 +1,5 @@
+# copilot adapter gotchas
+- Flags verified against GitHub Copilot CLI 1.0.107: `-p` (non-interactive), `-s` (response only), `--no-ask-user`, `--model`, `--reasoning-effort none|minimal|low|medium|high|xhigh|max`, `--allow-tool` / `--deny-tool` (deny always wins), `--add-dir`.
+- `-p` takes the prompt text, so the adapter passes a short instruction to read the prompt file (Windows command-line limit).
+- Vendor is derived from `--model` (claude/opus/sonnet → anthropic, gpt → openai, gemini → google). With `default` the vendor is `copilot-default`, which never matches another adapter, so set an explicit model when cross-family matters.
+- Model names are Copilot's (check `copilot --help` / `/model`), e.g. `claude-opus-5.5`; they are not Anthropic API ids.

package/adapters/copilot/run.mjs

@@ -0,0 +1,32 @@
+// GitHub Copilot CLI adapter. Vendor follows the model (Claude → anthropic, GPT → openai, Gemini → google).
+// The prompt is too large for a command line on Windows, so -p carries an instruction to read the file.
+import { dirname, join } from 'node:path';
+import * as A from '../lib/common.mjs';
+
+const o = A.parse(process.argv.slice(2));
+const bin = A.which(process.env.BRIDGIE_COPILOT_BIN || 'copilot');
+const vendorOf = m => {
+  m = String(m || '').toLowerCase();
+  if (/claude|opus|sonnet|haiku|fable/.test(m)) return 'anthropic';
+  if (/gpt|^o\d|codex/.test(m)) return 'openai';
+  if (/gemini/.test(m)) return 'google';
+  return `copilot-${m || 'default'}`;
+};
+if (o.probe) A.probe(bin, 'copilot CLI');
+if (o.caps) A.caps({ vendor: vendorOf(o.model), reads_repo: true, reads_web: process.env.BRIDGIE_COPILOT_WEB === '1' });
+A.requireRunArgs(o);
+if (!bin) { process.stderr.write('copilot CLI not found on PATH\n'); process.exit(1); }
+const instruction = `Read ${o.prompt} in full. It contains your role, inputs and output contract. Do the work now in ${o.cwd}, then end your reply with the fenced json block it asks for.`;
+const args = ['-s', '--no-ask-user', '--add-dir', dirname(o.prompt)];
+if (o.model !== 'default') args.push('--model', o.model);
+if (o.effort) args.push('--reasoning-effort', o.effort);
+if (o.mode === 'readonly') {
+  // Non-interactive mode denies anything not allowed; deny rules win over allow rules.
+  args.push('--allow-tool', 'shell(git diff:*)', '--allow-tool', 'shell(git log:*)', '--allow-tool', 'shell(git show:*)', '--allow-tool', 'shell(git status:*)', '--deny-tool', 'write');
+} else {
+  args.push('--allow-tool', 'write', '--allow-tool', 'shell', '--deny-tool', 'shell(git push)', '--deny-tool', 'shell(git remote)', '--deny-tool', 'shell(git checkout)', '--deny-tool', 'shell(git switch)');
+}
+args.push('-p', instruction);
+const reply = A.replyPath(o);
+const rc = A.run(bin, args, { cwd: o.cwd, stdout: reply, stderr: join(dirname(o.out), 'copilot.log') });
+A.finish(o, reply, rc);

package/adapters/fake/fixtures/buyer.2.json

@@ -0,0 +1 @@
+{ "card": "CARD", "buys": false, "purchases_first_month": 0, "reason_code": "price", "reason": "£39 a deal adds up when margins are thin", "objection": "my junior is cheaper", "would_flip": "a monthly cap", "prices": { "too_cheap": 5, "bargain": 15, "getting_expensive": 30, "too_expensive": 60 } }

package/adapters/fake/fixtures/buyer.3.json

@@ -0,0 +1 @@
+{ "card": "CARD", "buys": false, "purchases_first_month": 0, "reason_code": "trust", "reason": "never heard of you and lenders are fussy", "objection": "lenders will not accept it", "would_flip": "a monthly cap", "prices": { "too_cheap": 10, "bargain": 20, "getting_expensive": 50, "too_expensive": 100 } }

package/adapters/fake/fixtures/buyer.4.json

@@ -0,0 +1 @@
+{ "card": "CARD", "buys": true, "purchases_first_month": 1, "reason_code": "need", "reason": "worth a try on the next stuck deal", "objection": "setup time", "would_flip": "n/a", "prices": { "too_cheap": 8, "bargain": 20, "getting_expensive": 60, "too_expensive": 100 } }

package/adapters/fake/fixtures/buyer.5.json

@@ -0,0 +1 @@
+{ "card": "CARD", "buys": false, "purchases_first_month": 0, "reason_code": "price", "reason": "no", "objection": "price", "would_flip": "free tier", "prices": { "too_cheap": 20, "bargain": 10, "getting_expensive": 30, "too_expensive": 60 } }

package/adapters/fake/fixtures/buyer.json

@@ -0,0 +1 @@
+{ "card": "CARD", "buys": true, "purchases_first_month": 2, "reason_code": "need", "reason": "that is my Tuesday back", "objection": "will the lender take it", "would_flip": "n/a", "prices": { "too_cheap": 5, "bargain": 25, "getting_expensive": 60, "too_expensive": 120 } }

package/adapters/fake/fixtures/chair.json

@@ -0,0 +1,8 @@
+{
+  "seat": "chair",
+  "summary_md": "Vote: one fund, one fund if, one pass. Shared risk: lenders refuse a standard pack. Disagreement: the moat lens sees no defensibility; the product lens sees a whole product. Strongest version: broker-side pack builder sold per deal. First: ask three brokers what Tuesday costs them.",
+  "artifact_md": "## Shared risks\n- Lenders refuse a standard pack (offer, moat)\n\n## Conditions\n- [ ] Three brokers confirm the Tuesday cost (evidence)\n- [ ] Two lenders accept a standard pack (evidence)\n\n## Where the board disagrees\n- Moat: nothing stops a lender copying it. Product: the broker relationship is the product.\n\n## The strongest version\nA pack builder brokers pay for per deal, starting with the two lenders brokers use most.",
+  "questions": [
+    { "context": "How many applications a month does a typical target broker handle?", "options": ["under 5", "5–20", "over 20", "don't know"], "allow_free_text": true }
+  ]
+}

package/adapters/fake/fixtures/decomposer.json

@@ -0,0 +1,5 @@
+{
+  "seat": "decomposer",
+  "summary_md": "Two tasks; every requirement mapped.",
+  "artifact_md": "```json\n{\n  \"tasks\": [\n    {\n      \"id\": \"T1\",\n      \"title\": \"Create collection endpoint\",\n      \"requirements\": [\n        \"R1\"\n      ],\n      \"depends_on\": []\n    },\n    {\n      \"id\": \"T2\",\n      \"title\": \"Idempotent collection handling\",\n      \"requirements\": [\n        \"R2\",\n        \"P1\"\n      ],\n      \"depends_on\": [\n        \"T1\"\n      ]\n    }\n  ]\n}\n```"
+}

package/adapters/fake/fixtures/evidence.2.json

@@ -0,0 +1,6 @@
+{
+  "seat": "evidence",
+  "summary_md": "Three brokers interviewed; all three spend half a day per application; none measured it.",
+  "artifact_md": "## What the material shows\n- 3 of 3 brokers say a pack takes 4–6 hours (notes-broker-1.md, notes-broker-2.md, notes-broker-3.md)\n\n## What it does not show\n- No timesheet; the hours are recalled, not measured.\n\n## Sources\n- 3 interview notes, 3 independent firms",
+  "proposals": [ { "kind": "confidence", "text": "anecdote: three independent brokers, recalled not measured" } ]
+}

package/adapters/fake/fixtures/evidence.json

@@ -0,0 +1,5 @@
+{
+  "seat": "evidence",
+  "summary_md": "Two tasks. Both interviews; five brokers each; stop when three agree.",
+  "artifact_md": "### E1\nMethod: interview. Sample: 5 brokers from NACFB list. Stop: after 5 or when 3 give the same answer.\nQuestions:\n- When did you last prepare a statement pack, and how long did it take?\n- What did you do when a lender rejected the pack?\n\n### E-panel\nMethod: priced landing page. Sample: 100 visitors. Stop: 2 weeks.\n"
+}

package/adapters/fake/fixtures/framer.json

@@ -0,0 +1,9 @@
+{
+  "seat": "framer",
+  "summary_md": "Assumed: brokers pay per deal; the junior's time is the cost; lenders will accept a standard pack.",
+  "artifact_md": "## Who\nA UK commercial finance broker with two to five staff, preparing a lending application at month-end.\n\n## Problem in their words\n\"I spend most of Tuesday on statements for one client.\" (assumption: typical, from one broker)\n\n## Why now\nOpen banking data is now available for most SME accounts (changed 2025).\n\n## What they do today\nSpreadsheets and a junior who keys the numbers in.\n\n## Expected outcome\nStatement analysis per application from ~6 hours to under 1 hour, read from broker timesheets by 2027-03. Baseline unknown.",
+  "questions": [
+    { "context": "Who pays: the broker per deal, or the lender?", "options": ["broker per deal", "broker subscription", "lender", "don't know"], "allow_free_text": true },
+    { "context": "Is 6 hours per application measured or a guess?", "options": ["measured", "guess", "don't know"], "allow_free_text": true }
+  ]
+}

package/adapters/fake/fixtures/grill.json

@@ -0,0 +1,9 @@
+{
+  "seat": "grill",
+  "summary_md": "Strongest assumption: merchants notice failed collections. Weakest evidence: no interview quotes. Pre-mortem: finance already retries by hand; the metric cannot be read by the date.",
+  "questions": [
+    { "context": "Does anyone retry failed card payments manually today, and who?", "options": ["yes, finance does", "yes, support does", "no", "don't know"], "allow_free_text": true },
+    { "context": "Is the churn baseline (4.1%) measured or estimated?", "options": ["measured", "estimated", "don't know"], "allow_free_text": true }
+  ],
+  "proposals": [ { "kind": "confidence", "text": "anecdote: two tickets and one conversation; no measured data yet" } ]
+}

package/adapters/fake/fixtures/hardener.json

@@ -0,0 +1,8 @@
+{
+  "task": "T1",
+  "claims": [
+    { "text": "failure-modes: idempotency covered by the request-id unique index", "proof_cmd": "grep -n unique feature.txt", "proof_output": "1:unique request_id" }
+  ],
+  "files_changed": [],
+  "notes": "observability: no metric yet (gap)"
+}

package/adapters/fake/fixtures/implementer.json

@@ -0,0 +1,8 @@
+{
+  "task": "T1",
+  "claims": [
+    { "text": "Requirement satisfied; see the new file", "proof_cmd": "cat feature.txt", "proof_output": "implemented" }
+  ],
+  "files_changed": ["feature.txt"],
+  "notes": ""
+}

package/adapters/fake/fixtures/lens-moat.json

@@ -0,0 +1,5 @@
+{
+  "seat": "lens-moat",
+  "summary_md": "PASS — biggest risk from the moat lens.",
+  "artifact_md": "Score: 5 | Vote: PASS\n## What is strong\nThe person is specific.\n## What would kill it\n- Lenders refuse a standard pack (moat lens)\n- Brokers will not pay per deal\n## Questions for the owner\n- How many deals a month does a typical broker do?\n## Conditions\n- [ ] Three brokers confirm the Tuesday cost"
+}

package/adapters/fake/fixtures/lens-offer.json

@@ -0,0 +1,5 @@
+{
+  "seat": "lens-offer",
+  "summary_md": "FUND IF — biggest risk from the offer lens.",
+  "artifact_md": "Score: 7 | Vote: FUND IF\n## What is strong\nThe person is specific.\n## What would kill it\n- Lenders refuse a standard pack (offer lens)\n- Brokers will not pay per deal\n## Questions for the owner\n- How many deals a month does a typical broker do?\n## Conditions\n- [ ] Three brokers confirm the Tuesday cost"
+}

package/adapters/fake/fixtures/lens-product.json

@@ -0,0 +1,5 @@
+{
+  "seat": "lens-product",
+  "summary_md": "FUND — biggest risk from the product lens.",
+  "artifact_md": "Score: 8 | Vote: FUND\n## What is strong\nThe person is specific.\n## What would kill it\n- Lenders refuse a standard pack (product lens)\n- Brokers will not pay per deal\n## Questions for the owner\n- How many deals a month does a typical broker do?\n## Conditions\n- [ ] Three brokers confirm the Tuesday cost"
+}

package/adapters/fake/fixtures/packager.json

@@ -0,0 +1,5 @@
+{
+  "seat": "packager",
+  "summary_md": "All three criteria pass; one minor finding; no load test.",
+  "artifact_md": "# Acceptance \u2014 Net collection\n\n## The bet and its criteria\nMerchants' failed collections retried once, idempotently. R1, R2, P1.\n\n## What changed and blast radius\nTwo files; one new endpoint; no shared contracts changed.\n\n## Evidence matrix\n| Criterion | Tier | Result |\n|---|---|---|\n| R1 | adversarial | pass |\n| R2 | adversarial | pass |\n| P1 | adversarial | pass |\n\n## Findings\nOne minor wording finding, not fixed.\n\n## Rollback and alerts\nFlag net_collect off.\n\n## Residual uncertainty\n- No load test was run.\n- Property P1 checked by review, not by a property test.\n\n## Decision\nApprove, approve with conditions, or reject on the decision page.\n"
+}

package/adapters/fake/fixtures/profiler.json

@@ -0,0 +1,9 @@
+{
+  "seat": "profiler",
+  "summary_md": "Guesses: segment shares, the income ranges. Check: whether solo brokers exist in volume.",
+  "files": {
+    "customer.yaml": "business: Statement pack builder for brokers\ntarget: UK commercial finance broker, 2–5 staff, at month-end\nsegments:\n  - { name: solo broker, share: 0.5, income: [40000, 60000] }\n  - { name: small firm, share: 0.5, income: [60000, 120000] }\nbehaviours:\n  - { name: deadline buyer, share: 0.6, how: buys when a deal is stuck }\n  - { name: tool sceptic, share: 0.4, how: has been burned by software before }\nobjections:\n  - lenders will not accept it\n  - my junior is cheaper\nhabits:\n  - works Saturdays in Q4\n",
+    "pitch.md": "A pack builder that reads a year of bank statements and produces the lender's own template.\n£39 per application, pay as you go.\nUpload PDFs or connect open banking; the pack is ready in ten minutes.\nFirst three applications free.\n"
+  },
+  "proposals": [ { "kind": "confidence", "text": "data: the panel will tell us" } ]
+}

package/adapters/fake/fixtures/researcher.json

@@ -0,0 +1,6 @@
+{
+  "seat": "researcher",
+  "summary_md": "Crowded at the lender end, empty at the broker end. Prices £49–£299/month. Loudest complaint: categorisation errors. Gap: broker-side packs. Sources seen 2026-11.",
+  "artifact_md": "## Direct competitors\n| Name | What | Price | Link |\n|---|---|---|---|\n| StatementCo | PDF statement parsing for lenders | £299/month (seen 2026-11-07) | https://example.com/statementco |\n| Cashflow Ltd | Open-banking cash-flow reports | £49/month | https://example.com/cashflow |\n\n## Indirect and substitutes\n- Excel templates shared on broker forums https://example.com/forum/templates\n\n## Prices\n- Brokers pay nothing today; the junior's time is the cost (no public figure).\n\n## What their customers complain about\n- \"it miscategorises transfers as income\" https://example.com/reviews/statementco\n\n## Not found\n- No broker-side product with public pricing.\n\n## Sources\n- https://example.com/statementco\n- https://example.com/cashflow\n- https://example.com/forum/templates\n- https://example.com/reviews/statementco",
+  "questions": [ { "context": "Which geography counts: UK only, or UK and Ireland?", "options": ["UK only", "UK and Ireland", "don't know"], "allow_free_text": true } ]
+}

package/adapters/fake/fixtures/retro.json

@@ -0,0 +1,14 @@
+{
+  "seat": "retro",
+  "summary_md": "Two tasks, one retry. The first reviewer round found a missing status.",
+  "proposals": [
+    {
+      "kind": "test",
+      "text": "Add a test that a new collection starts in status pending."
+    },
+    {
+      "kind": "gotcha",
+      "text": "Collections must set status on create; reviewers caught it once."
+    }
+  ]
+}

package/adapters/fake/fixtures/reviewer-fail.json

@@ -0,0 +1,11 @@
+{
+  "task": "T1",
+  "criteria": [
+    { "id": "R1", "tier": "adversarial", "check": "status not set to pending", "result": "fail", "artifact": null }
+  ],
+  "findings": [
+    { "severity": "major", "file": "feature-1.txt", "line": 1, "criterion": "R1", "claim": "collection created without a status", "repro": "POST /collections then GET it", "disposition": null }
+  ],
+  "verdict": "fail",
+  "verifier": { "seat": "reviewer", "adapter": "fake", "model": "default" }
+}

package/adapters/fake/fixtures/reviewer.json

@@ -0,0 +1,13 @@
+{
+  "task": "T1",
+  "criteria": [
+    { "id": "R1", "tier": "adversarial", "check": "read the change; walked the event path", "result": "pass", "artifact": null },
+    { "id": "R2", "tier": "adversarial", "check": "duplicate request replayed against the handler", "result": "pass", "artifact": null },
+    { "id": "P1", "tier": "adversarial", "check": "unique index on request_id", "result": "pass", "artifact": null }
+  ],
+  "findings": [
+    { "severity": "minor", "file": "feature-1.txt", "line": 1, "criterion": null, "claim": "wording could be clearer", "repro": "read the file", "disposition": null }
+  ],
+  "verdict": "pass",
+  "verifier": { "seat": "reviewer", "adapter": "fake", "model": "default" }
+}

package/adapters/fake/fixtures/scout.json

@@ -0,0 +1,8 @@
+{
+  "seat": "scout",
+  "summary_md": "Sample: 3 support exports and 2 broker emails, Sept–Oct 2026. Over-represents brokers who already complain. Small.",
+  "artifact_md": "## O1 · Reading a year of bank statements by hand takes a day per application\nSources: 4 (tickets__t1.txt, tickets__t2.txt, emails__broker-a.txt, emails__broker-b.txt)\nSeverity: high — a day of a broker's time per deal, deals lost to faster lenders\nQuote: \"I spend most of Tuesday on statements for one client\" (emails__broker-a.txt)\nToday: spreadsheets and a junior who keys the numbers in\nSymptom or cause: symptom; the cause is that lenders each want a different pack\n\n## O2 · Lenders ask for the same documents in different formats\nSources: 2 (emails__broker-b.txt, tickets__t3.txt)\nSeverity: medium — rework, not lost deals\nQuote: \"every lender has its own template\" (tickets__t3.txt)\nToday: not said\nSymptom or cause: cause of O1",
+  "questions": [
+    { "context": "Are the two broker emails from the same firm?", "options": ["same firm", "different firms", "don't know"], "allow_free_text": true }
+  ]
+}

package/adapters/fake/fixtures/shaper.json

@@ -0,0 +1,15 @@
+{
+  "seat": "shaper",
+  "summary_md": "Event-driven retry with an idempotency key; one new table; flag off by default.",
+  "files": {
+    "requirements.md": "# Requirements \u2014 Net collection\n\n## R1 \u00b7 event \u00b7 risk: none\nWHEN a collection is created THE SYSTEM SHALL record it with status pending.\n\n## R2 \u00b7 event \u00b7 risk: none\nWHEN the same collection request arrives twice THE SYSTEM SHALL apply side effects exactly once.\n\n## P1 \u00b7 property\nFor any sequence of collection requests, the number of ledger entries SHALL equal the number of distinct request ids.\nValidates: R2\n",
+    "design.md": "# Design \u2014 Net collection\n\n## Context\nFailed card payments are retried by hand.\n## Options considered\nA: scheduled job. B: event-driven retry. B chosen.\n## Decision\nEvent-driven retry with an idempotency key per request id.\n## Consequences\nNeeds a dedup table.\n## Data changes\nNew table collection(request_id unique).\n## Interfaces and contracts\nPOST /collections.\n## Failure modes\nDuplicate delivery; partial failure between insert and publish.\n## Migration\nAdditive.\n## Rollout and flag\nFlag net_collect, default off.\n## Observability\nCounter collections_created.\n## Security notes\nMerchant-scoped.\n## Consumer trace\nNone changed.\n"
+  },
+  "questions": [],
+  "proposals": [
+    {
+      "kind": "task",
+      "text": "T1 create collection; T2 idempotency"
+    }
+  ]
+}

package/adapters/fake/gotchas.md

@@ -0,0 +1,2 @@
+# fake adapter gotchas
+- It ignores the prompt; tests that need prompt-dependent behaviour must assert on prompt.md directly.