qaas-python 0.1.0__tar.gz → 0.2.0__tar.gz

{qaas_python-0.1.0 → qaas_python-0.2.0}/.gitignore

@@ -13,3 +13,9 @@ node_modules/
 
 # Agent worktrees — real git worktrees, never repo content.
 .claude/worktrees/
+
+# Build artifacts.
+dist/
+
+# Agent evidence screenshots land here because put_artifact takes only text.
+target-app/*.png

{qaas_python-0.1.0 → qaas_python-0.2.0}/CLAUDE.md

@@ -80,6 +80,11 @@ adding an agent as a sign something is wrong. Constraints enforced in
 `config.py`: at most 6 MCP servers per agent (§5.3, tool-selection accuracy),
 and every `must_call` tool must name a server the agent actually has.
 
+Adding a discovery agent is a prompt file plus a YAML file and no Python --
+`_phase_discover` falls back to `tasks.discovery` for anything without a
+bespoke builder. VAULT and WARDEN were added that way and found that it was
+not true before them.
+
 `prompts/_shared.md` is appended to every agent prompt — house rules go there,
 not copy-pasted into six prompts. Prompts resolve through `Workspace.prompt_dirs`
 (`.qaas/prompts/` beats the packaged copy), **file by file and independently**, so

{qaas_python-0.1.0 → qaas_python-0.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: qaas-python
-Version: 0.1.0
+Version: 0.2.0
 Summary: A multi-agent QA system: finds real defects, reproduces them, files tickets, fixes them, and proves the fix
 Project-URL: Homepage, https://github.com/allaabdella2-us/qa-multi-agent-system
 Project-URL: Repository, https://github.com/allaabdella2-us/qa-multi-agent-system
@@ -22,7 +22,7 @@ Requires-Dist: claude-agent-sdk>=0.2.127
 Requires-Dist: pydantic>=2.9
 Requires-Dist: pyyaml>=6.0
 Requires-Dist: rich>=13.9
-Requires-Dist: typer>=0.15
+Requires-Dist: typer>=0.16
 Provides-Extra: dev
 Requires-Dist: pytest-asyncio>=0.24; extra == 'dev'
 Requires-Dist: pytest>=8.3; extra == 'dev'
@@ -43,7 +43,7 @@ Description-Content-Type: text/markdown
 [![Tests](https://img.shields.io/badge/tests-649%20offline-success)](#-contributing)
 [![Built on](https://img.shields.io/badge/built%20on-Claude%20Agent%20SDK-D97757)](https://docs.claude.com/en/api/agent-sdk/overview)
 
-[Quickstart](#-quickstart-in-60-seconds) · [What it costs](#-what-it-costs) · [Your repo](#-point-it-at-your-repository) · [Jira](#-file-into-jira) · [Architecture](ARCHITECTURE.md)
+[Quickstart](#-quickstart-in-60-seconds) · [Your repo](#-point-it-at-your-repository) · [Jira](#-file-into-jira) · [Architecture](ARCHITECTURE.md)
 
 </div>
 
@@ -51,7 +51,7 @@ Description-Content-Type: text/markdown
 
 Most "AI QA" tools generate tests. **This one behaves like a QA team.**
 
-Eight agents, each with its own context, tool allowlist and budget, coordinated by
+Ten agents, each with its own context, tool allowlist and budget, coordinated by
 a state machine that is ordinary Python — because a model cannot enforce a budget
 it is itself spending.
 
@@ -111,37 +111,6 @@ qaas run --repo https://github.com/you/your-app --dry-run
 
 ---
 
-## 💰 What it costs
-
-> [!IMPORTANT]
-> **Real runs spend real money.** Read this before your first one.
-
-Median cost per dispatch, **measured** across real runs — not estimated:
-
-| agent | median | what you get |
-|---|--:|---|
-| 🖱️ `SURFACE` | **$3.34** | broken flows, console errors, a11y, forms |
-| 🔧 `MENDER` | **$2.03** | the minimal fix, on a branch |
-| 🔌 `CONDUIT` | **$1.97** | API contract, authz and error-shape defects |
-| 🔨 `FORGE` | **$1.61** | a minimal repro + failing test — **per finding** |
-| ✅ `PROOF` / ⚖️ `ARBITER` | ~$1.00 | verification and adversarial review |
-| 📝 `CLERK` / 🗺️ `CARTOGRAPHER` | ~$0.70 | filing, and the map everything reads |
-
-A full discovery run over the demo app found **13 of 16** seeded defects for
-about **$15**. A `fix-cycle` pass costs **$5–7**.
-
-> `FORGE` runs **once per finding** in a fresh context, so cost scales with what
-> was found, not with how many agents exist.
-
-**The controls are real, not advisory:**
-
-- `max_budget_usd` per agent *and* per run mode; the governor checks before every dispatch and **stops the run** rather than overspending.
-- The cap survives a resume — `qaas run --run-id <existing>` carries forward what that run already spent.
-- `qaas validate` refuses a run mode whose agents could outspend its cap.
-- `--dry-run` on everything.
-
----
-
 ## 🎯 Point it at your repository
 
 ```bash
@@ -276,20 +245,20 @@ Every tool call, denial, verdict and escalation is on the record.
 
 ```console
 $ qaas trace run-20260908T182034-c6ed26
-    t+  agent         kind            detail                                    cost
-    0s  -             run_started     mode=nightly  agents=[5]  budget_usd=40
+    t+  agent         kind            detail
+    0s  -             run_started     mode=nightly  agents=[7]
     0s  CARTOGRAPHER  agent_started   model=claude-sonnet-5
     4s  CARTOGRAPHER  tool_call ×34   Read×25, Glob×6, ToolSearch×2
     6s  CARTOGRAPHER  denial          tool=Bash  reason=Bash is not in CARTOGRAPHER's
                                       tool allowlist (Read, Grep, Glob).
   146s  CARTOGRAPHER  system_map      version=20260907T233530  sections=[12]
-  156s  CARTOGRAPHER  agent_finished  subtype=success  num_turns=45            $0.60
+  156s  CARTOGRAPHER  agent_finished  subtype=success  num_turns=45
 ```
 
 ```bash
 qaas trace <run-id> --agent proof --kind verdict   # filter
 qaas trace <run-id> --json                         # export
-qaas show <run-id>                                 # mode, commit, cost, tickets, escalations
+qaas show <run-id>                                 # mode, commit, tickets, escalations
 qaas runs                                          # everything that ever ran
 ```
 
@@ -327,7 +296,6 @@ Two runs against the demo app, scored automatically:
 | 🎯 recall | **81%** — 13 of 16 | **69%** — 11 of 16 |
 | 🔇 precision | **100%** — 0 FP | **92%** — 1 FP |
 | 🏷️ severity agreement | **100%** | **100%** |
-| 💵 cost per accepted finding | $1.12 | $0.64 |
 
 **Both numbers are shown on purpose.** A single figure would be the flattering
 one, and it would not survive contact with a second run. These are stochastic
@@ -350,7 +318,8 @@ precision is measured rather than assumed.
 
 Honest about what exists:
 
-- ✅ **8 of the 16 agents** in the design are built — CARTOGRAPHER, CONDUIT, SURFACE, FORGE, CLERK, MENDER, ARBITER, PROOF. CONDUCTOR is the Python state machine rather than an agent. **Seven Phase-2 agents are designed, not written.**
+- ✅ **10 of the 16 agents** in the design are built — CARTOGRAPHER, CONDUIT, SURFACE, VAULT, WARDEN, FORGE, CLERK, MENDER, ARBITER, PROOF. CONDUCTOR is the Python state machine rather than an agent. The five that remain (KEYSTONE, PULSE, USHER, GAUGE, CHRONICLE) are additional discovery specialists, not missing parts of the loop.
+- ✅ **Adding an agent needs a prompt file and a YAML file — no Python.** VAULT and WARDEN were added exactly that way, which is how the claim finally got tested.
 - ✅ The fix loop has closed end to end on a real defect: `NOT_FIXED → MENDER → ARBITER APPROVE → VERIFIED`.
 - ✅ 30 skills, 7 in-process MCP servers, 649 offline tests.
 - ⚠️ Running the bundled demo needs `export CORVID_PASSWORD=password123` — credentials come from the environment, including the demo's.

{qaas_python-0.1.0 → qaas_python-0.2.0}/README.md

@@ -13,7 +13,7 @@
 [![Tests](https://img.shields.io/badge/tests-649%20offline-success)](#-contributing)
 [![Built on](https://img.shields.io/badge/built%20on-Claude%20Agent%20SDK-D97757)](https://docs.claude.com/en/api/agent-sdk/overview)
 
-[Quickstart](#-quickstart-in-60-seconds) · [What it costs](#-what-it-costs) · [Your repo](#-point-it-at-your-repository) · [Jira](#-file-into-jira) · [Architecture](ARCHITECTURE.md)
+[Quickstart](#-quickstart-in-60-seconds) · [Your repo](#-point-it-at-your-repository) · [Jira](#-file-into-jira) · [Architecture](ARCHITECTURE.md)
 
 </div>
 
@@ -21,7 +21,7 @@
 
 Most "AI QA" tools generate tests. **This one behaves like a QA team.**
 
-Eight agents, each with its own context, tool allowlist and budget, coordinated by
+Ten agents, each with its own context, tool allowlist and budget, coordinated by
 a state machine that is ordinary Python — because a model cannot enforce a budget
 it is itself spending.
 
@@ -81,37 +81,6 @@ qaas run --repo https://github.com/you/your-app --dry-run
 
 ---
 
-## 💰 What it costs
-
-> [!IMPORTANT]
-> **Real runs spend real money.** Read this before your first one.
-
-Median cost per dispatch, **measured** across real runs — not estimated:
-
-| agent | median | what you get |
-|---|--:|---|
-| 🖱️ `SURFACE` | **$3.34** | broken flows, console errors, a11y, forms |
-| 🔧 `MENDER` | **$2.03** | the minimal fix, on a branch |
-| 🔌 `CONDUIT` | **$1.97** | API contract, authz and error-shape defects |
-| 🔨 `FORGE` | **$1.61** | a minimal repro + failing test — **per finding** |
-| ✅ `PROOF` / ⚖️ `ARBITER` | ~$1.00 | verification and adversarial review |
-| 📝 `CLERK` / 🗺️ `CARTOGRAPHER` | ~$0.70 | filing, and the map everything reads |
-
-A full discovery run over the demo app found **13 of 16** seeded defects for
-about **$15**. A `fix-cycle` pass costs **$5–7**.
-
-> `FORGE` runs **once per finding** in a fresh context, so cost scales with what
-> was found, not with how many agents exist.
-
-**The controls are real, not advisory:**
-
-- `max_budget_usd` per agent *and* per run mode; the governor checks before every dispatch and **stops the run** rather than overspending.
-- The cap survives a resume — `qaas run --run-id <existing>` carries forward what that run already spent.
-- `qaas validate` refuses a run mode whose agents could outspend its cap.
-- `--dry-run` on everything.
-
----
-
 ## 🎯 Point it at your repository
 
 ```bash
@@ -246,20 +215,20 @@ Every tool call, denial, verdict and escalation is on the record.
 
 ```console
 $ qaas trace run-20260908T182034-c6ed26
-    t+  agent         kind            detail                                    cost
-    0s  -             run_started     mode=nightly  agents=[5]  budget_usd=40
+    t+  agent         kind            detail
+    0s  -             run_started     mode=nightly  agents=[7]
     0s  CARTOGRAPHER  agent_started   model=claude-sonnet-5
     4s  CARTOGRAPHER  tool_call ×34   Read×25, Glob×6, ToolSearch×2
     6s  CARTOGRAPHER  denial          tool=Bash  reason=Bash is not in CARTOGRAPHER's
                                       tool allowlist (Read, Grep, Glob).
   146s  CARTOGRAPHER  system_map      version=20260907T233530  sections=[12]
-  156s  CARTOGRAPHER  agent_finished  subtype=success  num_turns=45            $0.60
+  156s  CARTOGRAPHER  agent_finished  subtype=success  num_turns=45
 ```
 
 ```bash
 qaas trace <run-id> --agent proof --kind verdict   # filter
 qaas trace <run-id> --json                         # export
-qaas show <run-id>                                 # mode, commit, cost, tickets, escalations
+qaas show <run-id>                                 # mode, commit, tickets, escalations
 qaas runs                                          # everything that ever ran
 ```
 
@@ -297,7 +266,6 @@ Two runs against the demo app, scored automatically:
 | 🎯 recall | **81%** — 13 of 16 | **69%** — 11 of 16 |
 | 🔇 precision | **100%** — 0 FP | **92%** — 1 FP |
 | 🏷️ severity agreement | **100%** | **100%** |
-| 💵 cost per accepted finding | $1.12 | $0.64 |
 
 **Both numbers are shown on purpose.** A single figure would be the flattering
 one, and it would not survive contact with a second run. These are stochastic
@@ -320,7 +288,8 @@ precision is measured rather than assumed.
 
 Honest about what exists:
 
-- ✅ **8 of the 16 agents** in the design are built — CARTOGRAPHER, CONDUIT, SURFACE, FORGE, CLERK, MENDER, ARBITER, PROOF. CONDUCTOR is the Python state machine rather than an agent. **Seven Phase-2 agents are designed, not written.**
+- ✅ **10 of the 16 agents** in the design are built — CARTOGRAPHER, CONDUIT, SURFACE, VAULT, WARDEN, FORGE, CLERK, MENDER, ARBITER, PROOF. CONDUCTOR is the Python state machine rather than an agent. The five that remain (KEYSTONE, PULSE, USHER, GAUGE, CHRONICLE) are additional discovery specialists, not missing parts of the loop.
+- ✅ **Adding an agent needs a prompt file and a YAML file — no Python.** VAULT and WARDEN were added exactly that way, which is how the claim finally got tested.
 - ✅ The fix loop has closed end to end on a real defect: `NOT_FIXED → MENDER → ARBITER APPROVE → VERIFIED`.
 - ✅ 30 skills, 7 in-process MCP servers, 649 offline tests.
 - ⚠️ Running the bundled demo needs `export CORVID_PASSWORD=password123` — credentials come from the environment, including the demo's.

{qaas_python-0.1.0 → qaas_python-0.2.0}/pyproject.toml

@@ -2,7 +2,7 @@
 # Distribution name is `qaas-python` (`qaas` was taken); the import package and
 # the CLI are both `qaas`.
 name = "qaas-python"
-version = "0.1.0"
+version = "0.2.0"
 description = "A multi-agent QA system: finds real defects, reproduces them, files tickets, fixes them, and proves the fix"
 readme = "README.md"
 license = "MIT"
@@ -23,7 +23,12 @@ dependencies = [
     "claude-agent-sdk>=0.2.127",
     "pydantic>=2.9",
     "pyyaml>=6.0",
-    "typer>=0.15",
+    # typer>=0.16 is not cosmetic: click 8.2 changed Parameter.make_metavar()
+    # to require a ctx argument, and typer 0.15 calls it without one. The pair
+    # `typer 0.15.3` + `click 8.2.0` resolves happily and then crashes on
+    # `qaas <command> --help` with a TypeError -- which is how a user found it,
+    # on a package installed from PyPI.
+    "typer>=0.16",
     "rich>=13.9",
 ]
 

{qaas_python-0.1.0 → qaas_python-0.2.0}/src/qaas/conductor.py

@@ -272,18 +272,20 @@ class Conductor:
         if not discovery:
             return
 
+        # CONDUIT and SURFACE keep bespoke tasks because they name tools only
+        # they have. Everything else gets the generic discovery task, which is
+        # what makes "a new agent is a prompt plus a YAML" true: this used to be
+        # a closed dict, so a new discovery agent was skipped with `no task
+        # builder` -- it validated, it assembled, it showed up in `--dry-run`,
+        # and then it silently did nothing.
         builders = {
-            "CONDUIT": lambda: tasks.conduit(self.config, mode),
-            "SURFACE": lambda: tasks.surface(self.config, mode),
+            "CONDUIT": lambda spec: tasks.conduit(self.config, mode),
+            "SURFACE": lambda spec: tasks.surface(self.config, mode),
         }
         jobs = [
-            (spec, builders[spec.name]())
+            (spec, builders.get(spec.name, lambda sp: tasks.discovery(self.config, mode, sp))(spec))
             for spec in discovery
-            if spec.name in builders
         ]
-        unknown = [s.name for s in discovery if s.name not in builders]
-        if unknown:
-            store.log("skipped", reason="no task builder", agents=unknown)
 
         await self._gather(jobs, store, budget, report, map_version, self.config.run_modes[mode].max_concurrency)
 

qaas_python-0.2.0/src/qaas/defaults/config/agents/vault.yaml

@@ -0,0 +1,21 @@
+name: VAULT
+layer: discovery
+role: >
+  Database and data-integrity analyst. Finds schema constraints the application
+  assumes but the database does not enforce, migrations that lose or corrupt
+  data, missing indexes on paths the code queries, and cross-tenant reads that
+  the ORM makes easy to write. Reports what the schema actually says, never what
+  the model layer claims.
+prompt: VAULT.md
+model: claude-opus-5
+effort: high
+max_turns: 60
+max_budget_usd: 3.0
+mcp_servers: [envelope, env_control, defect_memory]
+builtin_tools: [Read, Grep, Glob]
+policy: {}
+
+skills: [authz-matrix-check, environment-pinning, severity-rubric, repro-minimisation]
+
+# No must_call: finding nothing is a valid outcome for a discovery agent, and
+# requiring an emission would manufacture findings to satisfy it.

qaas_python-0.2.0/src/qaas/defaults/config/agents/warden.yaml

@@ -0,0 +1,23 @@
+name: WARDEN
+layer: discovery
+role: >
+  Security and dependency auditor. Finds missing authorization, secrets committed
+  to the repository, dependencies with known advisories, and error paths that
+  leak internals to a caller. Reports a concrete exploit path or lowers its
+  confidence -- a security finding without one is a guess wearing a severity.
+prompt: WARDEN.md
+model: claude-opus-5
+effort: high
+max_turns: 60
+# Measured: WARDEN exhausted $3.00 on its first real run against the demo app
+# and was killed mid-audit. Building the endpoint-by-role matrix and actually
+# impersonating each role costs more than reading a spec does.
+max_budget_usd: 5.0
+mcp_servers: [envelope, env_control, defect_memory]
+builtin_tools: [Read, Grep, Glob]
+policy: {}
+
+skills: [authz-matrix-check, error-taxonomy, severity-rubric, routing-rules, repro-minimisation]
+
+# No must_call: see VAULT. Also: an auditor that must report something will
+# report something, and security noise is the fastest way to be ignored.

{qaas_python-0.1.0 → qaas_python-0.2.0}/src/qaas/defaults/config/system.yaml

@@ -33,11 +33,11 @@ run_modes:
 
   nightly:
     trigger: cron
-    agents: [CARTOGRAPHER, CONDUIT, SURFACE, FORGE, CLERK]
+    agents: [CARTOGRAPHER, CONDUIT, SURFACE, VAULT, WARDEN, FORGE, CLERK]
     # FORGE runs once per finding, so the deep sweep's budget scales with how
     # much discovery found, not with the number of agents. Measured: discovery
     # ~$6, then roughly $1-2 per finding reproduced.
-    max_budget_usd: 40.0
+    max_budget_usd: 50.0
     max_wall_clock_s: 7200
     max_concurrency: 3
 
@@ -63,7 +63,7 @@ run_modes:
   # expensive mode in the system and the only one that closes the loop.
   full-loop:
     trigger: on_demand
-    agents: [CARTOGRAPHER, CONDUIT, SURFACE, FORGE, CLERK, MENDER, ARBITER, PROOF]
-    max_budget_usd: 60.0
+    agents: [CARTOGRAPHER, CONDUIT, SURFACE, VAULT, WARDEN, FORGE, CLERK, MENDER, ARBITER, PROOF]
+    max_budget_usd: 70.0
     max_wall_clock_s: 10800
     max_concurrency: 3

qaas_python-0.2.0/src/qaas/prompts/VAULT.md

@@ -0,0 +1,59 @@
+You are VAULT, the database and data-integrity analyst.
+
+## Your domain
+
+The schema, and the distance between what it enforces and what the application
+assumes. Application code is full of invariants nobody wrote down; your job is to
+find the ones the database will not hold up.
+
+Detect:
+
+- **Constraints the code assumes and the schema does not enforce** — a field the
+  application treats as required with no `NOT NULL`, a relationship it treats as
+  unique with no unique index, an enum validated only in the model layer.
+- **Missing foreign keys**, or ones declared without a delete rule, so a parent
+  row can leave orphans behind.
+- **Cross-tenant reads** — a query filtered by id but not by the owning
+  organisation, on a table that has an owner column. The ORM makes this easy to
+  write and hard to see.
+- **Migrations that lose or corrupt data** — a column dropped and re-added, a type
+  narrowed without a backfill, a `NOT NULL` added without a default over existing
+  rows.
+- **Indexes the query patterns need and the schema lacks** — a column filtered or
+  joined on in application code with no index behind it. Say which query, not
+  just which column.
+- **Seed and fixture drift** — fixtures that no longer satisfy the constraints the
+  migrations now declare.
+
+## How you work
+
+1. Read the system map for the schema snapshot and the route inventory. Do not
+   rediscover them.
+2. Read the migrations in order. The current schema is the sum of them, and a
+   defect is often visible only in the sequence — a constraint added, then
+   dropped two migrations later to make a deploy pass.
+3. Read the model and query layer and compare its assumptions against what the
+   schema actually declares. The gap between the two is your finding.
+4. Where an environment is available, confirm the behaviour rather than inferring
+   it: insert the row the code believes is impossible, and see whether the
+   database refuses it.
+5. Pin the environment for anything you reproduce, so it runs the same way later.
+
+## What counts as evidence
+
+The schema text, the migration, and the query. A finding that says "this column
+should be indexed" without naming the query that scans it is an opinion. A
+finding that says "this insert succeeds and the model layer says it cannot" with
+the statement and the response is a defect.
+
+Where you could not observe the behaviour — no reachable database, no fixture
+that reaches the path — say so plainly and lower your confidence. An honest
+`unattempted` reproduction is worth more than a confident guess, because the next
+agent will treat your confidence as real.
+
+## What is not yours
+
+The HTTP surface is CONDUIT's, the UI is SURFACE's, and dependency advisories are
+WARDEN's. A cross-tenant read is yours when the defect is in the query, and
+CONDUIT's when the defect is in the missing authorization check. If both are true,
+report the one you can evidence.

qaas_python-0.2.0/src/qaas/prompts/WARDEN.md

@@ -0,0 +1,62 @@
+You are WARDEN, the security and dependency auditor.
+
+## Your domain
+
+The things that let someone do what they should not be able to do. You are the
+agent whose findings carry the most weight and therefore cost the most when they
+are wrong.
+
+Detect:
+
+- **Missing or wrong authorization** — an endpoint that mutates or reads data
+  without checking the caller's role, or that checks authentication and calls it
+  authorization. The presence of an auth dependency is not evidence that access
+  is checked.
+- **Cross-tenant access** — one organisation's data reachable by another's user.
+- **Secrets in the repository** — keys, tokens, passwords and connection strings
+  in source, fixtures, CI config or committed environment files.
+- **Dependencies with known advisories**, and dependencies pinned to a version
+  behind a security release.
+- **Internal detail leaking to a caller** — stack traces, SQL, file paths, library
+  versions in an error response.
+- **Mass assignment** — a handler that accepts fields the client should not
+  control, such as a role, a price, or a status.
+- **Weak or absent rate limiting** on authentication and password-reset paths.
+
+## How you work
+
+1. Read the system map for the route inventory and the role matrix. Do not
+   rediscover them.
+2. Build the endpoint-by-role matrix and look for the holes, rather than reading
+   handlers in file order and hoping to notice.
+3. Where an environment is available, **demonstrate the access** — impersonate the
+   lower-privilege role and make the call. A refusal you predicted and a refusal
+   you observed are different findings.
+4. For dependencies, name the advisory and the version that fixes it.
+
+## The bar for a security finding
+
+**A concrete exploit path, or lower your confidence.** Say which role, which
+endpoint, which field, and what they get. "This endpoint may be missing an
+authorization check" is a note to yourself; "a viewer can POST
+/v1/orders/3/refund and it succeeds" is a finding.
+
+This matters more here than anywhere else in the system. A security finding is
+routed to a restricted project, wakes people up, and is read as urgent. A false
+one spends that credibility, and the next real finding is read more slowly. If
+you cannot evidence it, report it with the confidence it actually deserves and
+say what you could not test.
+
+## Routing
+
+Security findings are routed to a restricted project, and the tracker will
+**refuse** to file one if no restricted project is configured rather than filing
+it somewhere the whole company can read. That refusal is correct; do not work
+around it by relabelling the finding as something else.
+
+## What is not yours
+
+Spec drift and error-shape inconsistency are CONDUIT's unless the leak has a
+security consequence. Schema constraints are VAULT's. A missing index is nobody's
+security problem. When a finding is genuinely both, report the security
+consequence and say which other surface it also touches.

{qaas_python-0.1.0 → qaas_python-0.2.0}/src/qaas/tasks.py

@@ -11,7 +11,7 @@ one repository's directory layout or one app's seeded users works exactly once.
 
 from __future__ import annotations
 
-from qaas.config import SystemConfig
+from qaas.config import AgentSpec, SystemConfig
 from qaas.envelope import DefectEnvelope
 from qaas.target import TargetProfile
 
@@ -218,6 +218,49 @@ fails, misleads, blocks or excludes someone. Do not report what you would have
 designed differently."""
 
 
+def discovery(config: SystemConfig, mode: str, spec: "AgentSpec") -> str:
+    """The task for a discovery agent with no hand-written builder.
+
+    The architecture's claim is that adding an agent needs a prompt file and a
+    YAML file and no Python. That was not true: `_phase_discover` dispatched
+    from a hardcoded dict of builders, so a new discovery agent was silently
+    skipped with `no task builder` -- it validated, it assembled, it appeared in
+    `--dry-run`, and then it did nothing. VAULT and WARDEN were added exactly
+    that way and this is the bug they found.
+
+    What an agent should be told is: which application, what it can reach, and
+    what its own prompt says its domain is. Everything specific to a domain
+    belongs in that agent's prompt, not here -- CONDUIT and SURFACE keep their
+    bespoke builders because they name tools (`diff_openapi`, the browser) that
+    only they have.
+    """
+    p = _profile(config)
+    reach = (
+        "The application is reachable, so prove what you report: observe the "
+        "behaviour and capture the evidence. A finding you have not observed is a "
+        "hypothesis, and its confidence should say so."
+        if p.environment.is_reachable
+        else "There is no reachable instance, so every finding is a reading of the "
+        "code. Quote the lines that support it and keep your confidence honest "
+        "about not having observed the behaviour."
+    )
+    return f"""Audit {p.name} for defects in your domain.
+
+Layout — {p.layout.described()}
+
+Your own instructions define what your domain is and what counts as evidence in
+it. Work within it and leave the other surfaces to the agents that own them.
+
+{reach}
+
+Emit one envelope per distinct defect with `emit_envelope`. Finding nothing is a
+valid outcome; inventing something to report is not. Deduplicate against
+`search_similar` before you emit, so a defect this system already knows about
+comes back as an occurrence rather than a new finding.
+
+Mode: {mode}."""
+
+
 def forge(envelope: DefectEnvelope, config: SystemConfig, flake_runs: int) -> str:
     p = _profile(config)
     evidence = "\n".join(f"  - {e.type.value}: {e.uri} {e.note}".rstrip() for e in envelope.evidence)

{qaas_python-0.1.0 → qaas_python-0.2.0}/tests/test_conductor.py

@@ -543,3 +543,21 @@ async def test_a_resumed_run_keeps_the_budget_it_already_spent(cfg, tmp_path, fa
 
     assert report.stopped_early, "the resumed run ignored what the run had already spent"
     assert "spend cap" in report.stopped_early
+
+
+async def test_a_new_discovery_agent_runs_without_a_hand_written_task(cfg, tmp_path, fake_agents):
+    """The architecture's central claim, finally tested.
+
+    `_phase_discover` used to dispatch from a closed dict of task builders, so a
+    discovery agent that was not in it got `skipped: no task builder` -- it
+    passed `qaas validate`, assembled correctly, appeared in `--dry-run`, and
+    then silently did nothing. VAULT and WARDEN were added as a prompt plus a
+    YAML each and found exactly that.
+
+    A silent skip is the worst shape this failure could take, which is why this
+    test asserts the agent RAN rather than asserting the config loaded.
+    """
+    calls, behaviour = fake_agents
+    report = await make_conductor(cfg, tmp_path).run("nightly")
+    ran = {name for name, _ in calls}
+    assert {"VAULT", "WARDEN"} <= ran, f"a config-only agent was skipped: {sorted(ran)}"

{qaas_python-0.1.0 → qaas_python-0.2.0}/tests/test_config.py

@@ -15,7 +15,12 @@ PROMPTS = REPO / "src" / "qaas" / "prompts"
 
 PHASE_1 = {"CARTOGRAPHER", "CONDUIT", "SURFACE", "FORGE", "CLERK", "PROOF"}
 PHASE_3 = {"MENDER", "ARBITER"}
-ROSTER = PHASE_1 | PHASE_3
+#: Added later, and the point of them is how they were added: a prompt file and a
+#: YAML file each, with no change to conductor, runner, registry or guardrails.
+#: That was the architecture's central claim and it went untested until someone
+#: actually tried it.
+PHASE_2 = {"VAULT", "WARDEN"}
+ROSTER = PHASE_1 | PHASE_3 | PHASE_2
 
 
 @pytest.fixture(scope="module")

{qaas_python-0.1.0 → qaas_python-0.2.0}/tests/test_paths.py

@@ -19,6 +19,9 @@ import pytest
 from support import CONFIG_SEARCH, PACKAGED_CONFIG, PACKAGED_PROMPTS, PACKAGED_SKILLS
 
 from qaas.config import load_config
+
+#: However many ship; the point is that shadowing one does not lose the rest.
+PACKAGED_AGENTS = sorted((PACKAGED_CONFIG / "agents").glob("*.yaml"))
 from qaas.paths import Workspace, find_project, package_root
 
 
@@ -115,7 +118,7 @@ def test_one_agent_file_can_be_shadowed_without_forking_the_rest(tmp_path):
 
     cfg = load_config(search=(tmp_path / "config", *CONFIG_SEARCH))
     assert cfg.agents["MENDER"].max_budget_usd == 9.0, "the override did not win"
-    assert len(cfg.agents) == 8, "shadowing one agent must not drop the others"
+    assert len(cfg.agents) == len(PACKAGED_AGENTS), "shadowing one agent must not drop the others"
     assert cfg.agents["CONDUIT"].max_budget_usd == 3.0, "an untouched agent changed"
 
 

{qaas_python-0.1.0 → qaas_python-0.2.0}/tutorial/01-code-structure.md

@@ -192,14 +192,24 @@ codebase is strict about it:
 | CARTOGRAPHER | map | services, routes, schema, ownership → `system-map.json` |
 | CONDUIT | discovery | API contract drift; ships a failing contract test |
 | SURFACE | discovery | drives the UI through real journeys |
+| VAULT | discovery | schema constraints the code assumes and the database does not enforce |
+| WARDEN | discovery | missing authorization, secrets, vulnerable dependencies, leaked internals |
 | FORGE | triage | reproduces, minimises, measures flake, commits a failing test |
 | CLERK | triage | dedupes, scores severity, routes, files — the only tracker writer |
 | MENDER | remediation | the minimal fix, on a `fix/*` branch |
 | ARBITER | remediation | adversarial review: APPROVE / REQUEST_CHANGES / ESCALATE |
 | PROOF | verify | re-runs the original test → VERIFIED / NOT_FIXED / REGRESSED |
 
-The design names 16. Eight are built; CONDUCTOR is the state machine rather than
-an agent; seven Phase-2 agents are designed and not written.
+The design names 16. Ten are built; CONDUCTOR is the state machine rather than an
+agent; five further discovery specialists are designed and not written.
+
+VAULT and WARDEN are worth noting for *how* they were added: a prompt file and a
+YAML file each, with no Python. That is the architecture's central claim, and it
+was not quite true until they tested it -- `_phase_discover` dispatched from a
+closed dict of task builders, so a new discovery agent validated, assembled,
+appeared in `--dry-run`, and was then silently skipped with `no task builder`.
+There is a generic discovery task now, and a test that asserts a config-only
+agent actually *runs*.
 
 ---
 

qaas_python-0.1.0/target-app/web/dist/assets/index-Cdavr0Eg.css

@@ -1 +0,0 @@
-:root{--ink: #202124;--ink-muted: #5f6368;--surface: #ffffff;--canvas: #f8f9fa;--line: #dadce0;--accent: #1a56b0;--accent-dark: #14428a;--danger: #b3261e;--radius: 6px;--font: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif}*{box-sizing:border-box}html,body{margin:0;padding:0}body{background:var(--canvas);color:var(--ink);font-family:var(--font);font-size:15px;line-height:1.5}#root{min-height:100vh}h1,h2{margin:0;font-weight:600}.app-shell{min-height:100vh;display:flex;flex-direction:column}.app-header{background:var(--surface);border-bottom:1px solid var(--line)}.app-header-inner{max-width:1040px;margin:0 auto;padding:12px 20px;display:flex;align-items:center;gap:24px}.app-brand{font-weight:700;font-size:16px;color:var(--ink)}.app-nav{display:flex;gap:16px;flex:1}.nav-link{color:var(--accent);text-decoration:none;padding:4px 2px;border-bottom:2px solid transparent}.nav-link:hover{text-decoration:underline}.nav-link-active{color:var(--ink);border-bottom-color:var(--accent)}.app-account{display:flex;align-items:center;gap:12px}.app-account-email{color:var(--ink-muted);font-size:13px}.app-account-role{margin-left:8px;padding:1px 6px;border:1px solid var(--line);border-radius:10px;text-transform:uppercase;font-size:11px;letter-spacing:.04em}.app-main{flex:1;width:100%;max-width:1040px;margin:0 auto;padding:24px 20px 48px}.page{display:flex;flex-direction:column;gap:16px}.page-header{display:flex;align-items:center;gap:12px}.page-title{font-size:22px}.section-title{font-size:15px;margin-top:8px}.card{background:var(--surface);border:1px solid var(--line);border-radius:var(--radius);padding:20px;display:flex;flex-direction:column;gap:16px}.muted{color:var(--ink-muted)}.link{color:#15c}.empty-state{align-items:flex-start;gap:8px}.toolbar{display:flex;align-items:flex-end;gap:16px;flex-wrap:wrap}.field{display:flex;flex-direction:column;gap:4px}.field-inline{min-width:180px}.field-grow{flex:1;min-width:180px}.field-narrow{width:110px}.field-label{font-size:13px;font-weight:600;color:var(--ink)}.field-input,.search-input{font:inherit;color:var(--ink);background:var(--surface);border:1px solid var(--line);border-radius:var(--radius);padding:8px 10px;width:100%}.search-input{min-width:260px}.field-input:focus-visible,.search-input:focus-visible,.btn:focus-visible,.table-row-clickable:focus-visible,.nav-link:focus-visible,.link:focus-visible{outline:2px solid var(--accent);outline-offset:2px}::placeholder{color:var(--ink-muted);opacity:1}.item-row{display:flex;align-items:flex-end;gap:12px;flex-wrap:wrap;padding-bottom:12px;border-bottom:1px solid var(--line)}.form-actions{display:flex;gap:12px;align-items:center}.btn{font:inherit;font-weight:600;border-radius:var(--radius);border:1px solid transparent;padding:8px 16px;cursor:pointer;text-decoration:none;display:inline-block}.btn:disabled{cursor:not-allowed;opacity:.6}.btn-primary{background:var(--accent);color:#fff;border-color:var(--accent)}.btn-primary:hover:not(:disabled){background:var(--accent-dark);border-color:var(--accent-dark)}.btn-secondary{background:#f1f3f4;color:#9aa0a6;border-color:var(--line)}.btn-secondary:hover:not(:disabled){background:#e8eaed}.btn-danger{background:var(--danger);color:#fff;border-color:var(--danger)}.table{width:100%;border-collapse:collapse;background:var(--surface);border:1px solid var(--line);border-radius:var(--radius)}.table th,.table td{text-align:left;padding:10px 12px;border-bottom:1px solid var(--line)}.table th{background:var(--canvas);font-size:13px;letter-spacing:.02em}.table tbody tr:last-child td{border-bottom:none}.table-row-clickable{cursor:pointer}.table-row-clickable:hover{background:#f1f6ff}.badge{display:inline-block;padding:2px 8px;border-radius:10px;font-size:12px;font-weight:600;text-transform:capitalize;color:var(--ink);background:#f1f3f4}.badge-draft{background:#eceff1}.badge-placed{background:#e8f0fe}.badge-paid{background:#e6f4ea}.badge-refunded{background:#fef7e0}.badge-cancelled{background:#fce8e6}.alert{margin:0;padding:10px 12px;border-radius:var(--radius);font-size:14px}.alert-error{background:#fce8e6;color:var(--danger);border:1px solid #f3c9c5}.alert-success{background:#e6f4ea;color:#0b6b34;border:1px solid #b7e0c4}.summary{margin:0;display:flex;flex-direction:column;gap:8px}.summary-row{display:flex;gap:12px}.summary-row dt{width:120px;color:var(--ink-muted);font-size:13px}.summary-row dd{margin:0}.total-line{margin:0;text-align:right;font-size:15px}.login-page{min-height:100vh;display:flex;align-items:center;justify-content:center;padding:24px}.login-card{width:100%;max-width:380px}.login-title{font-size:20px}.login-subtitle{margin:0;color:var(--ink-muted);font-size:14px}