toolgraph 0.1.0__tar.gz → 0.3.0__tar.gz

toolgraph-0.3.0/CHANGELOG.md

@@ -0,0 +1,772 @@
+# Changelog
+
+All notable changes to Toolgraph are documented here.
+
+## 0.3.0 - 2026-10-19
+
+Fourth alpha. Read **Upgrading from 0.2.0** below before installing: it lists
+twelve changes an existing setup, checkout or embedding program can notice.
+The one most installations meet first is that there is no implicit backend
+any more: a setup that relied on Neo4j being chosen when nothing named a
+backend now sets `TOOLGRAPH_BACKEND=neo4j` and installs `toolgraph[neo4j]`.
+
+### Upgrading from 0.2.0
+
+- There is no implicit backend any more (ADR-0017). 0.2.0 chose Neo4j at
+  `NEO4J_URI` whenever neither `TOOLGRAPH_BACKEND` nor a `backend` field in
+  the runtime config file named one; now a command that opens the graph
+  exits 1 with ``ERROR: no backend configured: run `toolgraph init`, or set
+  TOOLGRAPH_BACKEND to ladybug or neo4j``, `doctor` reports a Configuration
+  FAIL (`backend_source: none`), and `serve` and `python -m
+  toolgraph.server.app` exit 1 with that line before serving. `NEO4J_*`
+  variables alone no longer select Neo4j. A deployment that relied on the
+  fallback sets `TOOLGRAPH_BACKEND=neo4j` and installs `toolgraph[neo4j]`. A
+  `.env` copied from `.env.example` already sets it; a hand-written `.env`
+  with only `NEO4J_*` lines does not. A `--config` file without `backend`
+  also meant Neo4j; it is now refused with exit 1 and a line naming the file,
+  and the environment is still not consulted.
+  `serve` also refuses at startup, with one line and exit 1, an unknown
+  `TOOLGRAPH_BACKEND`, a backend variable that conflicts with the config file
+  and an unreadable or malformed config file found through
+  `TOOLGRAPH_CONFIG` or the working directory; 0.2.0 started in all four
+  cases and failed each tool call instead. (A malformed `--config` file was
+  already refused before serving, exit 2.) A missing `neo4j` extra or an
+  unreachable backend still surfaces per tool call.
+  Library callers that read `toolgraph.config.settings` with nothing
+  configured get `RuntimeConfigurationError`; `Settings(backend=...)` is
+  unaffected. `config.configure(path)` still reads the file at once but no
+  longer builds the settings: they are built when `config.settings` is next
+  read, so a file without `backend` is refused there, and a `NEO4J_*`
+  variable changed in between is the one used. Read `config.settings` right
+  after `configure()` to keep 0.2.0's timing.
+- The Neo4j driver is now the `neo4j` extra, and Ladybug is a core
+  dependency (ADR-0017). A Neo4j deployment installs `toolgraph[neo4j]`
+  (`uv sync --extra neo4j` from a checkout). Without it, a command that
+  opens a Neo4j graph exits 1 with one line: `ERROR: backend=neo4j requires
+  the 'neo4j' extra: install with ...`. `toolgraph[ladybug]` still resolves,
+  as an empty extra, so existing install recipes keep working. Installing
+  Toolgraph now installs Ladybug on every platform. Ladybug 0.18.3 ships
+  wheels for CPython 3.12–3.14 on macOS 15+ (arm64, x86_64), manylinux
+  (x86_64 glibc 2.27+, aarch64 glibc 2.26+), musllinux 1.2 and Windows
+  (amd64, arm64). Elsewhere, for example on macOS 14, pip builds Ladybug
+  from its source distribution; that build has not been verified.
+- Neo4j 5.7 is now the minimum server version. Toolgraph asks the server
+  not to send `UNRECOGNIZED` notifications, which needs Bolt 5.2; against an
+  older server every command that opens the graph exits 1 with `Neo4j
+  server is older than 5.7, the minimum Toolgraph supports; upgrade the
+  server` (measured on 5.6 and 4.4). Before this change the README
+  walkthrough ran on 5.6, with notification noise, and on 4.4 seven of its
+  eighteen commands already failed with Cypher syntax errors. Library
+  callers get `BackendConfigurationError` from the same seams as below.
+  Tooling that reads stderr no longer sees `Received notification from DBMS
+  server` lines for `UNRECOGNIZED` notifications; warning-level ones, such
+  as deprecations, still appear.
+- `toolgraph crawl` now requires the `servers` key in `servers.yaml`. A
+  blank file, a comment-only file or `{}` exits 1 with `Invalid servers
+  config` before any server is contacted; 0.2.0 treated those three as an
+  empty fleet and, under default pruning, retired every crawled server.
+  (`servers: null`, mapping or scalar values and unknown keys were already
+  rejected; a YAML set, `!!set {}`, is newly rejected instead of being
+  coerced to an empty list.) To retire the whole fleet on purpose, write `servers: []`
+  (library callers: `ServersConfig(servers=[])`); to crawl nothing and keep
+  the graph, add `--no-prune`. See ADR-0014.
+- Crawls behave differently in three smaller ways. A server whose catalog
+  exceeds 10,000 records or 4 MiB of retained record JSON now fails its
+  crawl instead of loading (ADR-0016); raise the limit with
+  `--max-catalog-records` / `--max-catalog-bytes`, or pass
+  `budget=CatalogBudget(...)` from a library call. A failed server's reason,
+  printed by `crawl` and returned in the failures of `crawl_all` /
+  `crawl_all_async`, names the underlying exceptions instead of
+  `ExceptionGroup: unhandled errors in a TaskGroup`. And `crawl_all` /
+  `crawl_all_async` raise `ValueError` for a `concurrency` that is not a
+  positive `int`: `0` used to hang, while a float or `bool` was accepted.
+- Read commands and MCP read tools no longer create a missing Ladybug
+  database: they exit 1 with one `BACKEND_UNAVAILABLE` line (MCP:
+  `backend_unavailable`). A Ladybug file without Toolgraph's schema, or with
+  another schema version, is refused by graph query commands, with a line
+  naming `init-schema` or the version mismatch, and by MCP read tools, as
+  `backend_misconfigured` with cause `schema_not_initialized` or
+  `schema_version_mismatch`; `check` tests connectivity only and still
+  prints `OK` for it. A script that let a read command
+  create the file runs `toolgraph init` or `init-schema` first. Writers
+  (`crawl`, `ingest-manifest`, `init-schema`) still create it.
+- `toolgraph check` and `doctor`'s Connectivity check now look in
+  `NEO4J_DATABASE`. With a database the server does not have, `check` exits
+  1 where 0.2.0 printed `OK`, and Connectivity FAILs where it passed. A
+  health check built on either now reports that setup as broken. Separately,
+  `doctor` and `doctor --json` report exactly one `Configuration` check;
+  when the settings could not be built, 0.2.0 reported a PASS or WARN
+  `Configuration` followed by a FAIL with the same name.
+- MCP tools report backend configuration problems as a new error kind,
+  `backend_misconfigured`, with `retryable: false` and a `cause`, where 0.2.0
+  returned the SDK's generic `Error executing tool <name>`. A client that
+  switches on `error_kind` should handle it; `backend_unavailable` is
+  unchanged. See the first bullet under Changed for the causes.
+- Code that calls `driver.session()`, `acquire_readonly_session()` or
+  `verify_connectivity()` and catches `neo4j.exceptions.AuthError` (or a
+  `ClientError` for expired credentials, the authentication rate limit or
+  `Database.DatabaseNotFound`) now gets
+  `toolgraph.graph.driver.BackendConfigurationError`, with the driver error
+  as `__cause__`; the CLI prints these as one `ERROR:` line, exit 1, instead
+  of a traceback. `BackendConfigurationError` now takes exactly one message
+  and an optional keyword `cause`, so constructing it with no arguments or
+  with two positional arguments raises `TypeError`.
+- More values are refused as credential-shaped (#130). Review notes, reviewer
+  names, review-report run ids and tool keys, every control-plan key and string
+  value, and artifact identities (now including `graph_instance_id` and
+  `graph_state.instance_id`) refuse `name=value` and `name: value` assignments
+  of a token, password, secret, credential, API key or authorization name,
+  `Bearer` values, and credential URIs and queries, also when the name is
+  percent-encoded (`?%74oken=`). A review note such as `token: rotated weekly`
+  now has to be rephrased. `policy propose` reports a refused export value
+  by its input line (`<file>:<line>: tool_key contains credential-shaped
+  data`) instead of its position in the proposal, and refuses it before any
+  other error could quote it. In the other direction, `policy impact` now
+  accepts single-colon names such as `srv::get_token:v2` and
+  `agent-token:blue`, which it refused before, and so does `policy propose`
+  where such a name reaches its output (an observed agent, an export
+  `candidate`, an ambiguous grant or a graph match in a note): 0.2.0 refused
+  the proposal with `additions[0].agent contains credential-shaped data`. An
+  export row's `tool_key` must still be `<server>::<tool>` with no other
+  colon.
+- Artifact evidence hides more around URI credentials (#204, #208, #209).
+  Policy bundles, preflight and control-plan artifacts, policy review plans,
+  `policy impact` and `policy check` output, and `ingest-manifest --dry-run
+  --json` scrub it in three steps. A quote, `%22` or `%27` after a URI's
+  scheme with an `@` anywhere after it hides everything after the scheme,
+  the rest of the evidence line included. Then a URI that cannot be parsed
+  (an unclosed `[`, an invalid bracketed host) stops the command as in
+  0.2.0, now with `artifact evidence holds a URI that cannot be parsed; it
+  is not shown` instead of the parser's message, which could quote the host
+  with its password. Then a `)` with an `@` after it in the same unbroken run
+  hides the rest of that run. So `https://u:"pw"@[::1/r` is shown as
+  `https://<redacted-uri>`, while `https://u:p[)ss@h/r` stops the command.
+  The hiding rules also hide values that hold no password. A consumer that
+  matches evidence paths should expect `<scheme>://<redacted-uri>`.
+- Error messages that go through Toolgraph's shared scrubber hide more (#201,
+  #202): CLI errors, crawl failures and the server's
+  backend-misconfiguration log record. A message drops what follows a quoted
+  or bracketed credential value; one whose scrubbed or percent-decoded form
+  is still credential-shaped reads `details withheld (they contain
+  credential-shaped data)` after the command's own context; and a
+  `<redacted` that the input itself contains is shown as `‹redacted`
+  (U+2039). CLI usage errors, `--dry-run` JSON warnings, MCP SDK errors
+  outside Toolgraph's own envelopes and logged tracebacks do not go through
+  the scrubber (#206).
+
+### Added
+
+- **A per-server catalog budget for crawls** (ADR-0016). Each server's
+  retained tools and resources are capped at 10,000 records and 4 MiB of
+  record JSON (UTF-8). A server over budget fails like any other crawl
+  failure, so no partial catalog is loaded and ADR-0007's failure
+  protection applies. Configure it with `toolgraph crawl
+  --max-catalog-bytes N --max-catalog-records N`, or `CatalogBudget` for
+  `crawl_server` / `crawl_all`. It bounds retained data, not transport
+  buffering or peak memory. Crawl failure messages now show the underlying
+  error instead of `ExceptionGroup: unhandled errors in a TaskGroup`.
+
+- **Read-only policy impact previews**: `toolgraph policy impact proposed.yaml
+  SERVER::TOOL... --agent AGENT` compares current and hypothetical selector
+  decisions with evidence. JSON reports bind graph state, governance digests,
+  profile and scoped catalog; `toolgraph policy check impact.json` checks the
+  proposed-side regression expectations against current state. Neither command
+  applies policy. See [the impact guide](docs/policy-impact.md).
+
+- **A first-run notebook**, `docs/notebooks/01-first-policy-bundle.ipynb`,
+  walks the Beginner Guide path one command at a time inside a temporary
+  directory. It goes from the quickstart to a compiled bundle, then removes and
+  restores an evidence pointer and runs the audits. It is committed with its
+  outputs. `tests/test_notebooks.py` checks those outputs row by row, and a new
+  CI `notebooks` job re-executes the notebook from a wheel install. The job
+  and the regeneration recipe install `nbclient` and `ipykernel` from a new
+  `notebook` dependency group in `uv.lock`, so Dependabot updates them.
+
+### Changed
+
+- MCP tools report a backend configuration problem as a typed result
+  instead of the SDK's generic `Error executing tool <name>`. Missing
+  `neo4j` extra, rejected Neo4j credentials, a missing `NEO4J_DATABASE`, a
+  Neo4j server older than 5.7, a missing Ladybug, a database file that is
+  not a Ladybug database and, on Ladybug, a graph without Toolgraph's schema
+  or with another schema version now return `isError: true` with
+  `structuredContent`
+  `{"error_kind": "backend_misconfigured", "retryable": false, "cause": ...,
+  "message": ...}`. The message is fixed per cause; the path or status code
+  goes to the server log as one WARNING record instead of a traceback. On
+  stdio, where the SDK's log handler wraps at 80 columns, a corrupt Ladybug
+  file went from 284 stderr lines to 9. A consumer that keys on
+  `error_kind == "backend_unavailable"` sees no change: these failures still
+  are not outages.
+
+- Toolgraph no longer chooses a backend when nothing names one (ADR-0017
+  step 3). The backend comes from `TOOLGRAPH_BACKEND` or the runtime config
+  file's `backend` field, or with `--config` from that file alone; otherwise
+  commands that open the graph, `serve` and the MCP server module refuse with
+  one line, and `doctor` reports a Configuration FAIL. Commands that do not
+  open the graph (help, `--version`, `example init`, `review-candidates`,
+  `init`) are unaffected.
+
+- Ladybug moves into the core dependencies and `neo4j` to a `[neo4j]`
+  extra (ADR-0017 step 2). No Ladybug path imports the Neo4j driver, so a
+  Ladybug install runs without it: a new required CI job, `no-extras`,
+  installs the wheel with no extras and checks that. Every Neo4j path
+  imports the driver through one guarded loader, so a missing extra is a
+  configuration error naming `toolgraph[neo4j]` rather than an
+  `ImportError`.
+
+- `toolgraph doctor` and `doctor --json` report exactly one
+  `Configuration` check. When the settings cannot be built (for example a
+  `TOOLGRAPH_BACKEND` that contradicts the config file), that check FAILs
+  with the error and keeps the file details such as `backend_source`;
+  previously a passing `Configuration` check was followed by a second,
+  failing one with the same name.
+
+- `servers.yaml` must declare its fleet explicitly (ADR-0014). The
+  `servers` key is required and must be a list; `servers: []` is the only
+  way to declare an empty fleet. A YAML set (`!!set {}`) is rejected too:
+  the field is strict, so it is not coerced to an empty list. The CLI
+  diagnostic, exit code and `--prune/--no-prune` semantics are unchanged.
+  ADR-0007's "empty file = empty fleet" consequence is amended accordingly.
+
+- The fleet duplicate checks (`duplicate_name_overrides`,
+  `duplicate_identities`) count names in one pass, so their cost grows
+  linearly with fleet size instead of quadratically. Results are unchanged.
+  `scripts/benchmark_duplicate_checks.py` is a repeatable size sweep; on one
+  developer machine an 8,000-server fixture went from about 300 ms to well
+  under a millisecond per check.
+
+- `SECURITY.md` now states that security fixes cover the latest published
+  Toolgraph release only, with no backports, instead of naming the `0.0.x`
+  line the policy was written for.
+
+### Fixed
+
+- Evidence in artifacts no longer shows a quoted URI password (#204). Policy
+  bundles, preflight and control-plan artifacts, policy review plans, policy
+  impact and `policy check` output, and `ingest-manifest --dry-run --json`
+  remove the userinfo the URI parser finds and keep the path of an ordinary
+  URI. A password in quotes (`neo4j://u:"p w"@host/db`) was kept whole when
+  it held a space or `)`, and partly when it held `@` and `/`. Resource URIs
+  read from the graph carry the quotes percent-encoded, and some of those
+  leaked too: `u:%22p@x/y%22@host` kept `x/y%22`. When a quote, `%22` or `%27`
+  follows a URI's scheme and an `@` follows it, everything after the scheme
+  is now hidden as `<redacted-uri>`, path and the rest of the evidence line
+  included, since where such a password ends cannot be found reliably. This
+  also hides values that hold no password, such as `https://host/a%22b@c`.
+  Other values are redacted as before; a password the parser cannot bound,
+  such as one holding an unquoted `/`, can still show (#212). Scanning a
+  long line of letters is also linear now: before this fix a
+  100,000-character line took 4.9 s.
+
+- Evidence in artifacts no longer shows a URI password holding an unquoted
+  `)` (#208). Evidence wraps a resource in graph notation,
+  `(tool) -READS-> (uri)`, so the scan stopped at the first `)` and kept
+  `neo4j://u:p)ss@host/r` whole. When a `)` follows a URI's scheme and an
+  `@` follows it with no whitespace between, the rest of that run is now
+  hidden as `<redacted-uri>`, the resource's closing `)` included. This also
+  hides values that hold no password, such as `file:///a(b)@c`.
+
+- Commands that write artifacts refuse evidence holding a URI that cannot be
+  parsed (#209), such as one with an unclosed `[` (`https://u:p@[::1/r`), an
+  invalid bracketed host or a character invalid under NFKC normalization.
+  `ingest-manifest --dry-run --json` failed with a traceback, and
+  `preflight` and `policy compile` printed the parser's message, which can
+  quote the host with its password (`netloc 'u:pw@h＃x' contains invalid
+  characters ...`). They now print `ERROR: artifact evidence holds a URI that
+  cannot be parsed; it is not shown`, exit 1 and write nothing. The rest of
+  the artifact is withheld as before: printing it around a hidden URI could
+  show other values the scan cannot bound.
+
+- An error message that would still show a credential after scrubbing is
+  withheld (#202). Scrubbing cannot recognise every credential, for example
+  a percent-encoded name such as `?%74oken=x` quoted in a validation error
+  about an observed export, a selector profile or a governance key. Such a
+  message now reads `details withheld (they contain credential-shaped
+  data)` after the command's own context, where it used to print the
+  value. A message whose percent-decoded form is credential-shaped is
+  withheld before scrubbing, since the scrubber does not decode: this also
+  withholds some messages it could have cleaned, such as an encoded URL with
+  a `?token=` query. A placeholder that the input itself contains is shown
+  as `‹redacted` (U+2039), so it cannot pass for one the scrubber inserted.
+  This covers every message that goes through the shared scrubber:
+  CLI errors, crawl failures and the server's backend-misconfiguration log
+  record. CLI usage errors, `--dry-run` JSON warnings, MCP SDK errors
+  outside Toolgraph's own envelopes and logged tracebacks do not go through
+  it yet (#206). Checking for credential URIs is also linear now: a long run of
+  `a:` with no `@` took seconds.
+
+- Error text no longer shows part of a quoted credential (#201). A password
+  in quotes inside a URI (`neo4j://u:"p@ss"@host`) was printed whole, and a
+  quoted value with a space (`password="hunter 2"`, `NEO4J_PASSWORD="s3cr3t
+  pass"`) kept its tail. A quoted or bracketed credential value, and a URI
+  with a quote in its authority and an `@` anywhere after it, are now hidden
+  through the end of the message: the message loses what follows them, and
+  never shows part of the value. A URI that is only wrapped in quotes, such
+  as `'neo4j://host:7687'` in a driver error, is shown as before unless an
+  `@` follows later in the message. Scrubbing a long message is also linear
+  now; before this fix a 100,000-character line of letters took 25 s.
+
+- Values that Toolgraph refuses rather than scrubs go through one rule,
+  `toolgraph.redaction.is_credential_shaped`, instead of five detectors that
+  disagreed (#130). Review notes accepted `Bearer abc123`, `"token"=abc` and
+  `mycredential=x`; control plans accepted `password=hunter2` and
+  credential-shaped keys, and echoed a forbidden or duplicate key in the
+  error; review metadata and graph instance ids were not checked; percent-encoded
+  query keys passed everywhere but `policy impact`. Errors name where a refused
+  value sits, never the value.
+
+- Commands and MCP tools that only read the graph no longer create a missing
+  Ladybug database. Pointed at a path that was never initialized, the read
+  commands (`check-access`, `unmapped-tools`, `audit-report`, `drift` and
+  the rest) created the file, then failed with a 72-98 line `Binder
+  exception: Table ... does not exist` traceback, and `check` printed `OK`
+  for that path from then on. Now a missing file gives one line, `ERROR:
+  backend unavailable: BACKEND_UNAVAILABLE: Ladybug database <path> does
+  not exist or is not a file`, and MCP returns `backend_unavailable`, as
+  `check` already did. A Ladybug file without Toolgraph's schema gives
+  `ERROR: backend schema is not initialized or incomplete — run 'toolgraph
+  init-schema'`, and MCP returns `backend_misconfigured` with cause
+  `schema_not_initialized` (`schema_version_mismatch` for another schema
+  version). A read that opened the database and fails this way closes it,
+  so `init-schema` from another process can repair it while `serve` runs
+  (a process that already wrote keeps its handle).
+  Neo4j is unchanged: an uninitialized Neo4j database still answers reads
+  with empty results.
+
+- `toolgraph check` and `doctor`'s Connectivity check now select
+  `NEO4J_DATABASE`. With a database the Neo4j server does not have, `check`
+  printed `OK` and the next graph command exited 1 with a
+  `DatabaseNotFound` traceback; now `check` and every graph command exit 1
+  with `ERROR: Neo4j database named by NEO4J_DATABASE does not exist on the
+  server`, and `doctor` reports that line as a Connectivity FAIL
+  instead of a Connectivity PASS followed by two raw driver errors.
+  Connectivity is checked with a read query in the configured database,
+  in READ access mode like the driver's own check (measured on Neo4j 5.7
+  and 5.26).
+
+- README's exit-code section no longer gives a rejected Neo4j password as an
+  error that ends in a Python traceback; since the authentication change
+  above it prints one `ERROR:` line. The exit-code table lists rejected
+  credentials, a missing Neo4j database and a server older than 5.7 among
+  the one-line errors, with the `init` and `serve` exceptions spelled out
+  for every class.
+
+- Neo4j no longer floods stderr with driver notifications. Every query that
+  named a label, property or relationship type the graph did not hold yet
+  logged a `Received notification from DBMS server` line, so a fresh
+  database printed eight before `doctor`'s diagnosis and a README walkthrough
+  printed 92. The driver now disables the `UNRECOGNIZED` classification.
+  Warning-level notifications, such as deprecations, still reach stderr;
+  performance notifications are still logged at INFO. Requires Neo4j 5.7
+  (see Upgrading).
+
+- `.env.example` no longer presents `TOOLGRAPH_SERVERS=servers.yaml` and
+  `TOOLGRAPH_GOVERNANCE=governance.yaml` as configured paths. Those values
+  are the built-in defaults, and the repository root has neither file, so
+  after `cp .env.example .env` a bare `toolgraph crawl` still exits 1 with
+  `Invalid servers config at servers.yaml` (`ingest-manifest`: `Invalid
+  governance config at governance.yaml`). The lines are now commented out
+  with a note to pass `--servers examples/servers.yaml` explicitly. They are
+  not pointed at `examples/`: a bare `crawl` would then reconcile the graph
+  against the example fleet and retire every other server.
+
+- A Neo4j that rejects the credentials no longer makes CLI commands print a
+  native traceback. Wrong, expired or rate-limited credentials and an
+  expired auth-provider token are typed at the driver seam as a
+  configuration error and reported as one line, `ERROR: Neo4j refused
+  authentication (<Neo4j status code>); check NEO4J_USER, NEO4J_PASSWORD
+  and the account's auth provider`, exit 1. The driver's own message is not
+  echoed. It is not treated as a retryable outage, and `doctor`'s schema
+  check keeps that message instead of rewording it as a catalog permission
+  problem. Query-level `Forbidden` errors are unchanged.
+
+- `crawl_all` and `crawl_all_async` now reject a `concurrency` that is not a
+  positive integer (`0`, negatives, `bool`, floats, strings, `None`) with a
+  `ValueError` before any server is contacted. Previously `concurrency=0` made
+  every crawl of a non-empty fleet wait forever, because the per-server
+  timeout only starts once the semaphore is held. The CLI always used the
+  default of 5 and is unaffected.
+
+## 0.2.0 - 2026-10-15
+
+Third alpha. Read **Upgrading from 0.1.0** below before installing: it lists
+seven changes an existing setup, checkout or embedding program can notice.
+The one most installations meet first is that commands now refuse a backend
+variable that disagrees with the config file.
+
+### Upgrading from 0.1.0
+
+- Commands stop with `ERROR: conflicting backend configuration (...)` when
+  `TOOLGRAPH_BACKEND` or `TOOLGRAPH_DB_PATH` disagrees with the config file;
+  0.1.0 silently let the variable win. Pass `toolgraph --config <file>
+  <command>`, or unset the variable; when it comes from `.env`, remove it
+  there, since the next command loads `.env` again. See the first bullet
+  under Changed.
+- `control-preflight` can now return `advisory_warn` for a plan that 0.1.0
+  answered with `advisory_allow`. See the `control-preflight` bullet under
+  Changed before relying on `advisory_allow` as a gate.
+- A program that imports Toolgraph no longer gets `./.env` loaded for it.
+- `toolgraph check` no longer creates a missing Ladybug database; it exits 1.
+  A script that used `check` to initialize should run `toolgraph init` or
+  `init-schema` first. See the `check` bullet under Changed.
+- Identities, control-plan strings and review notes that 0.1.0 accepted can
+  now be refused as credential-shaped (for example `?access_token=` or `?credential=`). See
+  the credential bullet under Changed.
+- `GraphState.token()` and `queries.with_generation()` are gone; use the
+  `graph_state` token results carry and `queries.with_graph_state()`. See
+  Removed.
+- The repository's `examples/policy_propose_from_observed.py`, which 0.1.0
+  introduced, is gone; use `toolgraph policy propose`, which reads the live
+  graph. See its entry under Removed for the provenance flags.
+
+### Added
+
+- **`toolgraph doctor [--json]`** checks the configuration file (present,
+  valid JSON, mode 0600, and `backend_source`: which of `--config`, the
+  environment, the config file or the built-in default decided the backend),
+  backend connectivity, the physical schema and the graph state. It opens the
+  database read-only — on Neo4j in READ access mode — and redacts what it
+  reports. It exits 1 when any check fails; a warning alone exits 0.
+- **`toolgraph ingest-manifest --dry-run [--json]`** reports the full delta a
+  manifest would apply without changing the database: agents, policies,
+  grants and `granted_by`, data access, `GOVERNED_BY`, expected exceptions,
+  provenance, and the resources and tools that would be pruned for having no
+  edges left. It exits 1 when the manifest would be rejected; with `--json` it
+  then prints a rejection report with the warnings, not a delta (the delta
+  lists stay empty because no comparison was made). A missing database file is refused and none is
+  created; an existing but uninitialized one is refused with `init-schema`
+  guidance. On Neo4j the preview
+  session opens in READ access mode, so the server rejects an auto-commit write
+  issued through it (verified on Neo4j 5.26 over `bolt://`); that is an access
+  mode, not access control, and does not replace a read-only role. `--json`
+  requires `--dry-run`. Plain `ingest-manifest` behaves as before.
+- **`toolgraph policy propose GOVERNANCE OBSERVED`** proposes grant changes to
+  `governance.yaml` from an observation-window JSONL export
+  (`contracts/observation-window.schema.json`), resolving grants against the
+  live graph. It never writes to the graph and never edits the manifest. The
+  export's envelope must match the manifest and the live graph;
+  `--revalidate-current` accepts a moved generation and re-checks additions
+  against the current strict catalog, and `--unverified-provenance` accepts a
+  missing envelope or identity and labels the result unverified. Other
+  options: `--min-would-block` (default 3), `--window-label`, `--json`, and
+  `-o/--output`, which cannot be combined with `--json` and writes the file
+  atomically with mode 0600 (a failure before the replacement leaves an
+  existing file as it was). It refuses, with a redacted `ERROR:` diagnostic
+  and exit 1 rather than a traceback: an export over 50,000,000 bytes,
+  malformed YAML, and any proposal row it would emit whose agent, tool or note
+  is credential-shaped — that error names the position (`additions[0].tool`),
+  never the value. See
+  `docs/policy-propose.md`.
+- **`control-preflight` reports potential exposures.** Each record in the new
+  `potential_exposures` field joins an eligible reader tool on one plan node to
+  an eligible writer tool on a distinct node downstream of it, with the
+  resources involved (redacted) and a shortest witness path. It marks a
+  potential path, not proof that data moved. The search is bounded: at most 100 records,
+  64,000 bytes of records and 65,536 traversal steps; when it stops early the
+  result also carries `truncated: true` and `total_candidate_pairs_found`,
+  which is then a lower bound. The fields are additive to the v1 schema. See
+  `docs/adr/0013-bounded-control-plan-preflight.md`.
+- `toolgraph audit-report` JSON now carries the graph state it was read at
+  (`graph_generation`, `graph_instance_id`, `graph_state`,
+  `graph_state_verified`), as the MCP `audit_report` tool already did. The
+  Markdown format is unchanged.
+- `toolgraph example init` now writes a `README.md` into the generated
+  directory. `docs/` is excluded from the sdist, so a user who installed from
+  PyPI previously landed in a directory of three unexplained files with only
+  the CLI's one-line hint to go on. The file lists the commands in order,
+  states the expected outcome, and links the full guide.
+- `examples/README.md` maps each `servers-*.yaml` / `governance-*.yaml` input
+  to the script or workflow that consumes it, and to what it needs to run.
+- A `Documentation` URL in the project metadata, so the beginner guide is
+  reachable from the PyPI sidebar.
+
+### Changed
+
+- **An environment variable can no longer silently override `toolgraph
+  init`'s config.** Without `--config`, `TOOLGRAPH_BACKEND` and
+  `TOOLGRAPH_DB_PATH` used to win over `.toolgraph/config.json` (or the file
+  `TOOLGRAPH_CONFIG` names). Copying `.env.example` (which sets
+  `TOOLGRAPH_BACKEND=neo4j`) into a directory where `toolgraph init` had
+  written a Ladybug config sent every command to Neo4j with no warning. When
+  the two disagree on the backend or on the resolved database path, commands
+  now stop with `ERROR: conflicting backend configuration (...)`. To migrate,
+  pass `--config <file>` to choose the file, or unset the variable (remove it
+  from `.env` if it is set there). Agreeing
+  values, or only one of the two being set, behave as before. `--help`,
+  `toolgraph init` and `toolgraph example init` are not blocked, so `init` can
+  still be used to rewrite the file. `toolgraph doctor` reports which input
+  decides the backend as `backend_source`.
+- **Only the CLI reads `./.env`.** Importing `toolgraph` as a library used to
+  load `.env` from the current directory into `os.environ`. The CLI, including
+  `toolgraph serve`, still does; a program that embeds Toolgraph and relied on
+  that should call `toolgraph.config.load_dotenv_file()` or set its own
+  environment.
+- **`toolgraph check` no longer creates a database.** It opens the configured
+  Ladybug file read-only, so a mistyped `db_path` exits 1 with
+  `BACKEND_UNAVAILABLE ... does not exist` instead of creating an empty
+  database and printing `OK`. Run `toolgraph init` or `init-schema` first.
+- **`control-preflight` can now warn where 0.1.0 allowed.** The decision is
+  `advisory_warn`, not `advisory_allow`, when the plan has a potential
+  exposure (see Added) or the exposure search was truncated. The exit code is
+  still 0 for both. A consumer that gates on `advisory_allow` will see more
+  warnings and should review them rather than treat them as failures of the
+  tool. A result without the exposure fields, from 0.1.0, means "not
+  checked", not "none found"; `total_candidate_pairs_found` on a truncated
+  result is a lower bound. ADR-0013 states the decision rules.
+- **Credential checks are shared and refuse more.** Agent and tool identities
+  (preflight, `policy compile`, `policy review-plan`, `policy propose`) and
+  strings in control plans are now also refused when they carry
+  `?access_token=` or `?credential=`, a password URI whose username is an
+  email address, or an `@` in a query after a port
+  (`https://user@host:443?x=a@b`). `review-candidates annotate` notes that
+  assign `secret=` or `access_token=` are now refused too. Error text
+  now also hides `credential=...` and a bare `Bearer <token>`. Nothing that
+  was refused or hidden before is accepted or shown now.
+- Control-plan validation errors give the location and the reason, never the
+  rejected value: `invalid control plan at nodes.0.principals.1: ...`, and a
+  bad candidate tool is reported by its index.
+- A file that is not a Ladybug database is now a non-retryable
+  `BACKEND_CORRUPT` configuration error, printed as one `ERROR:` line. In
+  0.1.0 it surfaced as a Python traceback from Ladybug.
+- `config.configure()` closes the cached backend connection, so a library
+  caller switching configs no longer keeps talking to the previous database.
+- Deny-path rows (`check-access` evidence, `unsafe-tools`, `audit-report`,
+  selector deny paths) now have a fixed order when rows used to tie: one
+  tool under several DENY policies, one tool name on several servers, or a
+  READS and a WRITES edge to the same resource. Ties are broken by tool key,
+  policy and mode; before, the backend chose, and Ladybug and Neo4j could
+  disagree. Rows that did not tie keep their order.
+
+### Removed
+
+- **`examples/policy_propose_from_observed.py`, the offline proposal script.**
+  There is no longer a graph-free way to produce a proposal. Its replacement,
+  `toolgraph policy propose` (see Added), applies the same row and proposal
+  rules but resolves grants against the live graph, and checks the envelope
+  against it too: an envelope the script accepted is refused when it lacks a
+  graph instance id, or when its generation, digest or instance differs from
+  the live graph. Re-export from the current graph, or pass
+  `--revalidate-current` (a moved generation) or `--unverified-provenance`
+  (missing identity) to proceed. The script was repository-only and never
+  shipped in the distribution. Its tests now drive `toolgraph.policy_propose`.
+- `GraphState.token()`, which nothing called and which returned a different
+  set of keys than the `graph_state` token that results actually carry.
+- `toolgraph.graph.queries.with_generation()`, which had no production
+  caller; use `with_graph_state()`. A strict read that sees the graph state
+  change on every retry now raises `queries.GraphStateChurnError`, a
+  `RuntimeError` subclass with the same message, so CLI output and exit codes
+  are unchanged.
+
+### Fixed
+
+- Backend configuration errors print one `ERROR:` line and exit 1 instead of
+  a Python traceback: an unknown `TOOLGRAPH_BACKEND`, a missing `ladybug`
+  extra, and a schema-version mismatch. `toolgraph doctor` is the exception
+  by design: it reports the same failure as a FAIL check (exit 1) in its text
+  or JSON report.
+- An outage or other backend error while reading the schema version is no
+  longer taken for a fresh database, so `init`, `init-schema`, `crawl` and
+  `ingest-manifest` stop before running any DDL.
+- `toolgraph check` no longer prints credentials embedded in `NEO4J_URI`, and
+  `ingest-manifest` warnings and notices are redacted.
+- `toolgraph init` with an unusable `--state-dir` prints `ERROR: cannot
+  initialize ...` and exits 1 instead of a traceback raised from its cleanup;
+  a partial database file it cannot remove is named in a warning.
+- `policy review-plan` refuses a graph with no instance id instead of writing
+  a plan its own schema rejects.
+- `policy compile` (and `build_policy_bundle()`) refuses a graph state whose
+  generation is not a non-negative integer — a boolean, float, string, null
+  or negative value — instead of writing it into the bundle. Toolgraph's own
+  graph always yields one; the check matters for a hand-edited graph or a
+  library caller's own graph reader.
+- Writing a private artifact — `init`'s config file, `preflight --out`,
+  `control-preflight --out`, `policy compile --output`, `policy review-plan
+  --output` — no longer reports failure when the new file was already in
+  place and only re-applying its 0600 mode afterwards failed. The file is
+  created 0600 before it is written, so that second step was redundant.
+- `crawl` no longer prints `vNone` for a server that reports no version.
+- The packaged quickstart governance manifest now carries `provenance` on
+  every authored edge. The beginner guide claimed every authored edge had an
+  evidence pointer and then told the reader to run `unbacked-edges`, which
+  listed all three of them. The audit now comes back empty, and both the
+  manifest and the guide say what that does and does not prove: the demo tools
+  return strings and touch no file, so the READS/WRITES effects are synthetic
+  and their evidence cites the fixture that defines them. An empty audit means
+  every edge has a pointer, not that a claim is true. As a result `read_note`'s
+  risk in the quickstart drops from 0.4 to 0.0 and the compiled bundle's
+  digest changes.
+- Beginner guide corrections (both languages): the shared-backend section
+  needs a repository clone, since `docker-compose.yml` and `.env.example` ship
+  with neither the package nor the generated quickstart; `policy compile`
+  prints metadata describing the bundle, not the bundle; and the gateway step
+  is labelled optional, matching the README's statement that memtomem-stm is
+  not a required dependency.
+
+### Dependencies
+
+- The `pyjwt` security floor rises from `>=2.13.0` to `>=2.15.0`. 2.13.0
+  carries 14 advisories published after it was set: PYSEC-2026-4140 through
+  PYSEC-2026-4152 and PYSEC-2026-4183. PYSEC-2026-4141 and PYSEC-2026-4183
+  are fixed in 2.15.0, PYSEC-2026-4146 lists 2.13.0 as its last affected
+  version, and the rest are fixed in 2.14.0. The floor is what keeps an
+  install from PyPI from resolving an older version; the lock already resolves
+  a newer one. Toolgraph does not import `pyjwt` itself; it is pinned for the
+  MCP SDK's dependency tree.
+
+### Testing
+
+- The `sse` crawl transport and auth-header forwarding on both HTTP
+  transports are now exercised end to end, at the lockfile version and at the
+  declared MCP SDK floor. Neither had been run before; no runtime behaviour
+  changed.
+
+## 0.1.0 - 2026-09-10
+
+Second alpha. Upgrading from `0.0.1` takes three things: the MCP SDK 2.x, one
+new field for anything caching MCP responses, and — if you generated a
+quickstart with `0.0.1` — a regenerated server script. Read **Upgrading from
+0.0.1** below before installing.
+
+### Upgrading from 0.0.1
+
+`toolgraph example init` copies a quickstart out of the package, so a
+quickstart generated by `0.0.1` holds its own copy of the demo MCP server —
+one that imports `mcp.server.fastmcp`. Upgrading that environment installs MCP
+2.x and the copied server stops importing, so the next `toolgraph crawl`
+fails. Updating the package does not touch a directory it already wrote, and
+`example init` refuses an existing destination on purpose.
+
+To migrate one, generate a fresh quickstart beside it and move your own work
+across:
+
+```bash
+toolgraph example init quickstart-0.1.0
+# Keep your authored governance and your database; take the new server script.
+cp quickstart/governance.yaml quickstart-0.1.0/governance.yaml
+cp -R quickstart/.toolgraph   quickstart-0.1.0/.toolgraph
+```
+
+`servers.yaml` records the interpreter that generated it, so regenerating from
+the upgraded environment already points at the right one; check it only if you
+generated from somewhere else. Re-run `toolgraph crawl`, then
+`toolgraph ingest-manifest`, to confirm.
+
+Verified end to end against the published `0.0.1`: a quickstart it generated
+crawls cleanly on `0.0.1`, fails on the next crawl once the environment holds
+`0.1.0`, and returns the same eligibility verdicts after these steps.
+
+### Added
+
+- `examples/policy_propose_from_observed.py` reads a window of observed gateway
+  usage and prints a reviewable `governance.yaml` grant proposal. It never
+  writes governance. Repository-only — it is not part of the distribution — and
+  deliberately narrow: it requires an envelope declaring the governance digest
+  it was recorded against and refuses a mismatch, refuses an export with no
+  envelope unless `--unverified-provenance` is given, proposes a removal only
+  from an explicit zero-call row, and accepts only fully qualified timestamps
+  with an explicit UTC offset. Digest equality shows what the export declared,
+  not where its rows came from; issue #89 tracks what a real binding needs.
+
+### Changed
+
+- **Breaking:** requires the MCP SDK 2.x (`mcp>=2.1.1,<3`). Environments pinned
+  to `mcp` 1.x can no longer install this package, because 2.x removed
+  `mcp.server.fastmcp` and renamed every model field from camelCase to
+  snake_case. The MCP *wire* format is unchanged, so a client that only speaks
+  the protocol needs no change; that is not the same as no change for anything
+  written against the Python SDK, caching responses, or reading exception text.
+- An unexpected exception raised by a toolgraph tool body no longer puts its
+  message on the wire. Only errors raised deliberately as `ContractError` (for
+  example an unknown selector profile) are forwarded; the rest keep the SDK's
+  generic text, so an internal failure cannot disclose filesystem paths, query
+  bodies or credential-shaped values. Argument-validation failures still echo
+  the caller's own input, as the MCP SDK intends.
+- **Breaking:** an MCP response whose read could not be bracketed by a stable
+  graph state now carries `graph_generation: null` and
+  `graph_state_verified: false` instead of the generation read afterwards.
+  Labelling such a result with a generation it was never read at let a verdict
+  from before a permission change be cached under the generation that changed
+  it, while the tool documentation tells callers to cache per generation.
+  Successful reads gain `graph_state_verified: true`, so the absence of a
+  guarantee is never inferred from a missing field. Null is not self-enforcing
+  — `None` is a usable cache key — so the obligation to skip the cache and to
+  attribute nothing to a graph state is the consumer's; ADR-0004 and the README
+  state it, and a consumer that turns eligibility into an authorization
+  decision should reject or retry such a response.
+
+## 0.0.1 - 2026-09-09
+
+Initial public alpha release.
+
+- Crawl MCP servers into a graph and combine factual capability data with
+  operator-authored governance.
+- Explain access, reachability, drift, blast radius, and tool selection.
+- Compile deterministic policy bundles for an independent runtime gateway.
+- Support local Ladybug and shared Neo4j backends.
+- Reject unknown manifest fields and duplicate YAML keys before graph mutation.
+- Persist only credential-free endpoint labels and redact failure diagnostics.
+- Ship an install-only, offline quickstart through `toolgraph example init`.
+- Give MCP consumers a typed `backend_unavailable` error envelope for transient
+  graph outages while keeping contract and internal errors fail-loud. Outages
+  are typed at the backend seam, and `retryable` follows the tool's advertised
+  read-only annotation instead of being assumed, failing closed to `false`
+  whenever that annotation cannot be read.
+- Add a body-free `toolgraph.control-plan` v1 contract and advisory
+  `control-preflight` command that lints bounded DAG topology and evaluates all
+  principal/tool surfaces under one collision-safe graph-state bracket.
+
+### Stabilization (2026-09-08)
+
+- Resolve ingest references in bounded batches shared with the selector,
+  preserving atomic application, diagnostic order and ambiguous identities.
+- Require explicit commit SHAs for ecosystem smokes and clean matching HEADs
+  for gateway smokes; record dependencies and artifact evidence.
+- Include Ladybug in `make install` and refresh the pip audit dependency lock.
+- Reconcile historical integration and selection roadmaps with shipped code.
+
+### Implementation review fixes (2026-09-07)
+
+- Handle empty crawl and manifest collections on both backends while retaining
+  stale-edge reconciliation, revocation, and transactional generation updates.
+- Preserve opaque artifact identities, including ordinary URI queries/fragments.
+  Reject credential-shaped identities and redact only evidence fields, including
+  policy review paths. Existing v1 schemas and decision vocabularies are unchanged.
+- Use collision-safe `v2:` exception keys and deterministic last-entry-wins
+  metadata for duplicate logical exceptions. No physical schema migration is
+  required: existing keys remain readable. Re-ingest the original governance
+  manifest to rebuild keys and restore entries lost to old key collisions, then
+  recompile policy bundles. Old artifacts cannot reconstruct lost manifest rows.
+- Require control-plan fields exactly as published: missing `maximum_cycles`,
+  explicit null bounds, string/boolean coercions are rejected. JSON integer
+  numbers such as `2.0` remain accepted; additive same-major fields remain valid.
+- Compile catalogs in chunks of at most 4,096 candidates under one strict state
+  bracket; reject incomplete or conflicting evaluations. Injected readers own
+  all state reads, and adapters without the optional `evaluate` capability work.
+- Load runtime settings lazily so help/version and explicit configuration remain
+  usable when the default config is broken. Invalid selected configs fail clearly.
+- Reject naive and sub-minute timestamps before preflight/review I/O; development
+  validation now installs and checks the RFC 3339 format validator.
+
+### Upgrade note for pre-release users (historical)
+
+`driver.session()` and `driver.verify_connectivity()` now raise
+`BackendUnavailableError` where they previously surfaced the raw `neo4j`
+`ServiceUnavailable`, `SessionExpired`, `ConnectionAcquisitionTimeoutError`, or
+`DatabaseUnavailable`. Code embedding toolgraph as a library and catching those
+types directly must catch `BackendUnavailableError` (or its `BackendLockedError`
+subclass); the original exception is preserved as `__cause__`. Configuration,
+authentication, and Cypher/client errors are unaffected, as is `get_driver()`.
+Successful CLI output and query result shapes are unchanged; an unhandled
+outage now names `BackendUnavailableError` in the traceback.
+
+`driver.is_backend_unavailable()` narrowed to match: it now recognizes only
+`BackendUnavailableError` (anywhere in the cause chain), not the raw `neo4j`
+types it previously classified. Code that pairs the still-raw `get_driver()`
+escape hatch with `is_backend_unavailable()` to decide whether to retry will
+silently stop matching — catch `BackendUnavailableError`, or route the call
+through `driver.session()` so the seam types it.
+
+Earlier snapshots could persist stdio arguments or complete HTTP URLs in graph
+endpoint/evidence properties. Reset and rebuild pre-release graphs after
+upgrading. If any argument or URL contained a credential, rotate it as well.