Commit graph Crafter.Debug/tests
Author SHA1 Message Date
2217037d72
feat(cli): pick the app of the project the launcher belongs to
Discovery is per user, so every project's launcher sees every debug app
running, and with two of them up every command refused to guess and
asked for -i, though the agent in one project meant its own app. Two
copies of the example, both named basic, could only be told apart by
id.

The launcher now exports CRAFTER_DEBUG_PROJECT, its project's directory
(flatpak-spawn gets it spelled out), and the client counts an app as
the project's when its executable lies under it, as `-r` runs it from
<project>/bin/<variant>/. With other projects' apps running, commands
pick this project's; two copies of it still need -i. Another project's
app is never picked by default: with this project's app not up yet, or
crashed, the agent would otherwise drive, or quit, someone else's. It
is listed instead, for -i.

`ls` lists this project's apps first, marked, `--json` says which are,
and `--wait` waits for one of them rather than any app; the exit code
says whether one runs. A client started without the launcher has no
project and picks the one app running, as before, and the command the
protocol check prints for an older app's own client now names it with
-i.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 18:33:59 +01:00
5d34df7afb
fix(cli): print lldb by its path, and report a held app as held
Tried on the example with the app's sandbox, two things were wrong:
- `debugger` printed `flatpak enter <id> lldb -p <pid>`, and flatpak
  enter runs a command without the sandbox's PATH, which /app/bin is
  on: "Failed to execute child process lldb". It now prints the path
  it found, /app/bin/lldb, which attached, listed threads and detached.
- With the app held by that debugger, the handshake times out, and
  the client said "cannot reach basic on port ..." and exited 3, no
  instance, though its lock said it lived. It now says the app is
  alive but not answering, stopped in a debugger or hung, prints how
  to attach, check and end it, and exits 5.

Environment::haveLldb becomes lldb, the path or empty. ShouldServeOverQuic
covers the second with an instance whose lock is held and whose port
is silent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 03:21:41 +01:00
2c77cdcfab
fix(session): report the frame a pause began at
Frames go on counting while paused, since Pump still runs, but status
gave no pause frame and `wait --paused` replied with the current one.
A wait that came minutes after `pause --on-check` caught a failure at
frame 237 said "paused at frame 5475", which an agent takes for when
the check failed.

The session records the frame in PauseNow, and status and the paused
wait report it as pausedAt. The client prints "PAUSED since frame N",
and "paused at frame N (now M)" when the two differ.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 03:21:41 +01:00
7574ad5c4e
fix(cli): take global flags after the command, and call --json again
Only --json was taken after the command, so `wait --paused --timeout
60s`, the order the guide's own synopsis and agents use, failed with
"unknown argument '--timeout'", and `-i` after the command did too.
And since that --json was taken as the output flag, `call X --json
'{...}'`, the documented way to pass arguments as JSON, never reached
call: its object then failed as "not key=value".

--timeout, -i and --json now go anywhere on the line; no command has
options of those names. After `call`, a --json followed by an object
is the arguments, and any other --json the output flag. A trailing
--timeout or -i without its value is a usage error that says so.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 03:21:41 +01:00
6987033d6b
fix(core): encode a float as its shortest decimal
A float was widened to double before encoding, so every float watch,
argument default and result read as the double nearest it: gravity
-9.81 showed as -9.8100004196167, a radius of 0.1 as 0.10000000149011612,
in `get`, `help` and call results alike. Agents read these, and copy
them back into `set` and `call`.

Codec<float> now encodes the shortest decimal that reads back as the
same float (std::to_chars), as a double. Decoding is unchanged, and so
is every value: 0.1 decodes to the float 0.1f came from. A double is
encoded as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 03:21:41 +01:00
fbb5809aea
test(transport): drive a listening session over loopback
ShouldServeOverQuic runs a Session that listens for real and reaches it
with the client's own transport. It covers:
- the discovery record, and its removal when the Session ends;
- status, call, a mistyped command, get and set, and log --follow
  streaming as entries are logged and ending on time;
- six clients at once, and eight streams at once on one connection;
- a wrong token, a request that isn't JSON, and one over 1 MiB, each
  answered with an error;
- ending the Session with a follower waiting, which is told "shutting
  down" and is not sent a backlog it asked none of;
- a client finding the port dead in milliseconds.

The shutdown reply went missing in 4 runs of 30 until Crafter.Network's
~ClientQUIC let stopped streams finish before closing; 50 of 50 pass
since.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 03:07:09 +01:00
34c60dcd34
feat(cli): crafter-debug, the client agents drive an app with
Cli::Run is the whole client but its transport, so tests run it in
process against a real Session; main will hand it the QUIC connection.

- Commands: ls, status, cmds, help, call, get, set, pause, resume,
  step, wait, trace, log (with --follow), screenshot, quit, debugger,
  guide and version, with -i, --json and --timeout.
- call types key=value from the command's argument schema (slot=1 is
  a number, name=1 a string, at.x=3 nests), so agents need no JSON
  for simple calls; a value that cannot fit its field is a usage
  error before anything is sent.
- The instance comes from discovery; one running instance needs no -i.
  The record's protocol is checked before any request: a mismatch
  exits 4, says which side is newer and prints the app's own client.
  Other sources only warn.
- A request with no reply is diagnosed: the app has exited, stopped
  in a debugger (with the attach command), or no longer calls Pump
  (with what to change). Exit codes: 2 usage, 3 no instance,
  4 version, 5 timeout, 6 the app's error.
- Output is short by default, arrays cut at 20 with a note; --json
  prints everything. `guide` is the full reference for agents.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 02:03:38 +01:00
37e0d9b9ae
feat(discovery): find live instances across sandboxes; build info
Each app writes $HOME/.cache/crafter-debug/instances/<id>.json (0600,
in a 0700 directory) and holds an flock on <id>.lock while it lives.
Every `flatpak run` is its own pid namespace, so pids collide and can
neither name an instance nor tell whether it lives; a random id names
it and the lock says it lives, since flock works across sandboxes on
the shared filesystem. An app killed without its destructor leaves its
files with the lock free, and the next listing removes them. The lock
is created under a temporary name, locked, then renamed, as in
Crafter.Review's LifetimeLock, so it is never seen unlocked while its
app lives. Inside a Flatpak the record also carries the sandbox's
instance id (for `flatpak enter`) and the host pid of its init (to end
the whole sandbox from the host).

project.cpp now generates :BuildInfo per core variant: a hash of the
library's sources and the path of the client built with it. App and
client share the core, so they report the same id exactly when built
from the same sources; this is the safety net under building the
client in the same pass as the app.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 01:56:31 +01:00
bd64c44fc5
feat(session): serve requests without blocking a thread, pause and step
The app's end of the connection, independent of the transport, which
reaches it through Detail::Exchange: one request's way back, safe from
any thread and never blocking.

- What needs only the session's own state (hello, status, list, log,
  pause, resume) is answered on the transport thread. What touches the
  app (calls, get, set, trace, screenshot) queues for the next Pump and
  runs on the app's thread between frames. Waits and steps reply from
  the Pump that sees them satisfied, events from the thread that emits
  them. No request ever holds a thread while it waits.
- A request whose timeout passes before Pump is answered Timeout and
  is guaranteed not to have run, so an agent never wonders whether a
  command went through; an Async result arrives on the Pump it is
  ready; requests left when the Session ends are answered ShuttingDown.
- Advance() for frame loops and Paused()/TakeSteps() for fixed-step
  loops hold what the app gates on them; pausing an app that gates
  nothing is refused instead of silently doing nothing. pause --when
  stops the frame its condition holds; pause --on-check stops before
  the next Advance in the frame a check fails, optionally raising
  SIGTRAP for an attached debugger.
- A wait that times out says what its condition's paths were; a log
  follower gets matching entries in order, from the backlog on, and is
  dropped when it cannot keep up.
- Request parameters go through Codec, so a mistyped request gets the
  same did-you-mean replies a mistyped command does. The session sets
  PR_SET_PTRACER_ANY so a debugger the agent starts can attach.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 01:53:15 +01:00
3500b73ecf
feat(registry): commands, watches, samples, the timeline and conditions
The app side of Crafter.Debug, minus the session:

- Command(name, description, handler): the handler takes nothing or
  one argument struct and returns void, a value, Result<T> or
  Async<T>. Arguments are decoded before it runs, so a mistyped call
  is answered with every mistake and never reaches it; a throwing
  handler becomes an Exception error instead of unwinding into the
  frame loop. This replaces 3DForts' three-file pattern per verb.
- Watch(path, variable | getter [, setter]) with bounds, help and
  read-only, and Sample(path, value) from any thread with a
  frame-stamped history. Paths resolve into values ("bodies[1].pos.y")
  and set the same way, writing a member writes the whole value back.
  Unknown paths and commands get a suggestion.
- Group: registrations under a prefix that end with their owner. Not
  movable, so neither is its owner: its watches refer to the owner's
  members, and a move would leave them pointing at the old object.
- Log, LogWarn, LogError, Check and Emit on one frame-stamped
  timeline, with source locations captured without macros. A check
  failing every frame is counted every time but stored sparsely, and
  the session's listener hears every failure.
- Condition: what `wait` and `pause --when` evaluate each frame. A
  condition that cannot be decided reports why instead of being false,
  which would make a wait hang.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 01:40:59 +01:00
89a12bcd61
feat(core): Value, reflection, codecs and results, shared by app and client
Crafter.Debug.Core is what the app side and the crafter-debug client
both need, without the session:

- Value: a JSON value with its own parser and printer, so no consumer
  interface sees a JSON library. Integers stay integers and whole
  doubles stay doubles; NaN and the infinities, which JSON cannot
  spell, are written as strings rather than hidden, because a debugger
  must not lose a NaN; invalid UTF-8 is replaced so output always
  parses. Objects keep insertion order, the order an agent reads.
- Reflect: aggregate members and names through structured-binding
  packs and source_location, taken from Crafter.DB, plus scoped-enum
  names by probing a range of values. Self-tested with static_asserts
  so a compiler that changes function_name()'s shape fails the build.
- Codec<T>: encode, decode and describe for bool, integers, floating
  point, strings, enums, optional, vector, array, string-keyed maps,
  Value, Required<T> and every aggregate. Decode reports every
  problem in a call with its path and, for a typo, the likely intended
  name, and never throws: a bad call cannot take an app down, which a
  wrongly typed parameter could in 3DForts' hand-written from_json.
- Result, Error and Async: handler results, including ones that take
  frames to arrive, so callers need no ready:false polling.

project.cpp now builds Crafter.Debug.Core as its own library, used by
both the app library and the client.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-01 01:31:58 +01:00