Skip to content

Alt 0.7: A01–A10, open audit items and Home dashboard (cloud, fixtures only) - #2

Merged
3d3dcanada merged 20 commits into
mainfrom
cloud/0.7-2026-10-11
Oct 11, 2026
Merged

3d3dcanada merged 20 commits into
mainfrom
cloud/0.7-2026-10-11

Conversation

@3d3dcanada

Copy link
Copy Markdown
Owner

Problem and resulting behavior

The 2026-10-08 PC trial showed small local and hosted models failing for two main reasons: thinking used up the whole output allowance before the model acted, and edits failed (Spark 4B 0/8 at 1,024 tokens; MiMo 9B 3/8; Kimi K3 4/8; GLM 5.3 3/8). This PR builds the 0.7 work orders on main (0.6.0 + Neon N01–N09). It follows docs/HANDOFF-0.7.md: A07 first, then A01–A06, then A09; A08 and A10 are prepared for the owner PC but not run; then the audit's open items and the Home dashboard harvest.

The three unmerged branches were harvested by hand, not merged:

  • review/2026-10-08: brought over the 0.7 work orders and the trial note.
  • audit/final-pass-2026-10-08: brought over the audit, its F01–F14 work orders and its small evidence. Its still-open items are numbered B01–B21 in docs/WORK_ORDERS_0.7.md.
  • build/final-pass-2026-10-08: only the Home dashboard was taken, re-implemented as H01 on main's theme engine. Its F01–F14 implementation (bb34872) conflicts with the Neon TUI throughout, so it stays a reference for the deferred B orders.

What users get:

  • Thinking control: one Thinking setting, sent in each server's own fields (llama.cpp, Ollama, vLLM/SGLang/NIM, OpenAI-style). When a response spends its whole allowance thinking, Alt offers a labelled recovery and never resends on its own.
  • Open parameters: penalties, seed, stop, typical-p, extra request JSON, Ollama num_ctx (with a read-back of what actually loaded) and named llama-server options.
  • Hosted providers: 13 presets including NVIDIA NIM, plus model search for long catalogs.
  • Resilience: retries for 429, 5xx and dropped connections, and an output guard that discards degenerate provider output once and resends.
  • Model probe: one short request, then suggestions that are applied only when chosen.
  • Edits: replace_file for small files, and an old_text replace that tolerates whitespace differences.
  • Security: terminal and external-tool actions need their own opt-in, separate from "Allow this session" and --allow-tools.
  • Docs and package: a README for new users and a package 96% lighter in docs.

Every new capability is in Settings tabs and Ctrl+P. Every Settings row is now also a palette entry. Mouse and keyboard both work, and screens stay usable at 80×24.

Validation

Every order commit ran cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test, the TUI smoke scripts relevant to it, screenshot captures at 80×24 and 120×40 (docs/screenshots/0.7/<order>/, via the new scripts/capture_screens.py), and scripts/secret_scan.py before each push. The final head passed the complete scripts/setup-cloud.sh (exit 0): 173 Rust tests, 72 Python unit tests, and every TUI, real-Goose (openai and ollama), compact-probe and stream-recovery smoke. No API keys, GPU or weights were used.

Order Commit Test Result Deferred
Board a9318a4 every_settings_row_is_in_the_palette; fixed stale workbench/harness smokes Board on main; settings reachable from palette —
A07 7bed166 scope/trust unit tests, integration --allow-terminal, runtime_ownership.rs, extension policy test, smoke_goose Pass persist-credentials: false on the 3 evidence jobs that git push. My hand-written token push was blocked by the safety classifier; those jobs are unchanged
A01 3508ee8 tests/thinking.rs, recovery-menu app test, smoke_goose cutoff via real Goose Pass —
A02 190a4c1 tests/params.rs (merge/refusal, Ollama num_ctx + /api/ps, golden argv, CLI) Pass Whether real Ollama honours num_ctx on /v1 is a PC measurement; Alt shows what loaded
A03 f801b52 tests/hosted.rs, presets unit, picker app test; smoke_goose proves the key reaches the provider and never Alt state Pass —
A04 9ba7be1 tests/faults.rs provider module; smoke_goose guarded garbage→clean via real Goose Pass —
A05 86c1ab8 tests/probe.rs (llama.cpp, Ollama, OpenAI-compatible); PTY probe against a fake llama-server Pass —
A06 76529ee tests/edits.rs; smoke_compact replace_file scenario (both providers, both compact formats, TUI) Pass Effect measured by A10
A09 f322b8f package-linux.py + smoke_package; check_repo_weight.py + secret scan in CI Pass (docs 9,267 files / 154.5 MB → 179 / 6.0 MB) —
A08 93b788c gpu unit tests, tests/cuda_runtime.rs (verify, install, argv/env), pin script determinism Prepared, not run Pin slot empty until a build is verified on the PC
A10 eb02759 scripts/test_pc_matrix.py; pc_matrix_0_7.sh --dry-run validates all 10 cells Prepared, not run Owner PC run (docs/PC_HANDOFF_0.7.md)
B01 518518c edit-keeps-settings app test (the audit's 4K profile) Pass —
B02 ccf9807 typed-while-connecting app test Pass Durable drafts across New/quit (B14)
B06 258374f runtime selection tests Pass —
B07 73812c4 Alt+M app test + PTY capture Pass —
B08 999b3f2 stalled-server cancel test (<1 s) Pass —
B09 6acd751 tests/gguf_header.rs Pass Import shard-set discovery (with B12)
B10 05ea611 oversized-paste app test Pass 60×18 connection form
H01 6042ae4 dashboard render test; smoke_verification fixed (it had never run on main) Pass H02 custom accent colour
Status 1787b4d — Board status table —
B12–B21 — — Not started F01, F02, F03 rest, F04, F05, F07 rest, F11, F13, F14

Not run here: anything needing a GPU, real hosted keys or weights (A08, A10, live evaluations). No release or tag was created.

Compatibility and recovery

  • Config: the new inference, runtime and preference fields all default, so existing config.toml and preferences.toml load unchanged. Old reasoning_tokens budgets keep their exact request shape.
  • Behaviour changes:
    • --allow-tools and "Allow this session" no longer approve terminal or external-tool requests. Use --allow-terminal or "Allow terminal this session". The live harness scripts that run in Full access pass --allow-terminal. Candidate exploration keeps --allow-tools, so its terminal requests are denied.
    • Headless extensions check requires --access trusted.
    • A selected but missing runtime or engine path no longer falls back to another binary.
    • Imported GGUF files need a valid 24-byte header.
    • Ollama requests now carry options.num_ctx unless "server default" is chosen.
    • The output guard holds responses from remote servers by default, which delays streamed text; set it to off to stream live.
  • AGENTS.md: live evaluations must now label checkpoints truthfully (--artifact-label), per the owner's 2026-10-08 direction; --uncensored remains an alias.
  • History was not rewritten.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx


Generated by Claude Code

claude added 20 commits October 11, 2026 13:20
…plan

Harvested by hand, not merged:
- review/2026-10-08: docs/WORK_ORDERS_0.7.md (A01-A10) and the
  2026-10-08 PC model trial note.
- audit/final-pass-2026-10-08: final audit, F01-F14 work orders and its
  small bounded evidence (about 190 KB, no weights).
- build/final-pass-2026-10-08: only planned (H01 Home dashboard); its
  F01-F14 implementation conflicts with the Neon TUI and stays a reference.

The board numbers the audit's still-open items B01-B21 and the harvest
H01-H02, and marks overlap with N01-N09.

Interface: every Settings row is now also a palette entry ("Setting: ..."),
so palette, settings tabs, keyboard and mouse reach the same actions.
smoke_workbench_tui.py and smoke_harness_tui.py still walked the pre-Neon
settings list and failed on main; they now use the palette.

Tooling: scripts/secret_scan.py (stdlib only) scans added lines before a
push; scripts/capture_screens.py records order screenshots at 80x24 and
120x40 through a real PTY with the fixture engine.

Tests: fmt, clippy -D warnings, cargo test (128 passed), smoke_tui,
smoke_task_tui, smoke_workbench_tui, smoke_harness_tui and the remaining
setup-cloud gates pass. New unit test every_settings_row_is_in_the_palette.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
1. Approval requests carry a scope tag. "Allow this session" and
   --allow-tools approve project actions only (checkpointed edits,
   configured checks, skill helpers). Terminal, external MCP tools and
   untagged engine requests are host scope and need their own opt-in:
   "Allow terminal this session" (TUI) or --allow-terminal (headless).
2. Project tools never list, read or edit Alt's data folder when the
   project contains it; a $HOME or / project root shows a warning (TUI
   notice and CLI stderr). Full access keeps arbitrary terminal capability.
3/4. More credential names excluded (.envrc, .netrc, .npmrc, .pypirc,
   .git-credentials, id_rsa*/id_ed25519*/id_ecdsa*, credentials*.json,
   .docker, .kube, .gnupg, *.p12, *.pfx), compared case-insensitively.
5. Plain-HTTP MCP connections with auth_env are accepted only on loopback,
   checked on save and on connect.
6. (B05) The owned llama-server gets a random per-launch key through a
   private --api-key-file (not argv); readiness needs an authenticated
   /v1/models; the key is attached to the in-memory profile, never saved.
7. CI: the release job checks out without persisted credentials;
   requirements-dev.txt is hash-pinned and installed with --require-hashes;
   SECURITY.md names the current line and the owner's reporting switch.
B04: headless `extensions check` refuses unless access is Full, sharing one
rule with the TUI.

Deferred: persist-credentials:false on the three evidence jobs that
`git push` (ci-diagnostics, lmstudio, model-matrix report) needs a
different push mechanism; left unchanged.

Tests: unit tests for scope classification, session trust, credential
names, data-folder exclusion, loopback HTTP rule; integration tests for
--allow-tools vs --allow-terminal and headless extension policy; new
tests/runtime_ownership.rs with tests/fixtures/fake_llama_server.py.
fmt, clippy -D warnings, cargo test (136 passed), smoke_tui,
smoke_task_tui, smoke_goose (openai, ollama), smoke_workbench_tui,
smoke_harness_tui, smoke_compact pass. Screenshots in
docs/screenshots/0.7/a07 (80x24, 120x40).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
- inference::Settings gains `thinking` (default | off | low | medium |
  high | budget:N) and `dialect` (auto | llama-cpp | ollama |
  chat-template | openai-effort). One choice becomes the fields each
  server understands: llama.cpp chat_template_kwargs.enable_thinking +
  reasoning_budget_tokens (budgets always leave action headroom); Ollama
  think + reasoning_effort; vLLM/SGLang/NIM chat_template_kwargs
  {enable_thinking, thinking} + reasoning_effort; OpenAI-style
  reasoning_effort. Default sends nothing; legacy reasoning_tokens keep
  their exact request shape. Every receipt records dialect and fields.
- A response that ends at the output limit with reasoning but no
  content and no tool call is a thinking cutoff: receipt flag, Stage
  ReasoningCutoff and a per-connection count. Tokens stay charged and
  nothing is resent.
- Recovery is an explicit choice: the TUI opens a labelled menu when the
  turn ends (thinking off / low for this connection with the last message
  back in the composer, or a larger allowance); Relay::set_thinking
  records evidence. Headless --json emits a typed reasoning_cutoff event.
- Reachable from Settings → Models → Thinking, Model output and sampling
  (Thinking, Server dialect), Ctrl+P (Model thinking, Recover from a
  thinking cutoff) and `alt inference --thinking/--dialect`.

Tests: tests/thinking.rs (outgoing body per dialect, budget validation,
cutoff detection, explicit live switch over a scripted upstream in
tests/fixtures/upstream.rs); app unit test for the recovery menu;
smoke_goose (openai, ollama) runs a reasoning-only fixture through real
Goose and asserts the typed event, no resend, and the off fields after
`alt inference --thinking off`. fmt, clippy -D warnings, cargo test
(141 passed), smoke_tui, smoke_task_tui, smoke_harness_tui,
smoke_allowance_tui pass. Screenshots: docs/screenshots/0.7/a01.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
- Request (per connection, inference::Settings): repeat_penalty
  (repeat_penalty on llama.cpp/Ollama, repetition_penalty on vLLM, refused
  on plain OpenAI-style APIs), presence/frequency penalty, seed, stop,
  typical_p, and extra_body: a JSON object merged last. Alt-owned keys
  (model, messages, tools, tool_choice, stream, stream_options, max_tokens,
  max_completion_tokens, n) are refused with a clear message.
- Ollama: options.num_ctx from the connection context (replacing any
  engine options) and optional keep_alive, unless "server default" is
  chosen. After a response the relay reads /api/ps and records the loaded
  context in the receipt and live status; the TUI says when Ollama ignored
  the request and how to fix it.
- Owned llama-server: mmproj, --n-cpu-moe, --override-tensor,
  --tensor-split, --main-gpu, RoPE/YaRN, --mlock, --no-mmap,
  --chat-template-file, --model-draft/--draft-max and extra_args; --host,
  --port, --api-key(-file), --alias, --model and --ctx-size are refused.
  The argv now comes from one pure function, runtime::server_args.
- One table (src/params.rs) drives `alt params set|unset|show` and the
  TUI's Advanced parameters (Settings → Models, Ctrl+P) with labelled
  fields and a raw JSON view of request fields and runtime argv.

Tests: tests/params.rs (merge order and refusals, per-server penalty
spelling, Ollama num_ctx and /api/ps read-back through the relay, golden
runtime argv, CLI round trip). smoke_goose for Ollama now requires options
to be exactly the selected num_ctx. fmt, clippy -D warnings, cargo test
(146 passed), smoke_tui, smoke_task_tui, smoke_goose (openai, ollama),
smoke_harness_tui, smoke_compact (ollama) pass. Screenshots:
docs/screenshots/0.7/a02.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
- src/presets.rs: NVIDIA NIM (integrate.api.nvidia.com/v1,
  NVIDIA_API_KEY, chat-template dialect), OpenRouter, Groq, Together,
  DeepSeek, Mistral, Gemini OpenAI-compatible, Cerebras, LM Studio,
  vLLM/SGLang, llama.cpp, Ollama and Custom. A preset fills endpoint, key
  variable name and thinking dialect; it never stores a key.
- `alt init --preset nvidia --model z-ai/glm-5.3` works headless;
  explicit --endpoint/--provider/--api-key-env win; a missing key
  variable is reported. `alt models --filter WORDS` searches catalogs.
- TUI: connection wizard row "A hosted API or another server preset" and
  Ctrl+P "Connect a hosted API…" open the preset list; the form carries
  the preset's dialect into the saved connection. The model picker has a
  search line that filters as you type (keyboard and mouse use the
  filtered rows).
- B03: the engine no longer refuses Ollama connections that name a key
  variable; the relay already sends the bearer key.

Tests: preset unit test; tests/hosted.rs (headless NVIDIA init stores only
the variable name and dialect; catalog listing/filtering with the key sent
to /models); app test for preset form + filtered pick; smoke_goose now runs
both providers with a random fixture key, asserts it reaches every model
request and that no file under Alt's state contains it. fmt, clippy
-D warnings, cargo test (150 passed), smoke_tui, smoke_task_tui,
smoke_goose (openai, ollama), smoke_workbench_tui pass. Screenshots:
docs/screenshots/0.7/a03.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
- Relay retries HTTP 429, 5xx and dropped connections with capped
  exponential backoff (0.5 s doubling, at most 30 s) that honours a
  numeric Retry-After. Settings.retries (default 2, 0-8). Each retry is a
  model call against the shared request allowance; receipts carry
  `retries` and `retry_log`; live status counts provider_retries.
- Output guard (Settings.output_guard: auto = on for non-loopback
  endpoints, on, off). When active the relay holds the complete response,
  then checks it: leaked chat-template tokens (<|name|>) or more than 30%
  of letters in scripts absent from the conversation, mixed with others.
  A degenerate response is discarded (raw bytes kept as
  <id>-discarded-N.raw), recorded as a provider fault with its charged
  tokens, and resent once; a second fault returns a 502 with a clear
  message, never model text. A whole answer in another single language
  is not flagged.
- TUI: Model output and sampling → Provider retries / Output guard,
  Ctrl+P → Hosted output guard; the allowance view shows retry and fault
  counts. CLI: `alt inference --retries N --output-guard auto|on|off`.

Tests: tests/faults.rs provider module (429+503 retries with budget
accounting, dropped connection, exhausted retries, single discard and
resend, repeated fault error, language-aware detector, automatic guard
scope) using tests/fixtures/upstream.rs (new hang-up reply). smoke_goose
(openai, ollama) runs a guarded garbage-then-clean exchange through real
Goose and checks the export never contains the discarded output.
smoke_harness_tui no longer races a stale "Tool focus saved" notice.
fmt, clippy -D warnings, cargo test (156 passed), smoke_tui,
smoke_task_tui, smoke_harness_tui, smoke_allowance_tui,
smoke_stream_recovery, smoke_compact pass. Screenshots:
docs/screenshots/0.7/a04.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
- src/probe.rs: server facts without inference (llama.cpp /props
  n_ctx and thinking template; Ollama /api/show context_length and
  capabilities; others report none), then one bounded chat request
  (output capped at 512, a trivial probe_echo tool, the model's default
  thinking, no sampling overrides). Records tool call, reasoning seen,
  finish reason and completion tokens.
- Suggestions (pure `probe::suggest`): thinking off/low, larger output
  allowance, compact tools, match native context. `probe::apply` changes
  only the chosen field on the same connection; never the model or
  provider. Reports are stored per model identity under probes/.
- Owned runtimes are started for the probe and stopped afterwards; the
  TUI drops an idle connection first so two owned models never share a
  small GPU.
- Reachable from Settings → Models → Probe this model, Ctrl+P → Probe
  this model, the post-save hint on new connections, and
  `alt probe [--apply ID]`.

Tests: tests/probe.rs fixtures for llama.cpp (reasoning cutoff →
thinking-off, more-output, native-context; apply changes only the chosen
fields), Ollama (/api/show capabilities, working tool call → no change)
and a plain OpenAI-compatible server (no tool call → compact tools).
scripts/capture_screens.py can now start tests/fixtures/fake_llama_server.py
so the TUI probe runs end to end in a PTY. fmt, clippy -D warnings, cargo
test (159 passed), smoke_tui, smoke_task_tui, smoke_harness_tui,
smoke_workbench_tui pass. Screenshots: docs/screenshots/0.7/a05.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
- replace_file(path, new_text) in the compact and compact-lines tool
  profiles (not in All/Coding): a checkpointed whole-file rewrite of a
  file of at most 300 lines. It is bound to the current revision's
  full-file read handle, so a partial read cannot drop unseen code;
  undo restores the original.
- old_text replace: exact unique match first; otherwise a
  whitespace-normalised line match. When only indentation differed,
  new_text is re-indented to the file's indentation, and the tool result
  says how the text was located (Change.matched). On a miss the error
  carries the closest current span (by similarity), its line numbers and
  a fresh handle usable with operation=handle; repeated or ambiguous
  text is refused instead of guessed.
- Compact prompts name replace_file; Tool focus describes it.

Tests: tests/edits.rs (complete-read binding, 300-line limit, undo;
whitespace match with re-indent; closest-span handle that edits exactly
that span; ambiguity refusal). smoke_compact adds a replace-file-and-undo
scenario through real Goose (openai, ollama; compact and compact-lines;
TUI). smoke_harness_tui polls saved allocation instead of trusting a
stale notice. fmt, clippy -D warnings, cargo test (162 passed),
smoke_tui, smoke_task_tui, smoke_harness_tui, smoke_goose,
smoke_practice_tui pass. Screenshots: docs/screenshots/0.7/a06. A10 is
the matched measurement of the effect.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
- README now opens with install, connect a model (Ollama, LM Studio,
  llama.cpp/vLLM, hosted presets such as NVIDIA NIM, or a GGUF file; keys
  as environment variables) and a first task, then keyboard/palette
  reachability. Results moved to "Status and evidence", including the
  2026-10-08 PC trial and the fact that the matched 0.7 re-run (A10) has
  not run yet.
- B11: scripts/package-linux.py ships guides, summaries and screenshots
  but not docs/evidence or non-Markdown research data; the package's
  docs/EVIDENCE.md names the source commit holding the raw evidence.
  Packaged docs go from 9,267 files / 154.5 MB to 179 files / 6.0 MB
  (96% smaller). History is not rewritten.
- scripts/check_repo_weight.py refuses tracked files over 1 MiB (one
  grandfathered receipt) and model weights or archives; the verify
  workflow runs it together with a full secret scan. CONTRIBUTING states
  the evidence rule.

Tests: package-linux.py on a fresh release build, then smoke_package.py
(archive and per-file integrity, repeated install, failed update keeps
the binary, state backup/restore, docs/licenses) pass; archive 14.3 MB.
fmt, clippy -D warnings, cargo test, smoke_tui, smoke_task_tui pass;
check_repo_weight and secret_scan --all report nothing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
Not run: the cloud has no GPU, and no CUDA build has been verified, so
the pin slot is deliberately empty and automatic installation refuses
with guidance.

- runtime: nvidia-smi parsing (name, VRAM, compute capability, driver)
  and cuda12_fit (compute 6.1+ for Pascal, driver 525.60+), labelled
  either way. Component::CudaInference with a `CUDA12: Option<Pin>` slot.
- runtime::install_pin: resumable download through
  models::download_verified (size + SHA256), unpack into
  tools/<directory>, keep the previous copy until the new one is in
  place; refuses an archive without <directory>/llama-server.
- Managed runtimes under Alt's tools folder get that folder prepended to
  LD_LIBRARY_PATH so bundled ggml/CUDA libraries load.
- Never auto-switched: installing does not select the runtime or change
  GPU layers. TUI offers "Use the CUDA runtime" explicitly; headless
  `alt install cuda-runtime [--use]`.
- TUI: Settings → Runtime → GPU runtime (CUDA 12) and Ctrl+P: detection
  summary, install when pinned and fit, build-from-source help, select an
  existing llama-server, or keep CPU.
- Owner tools: scripts/pin_cuda_runtime.py turns a verified build
  (scripts/build-pascal-runtime.sh) into a deterministic archive and
  prints the exact pin, with licence notes for bundled CUDA libraries;
  scripts/pc_cuda_check.sh checks devices, offload and a benchmark on
  the PC first.

Tests: runtime gpu_tests (parsing and fit); tests/cuda_runtime.rs
serves a fixture archive over HTTP: wrong checksum installs nothing,
correct archive installs and reinstalls in place, launch passes
--n-gpu-layers 99 and the bundled library folder on LD_LIBRARY_PATH (fake
llama-server records argv and env). pin_cuda_runtime.py verified
deterministic on a fixture build. fmt, clippy -D warnings, cargo test,
smoke_tui, smoke_task_tui, smoke_allowance_tui pass. Screenshots:
docs/screenshots/0.7/a08.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
Prepared and verified with --dry-run only; the owner PC runs it.

- scripts/live_acceptance.py: required --artifact-label
  uncensored|standard|hosted, recorded as given in configuration.json
  (artifact_label + scope); --uncensored remains an alias so existing
  callers keep working. Hosted runs need no local file or hash; a managed
  runtime still needs --model + --sha256. Brings over the review
  branch's --api-key-env, adds --thinking-level/--dialect (A01) and
  --check-args (validate, then exit before hashing or any request).
- scripts/pc_matrix_0_7.sh: suite v5, the 8 trial cases, 16K, compact,
  4,096 output, 3 repeats; Spark 4B (Ollama), MiMo 9B abliterated and
  standard (managed llama.cpp CUDA), kimi-k3 and glm-5.3 (NVIDIA NIM);
  arms "trial" and "thinking-off". Free-VRAM check before each GPU cell
  with recorded skips, Ollama unload between cells, stops every
  llama-server it started, ledger.jsonl of every cell. --dry-run prints
  and validates all ten commands through live_acceptance itself.
- docs/PC_HANDOFF_0.7.md (A08 + A10 steps), LIVE_EVALUATION.md labels;
  AGENTS.md now requires truthful labels per the owner's 2026-10-08
  direction (uncensored stays the default target).

Tests: scripts/test_pc_matrix.py (dry run lists and validates 10 cells
with the right labels and arms, writes nothing; labels required, hosted
rules, alias), now in setup-cloud.sh with the dry run; existing
test_evaluation_preflight, test_research_adapters and test_campaign pass.
fmt, clippy -D warnings, cargo test (164 passed), smoke_tui,
smoke_task_tui pass. No TUI surface: this is the PC evaluation harness.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
The connection form rebuilt a profile from preferences, so saving an
otherwise unchanged connection reset context (4K→8K), steps, output,
request/token allowances, sampling, thinking and the uncensored claim.
The draft now patches the existing profile: only name, provider,
address, key variable and dialect come from the form. The model picker
opens on the connection's current model.

Tests: app test editing_a_connection_keeps_its_custom_settings
reproduces the audit's customised 4K profile through edit → draft →
save and compares every inference field. fmt, clippy -D warnings, cargo
test, smoke_tui, smoke_task_tui pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
The first message sent before a connection was ready stayed in the
composer and was copied back over it on Ready, so anything typed while
connecting was lost; failure restored the old prompt over new text too.
Now the submitted message is held in pending_prompt and the composer is
cleared for the next message. Ready sends the held message directly.
Failure, cancel and a closed connection restore it exactly once, ahead
of any newer draft. Durable drafts across New/quit remain B14.

Tests: app test for failure, cancel (no duplicate) and closed
connection; smoke_tui and the TUI compact journey exercise Ready.
fmt, clippy -D warnings, cargo test (166 passed), smoke_tui,
smoke_task_tui, smoke_allowance_tui, smoke_compact --interface tui pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
find_runtime and find_engine returned the managed or PATH copy when the
executable the user had selected was missing, so a different binary ran
without notice. A selected path is now authoritative: missing or not
executable yields None, and the start errors name the selected path and
the recovery (choose it again or clear it in Settings). Unselected
discovery is unchanged; an explicit --engine was already exact.

Tests: runtime selection_tests (managed copies present, selected paths
deleted → no fallback; selected valid path used). fmt, clippy
-D warnings, cargo test (167 passed), smoke_tui, smoke_task_tui,
smoke_workbench_tui, smoke_allowance_tui, smoke_pc_recorder,
smoke_practice_tui pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
Ctrl+M is encoded as Enter by ordinary terminals, so the documented
shortcut ran the connection check instead. Manual entry is now Alt+M (a
real PTY sends ESC m) and a "Type model ID" button in the connection
form; editing an existing connection pre-fills its model. Allowance
recovery was already reachable from Ctrl+P (N06).

Tests: app test (Alt+M → model-ID input → saved connection uses the
typed id); PTY capture sends ESC m at 80x24 and 120x40
(docs/screenshots/0.7/b07). fmt, clippy -D warnings, cargo test (168
passed), smoke_tui, smoke_task_tui pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
The connection check ignored its cancellation token, so Esc left the
inventory request running and its result reopened the model picker
afterwards. The check now races the inventory against the token, and a
Tested/Inventory result from a cancelled job is discarded with a notice;
the connection form stays.

Tests: async app test with a stalled server: cancellation returns
control in under one second and the dialog remains the connection form.
fmt, clippy -D warnings, cargo test (169 passed), smoke_tui,
smoke_task_tui pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
verify_gguf read only the 4-byte magic, so a magic-only file imported.
It now requires the 24-byte header: magic, version 2 or 3, and plausible
tensor (<= 2^24) and metadata (<= 2^20) counts; truncated files fail
with a clear message. The same check runs after verified downloads and
for each split part. Shard-set discovery for imports and per-shard
integrity coverage remain B12/F08 work.

Fixtures that wrote "GGUF<text>" now write a minimal valid header
(tests/benchmark.rs, tests/model_downloads.rs, scripts/smoke_tui.py).

Tests: tests/gguf_header.rs (magic-only, truncated, v1, implausible
counts, metadata-only split; import refuses magic-only and keeps the
library unchanged). fmt, clippy -D warnings, cargo test (171 passed),
smoke_tui, smoke_task_tui, smoke_pc_recorder, smoke_native,
smoke_analysis pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
…21, part)

Editor::insert silently ignored input that would exceed 64 KiB. It now
returns whether the text was accepted (Editor::LIMIT), and paste() shows
the paste size, the limit and the file-path alternative when it is
refused; nothing is inserted. The palette query also accepts pastes.
The 60x18 connection-form overlap is still deferred.

Tests: app test (70 KiB paste refused with a reason, composer unchanged,
a later small paste works). fmt, clippy -D warnings, cargo test (172
passed), smoke_tui, smoke_task_tui, smoke_workbench_tui pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
Ported by hand onto main's Neon TUI, not cherry-picked: the final-pass
branch's Home layout predates the theme engine and rewrote view.rs.

- When the Home body has at least 96 columns and 26 rows, Home shows
  three status cards (MODEL: model, context and thinking; PROJECT:
  folder and configured checks; TOOLS: engine and access) and the eight
  Start-here actions as a two-column grid of cards with icons, drawn with
  the active theme's panels, borders and selection colour.
- Indexes, keyboard selection, Enter and mouse targets (Hit::Home) are
  unchanged; 80x24 and other small sizes keep the list.
- smoke_verification_tui.py still walked the pre-Neon settings list (it
  had never run on main because setup stopped earlier); it now uses the
  palette like the other smoke scripts.

Tests: view test (wide render shows the three cards and eight card
mouse targets; 80x24 shows the list). fmt, clippy -D warnings, cargo
test (173 passed), smoke_tui, smoke_task_tui, smoke_workbench_tui,
smoke_practice_tui, smoke_verification_tui, smoke_harness_tui pass.
Screenshots: docs/screenshots/0.7/h01 (80x24 list, 120x40 dashboard).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
Docs only. Adds the per-order status table to docs/WORK_ORDERS_0.7.md
after the final full scripts/setup-cloud.sh run (exit 0: 173 Rust tests,
72 Python unit tests, every TUI, Goose, compact and stream-recovery
smoke).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbobytt56pieYgqXeVGKrx
@3d3dcanada
3d3dcanada merged commit d2f9fc4 into main Oct 11, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants