8000 [pull] main from can1357:main by pull[bot] · Pull Request #105 · ericrihm/oh-my-pi · GitHub
[go: up one dir, main page]

Skip to content

[pull] main from can1357:main - #105

Merged
pull[bot] merged 24 commits into
ericrihm:mainfrom
can1357:main
Sep 21, 2026
Merged

pull[bot] merged 24 commits into
ericrihm:mainfrom
can1357:main

Conversation

@pull
@pull pull Bot commented Sep 21, 2026 •
Copy link
Copy Markdown

See Commits and Changes for more details.


Created by pull[bot] (v2.0.0-alpha.4)

Can you help keep this open source service alive? 💖 Please sponsor : )

…dges

- Capped eval `judge()`/`completion()` handles at 32 in flight via a process-wide semaphore so a cell spawning hundreds of judgments no longer floods every fallback in the role chain.
- ChainJudge skips a candidate for five minutes after its account rejects a request (401/402/403), avoiding a rejected call plus credential rotation on every judgment.
- Added regression tests for the in-flight ceiling and the rejection cooldown.
- Wrapped `hydrateCredentialScopedModelCaches` in `logger.time` so `PI_TIMING` startup profiles attribute the await.
… judgments

- Added `judge_batch(states, questions)` / `judgeBatch(...)`: the host judges every state under one bounded run that outlives the cell; cells pull settled items across turns with `await b.drain(timeout)` / `drain_iter`, plus `status()`, `results()`, `failed()`, `cancel()`, `close()`, and `attach(id)` after a reset.
- Per-item failures are recorded on the item and one retry re-walks the judge chain; `drain()` raises only when the run dies wholesale (no judge or `min_ok` unmet). Each batch registers as an async job so completion auto-delivers a summary and `hub wait` addresses it.
- Single `judge()` is now awaited and returns answers directly; removed `JudgmentHandle` and judgment support in `wait()`. Shared the eval request semaphore across completion handles, judge, and batch items.
- Rendered `judge_batch` progress on the eval card, documented the helpers in the eval prompt and workflow notice, and covered cursor, retry, min_ok, cancel, ownership, and cross-cell Python/JS prelude behavior with tests.
- Implemented Sugiyama-style rank assignment and barycenter-based crossing minimization to optimize flowchart readability.
- Added label word-wrapping logic for node text to prevent columns from stretching disproportionately.
- Standardized edge attachment points to ensure arrowhead alignment and correct edge grouping.
- Migrated flowchart layout engine from direct insertion to a robust DAG-based multi-pass layering system.
- Updated golden fixtures to reflect improved aesthetic quality and corrected edge-routing behavior.
…assification

- Added `jevify` keyword to trigger a hidden system notice for bulk `eval` kernel classification.
- Implemented `jevify` notice logic ensuring rubric freezing, pre-filtering, and escalation to human review.
- Added settings schema for `magicKeywords.jevify` with default-enabled state.
- Configured gradient highlighting for `jevify` in the TUI to maintain consistency with existing magic keywords.
- Added unit tests to verify notice injection is gated by the `eval` tool's availability.
- Deleted the `unexpectedStopSmart` condition from the settings UI configuration.
- Replaced XML-style `<SM:...>` edit tags with the new `*** SM:` format in several test suites.
- Modified inline edit recovery test to account for an explicit `*** End Patch` marker.
- Centralized magic keyword definitions in `modes/magic-keywords.ts`, removing fragmented per-keyword files and redundant logic.
- Implemented dynamic registration via `setMagicKeywords`, decoupling core TUI rendering from specific keyword implementations.
- Automated settings schema generation and notice type derivation directly from the centralized keyword table.
- Added protection against spelling autocorrect for registered magic keywords in the TUI editor.
- Simplified TUI gradient highlighters to use a generic, registry-driven system instead of per-keyword hardcoded logic.
…n model hub overlay

- Filtered role chips to only show roles that accept the model and added the fallback chip only when the default role supports chat models.
- Enabled Option+<-/-> on macOS terminals by treating "alt+b" and "alt+f" as equivalents to "alt+left" and "alt+right".
- Integrated OpenRouter's Decisions API via the existing System One native judgment interface.
- Updated discovery to selectively fetch models with `decisions` modality support for dynamic judgment routing.
- Extended `TypeSafeJudge` to handle non-TypeSafe providers and include billed USD costs in usage tracking.
- Unified judge role classification by abstracting API-based routing to `native` kind in the judgment chain.
- Refactored model hub to filter role-strip suggestions based on available model capabilities.
…lang compilation

- Added `darwin_sdk_root` function that resolves the macOS SDK path using `SDKROOT` or `xcrun`.
- Updated the build script to include `-isysroot` flag and watch `SDKROOT` for changes.
- Added tests confirming explicit `SDKROOT` usage and fallback behavior when empty.
- Replaced Ctrl+Right handling with Alt+Right in tests and component logic.
- Added support for macOS Option+-> key sequence in tests.
- Updated on-screen help text to show "Alt+<-/->" instead of "Ctrl+<-/->".
… directives

- Removed the prohibition on interpreting XML tags and clarified that tags may interrupt user messages while user content is sanitized.
- Simplified the role description to state the assistant is helpful and trusted in the Oh My Pi coding harness.
- Added `GlyphProtocol` client to stream custom Nerd Font icons via terminal APC escape sequences, enabling icon rendering without pre-patched fonts.
- Implemented automatic capability detection via startup probe in `ProcessTerminal` and deferred registration for supported terminals.
- Included bundled icon definitions and registration logic covering Private Use Area (PUA) codepoints.
- Added session-scoped suppression of protocol feedback sequences to prevent leakage into stdin.
- Configured `TUI` to perform a full repaint on successful icon registration, ensuring consistency for icons rendered during the probe phase.
- Added a `find` tool that performs multi-phase semantic searching using lexical scans, file ranking, and LLM-based passage verification.
- Integrated the `find` tool into the agent's core capability set, including updated system prompts, CLI support, and TUI visualization.
- Implemented a structured judgment-based search flow with cost tracking, keyword extraction, and parallelized LLM evaluation.
- Configured settings for the `find` tool enablement and added comprehensive unit and integration tests for search logic and rendering.
… resolution

- Extended #resolveCredentialTarget with allowStaleOAuthBearer flag and logic to fallback to durable OAuth credential.
- Updated usage-limit and account-policy flows to pass the flag, enabling proper credential selection.
- Added tests covering organization denial behavior and rotation with concurrent refresh scenarios.
- Queued background completions (finished async jobs, exited supervised processes) now cut interruptible `hub wait` calls short via a new `hasBackgroundCompletions` loop peek, instead of sitting in the yield queue until the wait's full window elapsed.
- Broker `wait` with `pattern` now returns when the process reaches a terminal state without printing the pattern; previously it blocked for the entire timeout and reported a bogus time-out.
- Tool output and TUI renderer state "Process exited before output pattern matched" for that case; hub prompt documents both behaviors.
…d tool

- Introduced `--tools` CLI flag and `tools` config field to specify allowed omp tools.
- Updated agent configuration to write `find` toggle based on the tool allowlist and pass `--tools` to omp.
- Added default tool list and propagated the tool allowlist through runner, CLI, and trial execution.
- New startup tip points at the `jevify` magic keyword so bulk classification runs through the eval kernel's `judge()` instead of manual scanning.
- `matches_golden_fixtures` read `fixtures/mermaid.json` at runtime via
  `CARGO_MANIFEST_DIR`, which under Bazel points into `bazel-out` where
  the file is compile_data, not a runfile; the test panicked with ENOENT
  on the CI `Validate Rust workspace (bazel)` job.
- Switched to `include_str!` like every other fixture-backed test in the
  crate, so the file is resolved at compile time on both cargo and bazel.
…urrent contracts

- `find` became a real builtin tool (936f976) and its `find -> glob`
  legacy alias was dropped; `--tools` and agent frontmatter tests still
  expected the alias and failed on CI.
- The user-message gradient test relied on a sibling test file having
  called `setMagicKeywords`; in CI's file chunk nothing registered any
  keyword, so "orchestrate" rendered as plain text. The test now registers
  `MAGIC_KEYWORDS` in `beforeAll` like the host does at startup and clears
  the registry afterwards.
- `find` declares `approval = "read"` but was never added to
  `READ_ONLY_TOOL_NAMES` when it landed (936f976); scout's frontmatter
  lists it, so `isReadOnlyAgent(scout)` flipped to false and the agent hub
  lost its "Read-only" badge for scout workers.
- Prompt fallback tests matched the bundled prompt on the old
  "Helpful, trusted assistant" role line; the wording moved to
  "You are a helpful, trusted assistant" (051d3e1). Marker is now the
  case-stable lowercase phrase.
@pull pull Bot locked and limited conversation to collaborators Sep 21, 2026
@pull pull Bot added the ⤵️ pull label Sep 21, 2026
@pull
pull Bot merged commit d716bcf into ericrihm:main Sep 21, 2026
5 checks passed
Sign up for free to subscribe to this conversation on GitHub. Already have an account? 6297 Sign in.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant

0