8000 [pull] main from can1357:main by pull[bot] · Pull Request #93 · ericrihm/oh-my-pi · GitHub
[go: up one dir, main page]

Skip to content

[pull] main from can1357:main - #93

Merged
pull[bot] merged 130 commits into
ericrihm:mainfrom
can1357:main
Sep 16, 2026
Merged

pull[bot] merged 130 commits into
ericrihm:mainfrom
can1357:main

Conversation

@pull
@pull pull Bot commented Sep 16, 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 : )

harshaygadekar and others added 30 commits September 15, 2026 22:04
…fresh

Online model refresh now clears the !command credential cache and
force-reloads provider keys so rotated tokens are picked up without
restarting omp.
… replay

History replay could leave overflowing transcript blocks unretired and tool calls collapsed to a single line until another keypress. This affects transcript rebuilds beyond /resume, including tree navigation, agent focus changes, display-setting changes, compaction, and handoff.

Scheduled the next render when Composer acknowledges a replay batch so pending transcript retirement continues. Kept regression coverage for overflowing transcripts and content that fits the viewport.
…resh

registerProvider now records !command header values so refreshProvider
and the 401 retry path re-run extension-backed header credentials, not
only models.yml entries.
The finding-level `optionalProperties` node was indented two columns too
deep, sitting inside `properties` instead of beside it.

jtd-to-json-schema.ts pushes every `properties` key into the emitted
`required` array, so each finding had to carry a field literally named
`optionalProperties`: valid findings were rejected and the yield retry
budget burned, while `anchor` and `remediation` were unreachable and the
agent had no way to report a remediation.

Dedent the node to sibling depth, matching locations.elements,
evidence.elements and deferred.elements in the same file.

Fixes #12195
Plan-included Devin models (SWE-2, SWE-1.7, GLM-5.2 High) and Kimi Code's
/coding/v1/models envelope report no token cost, so usage stats recorded $0.

Add a cost-fallback catalog axis that seeds a reviewed API-equivalent list
price only when every token dimension is zero. SWE-2 uses the 75%-off
enterprise promo through 2026-12-31 and switches to list on 2027-01-01.
Kimi Code rows use the public Moonshot list (kimi-for-coding at the K2.7
Code line rate; K2.8 Preview has no public SKU yet).

Fusion composites stop cost parsing at the Sidekick marker so the headline
card is not overwritten by a dispatched component, and harness-backed Fusion
lanes chat directly instead of taking the AssignModel router path.

Does not regenerate models.json.
Use the Moonshot K2 card for `kimi-k2` instead of the K2.6 rate, and apply
SWE-2's 2027 list change as a dated fallback overwrite rather than a
recurring off-peak tariff.
…pletions

`firstTokenTime` was only captured inside the text-delta and thinking-delta
handlers, so a stream whose first content delta is a structured
`tool_calls` entry (the standard agentic shape on OpenAI-compatible
gateways) produced assistant messages with `ttft` undefined for every
turn that ended in `toolUse`. The per-turn usage row then silently drops
the TTFT figure and the tok/s figure loses its first-byte anchor.

Stamp on the first tool-call delta branch as well; no-content streams
still leave `ttft` undefined (unknown, not zero).

Regression: packages/ai/test/openai-completions-ttft.test.ts
Removed duplicate entry for late non-blocking advisor notes in the changelog.
Online discovery no longer re-runs !command helpers. Only omp models refresh
and TUI F5 mint credentials. Track fetchDynamicModels header commands so F5
and 401 retries evict them.
Selecting a provider queues a catalog-only fetch. F5 during that
debounce or in-flight request now upgrades to a credential re-mint
instead of being dropped by the pending guard.
Hover navigation only cancels debounced catalog fetches. An explicit F5
queued behind an in-flight request still re-mints !command credentials
when that request settles.
…lies

A reviewed Cursor family inherits every unaggregated field from its first
member, so a roster that marks only the extended-context tiers maxMode
collapses to cursorMaxMode: false. The cursor-agent transport reads that
flag off the collapsed row while the wire id comes from effort routing,
so selecting the max tier sends max_mode: false on a max-mode wire id.

Aggregate the positive case like reasoning/input/contextWindow.
…nds the turn on stop

The Anthropic `/v1/messages` route mapped `stop_reason` from the turn's
`StopReason` alone, so a turn that ends with `stop` while carrying an
unpaired `tool_use` block was advertised as `end_turn`. Cursor's exec
protocol has no separate tool-use stop: when it hands a tool the caller
declared back for the caller to execute it ends the turn with `stop`. A
client driving the canonical Anthropic loop — run tools while
`stop_reason === "tool_use"` — therefore never ran the tool it was asked
for, and the conversation stopped on an unanswered call.

An assistant turn cannot carry an already-executed `tool_use` block on
this wire (results arrive in the next user message), so a present block
always means the client must run it. Map that case to `tool_use` on both
the JSON and SSE encoders, mirroring the OpenAI chat wire, which already
maps the same shape to `tool_calls`.
builtinCredentialSecretEntries() only knew the GitHub/GitLab/OpenAI/
Anthropic prefixes from SENSITIVE_TOKEN_RE, so with secrets.enabled on an
AWS, Google, Slack, npm, Stripe, Hugging Face, SendGrid, JWT, Bearer or
PEM private-key value in a tool result still reached the provider raw.
The shapes now live in secrets/patterns.ts as anchored vendor-prefix
regexes, each with a friendlyName so the model sees e.g.
$$AWSACCESSKEY_<hash>$$. No generic keyword or entropy rules: identifiers
like token_expiry_seconds and plain git SHAs must stay readable.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
…opic wire

Cursor stamps `kCursorExecResolved` on a `toolCall` its own exec channel
already ran server-side — `todo`, `web_fetch`, `connect_scm`, or a native
the gateway declined — and those turns also end with `stop`. The new
tool-use termination rule counted them, so a canonical Anthropic client
was told to execute an operation the server had already committed, for a
tool it never declared, and to post a `tool_result` for it.

Both encoders now treat only unresolved calls as handoffs: they alone
terminate the turn with `tool_use`, and resolved blocks are dropped from
the serialized content entirely — a client can neither run nor answer
one. The SSE encoder shifts the indices of blocks after a suppressed one
so index-addressed client snapshots stay aligned.
can1357 and others added 27 commits September 16, 2026 16:57
A DATABASE_URL-style value carries a credential but its variable name
rarely matches SECRET_ENV_PATTERNS, so the password reached the provider
verbatim whenever a tool echoed the URL. collectEnvSecrets() now also
scans every env value for scheme://user:password@host and registers the
password (raw and percent-decoded) as a plain obfuscate-mode secret.
Passwords under 8 chars are skipped, matching MIN_ENV_VALUE_LENGTH.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Redis/Postgres-style URLs like redis://:pw@host omit the username; the non-empty userinfo gate skipped them so the password leaked past secrets.enabled. Relax the gate and add a regression test.
- Dropped union-merge duplicates for #12224 and #12204, refiled #12200 under Fixed and #12233 under Added.
- Added missing entries for #12187 and #12213 and author credit for #12171.
…gistry signature

- The #12172 F5 tests read the third mock argument; the untyped mocks failed bun check with TS2493.
- Dropped the models.dev/api.json fallback fetch, hour-long memo, and test seams from #12227; the shared catalog.stencil.so payload now keeps reasoning_options, so fetchWellKnownModels (session-scoped, coalesced, ETag-revalidated, last-good on failure) is the single source.
- Indexed each catalog payload once via a symbol tag instead of a parallel TTL cache.
- An unreachable catalog leaves the discovery guess in place instead of failing the refresh.
…s early and trim the collapsed transcript's mid-turn head (@eggpeat)
…hrome growth

- Modified Composer to use a `belowFloor` variable for viewport calculations, preserving blank rows when persistent chrome grows.
- Added a test `tmp-compressed-repro.test.ts` that asserts blank rows remain after chrome expansion.
…-end

fix(coding-agent): defer TTSR AST matching to finalized tool calls
- The auth store now quarantines and recreates the database on SQLITE_CORRUPT
  (042028f), so open() legitimately succeeds after a synthesized CORRUPT
  and the "non-BUSY errors short-circuit retries" test failed on every OS.
- The contract under test (non-BUSY errors are surfaced on the first attempt)
  still holds; only the fixture's error code was obsolete. SQLITE_IOERR is
  neither retried nor recovered.
CJK miss sliced inside multibyte char and panicked at apply.rs:936 instead of reporting the miss.
fix(pi-edit): gate sloppy closest_fragment tail window on char boundary
@pull pull Bot locked and limited conversation to collaborators Sep 16, 2026
@pull pull Bot added the ⤵️ pull label Sep 16, 2026
@pull
pull Bot merged commit e220aab into ericrihm:main Sep 16, 2026
5 checks passed
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

0