Skip to content

[chore] Add GPT-6.1 Sol: bump Pi to 0.99.1 and refresh model catalogs - #7252

Merged
mmabrouk merged 2 commits into
release/v0.121.7from
agent/add-gpt-6.1-sol
Oct 1, 2026
Merged

mmabrouk merged 2 commits into
release/v0.121.7from
agent/add-gpt-6.1-sol

Conversation

@mmabrouk

@mmabrouk mmabrouk commented Sep 30, 2026 •

Copy link
Copy Markdown
Member

Caution

Release blocker: manual QA of the ChatGPT subscription flow is required before this ships.

Pi 0.99.0 reworked ChatGPT sign-in upstream. It added a new "Sign in with ChatGPT" flow on the openai provider and renamed the openai-codex provider that Agenta uses to "OpenAI Codex (legacy)". Our code-level and CLI checks pass (see "ChatGPT login check" below), but they do not cover the full product flow. Before the release that includes this PR, a person must run these steps on a test deployment (not oss-team) and confirm each one:

  • New sign-in: create a ChatGPT subscription connection in the UI and complete the device-code login. The connection reaches ready.
  • Run: run a Pi agent on that connection with gpt-6.1-sol and with one older model (for example gpt-6-sol). Both complete, and the trace shows the model and a cost.
  • Existing login: a connection signed in before this upgrade still runs without signing in again.
  • Token refresh: a run on a connection whose access token has expired refreshes it, and the connection stays ready afterward (no needs_login).
  • Re-login: sign out or remove the connection, then sign in again. It works.

Owner: TBD. Record the result as a PR comment with the deployment and commit SHA.

ChatGPT login check (done before review)

  • Pi's credential store and credential types are byte-identical between 0.87.1 and 0.99.1. Same auth.json file, same openai-codex key, same fields.
  • The token refresh function in the legacy openai-codex module is unchanged. The only change in that module is the interactive login's local callback server, which Agenta does not use.
  • The runner starts sign-in with loginOpenAICodexDeviceCode from its direct @earendil-works/pi-ai@0.80.6 dependency, which this PR does not change. The import resolves at runtime.
  • All 12 runner subscription test files pass against 0.99.1 (186/186).
  • Live on hetzner with pi 0.99.1 and an isolated copy of an existing login: an empty agent dir fails as expected; the copied login completes a gpt-6.1-sol run; the file afterward still passes the runner's validateSubscriptionLogin checks.
  • Not covered: a live token refresh (forcing one on a copy could rotate the shared refresh token), the hosted UI sign-in flow, and the Daytona sandbox path. The manual QA above covers these.

What

OpenAI released GPT-6.1 Sol (gpt-6.1-sol) at DevDay on 2026-09-29: $2/$10 per Mtok, $0.10 cached input, 272K context, near-Astra agentic coding at one fifth of Astra's price. This adds it everywhere the platform needs it, by bumping the pinned Pi harness to 0.99.1 (the direct successor of 0.87.1 — no releases exist between them) and refreshing the catalogs.

Sources: OpenAI DevDay launch, pi v0.99.1 release notes (adds GPT-6.1 Sol, makes it the Codex default, fixes /login with OpenAI), litellm price map (gpt-6.1-sol present).

Changes

Runner (Pi 0.87.1 → 0.99.1)

  • package.json: bump @earendil-works/pi-coding-agent and the pi-coding-agent-pi-ai alias to 0.99.1; lockfile regenerated.
  • pi-provider-cost-patch.json: piVersion 0.99.1, new CLI chunk openai-completions-XW2Q5HVC.js. Both anchors verified to bind exactly once in 0.99.1 (unbundled and minified copies).
  • PINNED_PI_VERSION (daytona.ts), PI_VERSION (build_snapshot.py), sandbox README.
  • pi-model-config.ts: dropped the Sonnet 5.5 runner-side registration entry; pi 0.99.1 carries the model (PI_MODELS_AHEAD_OF_CATALOG is now empty).
  • bash-tool.ts: mirror Pi 0.99's behavior change — a non-zero exit returns an isError result instead of throwing (parity test updated accordingly).

SDK / API catalogs

  • pi_models.generated.json regenerated from pi-ai 0.99.1 (541 → 555 models; adds openai/gpt-6.1-sol, openai-codex/gpt-6.1-sol, openrouter variants, anthropic/claude-sonnet-5-5).
  • pi_models.curated.json: overlay labels/descriptions/ratings for both gpt-6.1-sol ids; Sonnet 5.5 addition retired (generated entry supersedes it); sources note updated.
  • capabilities.py: gpt-6.1-sol first in PI_SUBSCRIPTION_MODELS["openai-codex"] (it is Pi's new Codex default); openai/gpt-6.1-sol in PROVIDER_DEFAULT_MODELS["openai"]; together_ai default Kimi-K2.7-Code → Kimi-K3 (renamed in the pi-ai catalog).
  • assets.py: gpt-6.1-sol in supported_llm_models["openai"].
  • enums.py (API): SUBSCRIPTION_PROVIDER_MODELS kept in agreement with the SDK copy.

Cost tracking: no code change needed. The API prices spans from litellm's live price map, which already carries gpt-6.1-sol. Verified through the API's own path: resolve_pricing_model("gpt-6.1-sol") resolves (bare and openai/-prefixed) and prices 5031 input / 10 output tokens to $0.010062 / $0.0001 — matching Pi's live billed numbers below. Note: running API containers only refresh the live map at import, so the model prices correctly after the next deploy/restart.

Deliberately NOT in this change: the Codex harness

The pinned Codex CLI 0.156.1 is rejected by the ChatGPT backend for this model (verified live on hetzner):

codex-cli 0.156.1
ERROR: {"type":"invalid_request_error","message":"The 'gpt-6.1-sol' model is not supported when using Codex with a ChatGPT account."}

Codex CLI 0.159.2 accepts it (verified live, reply SOL61-CODEX-NEW-OK), but wiring that in needs @agentclientprotocol/codex-acp 1.13.1 → 2.x (a major) plus runner image and Daytona snapshot rebuilds — a separate change. CODEX_MODELS and codex_models.curated.json are unchanged so the picker never offers an id that fails at run time.

Testing (hetzner: live; local: suites)

Live on hetzner (/data/mahmoud/sol61-test, pi 0.99.1, ChatGPT subscription login):

  • pi --provider openai-codex --model gpt-6.1-sol --print → completed: SOL61-OK.
  • JSON-mode run: events carry "model":"gpt-6.1-sol", usage input:5031, output:10, cost {"input":0.010062,"output":0.0001,"total":0.010162} — exactly 5031 × $2/M + 10 × $10/M, confirming the new catalog pricing end to end.
  • Codex 0.156.1 rejection and 0.159.2 acceptance (above).
  • The OpenAI API-key path could not be exercised live (no valid platform key on the box); the id is confirmed on OpenAI's API surface and in pi-ai's openai provider.

Suites:

  • Runner: pnpm run test:unit → 3914/3922 pass; the 8 failures (gateway/MCP/orchestration env-dependent tests) fail identically on a pristine release/v0.121.6 checkout with pi 0.87.1, so they are pre-existing and unrelated. pnpm run typecheck clean.
  • SDK: oss/tests/pytest/unit/agents → 1896 passed (includes the model-catalog coverage lock and capability guards).
  • API: subscription secrets + vault login route tests → 47 passed.
  • git diff --check clean; no GitButler-Conflict trailers.

Review notes

  • The curated ratings for GPT-6.1 Sol (cost 3 / intelligence 5 / speed 3) follow OpenAI's own positioning (near-Astra capability at Sol pricing) — judgment call, happy to adjust.
  • The together_ai default swap (Kimi-K2.7-Code → Kimi-K3) is a side effect of the pi-ai catalog rename; a saved connection with its own model list is unaffected.

OpenAI released GPT-6.1 Sol (gpt-6.1-sol) at DevDay on 2026-09-29:
$2/$10 per Mtok, $0.10 cached input, 272K context, near-Astra agentic
coding at one fifth of Astra's price. Pi 0.99.1 (the direct successor
of the pinned 0.87.1) carries it on the openai, azure, and openai-codex
providers and makes it the Codex default.

- Bump @earendil-works/pi-coding-agent and the pi-coding-agent-pi-ai
  alias to 0.99.1; update PINNED_PI_VERSION, the Daytona snapshot pin,
  and the provider-cost patch spec (new CLI chunk hash; both anchors
  verified to bind exactly once in 0.99.1).
- Regenerate pi_models.generated.json from pi-ai 0.99.1 (541 -> 555
  models; adds openai/gpt-6.1-sol, openai-codex/gpt-6.1-sol, the
  openrouter variants, and anthropic/claude-sonnet-5-5).
- Retire the Sonnet 5.5 curated addition and its runner-side
  registration entry; pi 0.99.1 carries the model itself.
- Add gpt-6.1-sol to supported_llm_models["openai"],
  PROVIDER_DEFAULT_MODELS["openai"], PI_SUBSCRIPTION_MODELS and the
  API's SUBSCRIPTION_PROVIDER_MODELS copy (first entry: it is Pi's new
  Codex default), plus curated overlay labels/descriptions/ratings.
- Update the together_ai default Kimi-K2.7-Code -> Kimi-K3 (renamed in
  the pi-ai catalog).
- Mirror Pi 0.99's bash-tool change in the in-process engine: a
  non-zero exit returns an isError result instead of throwing.

Not in this change: the Codex harness. The pinned Codex CLI 0.156.1 is
rejected by the ChatGPT backend for gpt-6.1-sol ("model is not
supported when using Codex with a ChatGPT account", verified live);
Codex 0.159.2 accepts it, but that needs a codex-acp 1.13.1 -> 2.x
major bump, tracked separately. CODEX_MODELS is unchanged.

Cost tracking needs no code change: the API prices spans from litellm's
live price map, which already carries gpt-6.1-sol (verified:
resolve_pricing_model("gpt-6.1-sol") prices 5031/10 tokens to
$0.010062/$0.0001, matching Pi's billed numbers).
@coderabbitai

coderabbitai Bot commented Sep 30, 2026 •

Copy link
Copy Markdown
Contributor

Review in Change Stack →

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Repository YAML (base), Organization UI (inherited)

Review profile: CHILL

Plan: Advanced

Run ID: be50099f-6efa-4ccf-b6ef-a6dc5ae601a7

📥 Commits

Reviewing files that changed from the base of the PR and between b76ebc1 and 9b236aa.

⛔ Files ignored due to path filters (2)
  • sdks/python/agenta/sdk/agents/data/pi_models.generated.json is excluded by !**/*.generated.*
  • services/runner/pnpm-lock.yaml is excluded by !**/pnpm-lock.yaml
📒 Files selected for processing (1)
  • sdks/python/agenta/sdk/utils/assets.py
📝 Summary

Summary by CodeRabbit

  • New Features
    • Added GPT-6.1 Sol to the available OpenAI and ChatGPT subscription model options, including Codex.
    • Updated the Together AI model options to use Kimi K3 instead of Kimi K2.7-Code.
  • Improvements
    • Failed Bash commands now return their exit status and available output details as tool results, rather than raising an error.

Walkthrough

Pi is updated to version 0.99.1. Model catalogs and provider lists add GPT-6.1 Sol, and the Together AI default changes to Kimi-K3. The Bash tool now returns non-zero command exits as error-marked results.

Changes

Pi runtime and model behavior

Layer / File(s) Summary
Pi runtime and catalog alignment
services/runner/package.json, services/runner/images/sandbox/daytona/*, services/runner/src/engines/sandbox_agent/daytona.ts, sdks/python/agenta/sdk/agents/data/pi_models.curated.json, services/runner/src/engines/sandbox_agent/pi-model-config.ts, services/runner/src/tools/pi-provider-cost-patch.*, services/runner/src/tools/pi-validation-patch.ts, services/runner/tests/unit/sandbox-agent-pi-model-registration.test.ts, sdks/python/oss/tests/pytest/unit/agents/connections/test_model_catalog.py
Pi pins change to 0.99.1, and the cost patch targets the updated CLI bundle. GPT-6.1 Sol and Sonnet 5.5 are checked as Pi-generated catalog entries; neither produces a registration plan.
Provider model lists and defaults
api/oss/src/core/secrets/enums.py, api/oss/tests/pytest/unit/secrets/test_subscription_provider.py, api/oss/tests/pytest/unit/vault/test_subscription_login_routes.py, sdks/python/agenta/sdk/agents/capabilities.py, sdks/python/agenta/sdk/utils/assets.py, sdks/python/oss/tests/pytest/unit/agents/connections/test_capabilities.py, sdks/python/oss/tests/pytest/unit/agents/connections/test_model_catalog.py
ChatGPT, OpenAI, and Codex model lists add GPT-6.1 Sol. The Together AI default changes to Kimi-K3, and the tests expect the updated lists and ordering.
Bash non-zero exit results
services/runner/src/engines/inprocess/tools/bash-tool.ts, services/runner/tests/unit/inprocess/bash-tool.test.ts
Non-zero command exits return an error-marked result with exit status and applicable truncation details. The test helper treats returned error-marked results as failures.

Priority: ⬇️ Low

Estimated code review effort: 3 (Moderate) | ~20 minutes

Merge Risk: 🔵 Low · up to b76eb

The Together AI prompt model list can omit the new Kimi-K3 default. Align that list with the default catalog; otherwise the change is mergeable with awareness of this bounded inconsistency.

Security Architecture Review

Security architecture risk: 🔵 Low · up to b76eb

The inspected changes preserve existing authorization checks and keep command failures visible. No introduced security issue was established, but incomplete dependency and rollout coverage leaves residual uncertainty.

Retained concerns
No architecture-level concerns identified.

Security review details

Security Blast Radius

  • inferred — The upgrade reaches both runner-side execution and Pi installations in Daytona sandboxes, so compatibility effects can propagate beyond catalog display. The inspected model-selection changes remain within existing authorized provider families; they do not establish new cross-provider credential authority.

Trust Boundaries and Controls

  • observed — The Bash result change continues to use the existing workspace command wrapper and propagates the caller's cancellation signal. Daytona setup rejects a missing permission extension when built-in tool gating is active, and rejects a missing extension when a model endpoint override requires it.

Resilience and Maintainability Implications

  • observed — Bash execution retains finally-based timer cleanup and output-tracking release for returned and thrown outcomes. Unknown command outcomes explicitly warn against blind repetition. Sandbox acquisition errors invoke environment teardown before returning failure, preserving the inspected failure-containment path.
🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name Status Explanation Resolution
Docstring Coverage ⚠️ Warning Docstring coverage is 0.00% which is insufficient. The required threshold is 60.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 9 functions across 15 files. (4 skipped: 4… Write docstrings for the functions missing them to satisfy the coverage threshold.
✅ Passed checks (4 passed)
Check name Status Explanation
Linked Issues check ✅ Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check ✅ Passed Check skipped because no linked issues were found for this pull request.
Title check ✅ Passed The title clearly summarizes the primary changes: adding GPT-6.1 Sol, upgrading Pi to 0.99.1, and refreshing model catalogs.
Description check ✅ Passed The description is directly related to the changeset and provides detailed context, scope, manual QA requirements, and test results.
Full details: Docstring Coverage

Explanation

Docstring coverage is 0.00% which is insufficient. The required threshold is 60.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 9 functions across 15 files. (4 skipped: 4 unsupported.)

  • Fix all pre-merge checks with AI
✨ Finishing Touches
📝 Generate docstrings
  • Commit to this branch
  • Create a new PR
🧪 Generate unit tests (beta)
  • Commit to this branch
  • Create a new PR

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@coderabbitai coderabbitai Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actionable comments posted: 1


ℹ️ Review info
⚙️ Run configuration

Configuration used: Repository YAML (base), Organization UI (inherited)

Review profile: CHILL

Plan: Advanced

Run ID: 6d24a845-94b1-458f-b16c-b4b56c97d3f1

📥 Commits

Reviewing files that changed from the base of the PR and between f5092b1 and b76ebc1.

⛔ Files ignored due to path filters (2)
  • sdks/python/agenta/sdk/agents/data/pi_models.generated.json is excluded by !**/*.generated.*
  • services/runner/pnpm-lock.yaml is excluded by !**/pnpm-lock.yaml
📒 Files selected for processing (19)
  • api/oss/src/core/secrets/enums.py
  • api/oss/tests/pytest/unit/secrets/test_subscription_provider.py
  • api/oss/tests/pytest/unit/vault/test_subscription_login_routes.py
  • sdks/python/agenta/sdk/agents/capabilities.py
  • sdks/python/agenta/sdk/agents/data/pi_models.curated.json
  • sdks/python/agenta/sdk/utils/assets.py
  • sdks/python/oss/tests/pytest/unit/agents/connections/test_capabilities.py
  • sdks/python/oss/tests/pytest/unit/agents/connections/test_model_catalog.py
  • services/runner/images/sandbox/daytona/README.md
  • services/runner/images/sandbox/daytona/build_snapshot.py
  • services/runner/package.json
  • services/runner/src/engines/inprocess/tools/bash-tool.ts
  • services/runner/src/engines/sandbox_agent/daytona.ts
  • services/runner/src/engines/sandbox_agent/pi-model-config.ts
  • services/runner/src/tools/pi-provider-cost-patch.json
  • services/runner/src/tools/pi-provider-cost-patch.ts
  • services/runner/src/tools/pi-validation-patch.ts
  • services/runner/tests/unit/inprocess/bash-tool.test.ts
  • services/runner/tests/unit/sandbox-agent-pi-model-registration.test.ts

Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 7 remain after this review.

Comment thread sdks/python/agenta/sdk/utils/assets.py Resolved
@github-actions

github-actions Bot commented Sep 30, 2026 •

Copy link
Copy Markdown
Contributor

Railway Preview Environment

Item Value
Status stopped
Commit 9b236aac226b72211493ae647a9fd8efdc692400
Manual expiry (UTC) Not granted
CI deadline (UTC) Not granted
Workflow https://github.com/Agenta-AI/agenta/actions/runs/36908940262

PR closed or converted to draft. Environment absence verified; retained storage requires separate provider accounting.

Post /preview as a new comment to request a one-hour preview. Repeating it while active does not extend the hour.
Preview data is disposable. Expiry is enforced by the next successful cleanup sweep, not an exact timer.
Environment deletion stops its compute. Provider-retained storage and permanent services may still be billed.

…mi K3

Pi 0.99.1 no longer lists the two Together AI Kimi models, and litellm's
pinned registry never had them, so test_model_exists_in_litellm failed.
Pi now lists together_ai/moonshotai/Kimi-K3 in their place.
@mmabrouk
mmabrouk changed the base branch from release/v0.121.6 to release/v0.121.7 October 1, 2026 18:45
@mmabrouk
mmabrouk merged commit 9e20fc4 into release/v0.121.7 Oct 1, 2026
64 of 67 checks passed
mmabrouk added a commit that referenced this pull request Oct 1, 2026
* feat(channels): inbound files on Slack, Telegram, and WhatsApp voice/video

Pass inbound files to the agent as session attachments on every channel,
reusing the WhatsApp media path:

- WhatsApp: audio (voice notes) and video join image and document as media
  parts; only stickers, locations, and contact cards keep the fixed reply.
- Telegram: photo, document, voice, audio, video, and video_note become
  media parts; fetch_media downloads through getFile with the bot token
  (deployment token on the hosted path); files.receive capability on at
  the Bot API's 20 MB cap.
- Slack: each entry of a message's files array becomes a media part;
  fetch_media downloads url_private_download with the bot token and
  refuses an HTML login-page response.

Attachment delivery downstream is unchanged: native images are inlined
into the model prompt where supported, everything else lands as a file in
the session working directory. OpenSpec change: channel-inbound-media.

* refactor(channels): resolve Slack files at download time; one Telegram token hook

Codex review follow-up on the inbound-media commit:

- Slack stores only the file id on the inbox event; fetch_media resolves
  the current size and private URL through files.info at download time,
  so no private URL or size is persisted in channel_inbox_events. The
  mapping moves to slack/mapping.py next to the other pure mapping code,
  and a file-only share no longer carries an empty text part.
- Telegram: one _api_token hook replaces _media_token; the base _call
  and fetch_media both read it, and the hosted adapter's _call override
  is deleted (it only swapped the token).
- Telegram photo selection picks the largest size by area (file_size as
  tiebreaker) instead of trusting list order.
- Tests: files.info resolution, no-URL refusal, file-only share shape,
  hosted deployment-token download, unordered photo sizes, and explicit
  project-scoping assertions on the dispatcher attachment path.
- OpenSpec wording matches the code; README gains the change entry.

* docs(changelog): Configure permissions for your agent

* v0.121.7

* docs(changelog): new-models-and-cost-tracking (#7271)

* docs(changelog): docs-on-wide-screens (#7277)

* fix(chart): render the sandbox provider registry on the web workloads (#7189)

The web entrypoint maps AGENTA_RUNNER_ENABLED_SANDBOX_PROVIDERS to the
browser's NEXT_PUBLIC_AGENTA_ENABLED_SANDBOX_PROVIDERS. The chart rendered
the registry on the runner, api and services only, so the browser fell back
to "local", new agents were minted with sandbox.kind "local", and on a
daytona-only install every send failed with "sandbox 'local' is not enabled
on this deployment". Seen on GKE staging on 2026-09-27.

The web and web-mobile Deployments now carry both registry variables through
agenta.webOptionalEnv, from the same agentRunner.providers values the other
workloads read. A rendered-chart test covers the default, a daytona-only
install, and that the entry is rendered once.

* fix(channels): keep the Telegram token out of download errors; name unnamed files

A failed Telegram file download raised httpx's HTTPStatusError, whose
message carries the request URL. That URL holds the bot token, and the
inbox dispatcher logs the message, so a file that failed to download put
a working token in the logs. Raise a status-only error instead.

Files a platform sends without a name (Telegram photos, voice notes,
video notes; WhatsApp images and voice notes) were stored as `image`,
`audio`, or `video` with no extension. Name them
`<kind>-<send time>.<ext>` from the message's send time and the
detected type, so a retried dispatch keeps the same name.

The spec now states that the platform attachment limit, not each
channel's advertised max_bytes, is the effective cap.

* fix(deps): update hono to 4.13.5 (#7249)

* docs(changelog): Configure permissions for your agent

* v0.121.7

* docs(changelog): new-models-and-cost-tracking (#7271)

* docs(changelog): docs-on-wide-screens (#7277)

* fix(chart): render the sandbox provider registry on the web workloads (#7189)

The web entrypoint maps AGENTA_RUNNER_ENABLED_SANDBOX_PROVIDERS to the
browser's NEXT_PUBLIC_AGENTA_ENABLED_SANDBOX_PROVIDERS. The chart rendered
the registry on the runner, api and services only, so the browser fell back
to "local", new agents were minted with sandbox.kind "local", and on a
daytona-only install every send failed with "sandbox 'local' is not enabled
on this deployment". Seen on GKE staging on 2026-09-27.

The web and web-mobile Deployments now carry both registry variables through
agenta.webOptionalEnv, from the same agentRunner.providers values the other
workloads read. A rendered-chart test covers the default, a daytona-only
install, and that the entry is rendered once.

* fix(deps): update hono to 4.13.5

---------

Co-authored-by: mmabrouk <4510758+mmabrouk@users.noreply.github.com>
Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* fix(deps): update fast-uri to 3.1.7 (#7245)

Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* [chore] Add GPT-6.1 Sol: bump Pi to 0.99.1 and refresh model catalogs (#7252)

* [chore] Add GPT-6.1 Sol: bump Pi to 0.99.1 and refresh model catalogs

OpenAI released GPT-6.1 Sol (gpt-6.1-sol) at DevDay on 2026-09-29:
$2/$10 per Mtok, $0.10 cached input, 272K context, near-Astra agentic
coding at one fifth of Astra's price. Pi 0.99.1 (the direct successor
of the pinned 0.87.1) carries it on the openai, azure, and openai-codex
providers and makes it the Codex default.

- Bump @earendil-works/pi-coding-agent and the pi-coding-agent-pi-ai
  alias to 0.99.1; update PINNED_PI_VERSION, the Daytona snapshot pin,
  and the provider-cost patch spec (new CLI chunk hash; both anchors
  verified to bind exactly once in 0.99.1).
- Regenerate pi_models.generated.json from pi-ai 0.99.1 (541 -> 555
  models; adds openai/gpt-6.1-sol, openai-codex/gpt-6.1-sol, the
  openrouter variants, and anthropic/claude-sonnet-5-5).
- Retire the Sonnet 5.5 curated addition and its runner-side
  registration entry; pi 0.99.1 carries the model itself.
- Add gpt-6.1-sol to supported_llm_models["openai"],
  PROVIDER_DEFAULT_MODELS["openai"], PI_SUBSCRIPTION_MODELS and the
  API's SUBSCRIPTION_PROVIDER_MODELS copy (first entry: it is Pi's new
  Codex default), plus curated overlay labels/descriptions/ratings.
- Update the together_ai default Kimi-K2.7-Code -> Kimi-K3 (renamed in
  the pi-ai catalog).
- Mirror Pi 0.99's bash-tool change in the in-process engine: a
  non-zero exit returns an isError result instead of throwing.

Not in this change: the Codex harness. The pinned Codex CLI 0.156.1 is
rejected by the ChatGPT backend for gpt-6.1-sol ("model is not
supported when using Codex with a ChatGPT account", verified live);
Codex 0.159.2 accepts it, but that needs a codex-acp 1.13.1 -> 2.x
major bump, tracked separately. CODEX_MODELS is unchanged.

Cost tracking needs no code change: the API prices spans from litellm's
live price map, which already carries gpt-6.1-sol (verified:
resolve_pricing_model("gpt-6.1-sol") prices 5031/10 tokens to
$0.010062/$0.0001, matching Pi's billed numbers).

* fix(sdk): replace retired Together AI Kimi K2.6 and K2.7-Code with Kimi K3

Pi 0.99.1 no longer lists the two Together AI Kimi models, and litellm's
pinned registry never had them, so test_model_exists_in_litellm failed.
Pi now lists together_ai/moonshotai/Kimi-K3 in their place.

* chore(deps): bump undici from 8.10.0 to 8.10.2 in /services/runner (#7230)

Bumps [undici](https://github.com/nodejs/undici) from 8.10.0 to 8.10.2.
- [Release notes](https://github.com/nodejs/undici/releases)
- [Commits](nodejs/undici@v8.10.0...v8.10.2)

---
updated-dependencies:
- dependency-name: undici
  dependency-version: 8.10.2
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* fix(deps): update image-size to 2.0.3 (#7246)

Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* fix(deps): update mermaid to 11.16.1

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: Agenta Product Agent <team@agenta.ai>
Co-authored-by: mmabrouk <4510758+mmabrouk@users.noreply.github.com>
Co-authored-by: Agenta Product Agent <agent@agenta.ai>
Co-authored-by: Depandabot Bot <dependabot@agenta.ai>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
mmabrouk added a commit that referenced this pull request Oct 1, 2026
* feat(channels): inbound files on Slack, Telegram, and WhatsApp voice/video

Pass inbound files to the agent as session attachments on every channel,
reusing the WhatsApp media path:

- WhatsApp: audio (voice notes) and video join image and document as media
  parts; only stickers, locations, and contact cards keep the fixed reply.
- Telegram: photo, document, voice, audio, video, and video_note become
  media parts; fetch_media downloads through getFile with the bot token
  (deployment token on the hosted path); files.receive capability on at
  the Bot API's 20 MB cap.
- Slack: each entry of a message's files array becomes a media part;
  fetch_media downloads url_private_download with the bot token and
  refuses an HTML login-page response.

Attachment delivery downstream is unchanged: native images are inlined
into the model prompt where supported, everything else lands as a file in
the session working directory. OpenSpec change: channel-inbound-media.

* refactor(channels): resolve Slack files at download time; one Telegram token hook

Codex review follow-up on the inbound-media commit:

- Slack stores only the file id on the inbox event; fetch_media resolves
  the current size and private URL through files.info at download time,
  so no private URL or size is persisted in channel_inbox_events. The
  mapping moves to slack/mapping.py next to the other pure mapping code,
  and a file-only share no longer carries an empty text part.
- Telegram: one _api_token hook replaces _media_token; the base _call
  and fetch_media both read it, and the hosted adapter's _call override
  is deleted (it only swapped the token).
- Telegram photo selection picks the largest size by area (file_size as
  tiebreaker) instead of trusting list order.
- Tests: files.info resolution, no-URL refusal, file-only share shape,
  hosted deployment-token download, unordered photo sizes, and explicit
  project-scoping assertions on the dispatcher attachment path.
- OpenSpec wording matches the code; README gains the change entry.

* docs(changelog): Configure permissions for your agent

* v0.121.7

* docs(changelog): new-models-and-cost-tracking (#7271)

* docs(changelog): docs-on-wide-screens (#7277)

* fix(chart): render the sandbox provider registry on the web workloads (#7189)

The web entrypoint maps AGENTA_RUNNER_ENABLED_SANDBOX_PROVIDERS to the
browser's NEXT_PUBLIC_AGENTA_ENABLED_SANDBOX_PROVIDERS. The chart rendered
the registry on the runner, api and services only, so the browser fell back
to "local", new agents were minted with sandbox.kind "local", and on a
daytona-only install every send failed with "sandbox 'local' is not enabled
on this deployment". Seen on GKE staging on 2026-09-27.

The web and web-mobile Deployments now carry both registry variables through
agenta.webOptionalEnv, from the same agentRunner.providers values the other
workloads read. A rendered-chart test covers the default, a daytona-only
install, and that the entry is rendered once.

* fix(channels): keep the Telegram token out of download errors; name unnamed files

A failed Telegram file download raised httpx's HTTPStatusError, whose
message carries the request URL. That URL holds the bot token, and the
inbox dispatcher logs the message, so a file that failed to download put
a working token in the logs. Raise a status-only error instead.

Files a platform sends without a name (Telegram photos, voice notes,
video notes; WhatsApp images and voice notes) were stored as `image`,
`audio`, or `video` with no extension. Name them
`<kind>-<send time>.<ext>` from the message's send time and the
detected type, so a retried dispatch keeps the same name.

The spec now states that the platform attachment limit, not each
channel's advertised max_bytes, is the effective cap.

* fix(deps): update hono to 4.13.5 (#7249)

* docs(changelog): Configure permissions for your agent

* v0.121.7

* docs(changelog): new-models-and-cost-tracking (#7271)

* docs(changelog): docs-on-wide-screens (#7277)

* fix(chart): render the sandbox provider registry on the web workloads (#7189)

The web entrypoint maps AGENTA_RUNNER_ENABLED_SANDBOX_PROVIDERS to the
browser's NEXT_PUBLIC_AGENTA_ENABLED_SANDBOX_PROVIDERS. The chart rendered
the registry on the runner, api and services only, so the browser fell back
to "local", new agents were minted with sandbox.kind "local", and on a
daytona-only install every send failed with "sandbox 'local' is not enabled
on this deployment". Seen on GKE staging on 2026-09-27.

The web and web-mobile Deployments now carry both registry variables through
agenta.webOptionalEnv, from the same agentRunner.providers values the other
workloads read. A rendered-chart test covers the default, a daytona-only
install, and that the entry is rendered once.

* fix(deps): update hono to 4.13.5

---------

Co-authored-by: mmabrouk <4510758+mmabrouk@users.noreply.github.com>
Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* fix(deps): update fast-uri to 3.1.7 (#7245)

Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* [chore] Add GPT-6.1 Sol: bump Pi to 0.99.1 and refresh model catalogs (#7252)

* [chore] Add GPT-6.1 Sol: bump Pi to 0.99.1 and refresh model catalogs

OpenAI released GPT-6.1 Sol (gpt-6.1-sol) at DevDay on 2026-09-29:
$2/$10 per Mtok, $0.10 cached input, 272K context, near-Astra agentic
coding at one fifth of Astra's price. Pi 0.99.1 (the direct successor
of the pinned 0.87.1) carries it on the openai, azure, and openai-codex
providers and makes it the Codex default.

- Bump @earendil-works/pi-coding-agent and the pi-coding-agent-pi-ai
  alias to 0.99.1; update PINNED_PI_VERSION, the Daytona snapshot pin,
  and the provider-cost patch spec (new CLI chunk hash; both anchors
  verified to bind exactly once in 0.99.1).
- Regenerate pi_models.generated.json from pi-ai 0.99.1 (541 -> 555
  models; adds openai/gpt-6.1-sol, openai-codex/gpt-6.1-sol, the
  openrouter variants, and anthropic/claude-sonnet-5-5).
- Retire the Sonnet 5.5 curated addition and its runner-side
  registration entry; pi 0.99.1 carries the model itself.
- Add gpt-6.1-sol to supported_llm_models["openai"],
  PROVIDER_DEFAULT_MODELS["openai"], PI_SUBSCRIPTION_MODELS and the
  API's SUBSCRIPTION_PROVIDER_MODELS copy (first entry: it is Pi's new
  Codex default), plus curated overlay labels/descriptions/ratings.
- Update the together_ai default Kimi-K2.7-Code -> Kimi-K3 (renamed in
  the pi-ai catalog).
- Mirror Pi 0.99's bash-tool change in the in-process engine: a
  non-zero exit returns an isError result instead of throwing.

Not in this change: the Codex harness. The pinned Codex CLI 0.156.1 is
rejected by the ChatGPT backend for gpt-6.1-sol ("model is not
supported when using Codex with a ChatGPT account", verified live);
Codex 0.159.2 accepts it, but that needs a codex-acp 1.13.1 -> 2.x
major bump, tracked separately. CODEX_MODELS is unchanged.

Cost tracking needs no code change: the API prices spans from litellm's
live price map, which already carries gpt-6.1-sol (verified:
resolve_pricing_model("gpt-6.1-sol") prices 5031/10 tokens to
$0.010062/$0.0001, matching Pi's billed numbers).

* fix(sdk): replace retired Together AI Kimi K2.6 and K2.7-Code with Kimi K3

Pi 0.99.1 no longer lists the two Together AI Kimi models, and litellm's
pinned registry never had them, so test_model_exists_in_litellm failed.
Pi now lists together_ai/moonshotai/Kimi-K3 in their place.

* chore(deps): bump undici from 8.10.0 to 8.10.2 in /services/runner (#7230)

Bumps [undici](https://github.com/nodejs/undici) from 8.10.0 to 8.10.2.
- [Release notes](https://github.com/nodejs/undici/releases)
- [Commits](nodejs/undici@v8.10.0...v8.10.2)

---
updated-dependencies:
- dependency-name: undici
  dependency-version: 8.10.2
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* fix(deps): update image-size to 2.0.3 (#7246)

Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* fix(deps): update mermaid to 11.16.1 (#7253)

* feat(channels): inbound files on Slack, Telegram, and WhatsApp voice/video

Pass inbound files to the agent as session attachments on every channel,
reusing the WhatsApp media path:

- WhatsApp: audio (voice notes) and video join image and document as media
  parts; only stickers, locations, and contact cards keep the fixed reply.
- Telegram: photo, document, voice, audio, video, and video_note become
  media parts; fetch_media downloads through getFile with the bot token
  (deployment token on the hosted path); files.receive capability on at
  the Bot API's 20 MB cap.
- Slack: each entry of a message's files array becomes a media part;
  fetch_media downloads url_private_download with the bot token and
  refuses an HTML login-page response.

Attachment delivery downstream is unchanged: native images are inlined
into the model prompt where supported, everything else lands as a file in
the session working directory. OpenSpec change: channel-inbound-media.

* refactor(channels): resolve Slack files at download time; one Telegram token hook

Codex review follow-up on the inbound-media commit:

- Slack stores only the file id on the inbox event; fetch_media resolves
  the current size and private URL through files.info at download time,
  so no private URL or size is persisted in channel_inbox_events. The
  mapping moves to slack/mapping.py next to the other pure mapping code,
  and a file-only share no longer carries an empty text part.
- Telegram: one _api_token hook replaces _media_token; the base _call
  and fetch_media both read it, and the hosted adapter's _call override
  is deleted (it only swapped the token).
- Telegram photo selection picks the largest size by area (file_size as
  tiebreaker) instead of trusting list order.
- Tests: files.info resolution, no-URL refusal, file-only share shape,
  hosted deployment-token download, unordered photo sizes, and explicit
  project-scoping assertions on the dispatcher attachment path.
- OpenSpec wording matches the code; README gains the change entry.

* docs(changelog): Configure permissions for your agent

* v0.121.7

* docs(changelog): new-models-and-cost-tracking (#7271)

* docs(changelog): docs-on-wide-screens (#7277)

* fix(chart): render the sandbox provider registry on the web workloads (#7189)

The web entrypoint maps AGENTA_RUNNER_ENABLED_SANDBOX_PROVIDERS to the
browser's NEXT_PUBLIC_AGENTA_ENABLED_SANDBOX_PROVIDERS. The chart rendered
the registry on the runner, api and services only, so the browser fell back
to "local", new agents were minted with sandbox.kind "local", and on a
daytona-only install every send failed with "sandbox 'local' is not enabled
on this deployment". Seen on GKE staging on 2026-09-27.

The web and web-mobile Deployments now carry both registry variables through
agenta.webOptionalEnv, from the same agentRunner.providers values the other
workloads read. A rendered-chart test covers the default, a daytona-only
install, and that the entry is rendered once.

* fix(channels): keep the Telegram token out of download errors; name unnamed files

A failed Telegram file download raised httpx's HTTPStatusError, whose
message carries the request URL. That URL holds the bot token, and the
inbox dispatcher logs the message, so a file that failed to download put
a working token in the logs. Raise a status-only error instead.

Files a platform sends without a name (Telegram photos, voice notes,
video notes; WhatsApp images and voice notes) were stored as `image`,
`audio`, or `video` with no extension. Name them
`<kind>-<send time>.<ext>` from the message's send time and the
detected type, so a retried dispatch keeps the same name.

The spec now states that the platform attachment limit, not each
channel's advertised max_bytes, is the effective cap.

* fix(deps): update hono to 4.13.5 (#7249)

* docs(changelog): Configure permissions for your agent

* v0.121.7

* docs(changelog): new-models-and-cost-tracking (#7271)

* docs(changelog): docs-on-wide-screens (#7277)

* fix(chart): render the sandbox provider registry on the web workloads (#7189)

The web entrypoint maps AGENTA_RUNNER_ENABLED_SANDBOX_PROVIDERS to the
browser's NEXT_PUBLIC_AGENTA_ENABLED_SANDBOX_PROVIDERS. The chart rendered
the registry on the runner, api and services only, so the browser fell back
to "local", new agents were minted with sandbox.kind "local", and on a
daytona-only install every send failed with "sandbox 'local' is not enabled
on this deployment". Seen on GKE staging on 2026-09-27.

The web and web-mobile Deployments now carry both registry variables through
agenta.webOptionalEnv, from the same agentRunner.providers values the other
workloads read. A rendered-chart test covers the default, a daytona-only
install, and that the entry is rendered once.

* fix(deps): update hono to 4.13.5

---------

Co-authored-by: mmabrouk <4510758+mmabrouk@users.noreply.github.com>
Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* fix(deps): update fast-uri to 3.1.7 (#7245)

Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* [chore] Add GPT-6.1 Sol: bump Pi to 0.99.1 and refresh model catalogs (#7252)

* [chore] Add GPT-6.1 Sol: bump Pi to 0.99.1 and refresh model catalogs

OpenAI released GPT-6.1 Sol (gpt-6.1-sol) at DevDay on 2026-09-29:
$2/$10 per Mtok, $0.10 cached input, 272K context, near-Astra agentic
coding at one fifth of Astra's price. Pi 0.99.1 (the direct successor
of the pinned 0.87.1) carries it on the openai, azure, and openai-codex
providers and makes it the Codex default.

- Bump @earendil-works/pi-coding-agent and the pi-coding-agent-pi-ai
  alias to 0.99.1; update PINNED_PI_VERSION, the Daytona snapshot pin,
  and the provider-cost patch spec (new CLI chunk hash; both anchors
  verified to bind exactly once in 0.99.1).
- Regenerate pi_models.generated.json from pi-ai 0.99.1 (541 -> 555
  models; adds openai/gpt-6.1-sol, openai-codex/gpt-6.1-sol, the
  openrouter variants, and anthropic/claude-sonnet-5-5).
- Retire the Sonnet 5.5 curated addition and its runner-side
  registration entry; pi 0.99.1 carries the model itself.
- Add gpt-6.1-sol to supported_llm_models["openai"],
  PROVIDER_DEFAULT_MODELS["openai"], PI_SUBSCRIPTION_MODELS and the
  API's SUBSCRIPTION_PROVIDER_MODELS copy (first entry: it is Pi's new
  Codex default), plus curated overlay labels/descriptions/ratings.
- Update the together_ai default Kimi-K2.7-Code -> Kimi-K3 (renamed in
  the pi-ai catalog).
- Mirror Pi 0.99's bash-tool change in the in-process engine: a
  non-zero exit returns an isError result instead of throwing.

Not in this change: the Codex harness. The pinned Codex CLI 0.156.1 is
rejected by the ChatGPT backend for gpt-6.1-sol ("model is not
supported when using Codex with a ChatGPT account", verified live);
Codex 0.159.2 accepts it, but that needs a codex-acp 1.13.1 -> 2.x
major bump, tracked separately. CODEX_MODELS is unchanged.

Cost tracking needs no code change: the API prices spans from litellm's
live price map, which already carries gpt-6.1-sol (verified:
resolve_pricing_model("gpt-6.1-sol") prices 5031/10 tokens to
$0.010062/$0.0001, matching Pi's billed numbers).

* fix(sdk): replace retired Together AI Kimi K2.6 and K2.7-Code with Kimi K3

Pi 0.99.1 no longer lists the two Together AI Kimi models, and litellm's
pinned registry never had them, so test_model_exists_in_litellm failed.
Pi now lists together_ai/moonshotai/Kimi-K3 in their place.

* chore(deps): bump undici from 8.10.0 to 8.10.2 in /services/runner (#7230)

Bumps [undici](https://github.com/nodejs/undici) from 8.10.0 to 8.10.2.
- [Release notes](https://github.com/nodejs/undici/releases)
- [Commits](nodejs/undici@v8.10.0...v8.10.2)

---
updated-dependencies:
- dependency-name: undici
  dependency-version: 8.10.2
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* fix(deps): update image-size to 2.0.3 (#7246)

Co-authored-by: Depandabot Bot <dependabot@agenta.ai>

* fix(deps): update mermaid to 11.16.1

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: Agenta Product Agent <team@agenta.ai>
Co-authored-by: mmabrouk <4510758+mmabrouk@users.noreply.github.com>
Co-authored-by: Agenta Product Agent <agent@agenta.ai>
Co-authored-by: Depandabot Bot <dependabot@agenta.ai>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* chore(deps): bump axios from 1.18.0 to 1.20.0 in /web/tests (#7279)

* chore(deps): bump axios from 1.18.0 to 1.20.0 in /web/tests

Bumps [axios](https://github.com/axios/axios) from 1.18.0 to 1.20.0.
- [Release notes](https://github.com/axios/axios/releases)
- [Changelog](https://github.com/axios/axios/blob/v1.x/CHANGELOG.md)
- [Commits](axios/axios@v1.18.0...v1.20.0)

---
updated-dependencies:
- dependency-name: axios
  dependency-version: 1.20.0
  dependency-type: direct:production
...

Signed-off-by: dependabot[bot] <support@github.com>

* chore(deps): update web lockfile for axios 1.20.0

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Agenta Product Agent <team@agenta.ai>

* fix(deps): update Pillow to 12.3.0

* fix(runner): install the sandbox lock with --no-deps so Pillow 12 can sit next to moviepy

moviepy 2.2.1 (the latest) caps pillow<12, so the hash-locked install refused
Pillow 12.3.0 and the runner image build failed. The lock is complete, so
--no-deps installs it as written. Also lists every Pillow 12.3.0 wheel hash
and drops the stale 11.3.0 ones.

---------

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: Agenta Product Agent <team@agenta.ai>
Co-authored-by: mmabrouk <4510758+mmabrouk@users.noreply.github.com>
Co-authored-by: Agenta Product Agent <agent@agenta.ai>
Co-authored-by: Depandabot Bot <dependabot@agenta.ai>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
mmabrouk pushed a commit that referenced this pull request Oct 2, 2026
#7252 pinned Pi 0.99.1 in the snapshot recipe but kept recipe version 1, so the
runner asked for a snapshot built from the old recipe and the fingerprint test
failed. Bump the version to 2, regenerate the fingerprint, and point the docs,
env examples, and Helm values at agenta-agent-sandbox-v2.
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.

1 participant