claude_sdk_otf: Claude SDK agent that encodes KB on-the-fly (DEV-1505) by ZmeiGorynych · Pull Request #9 · MotleyAI/bird-agents

ZmeiGorynych · 2026-05-29T14:23:24Z

Summary

Adds --framework claude_sdk_otf: a single Claude Agent SDK agent that encodes the relevant knowledge-base (KB) items into the per-task SLayer store on the fly — off the deterministic OTF cache (base models + KB pre-loaded as memories) — and then builds the final query off the named columns/measures it created instead of inlining everything. No forced stages or recursion (unlike pydantic_ai_otf_encode).

slayer-query-mode only; eval modes a-interact (mini-interact) + one-shot (LiveSQLBench); always --slayer-setup on-the-fly.
Reuses claude_sdk.agent's contextvar plumbing + native user-sim/submit/knowledge tools; adds the SLayer MCP write tools (create_model/edit_model/save_memory/query_nested/validate_models). One-shot drops ask_user (decides autonomously).
Prompts distilled from pydantic_ai_otf_encode's _STYLE_GUIDE (string-norm, cross-model access via declared joins, no invented joins, host choice, [kb=N] self-annotation) + claude_sdk's decompose/ask-user discipline — synthetic examples only.
Lifts a shared cache-only resolve_otf_task_storage_dir into slayer_otf/runtime.py so this Claude-SDK-only adapter doesn't import the pydantic_ai adapter packages. Recursive adapter left untouched (its tests pin its private copy).
Cloud: cache-only by construction — claude_sdk_otf already falls into the cache-only branches of gcs.slayer_artifact_name / driver._slayer_uploads_for / ray_app._slayer_artifacts_for; the reference upload-back/merge gates correctly skip it. No cloud source edits.
The save-time SLayer-write validator is intentionally deferred → DEV-1506.

Local runs must launch from a non-Claude-Code shell (the SDK spawns the claude CLI, which fails nested inside Claude Code). Cloud Ray actors don't nest, so cloud is unaffected.

Files

New: agents/claude_sdk_otf/{__init__,agent,prompts}.py
Modified: run.py (choices, _make_runner, _validate_slayer_setup, _validate_one_shot_framework, _maybe_force_wipe_otf, help text), slayer_otf/runtime.py + __init__.py (shared resolver), README.md
Tests: tests/test_claude_sdk_otf_agent.py, tests/test_claude_sdk_otf_run_wiring.py

Test plan

Full non-integration suite green (1491 passed, 94 skipped)
Mode gating (raw rejected, bad eval_mode rejected, on-the-fly required, non-Anthropic skip, one-shot⟹one-shot-benchmark)
Storage uses the cache resolver (NOT committed resolve_task_storage_dir); allowed_tools include write tools; one-shot omits ask_user
Cloud combo maps to slayer_otf_cache only (no reference / upload-back)
Import isolation (no pydantic_ai adapter packages), prompt hygiene + no-inlining, usage capture
Cloud smoke (bird-interact-cloud submit --framework claude_sdk_otf ...) — run manually
LiveSQLBench one-shot smoke

🤖 Generated with Claude Code

Summary by CodeRabbit

New Features
- Added a new "Claude SDK on-the-fly KB encode" framework selectable via CLI; supports a-interact and one-shot eval modes and a new --reasoning-effort option.
CLI / Cloud
- --dataset is now required for submit/run; cloud submit/resubmit propagate reasoning-effort and restricts cached artifacts to the OTF cache.
Documentation
- README updated with framework details, constraints, examples, and Agent Frameworks table entry.
Tests
- Extensive tests added for agent behavior, CLI/cloud wiring, cache handling, and token-usage aggregation.

…505) New Claude Agent SDK framework that encodes the relevant KB items into the per-task SLayer store off the deterministic OTF cache (KB pre-loaded as memories), then builds the final query off the named columns/measures it created instead of inlining everything. slayer-query-mode only; eval modes a-interact + one-shot; always --slayer-setup on-the-fly. - Reuses claude_sdk's contextvar plumbing + native user-sim/submit/knowledge tools; adds the SLayer MCP write tools (create_model/edit_model/save_memory/ query_nested/validate_models). One-shot drops ask_user. - Lifts a shared cache-only resolver (resolve_otf_task_storage_dir) into slayer_otf.runtime so the Claude-SDK-only adapter doesn't import the pydantic_ai adapter packages. - run.py: --framework choices, _make_runner branch, _validate_slayer_setup (on-the-fly required), _validate_one_shot_framework, _maybe_force_wipe_otf, Anthropic-lock help text. Cloud is cache-only by construction (no edits). - Save-time SLayer-write validator deferred to DEV-1506. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

linear · 2026-05-29T14:23:27Z

DEV-1505

coderabbitai · 2026-05-29T14:23:36Z

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

@coderabbitai resume to resume automatic reviews.
@coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

▶️ Resume reviews
🔍 Trigger review

📝 Walkthrough

Walkthrough

This PR adds a new claude_sdk_otf agent that performs on-the-fly KB encoding using a deterministic OTF cache and per-task SLayer write tools, and integrates prompts, runtime storage resolution, CLI/run/cloud wiring, and comprehensive tests.

Changes

Claude SDK OTF Agent Framework

Layer / File(s)	Summary
Documentation and module structure `README.md`, `src/bird_interact_agents/agents/claude_sdk_otf/__init__.py`	README adds on-the-fly KB encoding framework description, examples, and a new Agent Frameworks table row; package `__init__` re-exports `ClaudeSDKOtfAgent`.
Prompt templates for on-the-fly KB encoding `src/bird_interact_agents/agents/claude_sdk_otf/prompts.py`	New prompts module with a strict `submit_query` JSON contract, encode-then-query workflow, dependency/join/normalization constraints, and two system prompts (`a-interact` with `ask_user`, and `one-shot`).
On-the-fly task storage and cache resolution `src/bird_interact_agents/slayer_otf/runtime.py`, `src/bird_interact_agents/slayer_otf/__init__.py`	Adds `_otf_work_dir` and `resolve_otf_task_storage_dir()` to compute sorted `deleted_kb_ids`, ensure per-DB deterministic OTF cache, and materialize per-task scratch storage via `prepare_task_storage`; re-exported from package.
Agent helpers and SDK usage accumulation `src/bird_interact_agents/agents/claude_sdk_otf/agent.py`, `src/bird_interact_agents/agents/claude_sdk/agent.py`	Defines `SLAYER_MCP_TOOLS`, tool selection and turn-cap helpers, constructor validation for `slayer_setup` and `reasoning_effort`, and adds `accumulate_assistant_usage` plus tolerant usage extraction for Claude SDK streamed messages.
Agent task execution and result handling `src/bird_interact_agents/agents/claude_sdk_otf/agent.py` (`run_task`, error/success paths)	`run_task` validates modes, resolves per-task OTF storage, populates context, registers MCP servers (native tools + SLayer), constrains allowed tools, calls `ClaudeSDKClient.query` with streaming, aggregates token usage (including cache reads), enforces turn caps, and returns serialized trajectory/usage plus slayer context on success or error.
CLI and run.py framework wiring `src/bird_interact_agents/run.py`	Includes `claude_sdk_otf` in `--framework` choices; expands validation to require `--slayer-setup on-the-fly` for this framework where applicable; threads new `--reasoning-effort` into runner/agent construction; `_maybe_force_wipe_otf` wipes OTF cache for the framework; `_make_runner` dispatches to `ClaudeSDKOtfAgent`.
Cloud CLI/driver/ray_app wiring `src/bird_interact_agents/cloud/cli.py`, `src/bird_interact_agents/cloud/driver.py`, `src/bird_interact_agents/cloud/ray_app.py`	Cloud submit CLI accepts `--reasoning-effort`; driver persists `reasoning_effort` in manifest and appends CLI arg for submit/resubmit when present; `ray_app` threads `reasoning_effort` into `run_pool`, cached runners, and `_run_one_task_async`.
Agent unit tests `tests/test_claude_sdk_otf_agent.py`	Unit tests covering constructor gating, eval-mode tool selection, prompt hygiene/variants, `run_task` gating (mode/model/benchmark), stubbed resolver/materialize/client interactions, `reasoning_effort` behavior, LiveSQLBench specifics, token-usage aggregation, and import isolation.
Run/CLI integration and cloud tests `tests/test_claude_sdk_otf_run_wiring.py`	Tests verifying CLI `--framework` includes `claude_sdk_otf`, `run_evaluation` routes to the OTF agent with `slayer_setup` and `reasoning_effort`, validation guard behaviors, otf cache wipe behavior, CLI rejection of invalid `--slayer-setup` combinations, cloud artifact mapping (cache-only), and resubmit argument preservation.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

MotleyAI/bird-agents#2: Overlapping OTF runtime and cache/force-wipe behavior changes.
MotleyAI/bird-agents#5: Related one-shot/on-the-fly validation and wiring adjustments.

Poem

🐰 I hop through caches, prompts in tow,
On-the-fly I help the KBs grow,
SLayer tools scribble, models converse,
Tests hum a tune to check each verse,
Claude and I encode, then let queries flow.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 37.62% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title directly and clearly describes the main change: adding a Claude SDK agent (`claude_sdk_otf`) that encodes the knowledge base on-the-fly, which is the primary objective of this comprehensive PR.
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.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch egor/dev-1505-claude-code-sdk-as-agent-for-the-benchmarks

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

coderabbitai

Actionable comments posted: 2

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

src/bird_interact_agents/run.py (1)
1023-1030: ⚠️ Potential issue | 🟡 Minor | ⚡ Quick win

Refresh the CLI help for --slayer-setup.

Exposing claude_sdk_otf here makes --help misleading unless the --slayer-setup help text is updated too. It still says on-the-fly is only valid for pydantic_ai_recursive in a-interact, while _validate_slayer_setup() now accepts pydantic_ai_otf_encode, claude_sdk_otf, and one-shot.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@src/bird_interact_agents/run.py` around lines 1023 - 1030, Update the CLI
help text for the "--slayer-setup" option to reflect the expanded valid values
that _validate_slayer_setup() now accepts (pydantic_ai_otf_encode,
claude_sdk_otf, and one-shot) instead of only mentioning on-the-fly for
pydantic_ai_recursive; find the help string that documents "--slayer-setup" in
the CLI parser and revise its description to list or describe the new valid
frameworks (including claude_sdk_otf and pydantic_ai_otf_encode) and any
behavior differences so the --help output matches the validation logic in
_validate_slayer_setup().

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/bird_interact_agents/slayer_otf/runtime.py`:
- Around line 45-58: _otf_work_dir currently returns a stable path per
instance_id which causes concurrent runs to clash (see prepare_task_storage
deleting <work_dir>/<db>); change _otf_work_dir to create a unique
per-invocation scratch directory (e.g., append a short UUID or timestamp+PID or
use tempfile.mkdtemp/TemporaryDirectory) so each call returns a distinct path,
and update callers that assume a persistent path accordingly (refer to function
_otf_work_dir and prepare_task_storage to locate the usage).

In `@tests/test_claude_sdk_otf_run_wiring.py`:
- Line 76: Replace hardcoded "/tmp" test paths with the pytest tmp_path fixture:
find the call sites that pass data_path="/tmp/x.jsonl" and data_dir="/tmp" (and
the similar occurrences on the other two lines) and change them to use tmp_path
(e.g., data_path should be the stringified tmp_path / "x.jsonl" and data_dir the
stringified tmp_path); ensure the test function signature includes the tmp_path
parameter if it doesn't already so the fixture is available.

---

Outside diff comments:
In `@src/bird_interact_agents/run.py`:
- Around line 1023-1030: Update the CLI help text for the "--slayer-setup"
option to reflect the expanded valid values that _validate_slayer_setup() now
accepts (pydantic_ai_otf_encode, claude_sdk_otf, and one-shot) instead of only
mentioning on-the-fly for pydantic_ai_recursive; find the help string that
documents "--slayer-setup" in the CLI parser and revise its description to list
or describe the new valid frameworks (including claude_sdk_otf and
pydantic_ai_otf_encode) and any behavior differences so the --help output
matches the validation logic in _validate_slayer_setup().

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 52331b57-a8b5-4c4d-acbd-8c11d92a3f76

📥 Commits

Reviewing files that changed from the base of the PR and between 88e8b6f and 959a201.

📒 Files selected for processing (9)

README.md
src/bird_interact_agents/agents/claude_sdk_otf/__init__.py
src/bird_interact_agents/agents/claude_sdk_otf/agent.py
src/bird_interact_agents/agents/claude_sdk_otf/prompts.py
src/bird_interact_agents/run.py
src/bird_interact_agents/slayer_otf/__init__.py
src/bird_interact_agents/slayer_otf/runtime.py
tests/test_claude_sdk_otf_agent.py
tests/test_claude_sdk_otf_run_wiring.py

ZmeiGorynych · 2026-05-29T14:32:42Z

@CodeRabbit why does it say "review skipped" and yet there are review comments?

coderabbitai · 2026-05-29T14:33:09Z

Tip

For best results, initiate chat on the files or code changes.

@ZmeiGorynych Great question — those two things refer to different events:

"Review skipped" appears when the most recent push or trigger event did not generate a new review pass. Common reasons include: the PR was already reviewed on a prior commit, the incremental diff was very small, the branch was force-pushed/rebased, or a skip condition (e.g. draft, label) was in effect at the time of that push.
Existing review comments are from an earlier review pass (on a prior commit). GitHub keeps those inline comments attached to their diff lines permanently; they don't disappear just because a subsequent event was skipped.

So in short: the comments were posted during a previous review run, and the "review skipped" message only means that the latest push didn't trigger a fresh review — not that the old comments were removed or invalidated.

If you'd like me to do a fresh full review of the current state of the PR, just say the word!

…view fixes (DEV-1505) Addresses the Codex + CodeRabbit review on PR #9: - Pass model=native_model_id(model) to ClaudeAgentOptions so --agent-model actually takes effect instead of the claude CLI's configured default (Codex, pre-existing gap shared with claude_sdk). - Add --reasoning-effort {low,medium,high,max} mapped to ClaudeAgentOptions.effort, threaded through run.py (CLI + run_evaluation + run_one_task + make_runner) and the cloud path (cli/driver/ray_app, incl. manifest + resubmit). Default unset = SDK default. - Make the on-the-fly scratch dir unique per invocation (uuid suffix) so concurrent same-task / cross-framework runs can't rmtree each other's live per-task SLayer store (CodeRabbit). - Refresh the --slayer-setup help text to list the new on-the-fly frameworks (CodeRabbit). - Tests use tmp_path instead of hardcoded /tmp (CodeRabbit); add tests for model pinning, effort passthrough + validation, and resubmit preserving --reasoning-effort. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…submit nudge (DEV-1505) Root-caused from the PR #9 households_7 smoke session log: the agent encoded the KB correctly and called submit_query on its FINAL turn, but the answer was dropped. Three fixes: - The full Claude Code built-in toolset was present and our MCP tools were DEFERRED behind ToolSearch, so ~5 turns/run were wasted re-discovering our own tools. Set tools=[] + setting_sources=[] so only the ~16 MCP tools are exposed, directly (no ToolSearch, no built-ins, no CLAUDE.md bleed-through). - The manual MAX_MODEL_TURNS break severed the final turn's submit_query before it executed. Use the SDK's native max_turns (= 2x base = 120) instead, which lets the final turn's tool run before the run stops; drop the manual break. - Add a PostToolUse hook that injects a "submit now" nudge when within 3 turns of the cap. Tests: tools/setting restriction + max_turns + hook wiring; turn-budget hook. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…sage) (DEV-1505) The Claude Agent SDK delivers message.usage as a DICT, but both claude_sdk and claude_sdk_otf read it via getattr(...) (attribute access) — so every agent turn recorded 0 tokens and agent_cost_usd was $0 (only the user-sim's litellm cost showed up). Surfaced when an Opus high-effort run reported $0.035 while the SDK's ResultMessage.total_cost_usd was $1.02. - Add a shared accumulate_assistant_usage() helper in claude_sdk.agent that reads usage from either a dict (live SDK) or an attribute object (mocks), includes cache_write, and gates on AssistantMessage so the cumulative ResultMessage.usage is not double-counted. - Use it from both claude_sdk and claude_sdk_otf so SDK-backed agents report cost consistently with the litellm-tracked frameworks. - Regression test feeds a dict-shaped usage + a ResultMessage and asserts the tokens are captured once and agent_cost_usd > 0. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

coderabbitai

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/test_claude_sdk_otf_agent.py`:
- Around line 314-355: The test
test_accumulate_assistant_usage_dict_shaped_and_skips_result should monkeypatch
usage_mod._cost_per_token to return a deterministic non-zero per-token cost
before calling accumulate_assistant_usage so agent_cost_usd doesn't depend on
LiteLLM's pricing; specifically, in the test (which constructs accum =
usage_mod.TokenUsage() and calls accumulate_assistant_usage), set
usage_mod._cost_per_token = lambda model: 0.0001 (or similar) prior to the calls
and restore if needed so accumulate_assistant_usage → usage._safe_cost() →
usage._cost_per_token() produces a stable >0 agent_cost_usd for the final
assertion.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 8bca782d-86d5-436a-9544-6b90de2b3f13

📥 Commits

Reviewing files that changed from the base of the PR and between 31dd730 and 492b2c6.

📒 Files selected for processing (3)

src/bird_interact_agents/agents/claude_sdk/agent.py
src/bird_interact_agents/agents/claude_sdk_otf/agent.py
tests/test_claude_sdk_otf_agent.py

…act-cloud submit (DEV-1505) Surfaced after a smoke run was kicked off against mini-interact when the user had asked for livesqlbench: --dataset defaulted silently to mini_interact and --mode a-interact happened to be consistent with that default, so the validator stack saw no inconsistency and the wrong benchmark ran. Drop the implicit default and require --dataset on every local and cloud CLI call. The mode/framework validators still catch inconsistent combos; this closes the silent-correct-default trap. - run.py + cloud/cli.py: --dataset is required=True (no default), help text updated. - Tests: the two test helpers (_argv_base in test_slayer_setup_flag.py, _slayer_argv in cloud/test_cli.py) now inject --dataset; inline argv blocks in cloud/test_cli.py, test_pydantic_ai_otf_encode_run_wiring.py, and test_claude_sdk_otf_run_wiring.py also do. - The two tests that verified the old default (test_dataset_defaults_to_mini_interact, test_dataset_flag_default_is_mini_interact) are rewritten to assert that omitting --dataset is rejected. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

coderabbitai

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/cloud/test_cli.py`:
- Around line 68-82: Several negative submit tests are now passing for the wrong
reason because argparse rejects the missing required --dataset before the
intended logic runs; update the test argv lists for
test_require_audited_gold_default_on,
test_detach_and_allow_dirty_mutually_exclusive,
test_empty_instance_ids_string_rejected, and
test_empty_instance_ids_file_rejected to include the dataset flag (use
"--dataset", "mini_interact") so that cli.parse_args in those tests reaches the
intended validation paths rather than failing on the required-argument check.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

Push a commit to this branch (recommended)
Create a new PR with the fixes

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 1b797fbf-6a9c-4584-a802-cba20dea498d

📥 Commits

Reviewing files that changed from the base of the PR and between 492b2c6 and 949a7c2.

📒 Files selected for processing (7)

src/bird_interact_agents/cloud/cli.py
src/bird_interact_agents/run.py
tests/cloud/test_cli.py
tests/test_claude_sdk_otf_run_wiring.py
tests/test_one_shot_mode.py
tests/test_pydantic_ai_otf_encode_run_wiring.py
tests/test_slayer_setup_flag.py

…505) After making --dataset mandatory on the local and cloud CLIs, the existing README examples and the helper scripts continued to omit it and would exit during argument parsing. Add --dataset mini_interact to every bird-interact / bird-interact-cloud invocation in: - README.md: oracle, claude_sdk raw a-interact, claude_sdk slayer a-interact, claude_sdk_otf local mini-interact, claude_sdk_otf cloud (mini-interact). - scripts/run_haiku_slayer.sh - scripts/run_three_way.sh The README livesqlbench/one-shot examples already carry --dataset livesqlbench. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

CodeRabbit on PR #9: after making --dataset required, several test_cli.py negative tests still omitted --dataset and were therefore passing for the wrong reason — argparse rejected the missing required arg before the test's intended guard fired. Add --dataset mini_interact to the 5 16-space-indent argv blocks that test specific guards (audited-gold rule, --detach/--allow-dirty exclusivity, two empty --instance-ids paths, and the unknown-mode test which had the same gap). The test_dataset_is_required block stays untouched — it asserts that omitting --dataset is rejected. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

CodeRabbit on PR #9: the dict-shaped usage helper test asserts `accum.agent_cost_usd > 0`, which depended on litellm's cost-map being loaded for `anthropic/claude-opus-4-7`. Monkeypatch `usage_mod._cost_per_token` to a deterministic non-zero value (matches the pattern other usage tests already use), so the test is robust to litellm pricing-map state. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

coderabbitai Bot reviewed May 29, 2026

View reviewed changes

Comment thread src/bird_interact_agents/slayer_otf/runtime.py

Comment thread tests/test_claude_sdk_otf_run_wiring.py Outdated

ZmeiGorynych and others added 3 commits May 29, 2026 16:53

coderabbitai Bot reviewed May 30, 2026

View reviewed changes

Comment thread tests/test_claude_sdk_otf_agent.py Outdated

coderabbitai Bot reviewed May 30, 2026

View reviewed changes

Comment thread tests/cloud/test_cli.py

ZmeiGorynych and others added 3 commits May 30, 2026 13:19

ZmeiGorynych merged commit 0fb3ad5 into main May 30, 2026
1 check passed

This was referenced Jun 10, 2026

DEV-1546: wire distinct_dimension_values into slayer-mode OTF agents #41

Merged

DEV-1555 stage 1: discovery/main subagent split + context guards + autopsy squeeze #47

Merged

This was referenced Jun 21, 2026

DEV-1579: hermetic Claude SDK subprocess env + provider support for all claude_sdk agents #54

Merged

DEV-1605: version OTF references by encoder model + stamp consumed version #61

Open

This was referenced Jul 8, 2026

DEV-1649: persist + reuse the agent's edited slayer models per task #84

Merged

DEV-1666: slayer-only lean_introspection + readonly_mode flags #85

Merged

Navigation Menu

Uh oh!

claude_sdk_otf: Claude SDK agent that encodes KB on-the-fly (DEV-1505) - #9

claude_sdk_otf: Claude SDK agent that encodes KB on-the-fly (DEV-1505)#9
ZmeiGorynych merged 8 commits into
mainfrom
egor/dev-1505-claude-code-sdk-as-agent-for-the-benchmarks

ZmeiGorynych commented May 29, 2026 •

edited by coderabbitai Bot

Loading

Uh oh!

linear Bot commented May 29, 2026

Uh oh!

coderabbitai Bot commented May 29, 2026 •

edited

Loading

Reviews paused

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Poem

❌ Failed checks (1 warning)

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

Uh oh!

ZmeiGorynych commented May 29, 2026

Uh oh!

coderabbitai Bot commented May 29, 2026

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

ZmeiGorynych commented May 29, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Files

Test plan

Summary by CodeRabbit

Uh oh!

linear Bot commented May 29, 2026

Uh oh!

coderabbitai Bot commented May 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Reviews paused

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Poem

❌ Failed checks (1 warning)

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

ZmeiGorynych commented May 29, 2026

Uh oh!

coderabbitai Bot commented May 29, 2026

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

ZmeiGorynych commented May 29, 2026 •

edited by coderabbitai Bot

Loading

coderabbitai Bot commented May 29, 2026 •

edited

Loading