feat: add Planning capability with cache-safe plan reminders#266
Open
strawgate wants to merge 3 commits into
Open
feat: add Planning capability with cache-safe plan reminders#266strawgate wants to merge 3 commits into
strawgate wants to merge 3 commits into
Conversation
Gives agents a model-owned task plan via a single whole-plan-replacement `write_plan` tool, avoiding the index-bookkeeping fragility of incremental edits. The current plan is surfaced back to the model as an ephemeral reminder appended to the tail of each request in `wrap_model_request` (which runs after the durable history is persisted, so the reminder never enters `message_history`), with a `CachePoint` placed before it. The cached prefix therefore stays byte-identical across turns -- the plan is always fresh in the model's view without ever invalidating previously-cached tokens. The mutable plan is deliberately never injected into instructions/system prompt, which would bust the cache; only static guidance goes there. Includes a README, agent-spec serialization, and tests (100% branch coverage) covering the ephemerality and cache-breakpoint placement end-to-end.
… coverage The CachePoint-not-persisted test could never execute its assertion under TestModel (no list-content UserPromptParts survive in durable history), leaving line 300 uncovered. Ephemerality is already proven by the FunctionModel test asserting the reminder is absent from all_messages(); the CachePoint only ever rides that reminder, so the separate check was redundant.
Mutation testing surfaced 4 weak assertions that used substring checks where exact output mattered: the render_plan line separator and the write_plan in-progress note (empty vs present, wording, and wrapping). Assert exact strings / exact tails instead. Kill rate 30/36 -> 34/36; the 2 remaining survivors are equivalent mutants (add_function name= falls back to the method's __name__, which equals the explicit name).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds a
Planningcapability: a model-owned task plan that's surfaced back to the model every turn without ever invalidating the prompt cache.This is a from-scratch design (informed by reviewing #180), built on existing core primitives — no Pydantic AI core change required.
Design
One tool,
write_plan(items)— whole-plan replacement. The model passes the full ordered list every call. This eliminates the index/parent_indexbookkeeping that incremental-edit designs need (and the bugs that come with it), is concurrency-safe, and matches the most battle-tested agent todo tools. Status enum:pending/in_progress/completed/cancelled. Noget_plantool — the plan is already in context via the reminder.Cache-safe surfacing. The current plan is injected as an ephemeral tail reminder in
wrap_model_request, which runs after core persists the durable history (_agent_graph.py:911) and whose per-request message list is never written back. So the reminder reaches the model but never entersmessage_history— no accumulation, no stale copies. ACachePointis placed immediately before the reminder, so the cached prefix (tools + system + real conversation) stays byte-identical turn over turn and only the small reminder falls outside the cache.The mutable plan is never put in instructions/system prompt (that's the cache-busting path). Only static usage guidance goes there, which is cache-stable.
Why not dynamic instructions
Injecting the plan into the system prompt re-reads the whole conversation at full token price on every plan edit. Even with core's static/dynamic instruction split (which protects the tools+static-system prefix on Anthropic), the system precedes messages in cache order, so the message-history cache still busts. The ephemeral tail reminder avoids that entirely.
Files
pydantic_ai_harness/planning/{_capability,_toolset,__init__}.py+README.mdtests/planning/test_planning.pypydantic_ai_harnessTests
FunctionModel— that the reminder reaches the model with aCachePointin front of it and is absent fromresult.all_messages()(i.e. genuinely ephemeral, never persisted).Validation
make lintclean;pyrightclean on the package; full suite green.https://claude.ai/code/session_01LQ9NTr8q95A99UnVcq6Nzf
Generated by Claude Code