Files
David SFandGitHub 3ba9e2f9a5 docs: capability pages for the unified docs site + README/doc parity gate (#329)
* docs: publish capability docs to the unified site + add README/doc parity gate

Every capability shipped only a README (kept for GitHub/PyPI). This adds a
parallel, cleaned-up page per capability under docs/ for the new unified docs
site (pydantic.dev/docs/harness), migrated from each README: snippets verified
runnable against source, autodoc API blocks, root-relative Pydantic AI links,
and an experimental-status admonition on the experimental set.

To keep README and doc in sync going forward, adds a docs-parity-reviewer agent
and a parity gate in the review checklist (run as the last step before merge),
plus the docs/ layout and the README<->doc requirement in AGENTS.md and the
capability-authoring guide.

* docs: fix README<->doc<->source inconsistencies across capabilities

A parity audit against source found drift, mostly in the capability READMEs
(staler than the migrated docs). All fixes verified against source:

- Correctness: the "approval/deferred tools are excluded from the sandbox" claim
  (code_mode README + doc) was false -- those tools are sandboxed like any
  other; corrected in both. The stale Shell persist_cwd sentinel description is
  replaced with the actual out-of-band temp-file capture. filesystem protected
  default `.git/` -> `.git/*` (the bare form never matched).
- Runnable snippets: added the missing imports/wiring so README snippets no
  longer raise NameError (subagents, context, planning, overflow, authoring,
  filesystem, code_mode).
- Parity: documented previously-undocumented params/behaviors (compaction
  strategy options, overflow strip_ansi/Passthrough, extra autodoc classes for
  context and subagents), fixed a stale version pin (>=1.95.1 -> >=2.1.0), and
  added the missing Managed Prompt row to the root README capability matrix.
- Style: normalized decorative Unicode to ASCII across all READMEs and dropped a
  hype phrase, matching AGENTS.md writing style and the docs.

* docs: add nav.json to drive the unified-docs harness sidebar

The unified docs mount the harness docs under /docs/ai/harness (fed live from
this repo via the pydantic-ai 'Pydantic AI Harness' section). This nav.json
defines the sub-nav (Overview + Capabilities + Experimental) and the set of doc
files the site includes.

* docs: migrate "What goes where?" explainer into harness overview

Adds the core-vs-harness boundary section (anchor #what-goes-where) to the
canonical harness overview, so the pydantic-ai docs that link to it can point
here after the duplicated in-repo stub is removed.

* docs: address CodeRabbit review -- runnable snippets, accuracy, multi-class autodoc

* docs: flatten harness nav and align with graduated capabilities

Following the experimental-graduation refactor (#347), restructure the
unified-docs harness pages:

- Flatten docs/ (drop capabilities/ and experimental/ subdirs); the sidebar
  is now Overview + one flat list per Douwe's request.
- Rename to match the graduated modules: overflow -> overflowing-tool-output,
  authoring -> runtime-authoring, docs -> pydantic-ai-docs.
- Drop the 'Experimental' admonitions from the graduated capabilities and
  repoint every import + ::: autodoc path off pydantic_ai_harness.experimental.
- Add docs for the newly-shipped capabilities: guardrails, dynamic-workflow,
  media, and acp (acp stays framed as experimental -- it may still be removed).
- Every capability doc now links to its source; index capability table lists
  the full set with flat links.

* docs: apply team-sync authoring rules + enforce them in CI

From the 2026-07-10 docs review on #329:

- Purpose-first leads: drop hook names (before_model_request,
  after_tool_execute) from the opening paragraphs of compaction and
  overflowing-tool-output (doc + README); mechanism moves lower.
- Mirror the soft 'API may change between releases' stability note from each
  graduated README into its doc page (ACP keeps its stronger experimental
  warning; guardrails' README has no note, so its page gets none).
- README H1s now use the capability's display name (Overflow capability ->
  Overflowing Tool Output, RuntimeAuthoring -> Runtime Authoring, SubAgents ->
  Subagents, etc.).
- Extend tests/test_docs_parity.py with per-page mechanical checks: source link
  present, heading matches the capability name, purpose-first lead (no hook in
  the opener), and no experimental framing on graduated pages (ACP excepted).
- Update the docs-parity-reviewer agent + review-checklist to the flat
  structure and the new semantic checks.

* docs: add the stability note to guardrails (parity with sibling capabilities)

guardrails was the one graduated capability whose README and doc page lacked
the shared 'API may change between releases' note. Add it to both.

* fix: restore uv.lock to match pyproject (bad text-merge dropped 8 lines)

Merging origin/main did a git text-merge of the generated uv.lock, leaving it
inconsistent with pyproject.toml -- every CI job failed at 'uv sync --locked'.
pyproject.toml is identical to main here, so the correct lock is main's.

* docs: address CodeRabbit review on #329

Findings that failed to post inline (GitHub error) but were real:
- context/README.md, planning/README.md: two nested examples still imported
  from pydantic_ai_harness.experimental.* -- repoint to the graduated modules.
- guardrails/README.md: replace em dashes with '--' (repo style) and add the
  source-module link.
- docs/media.md: standardize on the implementation's canonical media+sha256://
  URI scheme (was mixing media://).
- tests/test_docs_parity.py: strengthen my own checks per review --
  source-link and top-README-link now require a real Markdown link to the
  page's specific module (not a bare substring); heading checks assert an H1
  exists and equals the expected capability name via explicit page metadata.

* fix: restore uv.lock [options.exclude-newer-package] block

The lock lost its [options.exclude-newer-package] manifest (pydantic-ai-slim
= false, ...) -- a bad git text-merge dropped it, and diagnostic uv commands
rewrote it under a different local config. Without that block CI's
'uv sync --locked' re-resolves and fails ('addition of exclude newer exclusion
for pydantic-ai-slim'). Restore origin/main's exact lock.

* fix: restore uv.lock [options.exclude-newer-package] block

A pre-commit hook was rewriting uv.lock under the local uv config, stripping
the [options.exclude-newer-package] manifest (pydantic-ai-slim = false, ...).
Without it CI's 'uv sync --locked' re-resolves and fails. Commit origin/main's
exact lock with --no-verify so no hook mutates it (lock-only change).

* test: cover the docs-parity helper edge cases (100% coverage)

The strengthened helpers added defensive branches (missing frontmatter close,
fenced code before the lead, missing/forbidden/ClassName H1, lead running to
EOF) that no real doc exercises. Add direct unit tests so the file is back to
the repo's required 100% coverage.

* docs: link every capability README to its source module + enforce it

CodeRabbit re-flagged planning/README.md for a missing source link. Only
guardrails had one, so add the source-module link to all 15 remaining
capability READMEs (matching the doc pages) and add a parity test so the
requirement is mechanical and cannot silently regress.

* docs(agents): drop stale folder tree; fix flat docs path + guard names

AGENTS.md's File-structure ASCII tree and capability-authoring's doc paths
still showed docs/capabilities// docs/experimental/ (flattened in this PR) and
the old /docs/harness URL. Delete the tree rather than redraw it -- the layout
is discoverable by listing the repo; keep only the non-obvious conventions
(flat docs/, the README<->doc parity requirement). Also fix the Vocabulary
guard examples (InputGuard/OutputGuard, not the nonexistent InputGuardrail/
CostGuard).

* test: statically validate doc snippets exist and parse

Every Python snippet in the capability READMEs and docs/*.md pages is now
checked for the two failures a reader hits immediately: it does not parse
(syntax), or it imports a pydantic_ai_harness symbol that does not exist (stale
module path or renamed name -- the class of bug behind the experimental.* import
drift). Static only: no model/network execution, so it needs no mocking. The
four illustrative API-signature blocks opt out with a {test="skip"} fence
(read by pytest-examples, stripped-safe for the unified-docs render).

* test: don't fail doc-snippet check on a missing optional extra

The static check imported capability modules to resolve their symbols, but in
the slim CI job (no extras) importing e.g. pydantic_ai_harness.experimental.acp
raises ModuleNotFoundError for the absent third-party 'acp' package -- the
harness module exists, its extra just isn't installed. Distinguish a genuinely
missing harness module (fail) from a missing extra (skip) by the ImportError's
module name.
2026-07-13 11:00:12 -05:00

11 KiB

Compaction

Note

Import these capabilities from their submodule -- there is no top-level pydantic_ai_harness re-export:

from pydantic_ai_harness.compaction import TieredCompaction

The API may change between releases. Where practical, breaking changes ship with a deprecation warning.

A menu of strategies for keeping an agent's conversation history within a model's context window. Each is a Pydantic AI Capability that edits the message history just before each request goes out; edits persist into the run's message history, so a trim/clear/summary carries forward to later steps (it is not recomputed from the full history every turn).

All strategies preserve tool-call / tool-return pairing -- core does not validate this, and a provider rejects an orphaned pair. The zero-LLM strategies never call a model.

Source

The menu

Capability Cost What it does Reach for it when
ClampOversizedMessages zero-LLM Head/tail-truncates a single oversized part (response text, tool-call args) One runaway generation blew past the context cap and no other strategy can reach it
SlidingWindow zero-LLM Drops the oldest whole messages down to a tail You only need the recent turns and can discard old context entirely
ClearToolResults zero-LLM Blanks the content of old tool results in place, keeping the last keep_pairs Tool outputs dominate context and can be re-fetched on demand (the cheap first tier)
DeduplicateFileReads zero-LLM Blanks every file read superseded by a newer read of the same file The agent re-reads files and only the latest version matters
SummarizingCompaction one LLM call Summarizes older messages into a structured summary, keeping the recent tail Old context still matters but must be compressed; use behind the cheap tiers
TieredCompaction escalates Runs cheap passes first, summarizes only if still over target_tokens You want a sensible default: spend the expensive summary only when needed
LimitWarner zero-LLM Injects an URGENT/CRITICAL warning as limits approach You want the agent to wrap up rather than have its history rewritten

Triggers

Every size-based strategy triggers on max_messages and/or max_tokens (estimated). Token counts use a ~4-chars-per-token heuristic by default; pass a tokenizer callable (e.g. tiktoken) for accuracy. DeduplicateFileReads runs on every request when no trigger is set (it is cheap and near-lossless). TieredCompaction triggers and stops on a single target_tokens budget. ClampOversizedMessages triggers per part (max_part_tokens / max_part_chars), not on the whole history -- the failure it targets is one oversized part, not a large total.

ClampOversizedMessages: surviving a runaway generation

A single model response of repeated whitespace, or a single tool call with a giant payload, can produce one part so large the next request exceeds the provider's context cap. None of the other strategies can reach it: SlidingWindow drops the oldest messages but the offender is the newest; ClearToolResults only touches tool results; LimitWarner never edits history; and feeding the history to SummarizingCompaction hits the same cap.

ClampOversizedMessages truncates the offending part in place, keeping a head slice and a tail slice with a [clamped: removed N of M characters] marker between them. Degenerate generations are low-entropy repetition, so a head/tail slice loses little.

from pydantic_ai import Agent
from pydantic_ai_harness.compaction import ClampOversizedMessages

agent = Agent(
    'openai:gpt-4o',
    capabilities=[ClampOversizedMessages(max_part_tokens=50_000, keep_head_chars=2_000, keep_tail_chars=2_000)],
)

A part is clamped only when it is oversized and the clamp actually shrinks it, so keep keep_head_chars + keep_tail_chars well below your per-part threshold.

It clamps two kinds of part inside each ModelResponse:

  • Response text (TextPart) -- the critical case, a runaway model-response text part.
  • Tool-call args (ToolCallPart), when clamp_tool_call_args=True (default) -- the same failure shape for a giant payload (e.g. a runaway write_plan). The args are replaced with a small JSON object {"_clamped": "<head>...<tail>"} so they stay valid function arguments; the original call already executed, so this only shrinks the history copy. Set clamp_tool_call_args=False to clamp response text only.

Request-side parts (user prompts, tool returns, system prompts) are deliberately out of scope: user input should not be silently rewritten, and oversized tool returns are the job of ClearToolResults.

Use it as the first tier of TieredCompaction, before ClearToolResults:

from pydantic_ai_harness.compaction import (
    ClampOversizedMessages,
    ClearToolResults,
    TieredCompaction,
)

TieredCompaction(
    tiers=[
        ClampOversizedMessages(max_part_tokens=50_000),
        ClearToolResults(max_tokens=1, keep_pairs=3),
    ],
    target_tokens=120_000,
)

SlidingWindow and ClearToolResults options

SlidingWindow keeps the last keep_messages down to a tail; pass keep_tokens instead for a token budget rather than a message count. By default preserve_first_user_message=True keeps the first user turn even when it falls outside the window, so the agent does not lose the original task.

ClearToolResults keeps the last keep_pairs intact. Set clear_tool_inputs=True to also blank the arguments of the cleared calls, and exclude_tools to a set of tool names whose results are never cleared.

LimitWarner thresholds

Warnings begin at warning_threshold (default 0.7, a fraction of the limit) and escalate to CRITICAL for iterations once the remaining request count drops to critical_remaining_iterations (default 3). It watches max_iterations, max_context_tokens, and max_total_tokens, warning on whichever are configured; narrow that with warn_on.

Cost: why summarization is the last resort

Summarization turns input tokens into output tokens, which are billed at a premium and generated serially -- so it is genuinely expensive. The zero-LLM strategies touch only the cheaper input side. The field consensus (Anthropic, OpenCode, Letta) is to clear/dedupe first and summarize only when that is not enough -- which is exactly what TieredCompaction encodes:

from pydantic_ai import Agent
from pydantic_ai_harness.compaction import (
    ClearToolResults,
    DeduplicateFileReads,
    SummarizingCompaction,
    TieredCompaction,
)

agent = Agent(
    'openai:gpt-4o',
    capabilities=[
        TieredCompaction(
            tiers=[
                DeduplicateFileReads(file_key=my_file_key),
                ClearToolResults(max_tokens=1, keep_pairs=3),
                SummarizingCompaction(max_messages=1, keep_messages=20),  # model inherits the run's
            ],
            target_tokens=120_000,
        )
    ],
)

A tier inside TieredCompaction is driven directly by the orchestrator, which re-measures after each and stops once under target_tokens -- so a tier's own max_* trigger is irrelevant there (set it to anything valid). Any object with async def compact(messages, ctx) -> list[ModelMessage] (CompactionStrategy) can be a tier, so you can plug in your own.

Cache tradeoff (read before using ClearToolResults)

Clearing or deduplicating rewrites message content, which invalidates the provider's prompt cache from the edit point onward -- the next request pays a cache-write. Use ClearToolResults' min_clear_tokens to skip clearing that reclaims too little to be worth busting the cache.

Model inheritance

SummarizingCompaction(model=...) accepts a model name or Model; when left None it inherits the running agent's model. No token caps are imposed on the summary call.

By default incremental=True extends an existing summary from a prior compaction rather than regenerating it from scratch, and preserve_first_user_message=True keeps the original task turn even when it falls outside the window. Pass keep_tokens to trim the retained tail to a token budget instead of keep_messages.

Usage accounting

The summary call is a real request to the model, so its full usage -- tokens and the request itself -- is folded into the run's ctx.usage. This is deliberate: it keeps cost honest, keeps the request count consistent (a model request that didn't count as one would be the surprise), and lets a UsageLimits request limit catch a runaway compaction. A run-request / iteration limiter will therefore see compaction calls among its requests.

DeduplicateFileReads.file_key

There is no default file_key: identifying a file read is agent-specific, and a wrong guess would drop live data. Supply a callable mapping a ToolCallPart to a stable file key, or None when the call is not a file read:

from pydantic_ai.messages import ToolCallPart

def my_file_key(call: ToolCallPart) -> str | None:
    if call.tool_name != 'read_file':
        return None
    args = call.args
    return args.get('path') if isinstance(args, dict) else None

Tracing

When core instrumentation is active (the Instrumentation capability, agent.instrument, or Agent.instrument_all()), each strategy emits a compact_messages span on the run's tracer the moment it actually compacts -- that is, in before_model_request, once the strategy's threshold is exceeded (ClampOversizedMessages emits only when a part is actually clamped). TieredCompaction emits a single span for the whole escalation rather than one per tier, because it drives each tier's compact directly. Without instrumentation the tracer is a no-op, so the span adds no overhead.

The span name is the static compact_messages; the strategy is an attribute, not part of the name, to keep span cardinality low. Attributes:

Attribute Type Meaning
gen_ai.conversation.compacted bool Always true; the OpenTelemetry GenAI convention's flag for a compacted context
compaction.strategy str Strategy class name (e.g. SlidingWindow, SummarizingCompaction)
compaction.messages_before int Message count before compaction
compaction.messages_after int Message count after compaction
compaction.tokens_before int Estimated token count before compaction
compaction.tokens_after int Estimated token count after compaction

gen_ai.conversation.compacted is the GenAI semantic convention's flag; the rest is harness-specific. Token counts use the strategy's tokenizer when set, otherwise the ~4-chars-per-token heuristic. Raw message content is not recorded.

Out of scope

These strategies compress or drop context inside the window. Moving large tool outputs out of the window -- overflowing them to a file the agent (or a subagent) can query on demand -- is a separate capability, not lossy truncation. Prefer it over capping individual tool outputs.