PolyCodeGraph 0.10.0

PolyCodeGraph 0.10.0 — selective workflow context

Version 0.10 adds an optional client workflow for selecting focused tool schemas,
collecting every required intent page and emitting one compact text response.
It preserves semantic evidence and ordinary server defaults across all ten
providers. The user authorized stable publication with the open experimental
acceptance criteria below; release does not mark those criteria as passed.

• The packaged Python workflow MCP relay selects a static task surface, fuses
  overlapping source windows without losing required records, and supports
  bounded cancellation/deadlines and explicit cursor recovery. Local workflows
  register no graph tools. The server does not automatically install the client
  into Codex or observe the provider's full prompt.
• Optional source_policy: intent selects locations for rename, contracts for
  signature changes and edit context for other intents; explicit views prevail.
  budget.max_source_chars separates source text from inventory and makes
  truncation visible. Optional tokenizer budgets apply only at the supported
  client insertion boundary, not to the whole model prompt.
• Lexical discovery applies scope before index caps, caches bounded snapshot
  indexes and exposes excluded files. Anchor grouping and binary-term BM25 are
  opt-in; the line/overlap baseline remains available. No new semantic edge is
  inferred from lexical ranking.
• Coverage separates known inventory, relevant observed diagnostics and unknown
  question coverage while preserving global health/freshness limits.
• Binding-owned retention validates exact identity/source/baseline and resets on
  compaction with self-contained rehydration. It requires actual retained-context
  acknowledgement. Generic MCP retention remains disabled; its model benefit
  has not been verified.
• Targeted edit-receipt guidance avoids full readbacks after successful edits
  while retaining source reads for concrete missing facts and authorized checks.
  Benchmark tooling preserves failures, frozen identities, unknown costs and
  explicit solver budgets.

Activation and rollback ·
Consumer harness ·
Codex subscription policy ·
Local release validation

Measured results and open acceptance

The dev.2 A/D screening
completed 24 solver turns on twelve known controlled TypeScript tasks. All
static outcomes passed, but D used 55.9% more total tokens than no-graph A
and 3.74% more uncached input. The historical savings gate did not pass.

The separately authorized T3 instruction diagnostic
completed eight turns on four known refactoring tasks, comparing old and
corrected guides with the same graph. All static outcomes passed. Same-hash
post-edit reads fell from 85 to zero and direct reads from 178 to 87. Total
tokens fell 15.5% versus the old guide, while uncached input rose 14.4%.
This passes the narrow correction criterion; it is not a fresh no-graph product
comparison or independent confirmation. These measurements identify development
artifacts; no final-version AI rerun is claimed.

Included Codex allowance savings and economic cost remain unknown. Raw total
tokens, weighted credits and included subscription usage are distinct. The
user's +15% total-token objective versus A belongs to 1.0 and is not a stable
0.10 requirement. Publishing 0.10 does not declare that objective achieved.

G5 has a prepared, untuned authored holdout
with twelve new tasks and a proposed 48-turn budget. It has not been run;
publication grants no additional AI budget. T6 long-task/post-compaction model
savings are unverified: the installed-client audit
does not provide a supported exact retained-set boundary. T4 ranking/discovery
benefit has not been isolated with the model and remains opt-in.

Static fixture oracles verify source bytes/findings, not application runtime.
No indexed application code, plugin, script, build hook or code generation is
executed. There is no quality or savings generalization to languages, clients,
models or repositories not actually tested.

Release validation and downloads

Final-version local checks cover Rust formatting/Clippy/tests, Python client and
measurement tests, native workflow/source-policy/retention checks and relocated
package/client smoke. The exact release commit's cross-platform CI, Android and
macOS UIKit checks, four native package results and archive SHA-256 digests are
recorded in the GitHub release after passing; earlier dev.2 CI is not substituted
for final artifact validation. Conditional skipped tests are not verified coverage.

Download the Linux x64, Windows x64, macOS Intel or macOS Apple Silicon package
from release v0.10.0
and verify it with SHA256SUMS. Keep the executable beside providers/ and
clients/; selected language adapters still require the runtimes documented in
README. Rust is required to build from source, not to run the native packages.

Verified publication identity

Source: a74ea963c35da7c55e27f46b7e791015289ad827. All 20 jobs passed for this exact source in release CI: Linux/Windows/macOS core and native intent checks, frozen-0.6 primitive differentials, Flutter, Android, macOS UIKit, query benchmark and all four native packages. Android SDK fixtures verify 5 symbols/6 edges; UIKit verifies 13 symbols/23 edges. Conditional platform steps skipped on other hosts are not additional verified coverage. Local checks passed 93 Rust and 104 Python tests. No additional AI runs were launched for stable publication.

Every archive was checked for manifest version/OS/architecture, native binary architecture, executable permissions where applicable, all eight client module hashes, compiled Dart/Go workers, exact release documentation, confined paths and absence of caches/environments. SHA-256 digests are in SHA256SUMS.