Keep one typed accumulation of observed decode facts instead of constructing
partial Evidence only after string projection. Record decoded client/target
before the post-command cancellation checkpoint, parsed metadata before
manifest reads, raw manifest before interpretation, and inventory prefix
before rethrowing its exact original cause. Omit unexamined fields.
Return inventory success or exact failure plus completed prefix through
Result. Keep resource-cap stops distinct from physical failures. Attach
truncation and inventory limitations at their producing stages. Validate
complete results with the named domain schema; centralize partial Evidence
assembly in the existing failure/cleanup boundary.
Validation: six authored stage controls fail unchanged34a88d6c1; 131 focused
checks across7files, check:fast, docs:check and discovery1556 pass. Covers
manifest bytes/XML, interrupted inventory, late cleanup cancellation,
denied removal/retry and existing table interruption. Two source reviewers
clear after moving the command handoff checkpoint. No separate runtime
qualification of that precise handoff race.
Public CLI manifest failure retains metadata/raw manifest with unknown
fields omitted and original failure framing unchanged. Twelve small normal
CLI tasks retain exact replies. No speedup or memory-saving claim; failure
reply estimate189 to610cl100k tokens preserves completed Evidence. No real
Apktool, large-case memory or whole-agent qualification. Heavy tests remain
stopped.
Fixes#4882
Separate physical string-table facts from per-selection output allocation.
Use one owner for legacy and batched projections, preserving caller labels
and normalized-directory reuse. Retain complete counts and at most one
bounded table prefix for later aliases. A skipped cached alias keeps its
known counts; unexamined tables keep explicit unknown counts.
Remove the provider's duplicated parsing and status branches. Charge the
primary legacy copy before selecting output entries and derive limitation
text from the actual emitted count. Reuse the named request input type and
clarify nullable count descriptions without changing their schema types.
Have the workspace report retained raw-prefix bytes, including a stripped
UTF-8 BOM, instead of reconstructing source cost from decoded text.
Validation: 145 focused checks across 8 files; baseline count control fails
unchanged fd6da7092; check:fast, docs:check and discovery 1556 pass. Twelve
small public CLI decode workflows preserve exact normal-path replies and
request/reply token estimates. Timing variation is noise-level; no speedup,
large-case memory or real Apktool qualification claim. Heavy tests/profiling
remain stopped.
Fixes#4874
Inventory observed entries through the owned artifact traversal, preserve other configuration names, and disclose non-regular tables with partial coverage. Share request alias normalization with exact table classification and preserve filesystem failures.
Adapt the disclosure policy from #4738. Correct the fixture command separator so all advertised tables are actually created.
Validation: 164 focused tests; check:fast; test discovery; docs:check. Two regression controls fail on the baseline. Twenty-four tiny authored CLI/MCP comparisons; no real Apktool or Windows verification.
Fixes#4612
Co-authored-by: Oke Iyanuoluwa E. <oke.iyanuoluwa12@gmail.com>
Replace raw descriptor and workspace bookkeeping with artifact file ownership, canonical chunk reads, private roots and cleanup leases. Retain failed owners, fence replacement acquisition, release successful peers once, and preserve completed Evidence and physical failures through final cleanup and cancellation.
Use the shared directory reader and enforce the examined-entry budget before special-file omission. Keep command and artifact retry ownership separate by exact lease identity; remove superseded reads, implicit traversal and suppressing closes.
Validation: 163 focused tests in 20 files, build, typecheck/lint, docs and discovery passed. Six public controls fail on clean baseline 11bcf4425. Tiny FIFO, descriptor denial/retry, late allocation and mixed outcomes are controlled fixtures. Twenty-four completed CLI/stdio MCP comparisons return byte-identical replies; no speedup claim. Actual Apktool/JVM, natural OS denial and native Windows remain unverified; heavy stress/profiling stayed stopped.
Closes#4812
morluto