Parity Report

PARITY_REPORT.md is the rebuild's self-audit: an item-by-item record proving the Python indusagi package reaches feature parity with the TypeScript source it was rebuilt from. It is prose, not importable code — but every claim cites concrete evidence (file:line into src/indusagi/ and tests/).

This page summarizes that audit. The Python build is an independent rebuild: it reuses no third-party application source and implements each subsystem from primary specifications. The parity check walks a 26-item milestone checklist against the rebuilt tree and records the result, plus eight deliberate waivers where the Python build intentionally diverges or where the original checklist miscounted.

Table of Contents

• Summary
• Gate results
• Area-by-area parity
• Deliberate waivers
• Version drift
• Follow-ups
• Relationship to the rest of the docs

Summary

The audit walks the M8 feature-parity checklist item by item against src/indusagi/, tests/, and dist/. All rows resolve to done.

	count
Checklist items	26
Done	26
Partial	0
Missing	0

Verdict: parity achieved. All 26 rows are implemented and evidenced. The three rows that were once partial (exit-transcript print, AWS_REGION registry routing, and the lineage scan plus a checked-in pyright config) have been closed.

Gate results

The report records the state of each release gate at audit time:

Gate	Result	Source
Test suite	755 passed (pytest, network-free; mock connector + respx)	tests/
Typecheck	pyright 1.1.405, standard mode, clean over a ratchet subset	pyrightconfig.json
Lineage scan	exit 0 — "LINEAGE GATE PASSED: no lineage markers in src/indusagi"	scripts/lineage_scan.py
Packaging	wheel + sdist build; fresh-venv install gate ([mcp,tui]) passes	dist/

The pyright gate is a coverage ratchet, not a blanket strict pass. pyrightconfig.json deliberately includes only the packages that check clean today (__init__.py, memory.py, _internal, interop, react_host, runtime) in standard mode. The report self-corrects an earlier ad-hoc claim of "strict, 0 errors over all of src" — that run had analyzed 0 files and did not reproduce (strict mode actually surfaces 907 errors). The ratchet policy is to expand the include list over time, never to add per-rule ignores or remove entries, which keeps the gate reproducible. See Testing.

Area-by-area parity

The checklist is grouped by the milestone that delivered each area. Every row is done; the evidence column below is condensed from the report.

Gateway

Area	Evidence
Connectors	CONNECTOR_REGISTRY maps all 11 ApiKinds; bedrock is a fail-fast unsupported stub (W-1). 41 gateway tests pass.
Emission protocol	8-variant Emission union with error-as-terminal-emission; re-iterable Channel; fold_reply orders thinking → text → tool calls and reassembles split JSON args.
Catalog + credentials	MODEL_CARDS (13, see W-5), fluent models() query, estimate_cost; env SECRET_TABLE; OAuth 2.0 + PKCE S256.
Wire framing	SSE (WHATWG edge cases) + NDJSON framers; HTTP status → error-kind mapping (401/403 → auth, 429 → rate_limit). 29 streaming tests.

See LLM Gateway.

Runtime and tools

Area	Evidence
Pure FSM	RunPhase = idle → invoking → streaming → dispatching → compacting → settled/faulted; 7 signals, 5 effects, 7 run events; reducer purity tested.
Steering + budget	steering folds into a live run; abort; 64-turn budget; compaction (trigger_ratio / keep_recent, cut-point never orphans a tool_result).
Session DAG	content-addressed SHA-256 DAG + append-only JSONL store + resume + branch; ledger pub/sub.
Tools	12 tools registered (read/write/edit/ls/grep/find/bash/process/todo_set/todo_read/websearch/webfetch — see W-6) with byte-budget clamp, Myers diffs, background process table.
Collections	read-only / coding / all collections; tool_box() one-call assembly.

See Runtime and Capabilities.

Bridges and connectors

Area	Evidence
MCP client	stdio + SSE endpoints, server__tool namespacing, schema normalization, per-server fault isolation. 48 interop/mcp tests.
MCP provider host	publishes the agent's ToolBox over tools/list / tools/call; in-memory round-trip tests.
SaaS control	enable / execute / connect / status control tools, OAuth polling state machine, identity-stable hydration cache.

See Interop and Connectors.

Swarm and builder

Area	Evidence
Kernel	JsonCell / JsonlLog with locking + atomic publish, ULID ids, 6 fault kinds.
Coordination	dependency-gated ready() + cycle rejection, cursor-exact mailbox, git-worktree isolation, run_round.
Builder	interview + profiles + blueprint validation + knowledge pack + transcript redaction.

See Swarm and Smithy.

CLI

Area	Evidence
Flag grammar	10-flag table and all 4 error strings byte-identical to the source; mode precedence (help / version / print / wire / repl) verbatim.
Wire + auth	NDJSON wire protocol byte-compatible with existing hosts; auth login/refresh/status over a compatible auth.json format (state dir is port-local, see W-2).
Settings	3-layer merge, idempotent upgrades, INDUSAGI_HOME override, exit codes returned (never a process.exit analogue).

See Shell App and CLI Reference.

TUI

Area	Evidence
Rendering	flicker-free streaming markdown (tables, highlighted fences, literal ~); word-level diffs; tool cards with clamp/expand.
Dialogs	all 12 dialogs map 1:1 to the source (model, scoped-models multi-select, session browser, startup picker, settings, theme with live preview, login, OAuth, tree, fork-from-message).
Chrome	footer stats with 70/90 % context thresholds, TaskPanel, StatusLine; Esc abort / Ctrl-C exit; queued-message restore.
Persistence	live theme switching; alternate-screen scrollback + exit transcript print.

See TUI and React-Ink.

Compat surface

Area	Evidence
Facades	indusagi.ai / .agent / .mcp shims export the original vocabulary (106 / 97 / 65 names) and pass 142 tests; indusagi.memory ships empty (W-4); a single model catalog seeds the ai registry; legacy v3 session JSONL is readable.

See ai, agent, mcp, and memory.

Cross-cutting

Area	Evidence
Cancellation	CancelToken honored end-to-end (model stream, tool runners, web tools); MCP invokes do a pre-flight check only (W-7).
Env registry	a single registry module (_internal/env.py) owns every env read — INDUSAGI_HOME, provider key probes, AWS_REGION, Vertex project/location; no config read touches os.environ directly.
Distribution hygiene	py.typed shipped; lineage scan green; NOTICE / CREDITS.md packaged into the wheel under dist-info/licenses/.

See Tracing for the cross-cutting observability seam.

Deliberate waivers

The report records eight intentional deviations or checklist corrections. These are decisions, not gaps:

ID	Waiver
W-1	Bedrock is a fail-fast unsupported stub — it assembles body/region, then raises, matching the source connector.
W-2	On-disk state is fully port-local: the Python build uses its own ~/.pindusagi state dir (override INDUSAGI_HOME) and never reads or writes the original ~/.indusagi. The auth.json format stays compatible, but the store is not shared.
W-3	react_host collapses to a backend-pinning shim; loader probing and the jsx-runtime subpath are removed and re-documented.
W-4	indusagi.memory ships intentionally empty (__all__ = []), guarded by a parity test.
W-5	MODEL_CARDS count is 13, not 14 — the checklist's "14" was a miscount; the source catalog also has 13 with identical ids.
W-6	Tool count is 12, not 11 — the checklist row said "11" but listed 12 names; the registry carries the same 12.
W-7	MCP invoke cancellation is pre-flight-only; in-flight calls are not interrupted, matching the source behavior.
W-8	TUI key decoding at runtime is Textual's; the full pure-Python key port (tui/keys.py) is kept for the key-id vocabulary and test corpus but is not in the live input path.

Version drift

The non-code meta files disagree on a few numbers because they were written at different points in the rebuild. When in doubt, pyproject.toml is the single source of truth — the runtime reads its version via importlib.metadata.version("indusagi"), and the version string is never hardcoded in code.

Fact	Authoritative value	Where drift appears
Package version	0.1.2 (pyproject.toml, dist/)	CHANGELOG.md is still headed 0.1.0; the parity report cites dist/indusagi-0.1.0*.
Test count	varies by snapshot	parity report: 755 passed; the changelog cites a different figure.
Model count	13 curated MODEL_CARDS (W-5)	a larger "models / providers" figure elsewhere counts the full routed catalog, not the curated cards — these count different things.
Tool count	12 registered tools (W-6)	the checklist row text says "11".

The example filenames listed in some READMEs are also stale relative to the on-disk names; trust the files in examples/. See Examples.

Follow-ups

The three formerly-partial rows are resolved with evidence in their checklist rows. One item remains outstanding and is not a checklist row: the actual PyPI publish has not been executed. The wheel and sdist in dist/ built and passed a fresh-venv install gate locally, but the release workflow (pytest, lineage scan, pyright ratchet, then publish) has not run.

Relationship to the rest of the docs

The parity report sits above every code subpackage and points down into all of them — it narrates the same tree this documentation set covers area by area, with file:line citations into src/indusagi/ and tests/. For the live runtime surfaces it audits, start at Getting Started and the Architecture overview; for the packaging metadata it relies on, see Package Exports.