Parity & Gaps

induscode is a clean-room Python rebuild of the TypeScript indusagi-coding-agent (lineage v0.1.62), running entirely on the indusagi framework. This page summarizes the area-by-area parity audit and the known gaps recorded against the TS ground truth. Read it via import induscode (the package barrel single-sources VERSION and re-exports all subsystems) or by running pindus --version.

The audit compares induscode-python-rebuild/src/induscode row by row against the TS source tree indus-code-rebuild/src. The verdict is full feature parity on the user-facing surface — every CLI flag, all 26 slash commands, every console overlay, settings, sessions, channels, addons, and the capability deck — with 13 deliberate, plan-cited divergences and a short list of known gaps that cluster in two subsystems (launch and insight). The tone here is neutral and factual: this is a rebuild, and this page records exactly where it matches and where it does not.

Table of Contents

• Provenance
• Parity by Area
• Deliberate Divergences
• Known Gaps
• Upstream Framework Defect
• Library Swaps
• Version & Doc-String Drift
• Source Files

Provenance

induscode is an independent implementation written against a behavioral spec plus the public indusagi API. No third-party application source was copied. Two release gates enforce this:

Gate	What it checks
scripts/lineage_scan.py	Word-boundary greps src/induscode (excluding tests and the separately-installed framework) for upstream pi / mariozechner markers; exits 1 on any hit.
public-API guard test	Pins the package's exported surface so it cannot silently drift.

python3 scripts/lineage_scan.py   # exits 1 if any upstream marker appears in src/induscode

The provenance and notice travel with redistributions: pyproject.toml embeds both NOTICE and CREDITS.md into the wheel via PEP 639 license-files. The package is MIT-licensed and built with hatchling. The underlying indusagi framework is a separate PyPI package with its own notice and is explicitly not scanned by the lineage gate — induscode is built on top of it and reuses it wholesale (LLM API abstraction, agent core, Textual TUI primitives, MCP client, observability), keeping only its own seams (transcript tree, auth vault, signal translation, capability deck).

Parity by Area

The port was delivered milestone by milestone (M0–M6). The M6 parity audit marked 106/106 checklist rows done (0 partial, 0 missing), with the test suite green. The table below summarizes the areas; each is a complete, tested layer.

Area	Parity	Notes
CLI flags + conventions	full (23/23)	All 17 flags (-p, --json/--rpc, -i, -m, --thinking, --list-models, --cwd, --resume, --continue, --tools, --no-tools, --mcp, --system, --append-system, --help, --version, --account) plus @file attachments, -- terminator, -pi clustering, and unknown-flag tolerance. See CLI reference.
Subcommands	full	signin/login, signout/logout, and install/remove/update/list/config package management. See auth.
Slash commands	full (26 static + dynamic + 10 aliases)	Transcript, workbench, and integrations groups, plus dynamic /skill:<name> and /<template> macros. See slash commands.
Overlays, theme, input, startup	full (16/16)	12 modal kinds routed through OVERLAY_HANDLERS, 4 theme schemes with live preview, 21-chord /keys, double-Ctrl+C exit, startup map and changelog splash. See dialogs and theming.
Channels	full (5/5)	SESSION_OPS ×6 with flat LinkSnapshot, dialog bridge ask/tell (90 s timeout→fallback), oneshot text + NDJSON framing, U+2028/29 escaping with split-rune reassembly, JSON-RPC error codes. See channels.
Addons + capability deck	full (8/8)	13 hook events with observe/transform/gate, interceptor onion, .indus/addons discovery via importlib, 12 framework built-ins, 5 app cards, and the content/ULID bridge ledger. See addons and capability deck.
Sessions, transcripts, export	full	Branchable NDJSON tree (indus/transcript@1), resume by id with deepest-leaf recovery, per-cwd --<slug>-- scoping, legacy import, and self-contained HTML /export. See sessions and transcript export.
Runtime-bridge	full (2/2)	3 dialects (claude-cli, codex-cli, indusagi-cli) → provider-neutral NormalizedEvent. Library-only in both builds (not wired into boot). See runtime bridge.
Window-budget + insight	full (wired)	window_budget condenser (0.75 trigger / 6000 keep-recent / 2048 reserve), manual /summarize-context, and the insight collector/replay wrapper over framework tracing. See window budget and insight.
Auth vault	full	Multi-account with OAuth refresh, default promotion, 0600 file mode. See auth.

For the boot pipeline that wires these together, see boot and launch.

Deliberate Divergences

These 13 differences are by-design ports of TS behavior onto Python / Textual / framework idioms. They are documented, plan-cited, and fully tested — not defects.

#	Divergence
W1	Composer deleted — the TS bespoke software-caret composer is replaced by the framework PromptEditor, folded into console/app.py + console/input/.
W2	Paste vault / burst-debounce deleted — replaced by framework EditorCore paste markers and native paste events.
W3	jiti → importlib — .py addons load via importlib.util.spec_from_file_location; the TS virtual-module/alias bridge collapses and BUNDLED_NAMESPACES is vestigial.
W4	insight wraps indusagi.tracing — the bulk tracing engine stays in the framework; only the collector sink, replay reader, and redaction wrapper are app code.
W5	repl placeholder → real Textual mount — the M0 cli.py stub is replaced by the real console mount in boot/runners/repl_runner.py (the seam stays injectable for tests).
W6	runtime-bridge barrel-only — a tested library layer in both builds, deliberately not wired into the boot path in either.
W7	/memory on/off cosmetic — flips a reporting flag only, never removes the memory card from the live deck. Matches TS.
W8	Two distinct compaction engines — window_budget (tokens, 0.75, 3.6 chars/token) is the wired conductor CondenseFn; the framework runtime.memory (turns, 0.8) is intentionally unwired.
W9	Chord-latch break semantics — timer-bounded (600 ms) under Textual since the editor consumes printables; behavioral parity with the TS latch.
W10	OAuth overlay on asyncio.Future — a parked-resolver replaces TS callback threading; provider threading and vault refresh are preserved.
W11	TypeBox → literal JSON-schema dicts — addon/tool schemas are plain dicts per the framework parameters: Mapping convention.
W12	Modal routing inversion — TS always-mounted overlays become awaited push_screen flows with typed outcomes via OVERLAY_HANDLERS.
W13	todo wire names — todo_read / todo_set (vs the TS internal names), honoring the Python framework's wire contract verbatim.

Known Gaps

The gap audit (2026-06-12 snapshot) records 7 gaps: 0 high, 1 medium, 6 low, clustered in launch and insight. Nothing in the core agent loop, tool execution, or console UI is affected. The estimate is ~98% behavioral parity.

#	Gap	Subsystem	Status	Severity
1	CLI --resume interactive picker never mounts on a TTY; the runner silently falls back to a fresh session	launch	broken	medium
2	GitHub Copilot browser sign-in provider absent (framework OAUTH_PROVIDERS carries only anthropic, openai)	launch	missing	low
3	Insight secret redaction over-redacts — whole value instead of the matched run	insight	divergent	low
4	Insight fail() emits no in-flight update signal before close	insight	divergent	low
5	Insight persisted/streamed fault loses the error name + stack	insight	partial	low
6	Insight redactor omit_keys option (drop keys wholesale) missing	insight	missing	low
7	Top-level package namespace re-export of all subsystems	root barrel	divergent	low

Notes on individual gaps:

• Gap 1 (medium, the most impactful): _DefaultResumeDeps.mount_picker in launch/pickers.py unconditionally raises, and boot/runners/repl_runner.py calls pick_resume_target with no injected deps, so pindus --resume on a real terminal starts a fresh session with a one-line stderr notice. It is medium rather than high because the capability is reachable: the in-console /resume slash command is fully wired and opens the exact same arbitrary-session picker (console/overlays/sessions.py via push_screen_wait). A user can run pindus --resume, land in a fresh session, type /resume, and reach the picker.
• Gaps 3–6 (insight): all four are wrapper-side costs of delegating to the framework SecretScrubber / SegmentError / transport (waiver W4). None affects the final close record's status or any end-user agent behavior; impact is limited to live trace dashboards and offline trace analysis. They are documented inline in insight/wrapper.py.
• Gap 7 (root barrel) is now closed in source. The gap report claims induscode/__init__.py re-exports only workspace and that the docstring still references the M0-era stub. The current source lazily re-exports all 17 subsystems via PEP 562 __getattr__ with __all__ = ["VERSION", "__version__", *sorted(_SUBSYSTEMS)], and the docstring describes the lazy barrel. Treat the gap report as a point-in-time snapshot; at least this row no longer holds.

import induscode
induscode.runtime_bridge      # lazy namespace re-export (PEP 562 __getattr__)
induscode.capability_deck     # any of the 17 subsystems
from induscode.workspace import BRAND, VERSION
print(BRAND.bin_names)        # ('pindus', 'induscode')

The recommended fix order, highest user-impact first: wire the CLI --resume Textual picker (gap 1), add the GitHub Copilot OAuth config (gap 2), then the four insight refinements (gaps 3–6).

Upstream Framework Defect

One known defect is not a port gap and is marked DO-NOT-FIX in the app:

• Symptom: the first streamed text chunk is dropped from -p / --json / NDJSON output ("pong" → "ong"). The whole first chunk is dropped, not a fixed character count.
• Root cause: the framework agent/agent.py _FacadeTranslator._on_delta emits the first text delta only inside the MessageStartEvent partial; only the second-and-later deltas surface as MessageUpdateEvent text deltas.
• App side is parity-correct: conductor/signal_hub.py maps message_start → [] (matching the TS translator table), so no TextSignal is produced for the first chunk. The oneshot text strategy accumulates only TextSignal deltas, so -p loses the first chunk.
• Blast radius: -p text output, the -p / --json text-signal stream. The interactive TUI masks it — the final message_end carries the complete assistant message, so the rendered result is correct.

The fix is filed upstream against the framework; the app intentionally does not work around it to avoid diverging from the TS translator table. See framework parity for the framework-side record.

Library Swaps

The TS dependencies were replaced with Python equivalents during the port:

TS dependency	Python replacement
marked	markdown-it-py>=3.0 (transcript-export markdown)
highlight.js	pygments>=2.18 (transcript-export highlighting)
ulid	python-ulid>=2.7 (bridge-ledger ULID keys)
Ink / React	Textual (interactive console)
jiti virtual modules	plain importlib addon loading

The sole framework dependency is indusagi[mcp,tui]>=0.1.2. The mcp and tui (Textual) extras are required because the framework's shell_app barrel imports mcp and textual eagerly, and the agent's MCP enrollment plus the Textual console need them. The process title is not set — the TS entry set process.title, but Python has no portable titling without a third-party dependency, so it is deliberately skipped (the same decision as the framework port).

Version & Doc-String Drift

The version is single-sourced: it lives only in pyproject.toml (currently 0.1.2), and workspace/brand.py reads it back at runtime via importlib.metadata.version("induscode"), with a "0.1.0.dev0" fallback that fires only on an uninstalled source checkout. This fixes the TS build, which duplicated the version as a literal.

import induscode
print(induscode.VERSION)        # e.g. '0.1.2' — from importlib.metadata, not a literal

Because the runtime version comes from package metadata, pindus --version reports whatever pyproject.toml declares (0.1.2), not the doc strings. Note that the README, CHANGELOG (header 0.1.0 — 2026-06-11), and PARITY_REPORT still reference 0.1.0/v0.1.0 from the first release. Test-count figures also drift across documents (the changelog and parity report cite 649 tests, the README cites 703); take the per-document figure as the snapshot at that document's date.

Source Files

File	Role
PARITY_REPORT.md	M6 row-by-row parity audit against the TS ground truth (106/106 rows, 13 waivers, live smoke results).
GAP_REPORT.md	The 7-gap audit (2026-06-12) with TS/Python evidence, user impact, and fix hints per gap.
CREDITS.md / NOTICE	Provenance and notice, embedded into the wheel via PEP 639 license-files.
CHANGELOG.md	Release log (first release header 0.1.0 — 2026-06-11).
pyproject.toml	The build manifest — version single-source, indusagi[mcp,tui] dependency, pindus/induscode console scripts.
scripts/lineage_scan.py	The clean-room provenance gate.
src/induscode/__init__.py	The lazy namespace barrel re-exporting all 17 subsystems.
src/induscode/workspace/brand.py	BRAND / Brand / VERSION — the single source of truth for identity literals.
src/induscode/entry.py	The pindus / induscode console-script entry (run → main → boot).

See also Package Exports for the full namespace map and Testing for the test suite layout.