v0.55.0 — Bounded analysis and reliable trace discovery

This release hardens local repository inspection and restores trace coverage under bounded traversal.

Changes

• Bound repository analysis, subprocess output, parser work, MCP framing, lifecycle payloads and persisted JSON decoding. Corrupt relation metadata produces explicit partial-index diagnostics.
• Isolate optional native PHP parsing with runtime and memory limits; retain bounded regex extraction when enforceable native isolation is unavailable.
• Protect repository reads and private metadata files, validate Git administrative paths and reject executable local Git configuration.
• Require a per-launch dashboard capability, use an OS-assigned port by default, and separate authenticated work from incomplete/public connections.
• Minimize exported signatures and encode untrusted fields in Markdown and plain-text reports.
• Avoid repeated traversal of the same file from one start. Across six real-repository queries, default trace results now match the expanded reference Top 10 without exhausting the 1,000-state budget.
• Allow large flat Python configuration literals while retaining expression-chain and nesting limits.
• Pin CI actions to immutable commits and use read-only repository permissions.

Compatibility and upgrade

• Git metadata may be unavailable for repositories with unsupported local configuration, linked/external Git administrative directories, or untrusted executable paths. Filesystem orientation remains available where supported.
• Secure metadata/filesystem operations require the supported POSIX ownership and descriptor controls. Native PHP extraction falls back when hard isolation is unavailable.
• Large analyses may return explicit partial results. Open the dashboard using its newly printed private link for each launch.
• After updating the installed package, run init-agent map and init-agent sync; restart the MCP host so it loads the new process. Existing running processes are not upgraded by publishing this release.

Validation

• 313 committed tests in the release CI suite, plus local benchmark and graph-quality guards and optional native PHP isolation checks on Linux/Python 3.11–3.13.
• 346 tests exercised in the complete local working copy, including 33 experimental tests outside this release (two platform/optional-dependency checks skipped locally).
• Real-repository map, incremental refresh, query, context, trace, export, estimate, reading plans, MCP stdio, and loopback dashboard checks passed. Six-query trace comparison retained 15/16 expected files in the Top 10; the remaining file ranked 28.

v0.54.0 — Local security and memory reliability

• Restrict the dashboard to loopback addresses and validate Host/Origin.

  • Minimize stored Python signatures.
  • Check dashboard memory freshness against live files.
  • Require explicit --revalidate to renew memory evidence.
  • Report unknown trajectory outcomes and outcome coverage.
  • Align agent instructions and privacy documentation.

  Validation: 238 tests passed, plus local benchmark and graph quality guards.

  After upgrading, run:
  init-agent map
  init-agent sync

  Metadata and exports are not guaranteed secret-free.

v0.53.1 - Fix resumed trajectory sessions

Resumed trajectory session fix

• Reopen a logical Codex session when SessionStart follows SessionEnd with the same session ID.
• Preserve repeated lifecycle events in chronological order.
• Keep correlated tool, turn and subagent event deduplication.
• Correct dashboard active/finished status after session resume.
• Add regression coverage for the Start -> End -> Start lifecycle.

All 230 tests pass.

v0.53.0 - Observable Agent Trajectory

Observable Agent Trajectory

This release adds optional, privacy-preserving trajectory observability for Codex sessions.

Added

• Optional Codex lifecycle hook installation:
  • init-agent trajectory status
  • init-agent trajectory install-codex
  • init-agent trajectory uninstall-codex
• Lightweight fail-open init-agent-hook collector.
• Local SQLite storage for sessions and observable events.
• Tool duration pairing, outcomes, errors and subagent lifecycle tracking.
• Deduplication, concurrent-write handling and bounded retention.
• New read-only Trajectory section in the web dashboard.
• Updated README, security documentation, command reference and Codex skill.

Privacy

Trajectory collection stores redacted metadata only.

It does not store prompts, full commands, patches, tool output, source snippets, secrets or model reasoning. Paths
outside the repository are redacted.

Usage

init-agent trajectory install-codex

Then review and trust the hooks through /hooks inside Codex and restart the session.

### Validation

- 229 tests passing
- Concurrent hook ingestion tested
- Payload and SQLite secret-redaction tested
- Existing MCP core surface remains unchanged

v0.52.0 - Fresher indexes and leaner MCP payloads

What's changed

• Added shared, read-only repository index health checks.
• MCP orientation tools now detect changed, deleted, unindexed files and extractor drift.
• Reading plans, traces, callers and related-file results exclude deleted paths.
• Added compact MCP responses for related-file and symbol-callers tools.
• Full contracts remain available through include_details=true.
• Session close now reports repository index freshness and avoids duplicate memory suggestions.
• Doctor now reuses the shared index-health logic.
• Updated the Codex skill and MCP documentation.
• Added focused tests for stale indexes, deleted files, memory deduplication and compact payloads.

Verification

• Full test suite passed.
• git diff --check passed.
• Bundled and source skill copies are synchronized.

v0.51.0 - Compact MCP workflow

Highlights

• Added a compact eight-tool MCP core profile for the daily agent workflow.
• Preserved all advanced and legacy tools through the optional full profile.
• Made reading-plan delegation explicit and opt-in.
• Simplified the default plan, verify, finish and session-close loop.
• Reduced MCP schema payload size and removed redundant plan construction.
• Shortened and aligned the Codex orientation skill.
• Cleaned obsolete task and changelog files.
• Updated README, MCP, workflow, feedback, roadmap and experiment documentation.
• Added documentation consistency and Markdown link tests.

Compatibility

Existing advanced integrations can retain the complete MCP surface with:

init-agent mcp install-codex --profile full --replace

Verification

• 217 tests passed
• Documentation links verified
• Source and bundled Codex skills verified as identical

v0.50.0 - Deeper structural graph and bounded reranking

Highlights

• Added qualified Python and PHP symbol identities.
• Added scoped symbol-to-symbol calls, inheritance and implementation relations.
• Improved relation resolution, trace, query and graph exports.
• Added a conservative one-hop structural reranker.
• Strongly matching test methods can promote directly related production code.
• Added fan-out limits and safeguards against generic test classes and inheritance noise.
• Expanded the deterministic flow-graph benchmark.

Validation

• 210 tests passed.
• 30/30 extraction expectations passed.
• 28/28 relation-resolution expectations passed.
• 8/8 ranking cases achieved a Top-3 hit.
• No ranking regressions observed in the tested direct-query sets.

The reranker is intentionally conservative. It provides an additional structural signal when the graph has
sufficient evidence; it is not a replacement for direct file verification.

Upgrade note

This release uses a newer graph extractor. After upgrading, rebuild existing repository indexes:

init-agent map
init-agent doctor

v0.49.1 - Fix Duplicate Codex Skill Discovery

Bug fix

Fixes Codex discovering skill backups as duplicate installed skills.

• Skill backups now live under ~/.codex/init-agent-backups/skills/.
• init-agent sync automatically migrates legacy *.bak-* directories out of
  ~/.codex/skills/.
• Locally modified skills are still backed up safely before replacement.
• Added regression coverage ensuring only the active skill remains discoverable.

After upgrading:

pipx upgrade init-agent
init-agent sync

v0.49.0 - Skill Sync and Update Diagnostics

What's new

init-agent now keeps copied Codex skills aligned with the installed package.

Codex skill synchronization

New command:

init-agent sync

It:

- installs or updates the bundled Codex skill;
- writes a local version and SHA-256 manifest;
- detects modified, outdated, or legacy-untracked skill copies;
- creates a timestamped backup before replacing uncertain local contents;
- checks whether the init-agent MCP server is registered;
- never modifies Codex MCP configuration.

### Doctor improvements

init-agent doctor now compares the installed Codex skill with the bundled version
locally and suggests init-agent sync when needed.

Remote release checks remain explicit:

init-agent doctor --check-updates

The check distinguishes:

- current version;
- local version ahead of the latest published release;
- update available;
- network unavailable.

Network failures remain non-fatal and normal doctor checks do not contact external
services.

### Upgrade workflow

After upgrading init-agent:

pipx upgrade init-agent
init-agent sync

Open a new Codex session when the skill or MCP installation changes.

### Validation

- 198 tests passing
- Real GitHub Releases API smoke test completed
- Documentation and bundled Codex skill updated

v0.48.0 - Orchestrated Subagent Delegation

What's new

init-agent can now help a parent coding agent decide whether a reading plan should
remain direct or be split into bounded delegated workstreams.

Delegation guidance

Reading plans can propose:

• direct, single_worker, or parallel strategies
• fast, balanced, or deep model tiers
• low, medium, or high reasoning effort
• bounded read-only scopes for worker agents

The advice is provider-agnostic: init-agent does not launch agents itself. The parent
orchestrator remains responsible for spawning workers and assigning available models.

Worker lifecycle

New CLI operations:

• init-agent plan report
• init-agent plan review

Workers can report:

• files read or modified
• tests performed
• verified findings
• risks
• remaining work

The parent orchestrator must explicitly accept, reject, or request rework. A reading
plan cannot finish while a delegated report is awaiting review or rework.

MCP support

Added:

• repo_workstream_report
• repo_workstream_review

Delegated reads, reports, and parent decisions are persisted in the local SQLite
metadata and remain visible through plan inspection.

Safety

• Proposed but unused workstreams do not block completion.
• Started workstreams require a report and parent review.
• Workers cannot approve their own work.
• Worker reports remain claims that the orchestrator must independently verify.

Documentation and tests

Updated CLI, MCP, agent workflow, README, and bundled Codex skill documentation.

All 190 tests pass, including delegation lifecycle, CLI, MCP, persistence, and finish-
gating coverage.