Source Atlas

scripts/rescue_voyage.py:448 · Identify the target, stages, probe and command overrides.

Guard

Target must contain .git; stage must be in STAGES

Output

Ctx and voyage.json with state=running

Failure

Return 2 for an invalid target or stage

Risk

Checks only for a .git path; repository trust is not verified

scripts/rescue_voyage.py:52 · Get the target name, scanner set, probe and owner mode from the registry.

Guard

--path takes precedence over the named target

Output

reg, target, scan_tools

Failure

An unknown name does not create a target

Risk

Stale configuration can diverge from the target

scripts/llm_config.py:34 · Assign write, verify and audit from config.toml or VIBE_LLM_CONFIG.

Guard

Merge DEFAULTS; do not read an API key from the file

Output

CLI command strings by role

Failure

A missing CLI fails at its stage

Risk

Arbitrary CLIs have host privileges; small models have not been fully evaluated

scripts/dialogue.py:58 · Record or ask for a decision; never approve automatically.

Guard

Random ID, nonce/binding, timeout and one-time answer consumption

Output

choice, note; QUESTION/ANSWER in file mode

Failure

Invalid answer or timeout raises an exception and blocks voyage

Risk

Anyone who can write the answer directory may impersonate the owner

scripts/rescue_voyage.py:444 · Run in fixed order even when --stages lists another order.

Guard

An rc or exception stops the loop

Output

voyage.json done/blocked/error

Failure

A failed stage returns 2 and records state

Risk

--stages permits stage omission; REPAIR checks only part of preflight itself

scripts/rescue_voyage.py:226 · An advisor or template creates the plan; plan_check runs before owner approval.

Guard

plan_check=0 plus accept required before saving hash/SHA

Output

approvals/PLAN-*.json

Failure

Rejection or three failed attempts stops

Risk

Approval file sits beside evidence and has no signature

scripts/rescue_voyage.py:285 · Check plan hash and HEAD; ask before execution and again before merge.

Guard

Hash and target HEAD must match

Output

Candidate branch or merge commit

Failure

Rejection retains/discards candidate without merging

Risk

Subprocesses have host privileges; the worktree is not a sandbox

scripts/rescue_voyage.py:403 · The owner decides whether to run drift checks when a target has rulesync.

Guard

Explicit accept/reject

Output

Drift result

Failure

Nonzero exit blocks when checks run

Risk

Accepting with no configuration leads to a skip

scripts/rescue_voyage.py:99 · Build a bounded context pack and target inventory.

Guard

Exit code must be 0

Output

session/map/

Failure

context_pack failure blocks MAP

Risk

The pack may contain private data; review before sending to an LLM

tools/meta/recon.py:117 · Count languages, large files, Git state and agent assets.

Guard

recon.json must exist

Output

session/map/recon.json

Failure

Missing artifact blocks MAP

Risk

Recon is inventory, not proof of runtime flow

scripts/rescue_voyage.py:113 · Select scanners and run them concurrently.

Guard

Fanout must exit 0 before merge

Output

prove/raw/*.json, run.summary.json

Failure

A failed tool blocks PROVE

Risk

A skipped scanner may still be accepted

scripts/fanout.py:63 · Launch one subprocess per tool in the manifest.

Guard

Thread pool; 1800 s per tool

Output

raw/<tool>.json

Failure

Missing output creates a failed record

Risk

Time and cost grow with tools; scanners may have side effects

tools/meta/scan.py:273 · Classify unavailable/skipped tools and construct pinned commands.

Guard

Configuration, manifest and tool must exist

Output

Command or skipped reason

Failure

Unknown tool is skipped

Risk

An unknown listed tool does not fail fanout

tools/meta/scan.py:357 · Normalize tool output into findings.

Guard

Parse failures or disallowed exits return an error

Output

Findings by tool and exit_code

Failure

Missing parser/report or fatal exit fails

Risk

Some parsers still need stricter schema and coverage-byte checks

scripts/merge_findings.py:37 · Redact text, normalize paths and deduplicate severity.

Guard

Drop records without a file

Output

findings.json/.md and SARIF

Failure

Malformed raw JSON is listed in errors

Risk

Pattern-based redaction cannot guarantee PII removal

scripts/rescue_loop.py:59 · Compare scanner lists/statuses and new findings.

Guard

Lost ok coverage blocks; delta key uses tool, file and rule

Output

Baseline or delta verdict

Failure

A new finding returns 2

Risk

An old finding on a different line is not new; skip-to-skip still passes

scripts/run_target.py:28 · An independent entry point scans a registry target.

Guard

MAP/recon/fanout/merge must exit 0, artifacts must be fresh, at least one scanner ok

Output

reports/targets/<name>/

Failure

A failed step blocks baseline save

Risk

Skipped scanners still need a required/optional policy

scripts/rescue_voyage.py:170 · Load the audit prompt through PROMPT_AGENT and invoke review_cmd.

Guard

Check README/REVIEW and PROPOSALS modification times

Output

review/{README,REVIEW,PROPOSALS}.md

Failure

CLI failure produces a template with a notice

Risk

mtime is weak evidence; the prompt may miss the domain

scripts/rescue_voyage.py:510 · Provide a fallback when the agent writes no fresh files.

Guard

Attach a no-LLM-audit notice

Output

Template plus scanner counts

Failure

Missing LLM output alone does not block REVIEW

Risk

A reader may mistake the template for a full audit

scripts/check_review_numbers.py:292 · Compare numerical claims with evidence and source.

Guard

Advisory result included in review

Output

Number-check table

Failure

Checker failure does not block REVIEW

Risk

Qualitative claims and external citations are not verified

scripts/collect_evidence.py:53 · Collect masked JSON for handoff.

Guard

Nonzero exit blocks REVIEW

Output

review/EVIDENCE/

Failure

Packaging failure blocks the stage

Risk

Masking is heuristic; review manually before sharing

scripts/rescue_voyage.py:148 · Have the verify role inspect wrong counts, contradictions and missing evidence.

Guard

Runs only without a custom review command

Output

cross-check.md if the CLI writes it

Failure

Missing/failing CLI fails open

Risk

Not a release gate or proof of independent audit

agents/orchestrator.md:1 · Specialized prompts describe roles and expected output.

Guard

Only auditor_architecture loads directly in voyage

Output

Manual handoff if orchestrator invokes them

Failure

No automatic runtime fan-out across all auditors

Risk

Prompt assets may be mistaken for executed checks

scripts/plan_check.py:81 · Check sections, verification per step, Scope and SHA.

Guard

Return 2 if the plan is missing or mismatched

Output

Valid PLAN-*.md

Failure

An invalid plan cannot be accepted

Risk

The checker cannot prove that Verification steps are sufficient

scripts/rescue_loop.py:75 · Check baseline, probe, clean target, HEAD and coverage.

Guard

Baseline SHA must match HEAD

Output

loop.json round 0

Failure

Violation returns BLOCK

Risk

_tracked_clean does not detect untracked target files

scripts/guards.py:42 · Capture probe rc and stdout/stderr hashes.

Guard

Every rc must be 0

Output

probe_before

Failure

A failing baseline probe returns BLOCK

Risk

Output hashes are sensitive to nondeterministic output

scripts/rescue_loop.py:161 · Create a candidate branch and invoke the writer.

Guard

Target HEAD and clean state must remain unchanged

Output

rescue/<plan-id> worktree

Failure

Worktree error returns BLOCK

Risk

The executor has host filesystem and network privileges

scripts/plan_check.py:129 · Require changed filenames to stay within Scope.

Guard

Check Scope before and after reviewer

Output

Allowed diff

Failure

Out-of-scope change returns REVISE

Risk

Malicious content in allowed files still needs reviewer scrutiny

scripts/rescue_loop.py:214 · Commit the candidate for a real merge and compare probes.

Guard

Probe rc and hash must match before

Output

candidate_sha

Failure

Commit failure or probe mismatch returns REVISE

Risk

A probe cannot cover every behavior

scripts/rescue_loop.py:243 · Scan candidate with the baseline scanner set.

Guard

No failures, lost coverage or new findings

Output

Round evidence

Failure

Violation returns REVISE

Risk

A scanner skipped in baseline can remain skipped

scripts/rescue_loop.py:38 · Have an independent verify role write a clear verdict.

Guard

Candidate must be clean; stale verdict removed

Output

APPROVE/REVISE/BLOCK

Failure

Unknown/REVISE repeats until max_rounds

Risk

A text verdict does not prove the reviewer ran tests

scripts/rescue_voyage.py:350 · Owner approves merge; compare SHA and run a post-merge probe.

Guard

Target HEAD and candidate SHA unchanged

Output

CHANGES.log and REVERT.md

Failure

Probe mismatch auto-reverts if HEAD is clean

Risk

Side effects outside Git are not rolled back

scripts/logic_atlas.py:46 · Constrain scope to the target and read only Git-tracked regular files.

Guard

Reject path escape and symlinks

Output

Language-specific file inventory

Failure

Bad scope raises error

Risk

Untracked code does not enter the graph

scripts/logic_atlas.py:69 · Build a static Python call graph with pinned Pyan.

Guard

max_files=400; uvx must be available

Output

python-calls.dot and HTML/SVG

Failure

unavailable/too_large/partial

Risk

Dynamic dispatch and callbacks remain uncertain

scripts/logic_atlas.py:136 · Build JS/TS import graph without executing target configuration.

Guard

--no-config; unresolved stays unknown

Output

frontend-dependencies.json

Failure

Partial for unresolved imports or Vue/Svelte

Risk

Event-to-HTTP/API edges are not linked

scripts/syntax_map.py:42 · Extract JS/TS/JSX/TSX functions, arrow variables and direct same-file calls.

Guard

Member/dynamic calls stay unknown; works without npx

Output

frontend-symbols.json and HTML/SVG

Failure

Partial for syntax/unknown/Vue/Svelte

Risk

Event, HTTP and cross-file calls are not linked

scripts/syntax_map.py:42 · Extract Rust/Go/Java/PHP/C# symbols and same-file calls.

Guard

Skip symlinks and files over 1 MB

Output

language-syntax.json

Failure

Unknown calls remain explicit

Risk

Cross-file calls, cfg and reflection remain unresolved

scripts/atlas_render.py:121 · Export offline HTML/SVG/PDF with a full node/edge/unknown HTML catalog.

Guard

SVG preview limited to 120 nodes

Output

*-map.{html,svg,pdf}

Failure

Truncated SVG marked partial

Risk

A static catalog does not explain end-to-end behavior

scripts/repo_explorer.py:38 · Index all functions, variables and calls in the harness without a 120-node cap.

Guard

Parse syntax; show unknown calls

Output

repo-explorer.html and five SVGs

Failure

Parse errors are listed

Risk

Static calls are not runtime traces

scripts/model_rescue_eval.py:148 · Smoke-test MAP/PLAN/REVIEW on synthetic fixtures through a budget model gateway.

Guard

Schema, plan_check and unsafe-verdict grader

Output

reports/model-eval/*.json

Failure

Contract failure returns 2

Risk

Not an end-to-end executor/owner test

scripts/public_snapshot.py:44 · Package fresh source without old Git history or reports.

Guard

Exclude reports/logs; reject symlinks

Output

tar.gz and SHA manifest

Failure

Unsafe path aborts

Risk

Still needs secret, license and manual content review

.gitleaks.toml:6 · Scan extracted source and verify scanned bytes are nonzero.

Guard

Do not exclude the entire root/public snapshot

Output

Gitleaks finding report

Failure

Zero scanned bytes is a false green and blocks release

Risk

Zero matched patterns does not prove absence of sensitive data

LICENSE:1 · Prepare source for MIT release.

Guard

Do not push the old branch/history

Output

Candidate public source snapshot

Failure

Missing review prevents publication

Risk

External tool licenses and data need separate review