Skip to content

PHASE2_5_BASELINE.md

Latest commit

 

History

History
88 lines (67 loc) · 3.7 KB

PHASE2_5_BASELINE.md

File metadata and controls

88 lines (67 loc) · 3.7 KB
Raw

Phase 2.5 Baseline

State of the tree at fb9b32a, the final Phase 2 commit, measured before any Phase 2.5 change.

Tests

suite count result
packages/shadowfetch-missions 252 OK (1 skip)
tools/tests (release gate) 143 OK
packages/shadowfetch-fireline 11 OK
every suite via make test 687 OK
make source-gate — SOURCE_GATE_PASSED

A fresh checkout of fb9b32a did not have 252 passing tests — it had 132. .gitignore carries *.log, and two of the provider conformance fixtures are .log files, so test_provider_conformance.py died with FileNotFoundError at import in any clone. The suite passed only in a working tree that happened to contain untracked files. This is measured, not inferred:

fb9b32a: Ran 132 tests  FAILED (errors=1)      <- Phase 2 final, fresh worktree
4351129: Ran 184 tests  FAILED (errors=1)      <- mid-Phase-2.5, same cause
5c16039: Ran 358 tests  OK (skipped=3)         <- after the .gitignore exception

Providers

Two, both shipped: codex (code_change, sourced_report) and offline-media (media_export). conformance-echo exists as a test fixture only.

Readiness on this host at baseline, unchanged since:

codex          available=False  Run shadowfetch-mission-account login for a
                                dedicated account, or save a user-owned 0600
                                CODEX_API_KEY environment file and restart the
                                idle worker. Credential presence does not verify
                                authentication.
offline-media  available=True

Conformance

23 assertions per provider, run against codex, offline-media and the conformance-echo fixture. Interface version 1.

The P1 this phase exists to close

PHASE2_REMAINING_RISKS.md §3: Phase 2 replaced the AST provider freeze with a JSON Schema, so any package that landed a schema-valid manifest in /usr/share/shadowfetch/providers became a provider, with whatever capabilities, credentials and network posture it declared. Nothing recorded that a human had ever agreed to run it.

Demonstrated at baseline, before the fix: a copy of offline-media.json with its id changed to helpful-agent and dropped into the discovery directory was loaded, activated, and offered for its declared capability.

Declared-but-unenforced at baseline

From PHASE2_REMAINING_RISKS.md §1–§2 and AGENT_ARCHITECTURE.md §6.6:

field status at baseline
egress_allowlist declared, validated, narrowable, reaches nothing
masked_paths declared, validated, narrowable, reaches nothing
cpu_seconds fixed during Phase 2; §6.6 still said it was unenforced
workspace_mode not in §6.6 at all — the field with a live enforcement hole was not on the known-gaps list
memory_mb passed and enforced — swap not considered
syscall profile not representable; not recorded anywhere

Executable trust at baseline

A path-prefix test. A program under /usr/bin/, /usr/sbin/, /usr/libexec/, /usr/lib/, /usr/local/lib/shadowfetch/, /bin/, /sbin/ or /opt/ was trusted, with no examination of the ownership or mode of the file or of any directory above it. trust: "user-runtime" additionally accepted any group-writable program, recorded as a residual risk.

Interface shape at baseline

Both shipped providers are batch-shaped: Codex emits one JSONL record per completed item, ffmpeg at loglevel=error emits almost nothing. No provider exercised incremental output, an explicit session, heartbeats, tool rounds, or a stream with no guaranteed terminal record. An interface validated against only those two had not been validated.